A survey of LLMs in drug discovery and precision medicine

1 Introduction

The pharmaceutical industry faces unprecedented challenges in developing safe, effective, and personalized therapeutics. Traditional drug discovery remains a lengthy and expensive endeavor, with average development timelines exceeding 10–15 years and costs surpassing $2.6 billion per approved drug (DiMasi et al., 2016; Wouters et al., 2020). The attrition rate in clinical trials remains alarmingly high, with approximately 90% of drug candidates failing to receive regulatory approval, often due to unforeseen toxicity or lack of efficacy that could not be predicted during preclinical development (Vamathevan et al., 2019; Paul et al., 2021). Simultaneously, the promise of precision medicine, which aims to tailor treatments to individual patient characteristics including genetic makeup, lifestyle, and environmental factors, remains partially unfulfilled due to the inherent complexity of integrating multiscale biomedical data across molecular, cellular, tissue, and organismal levels (Esteva et al., 2019; Rajkomar et al., 2018). The convergence of artificial intelligence and biomedical research offers emerging impact to address these longstanding challenges, with large language models emerging as particularly powerful and versatile tools capable of processing and synthesizing information across diverse data modalities (Figure 1) (Wang et al., 2023; Thirunavukarasu et al., 2023).

Large language models represent a paradigm shift in artificial intelligence, demonstrating remarkable capabilities in understanding and generating human language through deep learning on massive text corpora (Vaswani et al., 2017; Devlin et al., 2019; Brown et al., 2020). Models such as GPT (Generative Pre-trained Transformer), BERT (Bidirectional Encoder Representations from Transformers), and their derivatives have achieved unprecedented performance across diverse linguistic tasks, from text classification and question answering to creative writing and code generation (Radford et al., 2019; OpenAI, 2023). The fundamental insight driving their success is that self-supervised learning on vast amounts of text enables the extraction of rich semantic representations that capture not only syntactic patterns but also complex reasoning capabilities and world knowledge (Brown et al., 2020; Wei et al., 2022).

This insight has proven surprisingly transferable to biomedical domains, where diverse data types including molecular SMILES strings, protein amino acid sequences, genomic nucleotide sequences, and clinical narratives can be treated as specialized languages with their own syntax, grammar, and semantic relationships (Lee et al., 2020; Luo et al., 2022; Singhal et al., 2023). The ability of these models to learn meaningful representations from sequential data, regardless of whether that data represents English text or chemical structures, has opened new avenues for computational approaches to drug discovery and clinical decision making.

Recent years have witnessed explosive growth in applying LLMs to drug discovery and precision medicine, fundamentally reshaping how researchers approach complex biomedical problems (Schneider et al., 2020; Mak and Pichika, 2019). These models can process and integrate heterogeneous data types spanning multiple scales of biological organization, from atomic level molecular structures and protein sequences to patient level electronic health records and population level epidemiological data (Vamathevan et al., 2019; Shickel et al., 2018). In drug discovery, LLMs enable tasks that previously required extensive domain expertise or costly experimental validation, including target identification from disease biology, de novo molecular generation with desired properties, absorption, distribution, metabolism, excretion, and toxicity (ADMET) prediction, and systematic drug repurposing through knowledge synthesis (Stokes et al., 2020; Zhavoronkov et al., 2019; Bilodeau et al., 2022). In precision medicine, these models are transforming clinical decision support by analyzing unstructured clinical notes, enabling sophisticated patient stratification based on multi-modal data, personalizing treatment recommendations, and interpreting complex genomic variants (Singhal et al., 2023; Nori et al., 2023). The ability of LLMs to perform zero-shot and few-shot learning, making accurate predictions on novel tasks with minimal or no task-specific training examples, is particularly valuable in biomedical contexts where labeled data for specific applications may be scarce, expensive to obtain, or simply nonexistent for emerging diseases or novel therapeutic targets (Brown et al., 2020; Kojima et al., 2022).

The evolution of biomedical LLMs has followed a distinct trajectory from their general-purpose counterparts, with successive waves of domain-adapted text models (BioBERT, PubMedBERT, ClinicalBERT), generative biomedical models (BioGPT, Med-PaLM 2, MEDITRON-70B), and modality-specific foundation models for molecules, proteins, genomes, and single cells (ChemBERTa, ESM-2, DNABERT-2, scGPT) (Lee et al., 2020; Luo et al., 2022; Singhal et al., 2023; Gu et al., 2021; Alsentzer et al., 2019; Chithrananda et al., 2020; Lin et al., 2023; Zhou et al., 2023; Cui et al., 2024). A full technical treatment of these architectures, their pretraining strategies, and their multi-modal extensions is given in Section 2; here we highlight only that this diversification has produced an ecosystem of specialised models rather than a single general-purpose biomedical LLM, and that multi-modal architectures linking molecules, sequences, and natural language remain an active and unsettled research frontier (Cao et al., 2023; Liu S. et al., 2023; Taylor et al., 2022).

The emerging impact of LLMs in biomedicine extends beyond individual predictive tasks to encompass entire workflows and research paradigms. Recent work has demonstrated the feasibility of autonomous AI systems capable of designing, executing, and analyzing chemical experiments with minimal human intervention (Boiko et al., 2023). Multi-agent architectures employing specialized LLM agents for different aspects of drug discovery, from literature review and hypothesis generation to molecular design and virtual screening, are beginning to mirror the collaborative nature of real world research teams (Bran et al., 2024; Park et al., 2023). These developments suggest a future in which AI systems serve not merely as tools for specific tasks but as collaborative partners in the scientific discovery process, accelerating the pace of biomedical research while potentially democratizing access to sophisticated computational capabilities (Wang et al., 2023). At the same time, the evidence base for these capabilities remains uneven: most reported gains come from benchmark accuracy on retrospective datasets rather than prospective clinical or wet-lab validation, and statements such as “expert-level performance” or “transformative potential” should therefore be read as conditional on the specific dataset, task, and evaluation protocol. As discussed in Section 7, claims of expert-level performance on USMLE-style questions (Singhal et al., 2023; Nori et al., 2023) are not equivalent to expert-level performance in real clinical workflows, and current AI-designed drug candidates have begun Phase I/II trials but have not yet produced an approved therapeutic (Xu et al., 2025). However, realizing this potential requires careful attention to the limitations and risks associated with LLM deployment in high stakes biomedical applications, including concerns about hallucination, interpretability, data privacy, and algorithmic bias that could perpetuate or exacerbate health disparities (Thirunavukarasu et al., 2023; Obermeyer et al., 2019; Gerke et al., 2020).

This review provides a comprehensive examination of LLM applications in drug discovery and precision medicine, synthesizing findings from the rapidly evolving literature published between 2018 and 2025. We analyze the technical foundations underlying these approaches, including transformer architectures, pretraining strategies, multi-modal integration methods, and knowledge enhancement techniques. We survey key applications across the drug development pipeline, from early stage target identification through clinical trial design, as well as precision medicine applications spanning clinical decision support, patient stratification, and genomic interpretation. We critically assess current limitations, including challenges related to data quality and availability, model interpretability, hallucination risks, data privacy concerns, computational resource requirements, and ethical considerations. Finally, we discuss future directions that hold promise for advancing the field toward more reliable, transparent, and clinically applicable systems. Our goal is to provide researchers, clinicians, and industry practitioners with a thorough understanding of how LLMs are reshaping biomedical research and clinical practice, while highlighting both the tremendous opportunities and the substantial challenges that must be addressed to realize their full potential for improving human health. The remainder of this paper is organized as follows. Section 2 provides background on the evolution of LLMs and their adaptation to biomedical contexts, covering foundational architectures, pretraining strategies, and key design considerations for biological applications. Section 3 examines applications in drug discovery, spanning target identification, molecular generation, property prediction, drug repurposing, and integrated discovery pipelines. Section 4 discusses precision medicine applications, including clinical decision support, patient stratification, treatment personalization, and genomic interpretation. Section 5 presents technical methodologies and architectural innovations that enable these applications. Section 6 surveys current applications and case studies demonstrating real world impact. Section 7 addresses challenges, limitations, and risks associated with LLM deployment in biomedicine. Section 8 explores future directions and emerging opportunities. Finally, Section 9 concludes with a synthesis of key insights and recommendations for advancing the field.

2 Background and theoretical foundations

2.1 Evolution of large language models

Large language models represent the culmination of decades of progress in natural language processing and deep learning, building upon foundational advances in neural network architectures, optimization algorithms, and distributed computing infrastructure (Vaswani et al., 2017). The transformer architecture, introduced by Vaswani and colleagues in 2017, provided the critical innovation enabling modern LLMs through its self-attention mechanism, which allows models to capture long-range dependencies in sequential data without the vanishing gradient problems that plagued earlier recurrent architectures (Vaswani et al., 2017). The self-attention mechanism computes a weighted sum of all positions in a sequence, where the weights are determined by the relevance of each position to the current position being processed, enabling the model to directly attend to any part of the input regardless of distance (Hays, 2026a). This architectural innovation, combined with positional encodings to preserve sequence order information, has proven remarkably effective for modeling both natural language and biological sequences, establishing the foundation for essentially all modern large language models (Devlin et al., 2019; Brown et al., 2020).

The evolution of LLMs can be characterized by two complementary paradigms that have shaped the field: bidirectional encoder models exemplified by BERT and autoregressive decoder models exemplified by GPT (Devlin et al., 2019; Radford et al., 2019). BERT, introduced by Devlin and colleagues in 2018, employs masked language modeling as its pretraining objective, where the model learns to predict randomly masked tokens in a sequence by attending to context on both sides (Devlin et al., 2019). This bidirectional approach produces rich contextual representations that have proven particularly effective for understanding and classification tasks, where the full context is available at inference time. In contrast, the GPT family of models employs autoregressive language modeling, predicting the next token given all previous tokens, which naturally supports open-ended generation tasks but processes context in only one direction (Brown et al., 2020; Radford et al., 2019).

GPT-3, with its 175 billion parameters, demonstrated the remarkable emergent capabilities that arise from scaling model size and training data, including the ability to perform novel tasks through in-context learning without parameter updates (Brown et al., 2020). GPT-4 further extended these capabilities with improved reasoning, reduced hallucination, and multi-modal understanding incorporating both text and images (OpenAI, 2023). The PaLM family from Google demonstrated that further scaling to 540 billion parameters yields additional improvements in reasoning and multilingual capabilities (Chowdhery et al., 2023), while open-source models such as LLaMA and LLaMA-2 have democratized access to capable foundation models that can be fine-tuned for specific applications (Touvron et al., 2023a; Touvron et al., 2023b).

A particularly important development in recent years has been the emergence of instruction tuning and reinforcement learning from human feedback (RLHF) as techniques for aligning LLM behavior with human intentions and values. Rather than simply predicting the next token in training text, instruction-tuned models are trained to follow natural language instructions describing desired tasks, substantially improving their ability to generalize to novel tasks at inference time (Wei et al., 2022). RLHF further refines model outputs by training a reward model on human preferences and then optimizing the language model to produce outputs that maximize predicted reward. These alignment techniques have proven essential for creating models that are helpful, harmless, and honest in their interactions with users, considerations that are particularly important for high-stakes biomedical applications where misleading or harmful outputs could have serious consequences. The combination of massive pretraining, instruction tuning, and human feedback alignment has produced models with unprecedented capabilities for understanding complex queries, following multi-step instructions, and engaging in nuanced reasoning across diverse domains including medicine and biology.

Figure 2 organises the models surveyed here as an evolutionary tree, tracing how domain-adapted biomedical, molecular, protein, genomic, single-cell, and clinical models descend from a small set of general-purpose transformer foundation models along three architectural branches (encoder-only, encoder-decoder, and decoder-only). This lineage view emphasises which base model each framework is derived from, and complements the quantitative timeline in Figure 3, which represents model scale and the category.

2.2 Biomedical language models

Biomedical language models represent specialized LLMs trained on domain-specific data to capture the unique vocabulary, syntactic patterns, and semantic relationships that characterize biomedical domains (Figure 3) (Lee et al., 2020; Gu et al., 2021). The biomedical literature contains highly specialized terminology, acronyms, and naming conventions that differ substantially from general English text, and biomedical reasoning often requires understanding complex causal and mechanistic relationships between biological entities (Peng et al., 2019). Early work demonstrated that general-purpose language models pretrained on web text and books underperformed specialized models on biomedical natural language processing tasks, motivating the development of domain-adapted alternatives (Lee et al., 2020; Alsentzer et al., 2019). BioBERT, introduced by Lee and colleagues in 2020, pioneered the approach of continued pretraining of BERT on biomedical literature, using PubMed abstracts and PubMed Central full-text articles to adapt the model’s representations to biomedical language (Lee et al., 2020). This relatively simple approach of domain-adaptive pretraining yielded substantial improvements on biomedical benchmarks including named entity recognition, relation extraction, and question answering. PubMedBERT, developed by Gu and colleagues, demonstrated that pretraining from scratch on an in-domain corpus, rather than continued pretraining from a general-domain checkpoint, could yield additional improvements by avoiding the domain shift between general and biomedical vocabularies (Gu et al., 2021). ClinicalBERT adapted the approach specifically for clinical text, pretraining on clinical notes from the MIMIC-III database to capture the distinctive characteristics of clinical documentation including abbreviations, templated language, and references to specific clinical workflows (Alsentzer et al., 2019). These foundational models established biomedical NLP as a distinct subfield with its own benchmarks, datasets, and evaluation protocols.

More recent developments have extended biomedical language models to generative tasks and larger scales. BioGPT, developed by Luo and colleagues, applied the GPT architecture and autoregressive pretraining to biomedical literature, demonstrating state-of-the-art performance on biomedical text generation and question answering tasks (Luo et al., 2022). Med-PaLM and Med-PaLM 2 from Google demonstrated that appropriately prompted and aligned large language models can achieve benchmark-level performance on medical licensing examinations and clinical reasoning benchmarks (Singhal et al., 2023). MEDITRON-70B represents an open-source effort to create a large medical language model by continued pretraining of LLaMA-2 on carefully curated medical literature and guidelines (Chen et al., 2023). These generative models offer capabilities beyond the classification and extraction tasks that characterized earlier work, enabling interactive question answering, clinical decision support, and synthesis of complex medical information.

