Toward explainable and adaptive artificial intelligence systems for antimicrobial peptide discovery under resistance pressure

1 Introduction

Antimicrobial resistance (AMR) represents a critical and escalating global health challenge, driven by the rapid emergence and dissemination of multidrug-resistant pathogens (Murugaiyan et al., 2022; Murray et al., 2022; Tang et al., 2023). Despite sustained efforts in antimicrobial development, current strategies remain insufficient to address the adaptive and evolutionary nature of resistance, revealing fundamental limitations in how therapeutic agents are designed and optimised under sustained selective pressure (Kumavath et al., 2025).

Within this landscape, therapeutic peptides have emerged as a promising alternative due to their structural diversity, tunable physicochemical properties, and capacity to engage multiple microbial targets (Girdhar et al., 2024; Aghamiri et al., 2021). However, their development remains constrained by critical challenges, including toxicity, instability, and context-dependent loss of activity (Wu et al., 2022; Zheng et al., 2025). Moreover, contrary to early assumptions, antimicrobial peptides do not inherently prevent resistance, as microorganisms can evolve diverse adaptive mechanisms against them (Andersson et al., 2016; Lazzaro et al., 2020). These limitations highlight a persistent disconnect between peptide design strategies and the complex biological environments in which antimicrobial efficacy must be achieved.

Artificial intelligence (AI) has been widely adopted to address this gap by enabling large-scale exploration of peptide sequence space and accelerating candidate prioritisation (Mulat et al., 2025; Brizuela et al., 2025). Predictive and generative models have demonstrated strong performance under controlled benchmarking settings, but their practical impact in resistance-aware discovery remains limited (Ahvar et al., 2025). Many current systems continue to focus primarily on antimicrobial peptide (AMP) identification tasks, often framed as binary AMP versus non-AMP classification problems. While useful for early-stage screening, such predictions provide limited translational value in the absence of information regarding pathogen-specific activity, minimum inhibitory concentrations (MICs), antimicrobial spectrum, toxicity, stability, and efficacy under clinically relevant conditions (Wan et al., 2024). At the same time, many AI-driven pipelines primarily focus on predictive performance, whereas aspects such as interpretability, uncertainty quantification, and integration with experimental feedback remain less systematically incorporated, limiting their ability to adapt to the dynamic and non-stationary nature of antimicrobial resistance (Cui et al., 2025; Yakobi and Nwodo, 2025).

These limitations do not primarily reflect deficiencies in modelling capacity, but rather a structural mismatch between how AI methods are currently organised and the requirements of resistance-driven antimicrobial discovery (Roque-Borda et al., 2026). Prediction, generation, interpretation, and experimental validation are typically treated as loosely connected tasks, leading to fragmented workflows that limit the ability of computational systems to support adaptive optimisation, uncertainty-aware decision-making, and continuous learning from experimental evidence (Brizuela et al., 2025). Although important advances have demonstrated individual elements of iterative discovery, including explainable prediction, machine learning-guided optimisation, uncertainty-aware prioritisation, and design–test–learn strategies, these components are rarely integrated within coherent frameworks capable of supporting transparent, adaptive, and feedback-driven peptide development (Bae et al., 2025). Under such conditions, high-performing models often fail to translate into effective decision-support systems in real-world discovery settings. These challenges are synthesised in Figure 1, which contrasts the global burden of antimicrobial resistance, the promise and limitations of therapeutic peptides, and the fragmented nature of current AI-driven discovery pipelines.

In this Perspective, we propose that the next step for AI-assisted antimicrobial peptide discovery lies not in developing increasingly sophisticated predictive or generative models, but in integrating interpretability, uncertainty awareness, biological constraints, and experimental feedback into coherent discovery systems. While previous perspectives have primarily reviewed methodological advances (Brizuela et al., 2025; Cesaro et al., 2025; Wang J. et al., 2025; Wang B. et al., 2025), we focus on how these complementary elements can be combined into adaptive discovery workflows. We therefore present a conceptual framework that reframes antimicrobial peptide discovery as an iterative design–test–learn process, integrating explainable prediction, uncertainty quantification, biologically constrained generation, and feedback-driven refinement. By illustrating the practical integration of these components into transparent and experimentally grounded workflows, we show how existing methodologies can support more adaptive and informed decision-making under the evolving challenges of antimicrobial resistance. Finally, we demonstrate these principles through a resistance-aware antimicrobial peptide discovery example that integrates interpretability, uncertainty estimation, rule extraction, and experimental feedback for candidate prioritisation against multidrug-resistant pathogens.