Biomedical language models are typically pretrained on diverse data sources that provide complementary perspectives on biomedical knowledge:

• Scientific literature: PubMed contains over 35 million citations to biomedical literature, with abstracts providing concise summaries of research findings, while PubMed Central provides full-text access to millions of articles enabling richer contextual learning
• Clinical documentation: Electronic health records capture the longitudinal medical history of patients, including clinical notes, laboratory results, imaging reports, and medication records that reflect real-world clinical practice
• Biomedical databases: Structured resources such as DrugBank, ChEMBL, UniProt, and the Gene Ontology provide curated information about drugs, proteins, genes, and their relationships (Wishart et al., 2018; Mendez et al., 2019)
• Clinical guidelines: Evidence-based practice guidelines from organizations such as the National Comprehensive Cancer Network and the American College of Cardiology codify expert consensus on clinical decision making
• Regulatory documents: FDA drug labels, clinical trial protocols, and adverse event reports provide authoritative information on drug safety and efficacy

The choice of pretraining data significantly impacts model capabilities and biases. Models trained primarily on research literature may excel at answering scientific questions but struggle with practical clinical scenarios, while models trained on clinical notes may capture the nuances of clinical practice but lack access to cutting-edge research findings. Increasingly, developers are exploring strategies that combine multiple data sources to create models with broad biomedical competence.

2.3 Molecular and protein language models

A particularly exciting development in biomedical AI has been the recognition that molecular structures and biological sequences can be treated as specialized languages, enabling the application of transformer architectures and language modeling techniques to chemical and biological data (Ross et al., 2022; Rives et al., 2021). This insight has given rise to a family of molecular and protein language models that learn meaningful representations from sequence data without requiring labeled examples, analogous to how text language models learn from unlabeled corpora. For small molecules, the simplified molecular input line entry system (SMILES) notation provides a text-based representation of molecular structures that can be processed by language models (Segler et al., 2018). ChemBERTa and its successor ChemBERTa-2 apply BERT-style masked language modeling to SMILES strings, learning representations that capture chemical substructure patterns and molecular properties (Chithrananda et al., 2020; Ahmad et al., 2022). MolGPT applies autoregressive language modeling to molecular generation, learning to produce valid SMILES strings with desired properties (Bagal et al., 2022). Chemformer uses a sequence-to-sequence transformer architecture for tasks including reaction prediction and retrosynthesis (Irwin et al., 2022). More recent work has explored larger-scale pretraining, demonstrating that neural scaling laws observed in natural language also apply to chemical language models, with larger models exhibiting improved performance on property prediction and generation tasks (Ross et al., 2022; Frey et al., 2023).

Protein language models have achieved remarkable success by treating amino acid sequences as natural language and applying similar pretraining strategies. ESM-1b and ESM-2, developed by Meta AI, demonstrated that transformer models pretrained on millions of protein sequences through masked language modeling learn representations that capture evolutionary relationships, protein structure, and functional properties without any explicit structural supervision (Lin et al., 2023; Rives et al., 2021). The attention patterns learned by these models correlate with protein contact maps, suggesting that the models implicitly learn three-dimensional structural information from sequence data alone (Rives et al., 2021). ESMFold leverages these learned representations to predict protein structures with accuracy approaching AlphaFold2 for many proteins, while requiring orders of magnitude less computation at inference time (Lin et al., 2023). ProtTrans applied multiple transformer architectures to protein sequences, demonstrating that different architectures capture complementary aspects of protein biology (Elnaggar et al., 2021). ProtGPT2 uses autoregressive language modeling to generate novel protein sequences that fold into stable structures and exhibit desired functional properties (Ferruz et al., 2022). ProGen demonstrated that protein language models can generate functional enzymes across diverse protein families, with some generated sequences experimentally validated to have catalytic activity (Madani et al., 2023). Genomic language models extend similar approaches to DNA and RNA sequences, capturing regulatory patterns and sequence-function relationships. DNABERT applies BERT-style pretraining to DNA sequences, learning representations that improve performance on tasks including promoter prediction, splice site identification, and transcription factor binding site recognition (Ji et al., 2021). DNABERT-2 improves upon this approach with more efficient tokenization and larger-scale pretraining (Zhou et al., 2023). The Nucleotide Transformer represents a particularly large-scale effort, pretraining on genomes from thousands of species to learn representations that capture cross-species evolutionary patterns (Dalla-Torre et al., 2023). HyenaDNA introduces efficient long-range sequence modeling through the Hyena architecture, enabling modeling of sequences up to 1 million nucleotides at single-nucleotide resolution (Nguyen et al., 2024). These genomic language models are enabling new approaches to variant effect prediction, gene expression modeling, and regulatory sequence design.

2.4 Multi-modal integration

A critical advancement in biomedical LLMs is the integration of multiple data modalities within unified architectures (Cao et al., 2023; Liu S. et al., 2023). Drug discovery and precision medicine inherently involve diverse data types including molecular structures, protein sequences, genomic variants, clinical phenotypes, medical images, and natural language descriptions that must be jointly analyzed for comprehensive understanding. Multi-modal LLMs address this challenge by learning aligned representations across modalities, enabling reasoning that spans different data types. Multi-modal architectures typically employ separate encoders for different data types, with alignment mechanisms that map representations into a shared semantic space where meaningful cross-modal operations become possible (Liu S. et al., 2023). For example, a model might use a graph neural network to encode molecular structures, a transformer to encode protein sequences, and another transformer to encode natural language, with contrastive learning or cross-attention mechanisms to align these representations. The aligned representation space enables capabilities such as text-based molecular retrieval (finding molecules matching a natural language description), molecule captioning (generating natural language descriptions of molecular properties), and cross-modal prediction (predicting molecular properties from text descriptions). InstructMol represents a pioneering approach to multi-modal molecular understanding, aligning molecular structures with natural language through a two-stage instruction-tuning approach (Cao et al., 2023). In the first stage, separate encoders for molecules and text are pretrained on their respective data types. In the second stage, the combined system is fine-tuned on instruction-following tasks that require reasoning across both modalities, such as answering questions about molecules or generating molecules matching text descriptions. This approach enables the model to leverage both the structural information encoded in molecular representations and the semantic knowledge captured in language model representations, bridging the gap between chemical and natural language understanding.

Text-molecule translation represents another approach to multi-modal integration, treating the problem of connecting molecules and language as a machine translation task (Edwards et al., 2022a). Models trained for this task can translate between SMILES representations and natural language descriptions of molecular properties, mechanism of action, or therapeutic indications. MolT5 and related approaches use sequence-to-sequence transformer architectures to perform bidirectional translation between molecules and text, enabling both molecule captioning and text-conditioned molecule generation (Edwards et al., 2022a). More recent work has explored multi-modal molecule-text models that learn to jointly embed molecules and text in a shared space, enabling text-based retrieval and editing of molecules (Liu S. et al., 2023). ChatMol and similar systems provide interactive interfaces for molecular discovery, allowing users to explore chemical space through natural language conversation (Zeng et al., 2024). Beyond molecules, multi-modal integration in biomedicine increasingly encompasses images, structured data, and complex biological entities. Galactica represents an ambitious effort to create a language model capable of scientific reasoning across modalities including text, mathematical equations, chemical formulas, and protein sequences (Taylor et al., 2022). BiomedGPT aims for a unified generalist model capable of handling vision, language, and multi-modal biomedical tasks (Zhang et al., 2024). These efforts toward general-purpose biomedical AI systems suggest a future where a single model can seamlessly integrate information from literature, molecular databases, clinical records, and medical images to support complex biomedical reasoning.

2.5 Knowledge-guided approaches

Pure data-driven approaches, while powerful, may struggle with the inherent complexity and sparsity of biomedical data, particularly for rare diseases, novel therapeutic targets, or patient populations underrepresented in training data (Mohamed et al., 2021; Bonner et al., 2022). Knowledge-guided LLMs address these limitations by explicitly incorporating structured domain knowledge, including biomedical ontologies, knowledge graphs, and curated databases, to constrain and enhance model reasoning. Biomedical knowledge graphs provide structured representations of entities such as drugs, diseases, genes, proteins, and pathways along with their relationships (Mohamed et al., 2021; Bonner et al., 2022). Resources such as DrugBank catalog drug-target interactions and pharmacological properties (Wishart et al., 2018), ChEMBL aggregates bioactivity data from medicinal chemistry literature (Mendez et al., 2019), and disease ontologies like the Disease Ontology and Human Phenotype Ontology provide standardized terminologies for describing diseases and phenotypes. These knowledge resources encode decades of biomedical research in machine-readable form, providing a foundation of reliable information that can ground and constrain LLM reasoning. Integration of knowledge graphs with language models can take several forms. Graph neural networks can be used to encode knowledge graph structure, with the resulting representations incorporated into language model architectures through cross-attention or concatenation (Mohamed et al., 2021). Alternatively, knowledge graph information can be serialized into text and provided as context to the language model, leveraging the model’s text understanding capabilities to reason about structured relationships. Retrieval-augmented approaches retrieve relevant knowledge graph triples based on the input query and include them in the model context, combining the flexibility of language models with the reliability of curated knowledge sources.

Retrieval-augmented generation (RAG) represents a particularly promising approach for knowledge enhancement of LLMs (Lewis et al., 2020; Zakka et al., 2024). Rather than relying solely on knowledge encoded in model parameters during pretraining, RAG systems retrieve relevant documents or knowledge base entries at inference time and include them in the model’s context. This approach offers several advantages for biomedical applications: it enables access to current information beyond the model’s training cutoff, reduces hallucination by grounding outputs in retrieved evidence, provides transparency through citation of sources, and can be updated simply by changing the retrieval corpus without model retraining (Hays and Richardson, 2026a). Almanac demonstrated the effectiveness of RAG for clinical medicine, using retrieval over medical guidelines and literature to ground clinical recommendations in authoritative sources (Zakka et al., 2024). Benchmarking studies have confirmed that RAG substantially improves accuracy on biomedical question answering tasks while reducing factual errors (Xiong et al., 2024).

2.6 Zero-shot and few-shot learning

A distinguishing capability of modern LLMs is their ability to perform zero-shot and few-shot learning, making accurate predictions on novel tasks with minimal or no task-specific training examples (Brown et al., 2020; Kojima et al., 2022). This capability emerges from large-scale pretraining, which equips models with broad world knowledge and general-purpose reasoning abilities that can be applied to new situations through appropriate prompting. For biomedical applications, this capability is particularly valuable given the many scenarios where labeled data is scarce or expensive to obtain. Zero-shot learning enables LLMs to perform tasks they have never explicitly been trained on, guided only by natural language descriptions of the task (Kojima et al., 2022). For example, a model might be prompted to classify whether a clinical note indicates a diagnosis of diabetes, without ever having been trained on diabetes classification specifically. The model draws on its general understanding of medical terminology and the relationship between symptoms, tests, and diagnoses to perform the task. Studies have demonstrated that large language models can achieve impressive zero-shot performance on medical licensing examinations, clinical reasoning benchmarks, and biomedical question answering tasks (Singhal et al., 2023; Nori et al., 2023; Kung et al., 2023). Few-shot learning extends this capability by providing a small number of examples in the prompt that demonstrate the desired input-output mapping (Brown et al., 2020). These examples guide the model toward producing outputs in the desired format and help disambiguate task requirements. Few-shot prompting has proven particularly effective for structured prediction tasks where the output format might be ambiguous from the task description alone. In biomedical contexts, few-shot learning enables rapid adaptation to new prediction tasks without the need for large labeled datasets or model fine-tuning. Chain-of-thought prompting represents an important advancement that improves LLM reasoning on complex tasks by encouraging the model to produce intermediate reasoning steps before reaching a final answer (Wei et al., 2022). Rather than producing an answer directly, the model is prompted (through instructions or examples) to show its work, breaking down complex problems into smaller steps that can be verified. This approach has been shown to substantially improve performance on tasks requiring multi-step reasoning, mathematical calculation, or logical inference. For biomedical applications, chain-of-thought prompting can improve performance on clinical reasoning tasks and provide transparency into the model’s decision-making process, though the generated reasoning chains do not always faithfully reflect the model’s internal computations.

3 Large language models in drug discovery

Drug discovery is a complex, multistage process encompassing target identification, hit discovery, lead optimization, preclinical testing, and clinical trials, typically requiring over a decade of effort and billions of dollars in investment before a new therapeutic reaches patients (DiMasi et al., 2016; Wouters et al., 2020). The high attrition rate at each stage, with approximately 90% of candidates failing in clinical trials, underscores the need for improved computational methods that can more accurately predict drug efficacy, safety, and pharmacokinetic properties early in the discovery process (Vamathevan et al., 2019; Paul et al., 2021). Large language models are being applied across this entire pipeline, offering the potential to accelerate timelines, reduce costs, and improve success rates by leveraging patterns learned from vast amounts of biomedical data (see Figures 4–8 for a quantitative overview of architectures, application domains, benchmarks, openness, and reported performance across the 102-model corpus considered in this review). This section examines key applications of LLMs in drug discovery, from early-stage target identification through lead optimization and beyond.

3.1 Target identification and validation

Target identification, the process of determining which biological molecules should be modulated to treat a disease, represents a critical early step in drug discovery that fundamentally shapes the trajectory of a drug development program (Vamathevan et al., 2019; Santos et al., 2017). The ideal drug target is causally involved in disease pathophysiology, druggable with small molecules or biologics, and amenable to selective modulation without unacceptable side effects. Traditional approaches to target identification rely on genetic studies, differential expression analysis, and pathway databases, but the complexity of biological systems means that many potential targets remain undiscovered or poorly characterized.

LLMs facilitate target identification by analyzing vast amounts of biomedical literature, genomic data, and disease-gene associations to uncover novel target-disease linkages that may not be apparent from any single data source (Vamathevan et al., 2019; Stokes et al., 2020). Language models can synthesize information from thousands of publications to identify genes and proteins implicated in disease processes, even when these connections are only indirectly mentioned or must be inferred from multiple studies. By learning semantic relationships between biological entities from scientific text, these models can predict novel target-disease associations that extend beyond explicitly documented relationships. For example, a language model trained on biomedical literature might infer that a gene involved in a particular signaling pathway could be relevant to a disease affecting that pathway, even if no direct experimental link has been established. GeneGPT demonstrated the potential for augmenting language models with access to biomedical databases and tools, enabling more grounded and accurate reasoning about gene functions and disease associations (Jin et al., 2023).

Drug-target interaction (DTI) prediction represents a more focused application where the goal is to predict whether a given compound will bind to a specific target protein (Öztürk et al., 2018; Nguyen et al., 2021; Huang et al., 2021). Accurate DTI prediction is valuable both for identifying new uses for existing drugs and for prioritizing compounds in virtual screening campaigns. Traditional approaches rely on molecular docking simulations or ligand-based similarity methods, but these can be computationally expensive or limited by the availability of known ligands. Transformer-based approaches to DTI prediction have achieved state-of-the-art results by learning to embed both drug molecules and target proteins in a shared representation space (Huang et al., 2021; Chen et al., 2020). MolTrans uses a knowledge-inspired transformer to capture interaction patterns between molecular substructures and protein subsequences, outperforming previous methods on benchmark datasets including DAVIS and KIBA (Huang et al., 2021). TransformerCPI employs self-attention mechanisms to learn contextual representations of both compounds and proteins, achieving improved prediction accuracy on compound-protein interaction datasets (Chen et al., 2020). More recent work has combined transformer-based representations with graph neural networks and topological features, capturing complementary aspects of molecular structure to further improve prediction accuracy (Nguyen et al., 2021; Yazdani-Jahromi et al., 2022).

3.2 Molecular generation and design

De novo drug design, the generation of novel molecular structures with desired properties, represents one of the most exciting and potentially broad applications of LLMs in drug discovery (Bilodeau et al., 2022; Segler et al., 2018; Gómez-Bombarelli et al., 2018). Traditional medicinal chemistry relies heavily on intuition, iterative synthesis, and incremental modification of known active compounds. Generative models offer the possibility of exploring vast regions of chemical space more efficiently, potentially identifying structurally novel compounds with superior properties. By treating molecular representations as specialized languages, LLMs can learn the grammar and semantics of chemical space and generate novel, synthesizable molecules that satisfy multiple property constraints simultaneously. The SMILES representation provides a natural entry point for applying language models to molecular generation, as it encodes molecular graphs as linear strings amenable to sequence modeling (Segler et al., 2018; Bagal et al., 2022). Early work demonstrated that recurrent neural networks trained on SMILES strings could generate focused molecular libraries enriched for desired properties, with generated molecules often exhibiting structural novelty relative to training data while maintaining drug-like characteristics (Segler et al., 2018).

MolGPT applied the transformer decoder architecture to SMILES generation, demonstrating improved sample quality and property control compared to earlier recurrent approaches (Bagal et al., 2022). The model learns conditional generation, producing molecules that satisfy specified property constraints while maintaining validity and synthesizability. Variational autoencoders and related latent space approaches provide an alternative paradigm for molecular generation that enables smooth interpolation between molecules and property-directed optimization in a continuous latent space (Gómez-Bombarelli et al., 2018; Jin et al., 2018). These methods encode molecules into a low-dimensional latent representation where nearby points correspond to chemically similar molecules, then decode latent vectors back to molecular structures. Optimization in this continuous space can identify molecules with improved properties more efficiently than discrete search over molecular graphs. Junction Tree VAE addressed a key limitation of earlier VAE approaches by generating molecules as assemblies of valid substructures rather than atom-by-atom, substantially improving the validity of generated molecules (Jin et al., 2018).

More recent work has explored combining language models with other molecular representations and optimization techniques to achieve improved generation performance. DrugEx combines transformer-based generation with reinforcement learning to navigate chemical space toward molecules with desired properties, using multi-objective reward functions to balance potency, selectivity, and drug-likeness (Liu X. et al., 2023). PaccMann, an attention-based molecular generation framework that couples transcriptomic context with chemical language modelling, illustrates how such architectures can be directed toward kinase-inhibitor design tasks (Born et al., 2021), and attention-based molecular generators more broadly have been applied to EGFR inhibitor design, including efforts aimed at the 90M resistance mutation that limits the efficacy of first-generation tyrosine kinase inhibitors in non-small cell lung cancer. Similarly, transformer-based generative models have been applied to design selective JAK2 inhibitors for myeloproliferative neoplasms, producing novel scaffolds distinct from approved JAK inhibitors such as ruxolitinib while maintaining favorable pharmacokinetic properties (Born and Manica, 2023). Diffusion models have emerged as a powerful alternative paradigm for molecular generation, with approaches like DiffDock achieving state-of-the-art results for structure-based drug design by directly generating binding poses through learned diffusion processes (Corso et al., 2023; Hoogeboom et al., 2022; Xu et al., 2022). These methods can incorporate 3D structural constraints that are difficult to encode in sequence-based representations, enabling structure-based generation that accounts for the target binding site geometry.

The integration of natural language with molecular generation represents a particularly exciting recent development that promises to make generative models more accessible and controllable. By training models to generate molecules from natural language descriptions of desired properties, researchers can specify design criteria in human-readable terms rather than numerical property targets. Text-to-molecule models can translate descriptions like “a selective inhibitor of kinase X with good oral bioavailability” into molecular candidates, leveraging the semantic understanding encoded in language models (Liu S. et al., 2023; Edwards et al., 2022a). This capability could democratize access to computational drug design by reducing the barrier to entry for researchers without extensive computational chemistry expertise.

3.3 Property prediction and optimization

Predicting molecular properties, including absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles, is essential for prioritizing drug candidates and avoiding late-stage failures that account for substantial attrition in clinical development (Wu et al., 2018; Yang et al., 2019; Mayr et al., 2016). A molecule may show excellent target binding affinity in biochemical assays but fail as a drug due to poor solubility, rapid metabolism, or unexpected toxicity. The ability to predict these properties computationally from molecular structure alone can guide lead optimization and reduce the number of compounds requiring expensive experimental characterization. MoleculeNet established standardized benchmarks for evaluating molecular property prediction methods across diverse tasks including physical chemistry, biophysics, and toxicity prediction (Wu et al., 2018). This benchmarking effort revealed that while traditional fingerprint-based methods remained competitive on some tasks, graph neural networks and transformer-based approaches achieved state-of-the-art results on many prediction challenges (see Figure 6 for a corpus-wide view of which benchmarks, including the MoleculeNet sub-tasks discussed below, are most frequently used). Subsequent work has continued to improve performance through architectural innovations, pretraining strategies, and integration of multiple data modalities (Yang et al., 2019; Xiong et al., 2020).

Transformer-based molecular property prediction leverages the self-attention mechanism to learn representations that capture relevant substructure patterns for predicting specific properties (Chithrananda et al., 2020; Ross et al., 2022). ChemBERTa demonstrated that BERT-style pretraining on large SMILES corpora improves downstream property prediction by learning transferable representations of molecular structure (Chithrananda et al., 2020). Larger-scale pretraining efforts have shown that neural scaling laws apply to chemical language models, with larger models and more pretraining data consistently improving property prediction accuracy (Ross et al., 2022; Frey et al., 2023). These findings suggest that continued scaling of molecular language models may yield further improvements in predictive performance. Toxicity prediction represents a particularly important application of molecular property prediction, as unexpected toxicity accounts for a substantial fraction of drug candidate failures in clinical trials (Mayr et al., 2016). DeepTox established deep learning approaches for toxicity prediction, demonstrating that neural networks could outperform traditional quantitative structure-activity relationship (QSAR) methods on standardized toxicity prediction challenges (Mayr et al., 2016). More recent work has applied transformer architectures and language model pretraining to toxicity prediction, achieving improved accuracy particularly for rare toxicity endpoints where training data is limited. The ability of pretrained models to transfer knowledge from related tasks helps address data scarcity challenges that plague toxicity prediction for specific endpoints.

Lead optimization, the iterative process of modifying lead compounds to improve their property profiles, can be framed as a multi-objective optimization problem where the goal is to simultaneously optimize potency, selectivity, ADMET properties, and synthetic accessibility (Schneider et al., 2020; Bilodeau et al., 2022). LLMs enable optimization approaches that propose molecular modifications in natural language terms, generate molecules matching specified property profiles, or navigate property trade-offs through learned representations. By capturing the complex relationships between molecular structure and multiple properties, these models can suggest modifications more efficiently than exhaustive enumeration of chemical modifications.

3.4 Drug repurposing

Drug repurposing, the identification of new therapeutic uses for existing approved drugs, offers a faster and less expensive path to new treatments compared to de novo drug discovery (Pushpakom et al., 2019). Approved drugs have established safety profiles, manufacturing processes, and regulatory histories, substantially reducing the time and cost required to reach patients for a new indication. The COVID-19 pandemic highlighted the potential value of drug repurposing, as existing drugs were rapidly evaluated for potential activity against the novel coronavirus (Zhou et al., 2020; Zeng et al., 2020; Gysi et al., 2021). LLMs facilitate drug repurposing by analyzing diverse data sources to identify unexpected connections between existing drugs and new indications. By synthesizing information from scientific literature, clinical records, molecular databases, and knowledge graphs, these models can identify drugs whose mechanisms of action might be relevant to diseases beyond their original indications (Pushpakom et al., 2019; Zhou et al., 2020). For example, a language model might identify that a drug approved for one inflammatory condition shares mechanistic features with another inflammatory disease, suggesting potential repurposing even if no direct clinical evidence exists.

Network medicine approaches to drug repurposing leverage knowledge graphs representing relationships between drugs, targets, diseases, and biological pathways to identify repurposing opportunities through network proximity analysis (Gysi et al., 2021; Zitnik et al., 2018). Drugs whose targets are close in the interactome to disease-associated genes may have therapeutic potential for those diseases. LLMs can enhance these approaches by extracting relationships from unstructured text to populate knowledge graphs, resolving entity ambiguities, and integrating information from multiple sources. The combination of structured network analysis with language model capabilities for processing unstructured text enables more comprehensive identification of repurposing candidates. Adverse event databases and electronic health records provide another source of repurposing insights, as drugs sometimes produce unexpected beneficial effects in patient populations (Tatonetti et al., 2012). Data-driven analysis of side effect profiles has revealed unexpected drug-disease connections that have been validated in subsequent clinical studies. LLMs can process unstructured clinical notes to identify potential therapeutic effects that may not be captured in structured fields, expanding the scope of data-driven repurposing discovery. Notable successes in AI-driven drug repurposing include the identification of baricitinib, a JAK1/JAK2 inhibitor originally approved for rheumatoid arthritis, as a potential COVID-19 therapeutic through knowledge graph analysis of host-pathogen interactions (Richardson et al., 2020). This prediction was subsequently validated in clinical trials, leading to emergency use authorization. Similarly, network-based approaches identified sildenafil as a candidate for Alzheimer’s disease prevention, with a retrospective observational analysis of insurance claims data reporting an approximately 69% lower incidence of Alzheimer’s disease among individuals prescribed the drug (Fang et al., 2021); this association arises from non-randomized real-world data and is subject to residual confounding, and prospective randomized validation is still needed to establish a causal protective effect.

3.5 Multi-agent systems for integrated drug discovery

Recent advances have moved beyond single-task models toward multi-agent systems in which specialized LLM-based agents handle different stages of discovery (literature review, hypothesis generation, molecular design, property prediction, and experiment planning), coordinated by orchestration mechanisms that manage information flow across agents (Park et al., 2023; Wang et al., 2024). Tool-augmented agents such as ChemCrow (Bran et al., 2024), autonomous chemical research platforms that close the loop with robotic synthesis (Boiko et al., 2023), and industrial end-to-end campaigns exemplified by Insilico Medicine’s AI-driven kinase inhibitor programs (Zhavoronkov et al., 2019; Xu et al., 2025) illustrate how these ideas scale from individual tasks to integrated pipelines. A detailed treatment of these systems, together with representative case studies and their reported outcomes, is provided in Section 6.1.

3.6 Benchmarks and evaluation

Rigorous evaluation is essential for assessing progress in LLM-based drug discovery and identifying areas requiring further development. The field has developed standardized benchmarks covering diverse tasks including property prediction, molecular generation, and drug-target interaction prediction (Wu et al., 2018; Polykovskiy et al., 2020). MoleculeNet provides comprehensive benchmarks for molecular property prediction spanning physical chemistry, biophysics, physiology, and toxicology (Wu et al., 2018). The benchmark includes standardized data splits and evaluation protocols that enable fair comparison across methods. Performance on MoleculeNet tasks has improved substantially with the development of graph neural networks, transformers, and pretrained molecular representations.

The MOSES benchmarks evaluate molecular generation methods on metrics including validity (fraction of generated molecules that are chemically valid), uniqueness (diversity of generated molecules), novelty (fraction not present in training data), and similarity to reference compounds (Polykovskiy et al., 2020). These metrics capture different aspects of generation quality that are important for drug discovery applications. Distribution learning metrics assess whether generated molecules follow the same statistical patterns as training molecules, while goal-directed metrics evaluate the ability to generate molecules with specific desired properties. Evaluation of DTI prediction methods typically employs benchmark datasets such as DAVIS, KIBA, and BindingDB, with performance measured by area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), or mean squared error for binding affinity prediction (Öztürk et al., 2018; Huang et al., 2021). Cold-split evaluation protocols, where test sets contain compounds or targets not seen during training, provide more realistic assessments of model generalization to novel chemical or biological entities (Yazdani-Jahromi et al., 2022). Performance in these challenging settings is particularly relevant for drug discovery applications where the goal is often to identify novel compounds acting on previously undrugged targets.

4 Large language models in precision medicine

Precision medicine aims to tailor medical treatment to individual patient characteristics, including genetic makeup, molecular profiles, lifestyle factors, and environmental exposures, moving beyond one-size-fits-all approaches toward therapies optimized for specific patient subgroups or individuals (Esteva et al., 2019; Rajkomar et al., 2018). Realizing this vision requires integrating diverse data types across multiple scales of biological organization, from genomic variants and transcriptomic profiles to clinical phenotypes and treatment histories. The inherent complexity of this integration challenge, combined with the unstructured nature of much clinical data, makes precision medicine a natural application domain for large language models. LLMs are transforming precision medicine by enabling more sophisticated analysis of patient data, more accurate prediction of treatment responses, and more personalized clinical decision support systems that can synthesize information from multiple sources to guide therapeutic choices.

4.1 Clinical decision support

Clinical decision support systems (CDSS) assist healthcare providers in making evidence-based decisions by analyzing patient data in the context of medical knowledge, clinical guidelines, and population-level evidence (Esteva et al., 2019; Rajkomar et al., 2018; Shickel et al., 2018). Traditional CDSS rely on rule-based systems that encode explicit clinical logic, but these approaches struggle to incorporate the nuanced, context-dependent reasoning that characterizes expert clinical judgment. LLMs enhance CDSS capabilities by processing unstructured clinical narratives that contain rich information often absent from structured data fields, integrating diverse data sources including notes, laboratory results, imaging reports, and external knowledge bases, and providing contextualized recommendations that account for patient-specific factors. The ability of LLMs to process unstructured clinical text represents a particularly important capability for clinical decision support. Electronic health records contain extensive narrative documentation, including admission notes, progress notes, discharge summaries, and consultation reports, that capture clinical reasoning, contextual details, and subtle findings that may not be coded in structured fields (Shickel et al., 2018; Si et al., 2021). Studies have demonstrated that LLMs can extract clinically relevant information from these narratives, including diagnoses, medications, procedures, and clinical status, with accuracy approaching or exceeding that of human reviewers (Patel and Lam, 2023). This capability enables decision support systems that consider the full clinical picture rather than relying solely on structured data elements.