2 Limitations of current AI-driven peptide discovery pipelines

Despite substantial advances in predictive modelling and generative design for antimicrobial peptides (Wang et al., 2022; Hashemi et al., 2024; Wan et al., 2022; Tsai et al., 2024), important challenges remain regarding model interpretability, uncertainty quantification, biologically informed candidate generation, and the effective incorporation of experimental feedback into discovery workflows (Cesaro et al., 2025). These limitations reduce the ability of current pipelines to support reliable candidate prioritisation and iterative optimisation, particularly when translating computational predictions into experimentally actionable decisions.

2.1 Fragmented workflows and single-task optimisation

The current landscape of AI-driven peptide discovery is characterised by a proliferation of specialised tools targeting isolated objectives. Data resources such as APD6 (Wang et al., 2026), DRAMP 4.0 (Ma et al., 2025), DBAASP v3 (Pirtskhalava et al., 2021), and CAMPR4 (Gawde et al., 2023) provide experimentally validated sequences and expanded annotations on toxicity, stability, structure, and pathogen-specific activity. However, linking them consistently to unified, traceable peptide representations across different contexts may require additional curation, particularly when aiming to support dynamic modelling of antimicrobial resistance (Ghislat et al., 2024). At the modelling level, most predictive approaches remain focused on single endpoints, typically antimicrobial activity classification, with limited consideration of pharmacological properties, toxicity, strain-specific efficacy, or resistance-related constraints (Brizuela et al., 2025).

Antimicrobial peptide development extends far beyond AMP identification. Clinically relevant candidates must simultaneously satisfy multiple and often competing objectives, including antimicrobial potency, pathogen specificity, MIC, toxicity, stability, manufacturability, pharmacokinetic behaviour, and robustness against resistance emergence (Wan et al., 2024). Consequently, optimising a single endpoint in isolation provides only a partial representation of therapeutic potential and may lead to candidates that perform well computationally while failing during downstream experimental evaluation or translational development (Meng, 2025).

Recent advances illustrate both progress and limitations. Models such as deep-AMPpred and iAMP-SeE extend traditional classification frameworks towards activity-spectrum prediction, while ANIA enables MIC prediction against clinically relevant pathogens (Zhao et al., 2025; Chen Q. et al., 2026; Chiu et al., 2026). Similarly, generative frameworks based on variational autoencoders, generative adversarial networks, transformers, and diffusion models (e.g., (Zhao et al., 2025a), (Ying et al., 2025), AMPGen (Jin et al., 2025), Khan et al. (Khan et al., 2026), and deepAMP (Li et al., 2024)) have substantially expanded the exploration of peptide sequence space and introduced conditioning mechanisms based on physicochemical properties or activity-related targets. Nevertheless, these approaches remain largely optimised around individual objectives or surrogate metrics, limiting their ability to support integrated decision-making across multiple biological and pharmacological constraints.

More recent efforts have begun to approximate design–test–learn paradigms. Multi-objective optimisation frameworks such as ApexAmphion (Cao et al., 2025), ProDCARL (Sheng et al., 2026), and HMAMP (Wang et al., 2025) integrate predictive modelling with optimisation strategies coupled to reinforcement learning. In parallel, Diff-AMP (Wang R. et al., 2024) automates generation, identification, attribute prediction, and iterative optimisation. Mishra et al. (Mishra et al., 2025) demonstrated that machine learning-guided sequence optimisation combined with experimental validation can iteratively refine peptide variants for improved antimicrobial activity. These studies illustrate that elements of integrated discovery are already emerging within the field. However, most implementations remain focused on specific stages of the discovery process and generally rely on predefined objectives or static conditioning signals. Prediction, generation, optimisation, and validation are frequently developed as partially independent components, limiting the emergence of coordinated and continuously evolving discovery workflows. Consequently, experimental outcomes rarely propagate systematically across the entire discovery process to guide future candidate generation, prioritisation, or model refinement (Littmann et al., 2020).

As a consequence, current workflows often operate as sequential filtering pipelines, limiting the integration of biological, pharmacological, and resistance-related constraints within a unified decision-making framework. This fragmentation increases technical overhead, limits reproducibility, and complicates the incorporation of new experimental evidence into the design process.