Med-PaLM and its successor Med-PaLM 2 demonstrated that large language models can achieve expert-level performance on medical question answering and clinical reasoning tasks when appropriately adapted to the medical domain (Singhal et al., 2023). On the MedQA benchmark, which contains questions from medical licensing examinations, Med-PaLM 2 achieved accuracy exceeding that of average physicians, demonstrating the potential for LLMs to support complex medical reasoning. Importantly, the model’s responses were evaluated not only for accuracy but also for factors critical to clinical utility, including the quality of explanations, potential for harm, and alignment with medical consensus. These evaluations revealed both strengths and limitations, with the model occasionally producing plausible-sounding but incorrect information, highlighting the importance of human oversight in clinical applications. Retrieval-augmented approaches to clinical decision support address concerns about hallucination and currency of medical knowledge by grounding LLM outputs in retrieved evidence from authoritative sources (Zakka et al., 2024; Xiong et al., 2024). Almanac demonstrated this approach for clinical medicine, retrieving relevant passages from medical guidelines, textbooks, and literature to support clinical question answering (Zakka et al., 2024). By explicitly citing sources and constraining outputs to retrieved evidence, these systems provide transparency that enables clinicians to verify recommendations and maintain appropriate skepticism toward AI-generated content. Benchmarking studies have confirmed that retrieval augmentation substantially improves factual accuracy while reducing hallucinations on biomedical question answering tasks (Xiong et al., 2024). In cardiology, LLM-based systems have been deployed to interpret electrocardiograms and suggest diagnoses for arrhythmias, achieving performance comparable to cardiologists in detecting atrial fibrillation, left ventricular dysfunction, and hypertrophic cardiomyopathy (Attia et al., 2019; Noseworthy et al., 2020). Complementary computational approaches, such as mechanistic simulations of cardiac extracellular matrix remodeling, provide biophysical foundations that can inform and validate AI-driven predictions in cardiovascular applications (Hays and Richardson, 2026b). For diabetes management, AI-driven decision support systems analyzing continuous glucose monitoring data, insulin dosing history, and dietary patterns have demonstrated improved glycemic control in type 1 diabetes patients compared to standard care (Tyler et al., 2020).

4.2 Patient stratification and treatment personalization

Patient stratification, the process of grouping patients based on characteristics that predict treatment response or disease prognosis, is fundamental to precision medicine and guides both clinical trial design and therapeutic decision making (Esteva et al., 2019). Traditional stratification approaches rely on individual biomarkers or clinical characteristics, but the multifactorial nature of treatment response often requires integration of multiple data types to achieve optimal predictive performance. LLMs enable more sophisticated stratification by learning complex patterns from multi-modal patient data, including genomic profiles, transcriptomic signatures, clinical histories, and lifestyle factors, that together determine treatment outcomes. Single-cell analysis has emerged as a powerful approach for understanding cellular heterogeneity in disease and treatment response, but the high dimensionality and complexity of single-cell data present analytical challenges (Cui et al., 2024; Lotfollahi et al., 2019; Theodoris et al., 2023). scGPT applies transformer architectures to single-cell transcriptomics, learning representations that capture cell state, type, and perturbation responses from large-scale single-cell atlases (Cui et al., 2024). These learned representations enable prediction of cellular responses to drugs and genetic perturbations, identification of cell types and states, and integration of data across experimental batches. Geneformer demonstrated that transformer models pretrained on large-scale single-cell data can transfer learned representations to downstream tasks including disease state classification, drug response prediction, and network analysis (Theodoris et al., 2023). scGen showed that generative models could predict single-cell perturbation responses, enabling in silico exploration of how cells respond to drugs or genetic modifications (Lotfollahi et al., 2019).

Drug sensitivity prediction at the single-cell level represents a particularly promising application for precision oncology, where tumor heterogeneity often underlies treatment resistance and relapse. By analyzing gene expression patterns in individual cancer cells, models can identify resistant subpopulations that may persist despite initial treatment response and require alternative therapeutic strategies. Integration of single-cell transcriptomics with language models trained on drug mechanism information enables prediction of which cells will respond to which drugs, potentially guiding selection of combination therapies that address multiple resistant populations. Drug synergy prediction addresses the challenge of identifying combination therapies that exhibit synergistic effects, where the combined effect exceeds the sum of individual drug effects (Preuer et al., 2018; Menden et al., 2019). Predicting synergy is challenging because it depends on complex interactions between drug mechanisms, target pathways, and cellular context. DeepSynergy represented a breakthrough in this area, introducing a deep neural network architecture that integrates chemical descriptors of drug pairs with genomic features of cancer cell lines to predict synergy scores (Preuer et al., 2018). By learning joint representations of drug chemistry and cellular context, DeepSynergy achieved superior predictive performance compared to kernel methods, random forests, and gradient boosting approaches on the NCI-ALMANAC and O’Neil benchmark datasets. More recent work has explored transformer-based approaches that can incorporate natural language descriptions of drug mechanisms to improve synergy prediction, leveraging the semantic understanding of drug action encoded in language models.

4.3 Drug recommendation and dosage optimization

Personalized drug recommendation, selecting optimal medications and dosages for individual patients based on their specific characteristics and clinical context, requires integrating patient-specific factors with pharmacological knowledge and clinical evidence (Rajkomar et al., 2018; Si et al., 2021). Factors influencing drug selection include patient comorbidities, concurrent medications, genetic variants affecting drug metabolism, renal and hepatic function, and individual preferences and circumstances. LLMs can process electronic health records, drug databases, and clinical guidelines to generate evidence-based recommendations that account for these multiple factors. Drug-drug interaction (DDI) prediction represents a critical safety consideration in drug recommendation, as concurrent use of interacting medications can cause serious adverse events (Zitnik et al., 2018; Tatonetti et al., 2012). Polypharmacy, the use of multiple concurrent medications, is common especially in elderly patients and those with multiple chronic conditions, creating complex webs of potential interactions that are difficult to track manually. Graph-based approaches have proven effective for DDI prediction by representing drugs and their relationships as networks, with interactions predicted based on network structure and drug properties (Zitnik et al., 2018). Language models can enhance these approaches by extracting interaction information from unstructured literature and regulatory documents, expanding the scope of known interactions beyond those in structured databases.

Adverse drug event prediction aims to identify patients at elevated risk for specific adverse reactions, enabling proactive monitoring or alternative drug selection (Tatonetti et al., 2012; Huang et al., 2022). Data-driven analysis of electronic health records and adverse event databases has revealed patterns linking patient characteristics, drug exposure, and adverse outcomes that can inform risk prediction. Language models can process clinical notes to identify subtle signs of emerging adverse events that may not be captured in structured diagnosis codes, potentially enabling earlier intervention.

4.4 Genomic interpretation and variant analysis

Interpreting genomic variants, determining which genetic variations are clinically relevant and how they affect disease risk, prognosis, or treatment response, represents a major bottleneck in translating genomic sequencing to clinical practice (Zhou et al., 2023; Ji et al., 2021; Dalla-Torre et al., 2023). Clinical sequencing routinely identifies thousands of genetic variants in each patient, but only a small fraction have established clinical significance, and even fewer have actionable implications for treatment. Variants of uncertain significance (VUS) represent a particular challenge, as they cannot be definitively classified as pathogenic or benign based on current evidence. DNA language models trained on genomic sequences can predict the functional consequences of genetic variants by learning sequence patterns associated with regulatory elements, splice sites, and protein function (Zhou et al., 2023; Ji et al., 2021; Nguyen et al., 2024). DNABERT learns contextual representations of DNA sequences through masked language modeling, capturing patterns that predict regulatory function and variant effects (Ji et al., 2021). These learned representations improve performance on downstream tasks including transcription factor binding prediction, promoter detection, and splice site identification. Larger-scale genomic language models trained on diverse species have demonstrated improved generalization and zero-shot performance on novel prediction tasks (Dalla-Torre et al., 2023).

RNA language models extend similar approaches to RNA sequences, with applications including splice site prediction, RNA secondary structure prediction, and post-transcriptional regulation modeling. Protein language models can predict the effects of amino acid substitutions on protein structure and function, complementing DNA-level analysis to provide a complete picture of variant consequences (Lin et al., 2023; Rives et al., 2021). The attention patterns learned by protein language models correlate with protein contact maps, suggesting that these models capture structural information relevant to predicting mutation effects. Integration of variant effect predictions with clinical phenotype information and population genetics data enables prioritization of variants for clinical review. Language models can synthesize evidence from multiple sources, including variant databases, scientific literature, and clinical case reports, to support variant classification. By processing the vast literature on genetic variants and their clinical associations, these models can identify relevant evidence more efficiently than manual curation. In hereditary cancer syndromes, LLM-based variant interpretation has been applied to BRCA1/2 variants to improve classification of variants of uncertain significance, which affect approximately 10%–20% of tested patients and create substantial clinical uncertainty regarding cancer risk and management (Tavtigian et al., 2018). For pharmacogenomic variants affecting drug metabolism, including CYP2D6 and CYP2C19 polymorphisms that influence the efficacy of antidepressants, opioids, and antiplatelet agents, language models trained on clinical outcome data have improved prediction of drug response phenotypes from genotype data (Zhou et al., 2008). In pediatric populations, genomic language models have accelerated diagnosis of rare genetic diseases, reducing the diagnostic odyssey from years to weeks for conditions such as developmental epileptic encephalopathies and inborn errors of metabolism (Clark et al., 2019).

4.5 Clinical trial matching and outcome prediction

Matching patients to appropriate clinical trials and predicting trial outcomes are critical for advancing precision medicine by ensuring that new therapies are evaluated in the patients most likely to benefit and that trials are designed efficiently to detect true effects (Wang and Sun, 2022; Yuan et al., 2019). Clinical trial matching requires understanding complex eligibility criteria, often expressed in semi-structured or natural language, and evaluating patient characteristics against these criteria. Trial outcome prediction requires integrating information about trial design, patient population, drug mechanism, and historical precedent. Natural language processing enables automated extraction and interpretation of trial eligibility criteria from ClinicalTrials.gov and other registries (Yuan et al., 2019). Criteria2Query demonstrated that language models could translate natural language eligibility criteria into structured database queries, enabling automated patient matching at scale (Yuan et al., 2019). Trial2Vec learned representations of clinical trials from their text descriptions, enabling similarity search to identify related trials and predict trial outcomes based on similar precedents (Wang and Sun, 2022). These approaches can accelerate patient recruitment by automatically screening electronic health records against trial criteria and notifying clinicians of potential matches.

Trial outcome prediction can inform go/no-go decisions in drug development by estimating the probability of trial success based on available preclinical and early clinical data. Language models can synthesize information from trial protocols, preclinical data, mechanism of action descriptions, and historical trial results to predict success probability. While such predictions carry substantial uncertainty, they can help prioritize development resources toward programs with higher success probability.

4.6 Integration of multiscale data

A defining challenge of precision medicine is integrating data across multiple scales of biological organization, from molecular level genomics, proteomics, and metabolomics through cellular level single-cell analyses to organism level clinical phenotypes and imaging (Esteva et al., 2019; Wang et al., 2023; Hays et al., 2025). Each data type provides a different perspective on patient biology, and their integration can reveal patterns not visible in any single modality. Hierarchical molecular language models conceptualize cellular signaling as a specialized molecular language, where signaling molecules function as tokens, enabling the modeling of protein interactions and their functional consequences across multiple biological scales (Hays et al., 2025). Multi-modal LLMs are particularly well suited to this integration challenge, as they can learn aligned representations across diverse data types and reason about relationships between modalities. The integration of transcriptomic data with clinical information exemplifies the potential for multiscale analysis in precision medicine. Gene expression profiles reflect cellular state and pathway activity, providing molecular insight into disease processes and drug responses. By combining transcriptomic signatures with clinical phenotypes, laboratory values, and treatment history, models can develop richer representations of patient state that improve outcome prediction. Single-cell transcriptomics adds additional resolution by revealing cellular heterogeneity obscured in bulk tissue analysis (Hays et al., 2023). Imaging data, including radiological imaging and pathology slides, provides spatial information about disease that complements molecular and clinical data. Multi-modal models that integrate imaging with other data types can leverage complementary information for improved diagnosis and prognosis. For example, combining histopathology images with genomic profiles in cancer enables more precise tumor characterization than either modality alone.

Knowledge graphs provide a framework for integrating heterogeneous data types by representing entities and relationships in a unified structure (Mohamed et al., 2021; Bonner et al., 2022). Biomedical knowledge graphs connect drugs, targets, diseases, genes, pathways, and other entities through diverse relationship types, providing a backbone for data integration. Language models can enhance knowledge graph-based integration by extracting new relationships from unstructured text, resolving entity ambiguities, and reasoning about multi-hop relationships that connect distant entities through intermediate connections.

5 Technical methodologies and architectures

The success of LLMs in drug discovery and precision medicine depends on sophisticated technical approaches that adapt general-purpose language models to biomedical domains while addressing the unique challenges of biological and clinical data. This section examines key methodologies and architectural innovations enabling these applications, providing technical depth on the computational approaches underlying the applications discussed in previous sections.

Beyond cataloguing architectures, it is important to compare them critically and to identify the trade-offs that practitioners face when choosing between encoder-only, decoder-only, encoder-decoder, graph-based, and multi-modal designs. Encoder-only models such as BERT, BioBERT, and ChemBERTa excel at representation learning for classification and extraction tasks and are typically smaller and cheaper to fine-tune, but they cannot natively generate sequences and tend to underperform on open-ended question answering or de novo molecular design. Decoder-only models such as GPT-3/4, BioGPT, and MolGPT support flexible generation and in-context learning, yet they are more prone to hallucination on biomedical facts and are harder to constrain to validated knowledge. Encoder-decoder and sequence-to-sequence models (e.g., MolT5, Chemformer) bridge these regimes for translation-style tasks such as text-to-molecule mapping and retrosynthesis, at the cost of more complex training. Graph- and 3D-aware architectures (e.g., D-MPNN, equivariant networks, DiffDock) capture molecular topology and stereochemistry that linearised SMILES cannot, but they are less compatible with natural-language interfaces and require specialised data pipelines. Multi-modal systems (e.g., InstructMol, Galactica, BiomedGPT) unify these modalities at the price of substantially increased training data and compute. In short, no single architecture dominates across drug-discovery and precision-medicine tasks, and our subsequent discussion of representative models is framed around this trade-off space rather than as a ranking exercise. Tables 1, 2 make these comparisons explicit and tie reported metrics back to specific datasets and limitations.

TABLE 1. Expanded comparative summary of representative LLM-based frameworks used in drug discovery.

Columns capture model name, release year, base LLM/architecture, parameter count (“n.r.” indicates not reported), drug-discovery task, evaluation dataset (s), reported metric where stated, and documented limitations. Rows are drawn primarily from Tables 2–7 of the recent overview by Song et al. (2025) and cross-checked against the original publications cited where available. Entries marked with “-” for parameters or metric indicate values not reported in the source. This list is non-exhaustive.

TABLE 2. Expanded comparative summary of representative LLM-based frameworks used in precision medicine.

Columns capture model name, release year, base architecture, parameter count (“n.r.” indicates not reported), clinical task, evaluation dataset/cohort, reported metric where stated, and key limitations or clinical risks. Reported metrics are taken from the original publications; benchmark coverage cross-checked against Song et al. (2025) and the healthcare-LLM benchmarking review by Budler et al. (2025). This list is non-exhaustive.

5.1 Pretraining and fine-tuning strategies

Pretraining on large-scale corpora provides LLMs with foundational knowledge that can be transferred to downstream tasks through fine-tuning or prompting (Devlin et al., 2019; Brown et al., 2020; Lee et al., 2020). The choice of pretraining data, objective, and scale significantly impacts model capabilities and suitability for biomedical applications. For biomedical LLMs, pretraining typically combines general-purpose text corpora with domain-specific sources including scientific literature from PubMed and PubMed Central, clinical documentation from electronic health record systems, molecular databases such as ChEMBL, DrugBank, and UniProt, and structured knowledge resources including ontologies and knowledge graphs (Lee et al., 2020; Gu et al., 2021; Wishart et al., 2018; Mendez et al., 2019). Pretraining objectives vary based on the intended application and model architecture. Masked language modeling (MLM), used in BERT-style encoder models, trains the model to predict randomly masked tokens from bidirectional context, producing representations suited for classification and extraction tasks (Devlin et al., 2019; Lee et al., 2020). Autoregressive language modeling, used in GPT-style decoder models, trains prediction of the next token given previous tokens, naturally supporting generation tasks (Radford et al., 2019; Luo et al., 2022). Sequence-to-sequence objectives, used in encoder-decoder architectures like T5 and BART, train mapping from input sequences to output sequences, supporting translation and transformation tasks (Edwards et al., 2022a). Domain-specific pretraining objectives have been developed for molecular and biological sequences, including span prediction for proteins and contrastive learning across modalities (Ross et al., 2022; Rives et al., 2021).

Fine-tuning adapts pretrained models to specific downstream tasks using labeled training data. Standard fine-tuning updates all model parameters on the downstream task, which can achieve high performance but requires substantial computational resources and risks catastrophic forgetting of pretrained knowledge. Parameter-efficient fine-tuning methods, including Low-Rank Adaptation (LoRA) and adapter modules, update only a small fraction of parameters while keeping pretrained weights frozen, reducing computational requirements and maintaining more of the pretrained knowledge (Touvron et al., 2023b). Instruction tuning trains models to follow natural language task descriptions, improving generalization to novel tasks specified at inference time (the scaling and openness of these instruction-tuned models in our corpus are visualised in Figures 4, 7) (Wei et al., 2022; Fang et al., 2023). Reinforcement learning from human feedback (RLHF) further refines model outputs by training on human preference judgments, improving alignment with human intentions and values.

Table 3 provides a comprehensive overview of key biomedical language models across different application domains, including their architecture, training data, and primary applications.

TABLE 3. Overview of key large language models for biomedical applications. Parameter counts and training data verified from original publications (Lee et al., 2020; Luo et al., 2022; Singhal et al., 2023; Gu et al., 2021; Alsentzer et al., 2019; Lin et al., 2023; Zhou et al., 2023; Cui et al., 2024; Ahmad et al., 2022).

To complement the architectural overview above, Table 1 summarises representative LLM-based frameworks for drug discovery alongside their reported performance metrics, evaluation datasets, and documented limitations, and Table 2 provides the analogous comparison for precision-medicine applications. These tables synthesise data drawn from the source publications cited in Sections 3, 4, together with the recent overview by Song et al. (2025), and are intended to facilitate side-by-side comparison of strengths, weaknesses, and trade-offs rather than to repeat the descriptive narrative.

5.2 Transformer architectures for biomedical data

The transformer architecture provides the foundation for essentially all modern large language models, with variants adapted for different data types and application requirements (Vaswani et al., 2017; Hays, 2026a). The core self-attention mechanism enables each position in a sequence to attend to all other positions, capturing long-range dependencies without the sequential processing constraints of recurrent architectures. Multi-head attention allows the model to jointly attend to information from different representation subspaces, capturing diverse relationship types. Position encodings, whether learned or sinusoidal, provide sequence order information that the permutation-invariant attention mechanism would otherwise lack.

For molecular data, transformer architectures process SMILES strings as character or subword sequences, with tokenization strategies that can influence model performance (Chithrananda et al., 2020; Bagal et al., 2022). SMILES-level tokenization treats individual characters as tokens, while subword tokenization methods like byte-pair encoding (BPE) can learn meaningful chemical substructure tokens from data. Graph transformers extend the architecture to process molecular graphs directly, using attention mechanisms that respect graph structure by restricting attention to connected nodes or incorporating edge features into attention computation (Xiong et al., 2020). Hybrid approaches combine sequence-based transformers with graph neural networks to capture complementary aspects of molecular structure.

For protein sequences, protein language models apply standard transformer architectures to amino acid sequences, with position encodings that can extend to the longer sequence lengths typical of proteins compared to natural language sentences (Lin et al., 2023; Rives et al., 2021). ESM-2 demonstrated that scaling protein language models to 15 billion parameters substantially improves learned representations, with the largest models approaching AlphaFold2-level structure prediction accuracy (Lin et al., 2023). ProtGPT2 applied autoregressive language modeling to protein generation, producing novel sequences that fold into stable structures (Ferruz et al., 2022). Specialized architectures for proteins have explored incorporating evolutionary information through multiple sequence alignments or structure-aware positional encodings.

For genomic sequences, DNA language models adapt transformer architectures to nucleotide sequences with appropriate tokenization and positional encoding strategies (Zhou et al., 2023; Ji et al., 2021). DNABERT uses k-mer tokenization to create tokens representing short nucleotide subsequences, reducing sequence length while capturing local context (Ji et al., 2021). Long-context architectures are particularly important for genomics, where regulatory relationships can span kilobases or megabases. HyenaDNA achieves single-nucleotide resolution over sequences up to 1 million bases through efficient attention approximations (Nguyen et al., 2024).

5.3 Multi-modal architectures

Multi-modal architectures integrate information from different data types within unified model frameworks, enabling reasoning across modalities that is essential for drug discovery and precision medicine applications (Cao et al., 2023; Liu S. et al., 2023; Taylor et al., 2022). Design choices for multi-modal systems include the selection of modality-specific encoders, alignment mechanisms to create shared representations, and fusion strategies to combine information for downstream tasks. Modality-specific encoders process each data type using architectures suited to its structure. For molecules, encoders may be graph neural networks, transformers processing SMILES strings, or 3D convolutional networks processing molecular geometry. For proteins, transformers processing amino acid sequences leverage pretrained protein language model representations. For text, standard transformer language models provide encodings. For images, vision transformers or convolutional networks extract features. The choice of encoder architecture affects what information is captured and how effectively it transfers to downstream tasks.

Alignment mechanisms create shared representation spaces where entities from different modalities can be compared and combined. Contrastive learning aligns representations by training the model to produce similar embeddings for matched pairs (such as molecules and their text descriptions) while producing dissimilar embeddings for mismatched pairs (Liu S. et al., 2023). This approach learns an embedding space where semantically related entities cluster together regardless of modality, enabling cross-modal retrieval and comparison. Cross-attention mechanisms provide another alignment approach, where queries from one modality attend to keys and values from another, allowing flexible information exchange between modalities. Fusion strategies combine aligned representations for downstream prediction or generation. Early fusion concatenates or combines modality-specific representations before processing through shared layers. Late fusion processes each modality independently through most of the model, combining only at the final prediction stage. Cross-modal attention enables dynamic fusion where the degree of information exchange is learned based on the task. The optimal fusion strategy depends on the specific task and the nature of relationships between modalities. InstructMol exemplifies multi-modal integration for molecular understanding, combining a molecular encoder with a language model through instruction tuning (Cao et al., 2023). The two-stage training approach first pretrains encoders separately, then fine-tunes the combined system on instruction-following tasks that require cross-modal reasoning. This approach enables the model to answer questions about molecules, generate molecules from text descriptions, and perform other tasks requiring integration of molecular and language understanding.

5.4 Graph neural networks and molecular representations

Graph neural networks (GNNs) process molecular structures as graphs where atoms are nodes and bonds are edges, learning representations that capture topological and chemical properties (Nguyen et al., 2021; Xiong et al., 2020; Hays, 2026b). Unlike sequence-based representations that require imposing an artificial ordering on atoms, graph representations naturally preserve molecular symmetry and can incorporate rich atomic and bond features. GNN architectures differ in their message-passing schemes that aggregate information from neighboring nodes and their readout functions that produce graph-level representations from node embeddings. Message passing neural networks (MPNNs) provide a general framework encompassing many GNN variants, where node representations are iteratively updated by aggregating messages from neighboring nodes (Yang et al., 2019; Xiong et al., 2020). The Attentive FP architecture applies graph attention to molecular property prediction, using attention weights to differentially weight contributions from neighboring atoms based on their relevance (Xiong et al., 2020). Graph convolutional networks aggregate neighbor information through convolution-like operations, while graph transformers extend self-attention to graph-structured data.

The integration of GNNs with language models enables molecular understanding that leverages both structural and semantic information. Graph-enhanced language models incorporate GNN-derived representations into transformer architectures through cross-attention or concatenation. Conversely, language model representations can provide semantic features that augment graph-based molecular representations, for example, by encoding textual descriptions of molecular properties or mechanism of action. Three-dimensional molecular representations capture spatial geometry that sequence-based and two-dimensional graph representations cannot encode (Corso et al., 2023; Hoogeboom et al., 2022; Xu et al., 2022). Equivariant neural networks process 3D molecular structures while respecting their rotational and translational symmetry, ensuring that predictions are independent of the arbitrary orientation of input molecules. These architectures are particularly important for structure-based drug design, where the geometry of protein-ligand interactions determines binding affinity and specificity.

5.5 Diffusion models for molecular generation

Diffusion models have emerged as a powerful paradigm for generative modeling, achieving state-of-the-art results in image generation and increasingly being applied to molecular generation and structure prediction (Corso et al., 2023; Hoogeboom et al., 2022; Xu et al., 2022; Schneuing et al., 2022). The diffusion framework learns to reverse a gradual noising process, training a neural network to predict the noise added at each step so that sampling can proceed by iteratively denoising from random noise to generate realistic samples.

For molecular generation, diffusion models can operate on various molecular representations including atom coordinates in 3D space, molecular graphs, or sequences (Hoogeboom et al., 2022; Xu et al., 2022). Equivariant diffusion models generate 3D molecular structures while respecting rotational and translational invariance, producing molecules whose properties are independent of coordinate frame choice (Hoogeboom et al., 2022). GeoDiff applied diffusion to molecular conformation generation, predicting the 3D geometry of molecules from their 2D graphs (Xu et al., 2022). These approaches can generate molecular structures that satisfy complex geometric constraints relevant to drug binding. Structure-based diffusion models generate molecules that bind to specific protein targets by conditioning the diffusion process on target structure (Corso et al., 2023; Schneuing et al., 2022). DiffDock applied diffusion to molecular docking, directly generating binding poses rather than scoring enumerated poses, achieving state-of-the-art docking performance (Corso et al., 2023). Pocket-conditioned generation enables design of molecules complementary to target binding sites, integrating target structure information into the generation process.

5.6 Retrieval-augmented generation

Retrieval-augmented generation (RAG) enhances language model capabilities by retrieving relevant information from external knowledge sources at inference time, combining the flexibility of neural generation with the reliability of curated knowledge bases (Lewis et al., 2020; Zakka et al., 2024; Xiong et al., 2024). Rather than relying solely on knowledge encoded in model parameters during pretraining, RAG systems query external databases or document collections to find relevant information that is then included in the model’s context for generating responses. RAG architectures typically consist of a retriever that identifies relevant documents or passages and a generator that produces outputs conditioned on retrieved context. Dense passage retrieval uses neural embeddings to find semantically relevant passages, while sparse retrieval based on term matching can efficiently search large collections. The generator may be a standard language model with retrieved passages prepended to the input, or specialized architectures that explicitly model the relationship between query, retrieved documents, and output.

For biomedical applications, RAG offers several advantages over purely parametric models. It enables access to current information beyond the model’s training cutoff, which is particularly important given the rapid pace of biomedical research. It reduces hallucination by grounding outputs in retrieved evidence, addressing a major concern for clinical applications. It provides transparency through citation of sources, enabling verification of model outputs. And it can be updated by changing the retrieval corpus without retraining the model, facilitating rapid incorporation of new findings.

5.7 Evaluation frameworks and benchmarks

Rigorous evaluation is essential for assessing LLM performance and identifying areas for improvement. Evaluation frameworks for biomedical LLMs encompass diverse tasks and metrics reflecting the range of applications in drug discovery and precision medicine (Singhal et al., 2023; Wu et al., 2018; Polykovskiy et al., 2020). For molecular property prediction, MoleculeNet provides standardized benchmarks covering physical chemistry, biophysics, physiology, and toxicology, with metrics including AUROC, RMSE, and MAE depending on the specific task (Wu et al., 2018). For molecular generation, MOSES benchmarks evaluate validity, uniqueness, novelty, and distribution learning metrics (Polykovskiy et al., 2020). For drug-target interaction prediction, benchmarks on DAVIS, KIBA, and BindingDB datasets evaluate prediction accuracy with particular attention to cold-split scenarios that test generalization to novel compounds or targets.

For clinical applications, medical question answering benchmarks including MedQA, PubMedQA, and MedMCQA evaluate knowledge and reasoning capabilities (Singhal et al., 2023; Kung et al., 2023). Beyond accuracy, evaluation of clinical systems must consider factors including explanation quality, potential for harm, and alignment with clinical consensus. Human evaluation remains essential for assessing whether model outputs would be useful and safe in real clinical contexts.

6 Current applications and case studies

The theoretical potential of LLMs in drug discovery and precision medicine is being realized through diverse real-world applications that demonstrate practical impact across the drug development pipeline and clinical practice. This section examines specific case studies and applications that illustrate how these technologies are being deployed to address concrete biomedical challenges, providing evidence of both the current capabilities and the remaining gaps between research advances and clinical utility.