2.2 Black-box predictions and limited biological interpretability

A second critical limitation concerns the interpretability of many AI-driven models. Deep learning architectures achieve strong predictive performance, but often function as black boxes, offering limited insight into the biological determinants underlying antimicrobial activity (Alizadehsani et al., 2024; Orsi and Reymond, 2024). Although interpretable models and explainability frameworks have been proposed, most studies continue to rely on post hoc approaches such as SHAP (Lundbe et al., 2017) or LIME (Ribeiro et al., 2016) to estimate feature importance in predictive models, including E-CLEAP (Wang, 2024), AMP-meta (Tsai et al., 2024), pAV-PSSMDWT-EnC (Akbar et al., 2022), XAI-INVENT (Sharma et al., 2025), the explainable machine learning framework proposed by Bhatnagar et al. (Bhatn et al., 2024), and related approaches. These methods provide useful feature-level attributions, yet their explanations remain strongly influenced by the model architecture and input representation employed. Similarly, gradient-based approaches have been used to analyse residue-level contributions and guide sequence optimisation (Wang B. et al., 2025). While such methods can provide valuable local explanations, they are often sensitive to model architecture, training conditions, and input perturbations, and rarely reveal stable global patterns or mechanistic determinants of antimicrobial activity (Alvarez-Melis and Jaakkola, 2018; Laugel et al., 2019).

In practice, this limitation reduces the utility of predictive models for rational peptide engineering. Current machine learning approaches can effectively identify antimicrobial peptide candidates, yet they often provide limited mechanistic understanding of why particular sequences are predicted to be active. DLFea4AMPGen (Gao et al., 2025) and KMLEE (Lee et al., 2026) apply SHAP and attention mechanisms to quantify amino acid contributions, but these analyses remain largely retrospective and are not incorporated into iterative design cycles. Bhangu et al. (Bhangu et al., 2025) and Padi et al. (Padi et al., 2025) identify sequence motifs and relate them to physicochemical properties to rank candidates without integrating these insights into the generative process. Similarly, CoLPAT-AMP (Salimi and Lee, 2026) combines candidate generation with SHAP-based attribution, but the resulting explanatory signals are primarily used for post-generation interpretation and are not systematically integrated into sequence redesign, candidate optimisation, or hypothesis generation.

An additional challenge arises from the widespread use of complex sequence representations, particularly embedding-based approaches derived from deep neural networks (Weissenow and Rost, 2025; Hunklinger and Ferruz, 2025). Although these representations frequently improve predictive performance (Liang et al., 2025), they introduce an additional layer of abstraction that decouples model outputs from interpretable biochemical features (Medina-Ortiz et al., 2025). Consequently, even when attribution methods are applied, the resulting explanations can be difficult to translate into actionable design decisions because the underlying feature space lacks direct biological meaning.

As a result, current explainability strategies remain disconnected from the discovery loop (Hunklinger and Ferruz, 2025). Interpretability is rarely operationalised as a component of candidate generation (Bhangu et al., 2025; Padi et al., 2025), candidate selection (Rashid et al., 2026), or optimisation (Sheng et al., 2026), and is instead commonly treated as a secondary diagnostic layer applied after prediction. This limits the ability of AI-driven systems to support mechanistically informed design, constrains their utility in experimental settings, and reduces confidence in model-guided prioritisation under resistance-aware conditions. Future discovery systems should therefore treat interpretability not merely as a post hoc explanatory tool, but as an active source of design constraints, optimisation objectives, candidate prioritisation criteria, and experimentally testable hypotheses throughout the discovery cycle.

2.3 Static pipelines and decoupling from experimental feedback

Beyond fragmentation and limited interpretability, most AI-driven peptide discovery pipelines remain fundamentally static (Chang et al., 2022). Models are typically trained on in vitro datasets and subsequently deployed as fixed predictors, with limited incorporation of experimental feedback for iterative refinement (Wu et al., 2025).

This limitation is particularly pronounced in antimicrobial peptide discovery, where discrepancies between in vitro and in vivo performance are common (Cesaro et al., 2025). Peptides predicted to be highly active may fail under physiological conditions because of degradation, serum binding, immune interactions, clearance effects, or altered pharmacokinetics (Meng, 2025). Such failures highlight a critical disconnect between computational prediction and biological reality that remains insufficiently addressed by current modelling frameworks (Cesaro et al., 2025).