6.1 End-to-end drug discovery pipelines

The vision of fully integrated, AI-driven drug discovery pipelines is beginning to be realized through systems that combine multiple LLM capabilities within unified computational frameworks. These end-to-end approaches aim to automate or accelerate the entire discovery process, from initial target identification through lead optimization and preclinical characterization. ChemCrow demonstrated an integrated approach to chemical research by augmenting a large language model with access to specialized chemistry tools (Bran et al., 2024). When presented with drug discovery objectives, the system can autonomously decompose complex tasks into subtasks, invoke appropriate computational tools for property prediction and reaction planning, search the literature for relevant precedents, and synthesize results into actionable recommendations. In the original report, ChemCrow demonstrated promising performance on multi-step synthesis planning benchmarks and property optimization objectives, illustrating the potential for LLM-orchestrated drug discovery workflows; independent evaluations, however, have been mixed, with subsequent studies highlighting sensitivity to tool choice, prompting strategy, and benchmark construction and cautioning against extrapolating from curated demonstrations to routine medicinal-chemistry practice.

Autonomous chemical research systems have extended this paradigm to include robotic experimentation, creating closed-loop systems that integrate computational design with physical synthesis and characterization (Boiko et al., 2023). These systems use LLMs to interpret experimental results, propose next experiments based on observed outcomes, generate detailed experimental protocols, and coordinate robotic platforms to execute synthesis and testing. While still primarily demonstrated in academic settings, such systems point toward a future where drug discovery proceeds through continuous automated cycles of design, synthesis, testing, and optimization. The development of potent DDR1 kinase inhibitors by Insilico Medicine represents one of the most publicized examples of AI-accelerated drug discovery (Zhavoronkov et al., 2019). This landmark study employed generative tensorial reinforcement learning (GENTRL) to design novel DDR1 kinase inhibitors, completing the entire cycle from target selection to synthesis of lead candidates in just 21 days. The generated molecules demonstrated nanomolar potency against DDR1 and exhibited favorable pharmacokinetic properties in mouse studies, providing compelling proof-of-concept that AI-driven molecular generation can produce experimentally validated drug candidates with activity comparable to traditionally discovered compounds. While this work predates the current generation of LLMs, it established the potential for AI-driven molecular generation to produce experimentally validated drug candidates. Building on this foundation, Insilico Medicine subsequently advanced INS018_055, a first-in-class inhibitor of TNIK (TRAF2 and NCK-interacting kinase) for idiopathic pulmonary fibrosis, through Phase II clinical trials, representing one of the first AI-designed drugs to reach this stage (Xu et al., 2025). The drug advanced from initiation of target discovery to completion of phase 0/1 clinical testing in under 30 months, substantially faster than traditional timelines. Another landmark achievement was the discovery of halicin, a structurally novel antibiotic identified through deep learning screening of the Drug Repurposing Hub, which demonstrated broad-spectrum activity against pan-resistant Acinetobacter baumannii infections in mouse models (Stokes et al., 2020). The halicin discovery exemplified how AI could identify compounds with mechanisms of action distinct from existing antibiotics, addressing the critical challenge of antimicrobial resistance. More recently, Recursion Pharmaceuticals has applied transformer-based models to identify novel treatments for rare diseases, including cerebral cavernous malformation and neurofibromatosis type 2, leveraging phenotypic screening data integrated with multi-modal embeddings (Stokes et al., 2020; Chon et al., 2017).

It is important to put these case studies in proper perspective with respect to clinical-trial integration. As of late 2025, the number of AI-designed small molecules that have actually entered human studies remains small (on the order of 20–30 publicly disclosed Phase I or Phase I/II candidates across companies such as Insilico Medicine, Exscientia, Recursion, BenevolentAI, and Schrödinger), and none has yet received marketing authorisation. Reported Phase I success rates of ∼80%–90% for the first wave of AI-discovered candidates (Jayatunga et al., 2024) (compared to historical industry averages of ∼50%–60%) are based on small cohorts and short follow-up, so they should be treated as encouraging but preliminary signals rather than as established attrition reductions. The strongest evidence to date that AI-based discovery can compress timelines is completion of phase 0/1 clinical testing in under 30 months from the initiation of target discovery reported for INS018_055 (versus ∼4–6 years traditionally) (Xu et al., 2025) and the rapid antibiotic discovery exemplified by halicin (Stokes et al., 2020); whether these accelerations translate into improved Phase II/III success and approved therapies will only become clear over the next 3–5 years. Throughout this review we therefore distinguish carefully between (a) benchmark performance, (b) wet-lab or preclinical validation, and (c) prospective clinical evidence, and we recommend the same distinction when interpreting individual claims.

6.2 Precision oncology applications

Cancer treatment exemplifies the promise and complexity of precision medicine, as tumors exhibit substantial molecular heterogeneity that requires personalized therapeutic strategies. LLM applications in precision oncology span molecular tumor characterization, treatment recommendation, and clinical trial matching. The integration of large language models with molecular tumor boards represents an emerging application that could enhance clinical decision making for cancer patients. By synthesizing information from tumor genomic profiles, scientific literature, clinical guidelines, and trial databases, LLMs can identify relevant biomarkers, suggest targeted therapies, and match patients to appropriate clinical trials. Studies have demonstrated that LLM-based systems can identify treatment options with high concordance with expert tumor board recommendations while potentially identifying additional options not considered by human experts (Singhal et al., 2023).

Single-cell analysis powered by foundation models enables unprecedented resolution in understanding tumor heterogeneity and treatment resistance. scGPT and related models can analyze single-cell transcriptomes to identify resistant subpopulations, predict responses to perturbations, and suggest combination therapies targeting multiple cell states (Cui et al., 2024). For triple-negative breast cancer, which lacks targetable hormone receptors, transformer models analyzing tumor transcriptomes have identified patient subgroups with differential responses to immune checkpoint inhibitors, enabling more precise patient stratification for immunotherapy (Jiang et al., 2018). Drug synergy prediction for cancer combination therapy addresses the challenge of identifying drug pairs that work better together than individually (Preuer et al., 2018; Menden et al., 2019). DeepSynergy, a pioneering deep neural network developed by Preuer and colleagues, demonstrated the power of integrating chemical structure features with genomic information from cancer cell lines to predict synergistic drug combinations (Preuer et al., 2018). The model substantially outperformed other machine learning methods on benchmark datasets, establishing deep learning as a promising approach for combination therapy discovery. The subsequent DREAM Challenge on cancer drug combinations further validated these approaches and established benchmarks that continue to drive methodological advances in this area.

6.3 Clinical decision support implementations

Clinical decision support systems incorporating LLMs are being piloted in various healthcare settings, with applications spanning diagnostic support, treatment recommendation, and documentation assistance. Med-PaLM and related medical language models have demonstrated high-level performance on clinical reasoning benchmarks, achieving accuracy on medical licensing examinations comparable to or exceeding that of practicing physicians (Singhal et al., 2023; Nori et al., 2023). While performance on standardized exams does not guarantee clinical utility, these results establish a baseline of medical knowledge and reasoning capability that supports more applied clinical applications.

Discharge summary generation represents an early practical application of LLMs in clinical workflows. Generating comprehensive discharge summaries is time-consuming for clinicians but important for continuity of care. Studies have demonstrated that LLMs can generate high-quality draft summaries from clinical notes, potentially reducing documentation burden while maintaining or improving summary quality (Patel and Lam, 2023). Human oversight remains essential to verify accuracy and appropriateness for specific patients. Diagnostic decision support systems use LLMs to synthesize patient information and suggest differential diagnoses or next diagnostic steps. By processing clinical notes, laboratory values, and imaging reports, these systems can identify patterns consistent with specific diagnoses and flag findings requiring urgent attention. Integration with electronic health record systems enables deployment within existing clinical workflows, though careful attention to alert fatigue and workflow integration is essential for adoption.

6.4 Drug repurposing case studies

Drug repurposing accelerated by AI methods has demonstrated real-world impact, particularly during the COVID-19 pandemic when rapid identification of treatment options was critical (Zhou et al., 2020; Zeng et al., 2020; Gysi et al., 2021). While specific LLM-driven repurposing successes are still emerging, the methodological foundations established by these efforts support ongoing applications. Network medicine approaches combining knowledge graphs with machine learning identified candidate COVID-19 therapeutics by analyzing the proximity of drug targets to disease-associated genes in the interactome (Gysi et al., 2021). Several computationally identified candidates subsequently entered clinical trials, demonstrating the translational potential of AI-driven repurposing. Language models enhance these approaches by extracting relationships from literature to populate knowledge graphs and by synthesizing evidence to prioritize candidates.

Systematic mining of electronic health records for drug repurposing opportunities continues to yield candidates for clinical investigation. By analyzing patterns of drug exposure and disease outcomes across large patient populations, researchers can identify unexpected therapeutic effects that warrant prospective validation. LLMs enable more comprehensive analysis of these data by processing unstructured clinical notes to capture therapeutic effects not coded in structured fields. A compelling example is the repurposing of metformin, the widely prescribed diabetes medication, for cancer prevention and treatment. Analysis of large-scale electronic health records combined with mechanistic modeling revealed reduced cancer incidence in diabetic patients taking metformin, leading to numerous clinical trials investigating its anticancer properties across tumor types including breast, colorectal, and prostate cancers (Zaidi et al., 2019; Yu and Suissa, 2023). In neurodegenerative diseases, LLM-based analysis of clinical records has identified potential benefits of beta-2 adrenergic agonists for Parkinson’s disease risk reduction, with retrospective cohort studies supporting a protective association (Mittal et al., 2017).

6.5 Literature synthesis and knowledge extraction

The ability of LLMs to process and synthesize vast scientific literature represents a capability with broad applications across drug discovery and precision medicine. With biomedical publication rates exceeding 1 million articles annually, no human researcher can comprehensively survey even their specific field. Automated literature review using LLMs can rapidly synthesize evidence on specific topics, identify consensus and controversies in the field, and highlight knowledge gaps requiring further investigation. While current systems cannot fully replace expert literature review, they can substantially accelerate initial surveys and ensure comprehensive coverage of relevant publications.

Knowledge graph population using information extraction from text enables creation of structured knowledge resources from unstructured literature. Named entity recognition identifies mentions of drugs, diseases, genes, and other biomedical entities, while relation extraction identifies the relationships between them. LLMs have demonstrated improved performance on these extraction tasks, enabling more complete and accurate knowledge graphs that support downstream reasoning and prediction. Hypothesis generation from literature synthesis represents an intriguing emerging application. By identifying patterns across large bodies of literature, LLMs may suggest novel hypotheses that connect findings from disparate research areas. While validation of such hypotheses requires experimental investigation, computational hypothesis generation could accelerate scientific discovery by directing research attention to promising directions.

6.6 Comparative performance analysis

Table 4 summarizes the performance of selected LLM applications across different biomedical tasks, illustrating the current state of the art and areas of particular strength or remaining challenge.

TABLE 4. Performance comparison of large language model applications in drug discovery and precision medicine.

Performance metrics verified from original publications: Med-PaLM, 2 (Singhal et al., 2023), ESMFold (Lin et al., 2023), ChemBERTa-2 (Ahmad et al., 2022), DiffDock (Corso et al., 2023), scGPT (Cui et al., 2024), MolGPT (Bagal et al., 2022).

These results demonstrate that LLMs can achieve strong performance across diverse biomedical tasks (see Figure 8 for the corpus-wide view of reported metrics), often matching or exceeding specialized methods while offering greater flexibility and generalization. However, performance varies substantially across specific subtasks and evaluation conditions, and real-world utility depends on factors beyond benchmark performance including robustness, interpretability, and integration with existing workflows.

7 Challenges and limitations

Despite remarkable progress, the application of LLMs in drug discovery and precision medicine faces significant challenges that must be addressed to realize their full potential for improving human health. These challenges span technical, regulatory, ethical, and practical dimensions that together determine whether research advances will translate into clinical and commercial impact. Importantly, these challenges manifest differently across application domains, with distinct priorities emerging for drug discovery, clinical decision support, genomics and variant analysis, and clinical trial design. This section examines key limitations and obstacles, providing a balanced assessment of the current state and identifying priorities for future development.

7.1 Domain-specific challenge profiles

The relative importance of implementation challenges varies substantially across biomedical AI application domains, reflecting fundamental differences in data characteristics, regulatory environments, and end-user requirements (Gerke et al., 2020; Char et al., 2018). Understanding these domain-specific profiles is essential for prioritizing research investments and deployment strategies.

Drug discovery presents a unique challenge landscape where interpretability and generalization emerge as primary concerns. Interpretability is critical because understanding why a model predicts a molecule will be active against a target is essential for guiding medicinal chemistry optimization and generating testable mechanistic hypotheses (Jiménez-Luna et al., 2020; Preuer et al., 2019). Without mechanistic insight, researchers cannot rationally improve predicted candidates or identify potential liabilities. Generalization poses particular difficulties because drug discovery inherently requires predictions for novel chemical entities and previously unexplored targets; cold-start performance on entities absent from training data directly determines practical utility (Huang et al., 2021; Yazdani-Jahromi et al., 2022). Hallucination risk is moderate but consequential, as fabricated mechanisms of action or pharmacokinetic properties could misdirect expensive experimental programs. Data quality concerns center on inconsistencies across assay conditions in public databases rather than demographic bias, while privacy considerations are less acute given that molecular data typically lacks direct patient identifiers.

Clinical decision support faces the broadest array of critical challenges due to direct patient safety implications. Hallucination risk is paramount because fabricated symptoms, drug interactions, or treatment recommendations could cause immediate patient harm (Thirunavukarasu et al., 2023; Lee et al., 2023). Interpretability is equally critical, as clinicians must understand the reasoning behind AI recommendations to appropriately trust and act upon them, and regulatory agencies require explanations for AI-driven decisions in patient care (Gerke et al., 2020). Privacy and security concerns are severe given the sensitivity of protected health information under HIPAA, GDPR, and similar regulations, with additional risks that language models may memorize and reveal patient data (Rieke et al., 2020; Dayan et al., 2021).

Regulatory compliance requirements are stringent for medical devices, and workflow integration must address alert fatigue and clinician acceptance (Rajkomar et al., 2018; Shickel et al., 2018). Data quality challenges include demographic underrepresentation that can produce biased predictions across patient populations (Obermeyer et al., 2019).

Genomics and variant analysis applications face distinctive challenges centered on privacy, generalization, and interpretability. Privacy concerns are particularly acute because genetic data is inherently identifiable and has implications extending to family members, raising unique consent and re-identification risks. Generalization across populations represents a critical concern because genomic models trained predominantly on European ancestry populations may perform poorly for underrepresented groups, potentially exacerbating health disparities (Obermeyer et al., 2019). Interpretability matters for variant classification because clinicians and genetic counselors must explain pathogenicity assessments to patients in the context of clinical decision-making. Hallucination risk is moderate but concerning when models confidently misclassify variants of uncertain significance. Data quality issues arise from heterogeneous variant annotation standards across databases and the challenge of rare variant interpretation with limited population frequency data.

Clinical trial design requires careful attention to data quality, regulatory compliance, and workflow integration. Data quality concerns focus on selection bias in historical trial data that could propagate to eligibility criteria optimization or endpoint selection. Regulatory compliance is critical because clinical trials operate under stringent ICH guidelines, and AI-assisted design decisions must satisfy regulatory scrutiny across multiple jurisdictions. Workflow integration challenges arise from the need to coordinate AI tools across sponsors, clinical research organizations, and trial sites with heterogeneous information systems. Interpretability requirements ensure that sponsors can justify AI-driven design decisions to regulators and ethics committees. Privacy considerations involve protecting participant data while enabling multi-site collaboration, though formal data sharing agreements typically address these concerns. Generalization challenges emerge when applying models trained on historical trials to novel therapeutic areas or patient populations.

7.2 Data quality, availability, and bias

The performance of LLMs depends critically on the quality, quantity, and representativeness of training data, and biomedical domains face particular challenges in each of these dimensions (Vamathevan et al., 2019; Obermeyer et al., 2019). Data scarcity presents a fundamental obstacle for rare diseases, novel therapeutic targets, and emerging pathogens where limited historical data is available for model training. Even for common diseases, labeled datasets for specific prediction tasks may be small, expensive to curate, or of uncertain quality due to noise in experimental measurements or inconsistencies in clinical documentation. Data bias represents a critical concern with implications for both scientific validity and health equity (Obermeyer et al., 2019; Char et al., 2018).

Training datasets may underrepresent certain patient populations, disease subtypes, or molecular classes, leading to models that perform poorly for these underrepresented groups. Historical biases encoded in clinical data, such as differential access to care or documentation practices that vary across providers and institutions, can propagate through model training to produce biased predictions. Studies have demonstrated that algorithms used in healthcare can exhibit racial bias, recommending different levels of care for Black and white patients with similar health needs (Obermeyer et al., 2019). Addressing these biases requires careful attention to training data composition, evaluation across demographic subgroups, and ongoing monitoring of deployed systems.

Bias in biomedical training corpora is especially acute for genomic and clinical models. Genome-wide association studies and reference panels remain dominated by individuals of European ancestry, so genomic LLMs such as DNABERT-2 and the Nucleotide Transformer inherit ancestry-skewed variant frequency distributions and may produce miscalibrated pathogenicity scores for African, Hispanic, and South Asian populations. Clinical LLMs trained on MIMIC-III/IV inherit the demographics of a single urban ICU and the documentation conventions of one institution, which limits their generalisability to outpatient, paediatric, and non-U.S. settings. Similarly, molecular databases such as ChEMBL over-represent kinases and GPCRs relative to less tractable targets, biasing generative chemistry models toward chemotypes that have already been thoroughly explored. Mitigation strategies include explicit re-weighting of under-represented strata, federated training across institutions to broaden patient representation, and routine subgroup evaluation as part of model release. These mitigations are necessary but not sufficient: most current biomedical LLM releases do not report demographic or chemotype-stratified performance, making it difficult to audit residual bias from the published evidence alone.

Closely related to bias are concerns about reproducibility and generalisability. Many recent LLM-based drug-discovery studies report state-of-the-art results on a single benchmark using a single random seed, with limited information about training data overlap, hyperparameter search, or held-out cold-split protocols. Performance on temporally split test sets (i.e., training on pre-2023 data and testing on 2024 entries) tends to be substantially lower than on standard random splits, indicating that some published gains partly reflect dataset leakage rather than genuine generalisation. The community has begun to converge on more rigorous evaluation through benchmarks such as Therapeutics Data Commons (TDC) 2.0, ChemBench, MOSES, and GuacaMol, but reproducibility remains uneven: closed-source models such as GPT-4, Med-PaLM 2, and Tx-LLM cannot be independently audited, and even open-source releases often omit the exact training-data snapshots and prompts used to obtain the reported numbers. Reviewers and readers should therefore interpret single-metric headlines with caution and prefer multi-benchmark, multi-seed comparisons whenever available.

The quality of biomedical data is often uncertain and variable. Experimental measurements contain noise from technical variability and biological heterogeneity. Clinical documentation may be incomplete, inconsistent, or ambiguous. Drug activity annotations in public databases may reflect different assay conditions that are not comparable. Label noise in training data can degrade model performance and lead to unreliable predictions. Careful curation, data cleaning, and uncertainty quantification are essential for training reliable models, but comprehensive quality assessment remains challenging at the scale of data used for modern LLMs. Privacy constraints limit access to clinical data needed for training and validating precision medicine applications. Electronic health records contain sensitive patient information protected by regulations including HIPAA in the United States and GDPR in Europe. Sharing clinical data across institutions for model development raises privacy concerns even when data is de-identified, as language models may memorize and reveal sensitive training information. These constraints create tension between the need for large, diverse training datasets and the imperative to protect patient privacy.

7.3 Model interpretability and explainability

LLMs are often characterized as black boxes whose decision-making processes are opaque and difficult to understand (Jiménez-Luna et al., 2020; Preuer et al., 2019). In healthcare and drug discovery, where decisions have significant consequences for patient outcomes and resource allocation, interpretability is not merely desirable but essential for multiple stakeholders. Clinicians must understand why a model makes specific recommendations to trust and act on them appropriately. Regulatory agencies require explanations for AI-driven decisions in drug development and clinical care. Researchers need to understand model reasoning to generate biological hypotheses and design experiments. And patients deserve transparency about how AI systems influence their care. Current approaches to interpretability provide partial solutions but fall short of fully transparent reasoning. Attention visualization shows which input elements the model attends to, but attention weights do not necessarily indicate causal importance for predictions (Rodríguez-Pérez and Bajorath, 2020). Feature attribution methods like SHAP values quantify the contribution of input features to predictions, but may not capture the complex interactions learned by deep models. Chain-of-thought prompting can elicit natural language explanations of model reasoning, but these explanations may not faithfully reflect the model’s internal computations. The gap between model explanations and actual reasoning processes limits the utility of current interpretability methods for building justified trust in model outputs.

Knowledge-grounded approaches offer improved transparency by explicitly connecting model outputs to external knowledge sources (Lewis et al., 2020; Zakka et al., 2024). Retrieval-augmented systems can cite the sources used to generate responses, enabling verification of factual claims. Knowledge graph-enhanced models can trace reasoning through explicit relationship chains. These approaches provide more auditable decision processes than purely parametric models, though they introduce additional complexity and potential failure modes in the retrieval and reasoning components.

7.4 Hallucination and factual reliability

Hallucination, the generation of plausible-sounding but factually incorrect information, represents a fundamental limitation of current LLMs with particularly serious implications for biomedical applications (Thirunavukarasu et al., 2023; Lee et al., 2023). In drug discovery, hallucinated mechanisms of action or pharmacokinetic properties could lead to misguided research directions. In clinical care, fabricated symptoms, drug interactions, or treatment recommendations could cause patient harm. Even citation of nonexistent studies or misrepresentation of research findings undermines scientific integrity and clinical decision making. The frequency and characteristics of hallucination vary across model architectures, training approaches, and application domains. Studies have documented that even state-of-the-art medical LLMs produce factually incorrect responses at non-negligible rates, including on high-stakes clinical questions (Singhal et al., 2023; Nori et al., 2023). Hallucinations can be particularly difficult to detect because they are often fluent and superficially plausible, potentially fooling both automated evaluation systems and human reviewers unfamiliar with the specific topic.

Approaches to mitigate hallucination include retrieval-augmented generation that grounds outputs in retrieved evidence, knowledge graph constraints that limit responses to supported facts, and uncertainty quantification that flags low-confidence predictions (Zakka et al., 2024; Xiong et al., 2024). However, these approaches add complexity and do not completely eliminate hallucination risk. Overconfident predictions, where models express high certainty about incorrect statements, represent a particular challenge because they provide false assurance that can override appropriate skepticism. Given current limitations, human oversight remains essential for biomedical applications of LLMs. Critical outputs should be verified by domain experts before acting on them. Systems should be designed to augment rather than replace human judgment, providing information and suggestions that clinicians or researchers evaluate in context. Clear communication about model limitations and uncertainty is essential for appropriate use.

7.5 Privacy and security

Healthcare data is among the most sensitive personal information, and its use in training and deploying LLMs raises significant privacy and security concerns (Gerke et al., 2020; Rieke et al., 2020; Dayan et al., 2021). Language models trained on clinical text can memorize and potentially reveal sensitive patient information, even from de-identified data. Adversarial attacks could potentially extract training data from models or manipulate model outputs in harmful ways. Compliance with healthcare privacy regulations adds complexity to model development and deployment. Federated learning offers a promising approach to training models on distributed clinical data without centralizing sensitive information (Rieke et al., 2020; Dayan et al., 2021; Pati et al., 2022). In federated learning, model parameters rather than raw data are shared across institutions, with updates computed locally and aggregated centrally. This approach enables collaborative model development while keeping patient data within institutional boundaries. Successful demonstrations in COVID-19 outcome prediction and cancer imaging show the feasibility of federated approaches for healthcare AI (Dayan et al., 2021; Pati et al., 2022).

Differential privacy provides formal guarantees about information leakage by adding calibrated noise to model training or outputs. However, the privacy-utility tradeoff means that stronger privacy guarantees typically come at the cost of reduced model performance. Finding appropriate balance between privacy protection and model utility requires careful analysis of specific application requirements and risk tolerances.

7.6 Computational resources and accessibility

Training and deploying large language models requires substantial computational resources that may limit accessibility (Touvron et al., 2023a). Pretraining large models requires expensive GPU clusters that are available only to major technology companies, well-funded research institutions, and pharmaceutical companies. Even inference with large models may require specialized hardware, limiting deployment in resource-constrained settings. The environmental impact of training and running large models raises sustainability concerns. Efforts to democratize access to LLM capabilities include release of open-source model weights, development of efficient inference methods, and cloud-based APIs that amortize infrastructure costs across many users (Touvron et al., 2023a; Touvron et al., 2023b). Parameter-efficient fine-tuning methods reduce the computational cost of adapting pretrained models to specific tasks. Model compression techniques including quantization and distillation create smaller, faster models suitable for deployment on limited hardware. However, fundamental tradeoffs between model size, performance, and efficiency mean that the most capable models remain computationally expensive. Accessibility concerns extend beyond computational resources to include expertise required for effective use of LLM technologies. Prompt engineering, model selection, fine-tuning, and evaluation all require skills that may not be present in all research groups or clinical settings. Democratizing access requires not only making models available but also developing tools, documentation, and training that enable effective use by non-experts.

7.7 Generalization and domain shift

LLMs trained on specific datasets or domains may not generalize well to new contexts, a phenomenon known as distribution shift or domain shift (Yazdani-Jahromi et al., 2022). Models trained on one patient population may perform poorly on others with different demographic characteristics, disease prevalence, or clinical practices. Models trained on historical data may become outdated as medical knowledge and treatment practices evolve. Models trained for one therapeutic area may not transfer to others with different molecular mechanisms or patient presentations. Cold-split evaluation scenarios, which assess performance on entities entirely absent from training data, reveal generalization challenges that random splits may obscure (Huang et al., 2021; Yazdani-Jahromi et al., 2022). For drug-target interaction prediction, performance on novel drugs or novel targets typically falls substantially below performance on the full test set. For clinical prediction, performance on patients from new institutions or time periods may differ from performance on held-out patients from the original data. These generalization gaps are particularly concerning for applications where the goal is to make predictions about novel entities.

Continual learning and periodic retraining can help maintain model currency as new data becomes available, but raise questions about validation requirements and change management. Online learning approaches that update models continuously based on new observations must balance adaptation to changing conditions with stability and safety requirements. Determining when model updates require revalidation and regulatory review remains an open question with significant practical implications.

7.8 Ethical and regulatory considerations

The deployment of AI in healthcare raises profound ethical questions about autonomy, accountability, equity, and the appropriate role of technology in medical decision making (Gerke et al., 2020; Char et al., 2018). Patients have the right to understand when AI systems influence their care and to make informed decisions about accepting AI-derived recommendations. When AI systems contribute to clinical decisions, determining accountability for errors becomes complex. The potential for AI to exacerbate health disparities through biased predictions or unequal access to AI-enhanced care must be actively addressed.

Regulatory frameworks for AI in healthcare are evolving rapidly but remain incomplete (Gerke et al., 2020). Traditional regulatory paradigms designed for static medical devices may not adequately address continuously learning AI systems. Requirements for clinical validation must balance the need for evidence of safety and efficacy with the pace of technological development. International variation in regulatory requirements creates challenges for global deployment of AI-based tools. Addressing these ethical and regulatory challenges requires collaboration among diverse stakeholders including AI developers, healthcare providers, patients, ethicists, regulators, and policymakers. Principles for responsible AI development, including transparency, fairness, accountability, and privacy, must be operationalized in specific technical and organizational practices. Ongoing monitoring and evaluation of deployed systems is essential to identify and address problems that emerge in real-world use.

In practice, LLM-based drug-discovery and clinical-decision tools must navigate multiple, partially overlapping regulatory regimes. In the United States, software intended to inform clinical management can fall under the U.S. Food and Drug Administration’s (FDA) Software as a Medical Device (SaMD) framework, with additional guidance for Predetermined Change Control Plans for adaptive AI/ML models. In the European Union, equivalent tools are regulated under the Medical Device Regulation (MDR) and increasingly under the EU AI Act, which classifies most clinical AI systems as high-risk and imposes documentation, post-market surveillance, and human-oversight obligations. For drug-discovery applications, evidence generated by LLM-based tools that contributes to investigational new drug (IND) submissions or marketing authorisations must satisfy ICH guidelines (e.g., ICH E9, E14, M11) and is subject to scrutiny by the FDA Center for Drug Evaluation and Research and the European Medicines Agency (EMA). LLM-assisted patient stratification and trial-eligibility tools additionally interact with HIPAA, GDPR, and emerging national health-data regulations. These regimes generally require traceable training data provenance, lock-in of model versions used for regulatory submissions, documented evaluation across relevant subpopulations, and demonstration that human decision-makers retain meaningful oversight, all of which are challenging to provide for very large, closed-source foundation models.

Beyond regulatory compliance, the clinical risks of deploying LLMs in patient-facing settings warrant explicit consideration. Fluent but factually incorrect outputs (hallucinations) can produce confidently wrong diagnoses, drug-dose recommendations, or interaction warnings that propagate downstream into orders and prescriptions. Subtle distributional shifts (e.g., a new electronic health-record vendor, a new ICD coding standard, or a change in laboratory reference ranges) can silently degrade performance long after deployment. Automation bias, in which clinicians over-trust LLM-generated suggestions, may also lead to harm even when the underlying model is well-calibrated on average. Mitigations include systematically restricting LLM outputs to advisory rather than autonomous roles, embedding clear uncertainty signals, requiring explicit clinician confirmation for high-risk actions, performing prospective silent-mode evaluation before go-live, and instituting post-deployment monitoring with predefined alert thresholds for drift in calibration, demographic subgroup performance, and incident rates.

7.9 Integration with clinical workflows

Successful deployment of LLMs in clinical settings requires seamless integration with existing workflows, electronic health record systems, and clinical practices (Rajkomar et al., 2018; Shickel et al., 2018). Technical interoperability with healthcare IT infrastructure must be achieved, including compliance with data standards and security requirements. User interfaces must be intuitive for clinicians with varying levels of technical sophistication. AI recommendations must be presented at appropriate points in clinical workflows without causing alert fatigue or disrupting established practices. Human factors in AI adoption are at least as important as technical capabilities. Clinicians may distrust AI recommendations due to concerns about accuracy, accountability, or erosion of professional autonomy. Conversely, over-reliance on AI could lead to automation complacency where clinicians accept recommendations without appropriate scrutiny. Achieving appropriate calibration of trust requires transparent communication about AI capabilities and limitations, training in effective AI use, and ongoing feedback about AI performance.

Workflow disruption represents a significant barrier to adoption even for systems with demonstrated accuracy. Implementation of new technologies in healthcare settings is complex and often unsuccessful, with many clinical decision support systems failing to achieve widespread adoption despite evidence of effectiveness. Success requires careful attention to implementation science principles including stakeholder engagement, pilot testing, iterative refinement, and sustained organizational support.

8 Future directions

The rapid evolution of large language models in drug discovery and precision medicine suggests several promising directions for future research and development that could substantially expand the impact of these technologies. This section outlines key opportunities and emerging trends that are likely to shape the field in the coming years, while acknowledging the uncertainties inherent in forecasting technological progress.

8.1 Advanced foundation models for biomedicine

The development of increasingly capable foundation models specifically designed for biomedical applications represents a primary direction for advancing the field (Wang et al., 2023; Singhal et al., 2023). Current biomedical language models are substantially smaller than the largest general-purpose models, and scaling to larger sizes may yield significant performance improvements if appropriate training data and computational resources are available. The development of truly multi-modal foundation models that seamlessly integrate text, molecular structures, protein sequences, genomic data, medical images, and clinical records within unified architectures could enable more comprehensive biomedical reasoning than current single-modality or loosely coupled multi-modal systems. Specialized pretraining objectives and architectures tailored to the unique characteristics of biomedical data may yield further improvements beyond simple scaling. For molecular data, pretraining objectives that explicitly capture chemical constraints, stereochemistry, and reaction feasibility could improve generation quality. For protein sequences, incorporating evolutionary and structural information into the pretraining process could enhance representation learning. For clinical text, pretraining that captures the temporal structure of patient trajectories and the hierarchical organization of medical knowledge could improve clinical prediction tasks. The emergence of models capable of generating not only text but also experimental protocols, simulation code, and robotic commands points toward increasingly autonomous research systems (Boiko et al., 2023). Future foundation models may serve as central hubs coordinating complex research workflows that span computational prediction, experimental design, data analysis, and hypothesis generation. Realizing this vision will require advances in reliability, transparency, and safety that ensure appropriate human oversight of autonomous research processes.

8.2 Federated and privacy-preserving approaches

Privacy-preserving methods for training models on sensitive healthcare data represent a critical enabling capability for clinical applications (Rieke et al., 2020; Dayan et al., 2021; Pati et al., 2022). Federated learning approaches that train models across distributed clinical datasets without centralizing patient information are maturing from proof-of-concept demonstrations to production-ready systems. Future work will likely explore more sophisticated federated architectures including hierarchical federation across institutions, asynchronous updates that accommodate varying institutional participation, and personalized models that capture institution-specific patterns while sharing common structure. Differential privacy and other formal privacy frameworks may become standard components of biomedical LLM development, providing rigorous guarantees about information leakage that enable deployment in regulated healthcare environments. Advances in efficient privacy-preserving computation, including secure multi-party computation and homomorphic encryption, could enable new forms of privacy-preserving inference and training that are currently computationally prohibitive. Balancing privacy protection with model utility remains an active research area with substantial room for improvement.

The integration of privacy-preserving approaches with federated learning could enable collaborative model development across international boundaries, overcoming regulatory barriers that currently fragment healthcare AI development. Such collaboration could produce models trained on more diverse patient populations, addressing representativeness concerns that limit current models trained primarily on data from high-income countries.

8.3 Explainable and trustworthy artificial intelligence

The development of more interpretable and trustworthy AI systems represents an essential priority for clinical adoption (Jiménez-Luna et al., 2020; Preuer et al., 2019). Future work must move beyond post hoc explanation methods toward architectures that are inherently more interpretable, perhaps through modular designs with explicit reasoning components or hybrid systems that combine neural networks with symbolic reasoning. Uncertainty quantification that accurately captures model confidence and identifies out-of-distribution inputs could help clinicians calibrate their trust in AI recommendations appropriately. Current LLMs often exhibit poor calibration, expressing unwarranted confidence in incorrect predictions. Developing models that reliably know what they do not know would substantially improve their utility for high-stakes decisions.

Formal verification methods that can prove properties of neural network behavior may eventually provide stronger guarantees about model safety and reliability. While current verification techniques scale poorly to large models, advances in this area could enable deployment of AI systems in applications where empirical testing alone is insufficient assurance.

8.4 Multi-agent systems and autonomous research

The development of multi-agent LLM systems that can collaborate on complex biomedical problems represents an exciting frontier (Boiko et al., 2023; Bran et al., 2024; Park et al., 2023; Wang et al., 2024). Future systems may employ diverse specialized agents for different aspects of drug discovery and clinical research, including literature synthesis, hypothesis generation, experimental design, data analysis, and manuscript preparation. Coordination mechanisms that enable effective collaboration among agents while maintaining scientific rigor and ethical constraints will be essential. The integration of LLM-based reasoning with physical laboratory systems through robotic experimentation could create closed-loop discovery processes that iterate between computational prediction and experimental validation with minimal human intervention (Boiko et al., 2023). Such systems could accelerate scientific discovery by enabling rapid exploration of hypothesis spaces that would be impractical for human researchers. However, ensuring appropriate human oversight and maintaining scientific standards in increasingly autonomous research processes poses significant governance challenges.

Agent architectures that can learn and improve through experience, accumulating knowledge and refining strategies across research projects, could exhibit a form of scientific creativity that goes beyond pattern matching on training data. Whether such systems can generate genuinely novel scientific insights or are fundamentally limited to recombining existing knowledge remains an open question with profound implications for the future of scientific research.

8.5 Personalized and adaptive models

Future LLMs may be capable of continuous learning and adaptation to individual patient characteristics, enabling truly personalized medicine approaches that evolve with each patient interaction (Cui et al., 2024; Lotfollahi et al., 2019). Patient-specific models that incorporate an individual’s complete medical history, genomic profile, and lifestyle factors could provide highly tailored predictions and recommendations. Such personalization raises questions about validation requirements, as standard clinical trial methodologies may not be applicable to interventions that differ for each patient. Adaptive models that update based on treatment outcomes could learn which interventions work best for specific patient types, accumulating institutional knowledge that improves care quality over time. However, ensuring that adaptive systems remain safe and effective as they evolve requires novel approaches to monitoring and quality assurance that are not yet well established.

The integration of patient preferences and values into AI-driven decision support could enable shared decision making that respects patient autonomy while leveraging AI capabilities. Systems that can explain treatment options in terms patients understand and incorporate patient preferences into recommendations would represent a significant advance beyond current clinical decision support approaches.

8.6 Real-time clinical integration

Integration of LLMs into clinical workflows with real-time processing capabilities and seamless interoperability with electronic health record systems represents a near-term priority for translation (Rajkomar et al., 2018; Shickel et al., 2018). Efficient inference methods that enable responsive interaction without specialized hardware could facilitate deployment across diverse clinical settings. Standardized interfaces between AI systems and clinical information technology infrastructure would reduce implementation barriers and enable more rapid adoption. Edge deployment of compact, efficient models could bring AI capabilities to resource-limited settings including community health centers, rural hospitals, and low-income countries where centralized cloud computing may not be feasible or affordable. Developing models that maintain acceptable performance while fitting within severe computational constraints represents an important technical challenge with significant equity implications.

Real-time monitoring of patient status through continuous analysis of physiological data streams, laboratory results, and clinical documentation could enable earlier detection of deterioration and more timely intervention. Integration of such monitoring with clinical alerting systems requires careful attention to alert fatigue and workflow integration to ensure that AI-generated alerts are actionable and effective.

8.7 Regulatory science and validation frameworks

The development of appropriate regulatory frameworks and validation methodologies for AI in healthcare represents a critical enabler for clinical deployment (Gerke et al., 2020). Standardized benchmarks that reflect clinically meaningful outcomes, rather than purely technical metrics, would provide more relevant assessments of clinical utility. Prospective clinical trials of AI-based interventions, while resource-intensive, may be necessary to establish evidence of benefit for high-stakes applications. Regulatory approaches that accommodate continuously learning systems while maintaining safety standards are needed to enable deployment of adaptive AI. Current regulatory frameworks largely assume static devices that are validated once and deployed without change, an assumption that does not apply to systems designed to improve with experience. Developing appropriate oversight mechanisms for dynamic AI systems is an active area of regulatory science.

International harmonization of regulatory requirements could reduce barriers to global deployment of AI-based medical technologies. Currently, varying requirements across jurisdictions create complexity for developers and may delay access to beneficial technologies in some regions. Efforts toward convergence on core principles while respecting legitimate national differences could accelerate safe adoption worldwide.

8.8 Interdisciplinary collaboration and workforce development

Realizing the potential of LLMs in drug discovery and precision medicine requires sustained interdisciplinary collaboration among computer scientists, biologists, chemists, clinicians, ethicists, and policymakers (Wang et al., 2023). Building teams with the diverse expertise needed to develop, validate, and deploy biomedical AI systems represents an organizational challenge that extends beyond purely technical considerations. Workforce development programs that train researchers and clinicians in AI methods, and conversely train AI researchers in biomedical science, could expand the talent pool available to work on these problems. Educational curricula that prepare students for careers at the intersection of AI and biomedicine are beginning to emerge but require continued development and scaling. Community resources including shared datasets, benchmark tasks, and open-source software tools accelerate progress by enabling researchers to build on each other’s work rather than duplicating effort. Continued investment in such resources, along with incentive structures that reward sharing and collaboration, could substantially accelerate the field.

9 Conclusion

The body of work surveyed here shows that large language models have moved from proofs of concept to a set of practical instruments that intersect nearly every stage of drug discovery and precision medicine. The more interesting question is no longer whether these models can learn from molecules, sequences, and clinical text, but how the field should invest to make their outputs dependable enough for decisions that affect patients.

Several priorities follow from the evidence assembled in this review. First, evaluation must shift from in-distribution benchmark accuracy toward prospective, cold-split, and real-world validation, particularly for drug-target prediction, generative chemistry, and clinical decision support, where reported gains frequently shrink under realistic conditions. Second, grounding mechanisms such as retrieval augmentation, knowledge-graph constraints, and calibrated uncertainty should be treated as first-class design requirements rather than optional add-ons, given how consequential hallucination and overconfidence can be in biomedical use. Third, privacy-preserving training regimes such as federated and differentially private learning need to mature alongside governance frameworks that can keep pace with continually updated models.

Looking ahead, the most productive trajectories appear to be multi-agent systems that couple LLM reasoning with experimental automation, multi-modal foundation models that integrate molecular, cellular, imaging, and clinical data, and personalized models that adapt over a patient’s trajectory. Realizing these trajectories will require sustained collaboration among computational scientists, chemists, clinicians, regulators, and ethicists, and an honest accounting of both achievements and failure modes. If pursued with that balance of ambition and rigor, LLMs can plausibly become durable partners in biomedical discovery rather than a passing wave of benchmark results.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Glossary

Artificial intelligence and machine learning

AI Artificial Intelligence API Application Programming Interface AUROC Area Under the Receiver Operating Characteristic Curve AUPRC Area Under the Precision-Recall Curve BERT Bidirectional Encoder Representations from Transformers BPE Byte-Pair Encoding CNN Convolutional Neural Network GNN Graph Neural Network GPT Generative Pre-trained Transformer GPU Graphics Processing Unit LLM Large Language Model LoRA Low-Rank Adaptation MAE Mean Absolute Error ML Machine Learning MLM Masked Language Modeling MPNN Message Passing Neural Network NER Named Entity Recognition NLP Natural Language Processing OOD Out-of-Distribution QA Question Answering RAG Retrieval-Augmented Generation RLHF Reinforcement Learning from Human Feedback RMSE Root Mean Squared Error RNN Recurrent Neural Network SOTA State-of-the-Art VAE Variational Autoencoder XAI Explainable Artificial Intelligence

Drug discovery and chemistry

ADMET Absorption, Distribution, Metabolism, Excretion, and Toxicity DDI Drug-Drug Interaction DDR1 Discoidin Domain Receptor 1 DTI Drug-Target Interaction EGFR Epidermal Growth Factor Receptor GENTRL Generative Tensorial Reinforcement Learning GPCR G Protein-Coupled Receptor JAK Janus Kinase QSAR Quantitative Structure-Activity Relationship SMILES Simplified Molecular Input Line Entry System TNIK TRAF2 and NCK-Interacting Kinase

Clinical and regulatory

AML Acute Myeloid Leukemia CDSS Clinical Decision Support System ECG Electrocardiogram ECM Extracellular Matrix EHR Electronic Health Record EMA European Medicines Agency FDA Food and Drug Administration GDPR General Data Protection Regulation HIPAA Health Insurance Portability and Accountability Act ICD International Classification of Diseases ICH International Council for Harmonisation IND Investigational New Drug MDR Medical Device Regulation SaMD Software as a Medical Device USMLE United States Medical Licensing Examination VUS Variants of Uncertain Significance

Genomics and molecular biology

BRCA1/2 Breast Cancer gene 1 and 2 CYP2C19 Cytochrome P450 2C19 CYP2D6 Cytochrome P450 2D6 DNA Deoxyribonucleic Acid ESM Evolutionary Scale Modeling mRNA Messenger Ribonucleic Acid pLDDT Predicted Local Distance Difference Test RNA Ribonucleic Acid scGen single-cell Gene expression Generator

Databases and resources

ChEMBL Chemical European Molecular Biology Laboratory database MIMIC Medical Information Mart for Intensive Care PubMed Public Medicine database TDC Therapeutics Data Commons UniProt Universal Protein Resource

Footnotes

Edited by: Rodolpho C. Braga, InsilicAll, Brazil

Reviewed by: Ruchi Yadav, Amity University, India

Kumararaja K., New Prince Shri Bhavani College of Engineering and Technology, India