This decoupling is not only methodological but also organisational. Computational and experimental efforts are frequently conducted by separate groups operating under different constraints, timescales, and throughput capacities (Littmann et al., 2020). AI systems can generate thousands or even millions of candidate sequences, whereas only a small fraction can realistically be synthesised and experimentally evaluated, making prioritisation under uncertainty a central bottleneck in antimicrobial peptide discovery (Ji et al., 2024; Meng, 2025).

Under these conditions, candidate prioritisation is best viewed as a decision-making problem that requires the integration of predictive, biological, and experimental evidence. Uncertainty quantification methods such as conformal prediction, Bayesian approaches, and deep ensembles provide complementary mechanisms for estimating predictive confidence, identifying regions of model uncertainty, and selecting candidates that maximise expected information gain under limited experimental budgets (Herrmann et al., 2021; Hie et al., 2021; He et al., 2026). These approaches are particularly relevant in antimicrobial peptide discovery, where positive examples are often scarce, negative datasets may be imperfectly curated, and experimental throughput remains substantially lower than computational generation capacity.

Although tools such as AutoPeptideML facilitate model development and accessibility (Fernández-Díaz et al., 2024), they do not resolve the underlying challenge of weak feedback integration. More fundamentally, most pipelines lack mechanisms for continuously updating predictive or generative models based on newly acquired experimental evidence. While active learning (Barrett and White, 2020), reinforcement learning (Pandey et al., 2025), adaptive retraining strategies (Soto-Garcia et al., 2026), and uncertainty-guided acquisition functions (Chen L. et al., 2026) have been extensively explored in the broader machine learning literature, their adoption within antimicrobial peptide discovery remains relatively limited (Meng, 2025). In practice, experimental results are more commonly used for retrospective validation than for guiding model updates, candidate selection, or generative refinement across successive design cycles, reinforcing the static nature of current discovery pipelines.

These limitations reflect a fundamental mismatch between current AI-driven pipelines and the dynamic nature of antimicrobial resistance. Under sustained selective pressure and continuously evolving resistance mechanisms, static and weakly integrated workflows are insufficient to support effective peptide design, prioritisation, and optimisation (Choudhury et al., 2025). Addressing this challenge will likely require a more systematic integration of computational and experimental components through iterative design–test–learn cycles, uncertainty-aware candidate selection strategies, active learning frameworks that maximise experimental information gain, and adaptive retraining procedures capable of incorporating newly acquired biological evidence.

3 Methodological pillars for next-generation peptide AI systems

The limitations discussed in the previous section indicate that advancing resistance-aware antimicrobial peptide discovery requires more than increasingly sophisticated predictive models. It demands discovery systems that combine decision transparency, biologically grounded exploration, and feedback-driven learning. In this Perspective, we frame these requirements as a set of complementary design principles governing the interaction of predictive, generative, and interpretative components under evolving antimicrobial resistance pressures.

This framework is organised around three interdependent pillars, each addressing a distinct but interconnected limitation of current AI-driven peptide discovery pipelines. They promote discovery systems that are transparent, adaptable, and responsive to experimental feedback. Figure 2 summarises these pillars and their interactions. In a representative workflow, predictive models identify promising yet uncertain regions of sequence space, constrained generative models explore these regions under biological and physicochemical constraints, and experimental observations are incorporated through iterative design–test–learn cycles to refine subsequent predictions and designs. These pillars provide a foundation for integrated and feedback-driven antimicrobial peptide discovery.

3.1 Explainable and uncertainty-aware predictive modelling

In many AI-driven antimicrobial peptide discovery pipelines, predictive models are primarily used as ranking engines that assign scores or class probabilities to large candidate sets (Das et al., 2021; Ma et al., 2022; Torres et al., 2025; Zhao et al., 2025b). While effective for coarse filtering, such outputs provide limited insight into the biological or physicochemical determinants underlying antimicrobial activity against resistant pathogens.

Explainable predictive modelling addresses this limitation by providing interpretable signals that connect model outputs to sequence-level and physicochemical determinants (Wang L. et al., 2024). In practice, this requires combining local attribution methods (e.g., gradient-based or perturbation-based approaches) with global interpretability strategies that characterise model behaviour across sequence space (Medina-Ortiz et al., 2025). These signals can be operationalised to guide sequence modification, identify motifs associated with activity or toxicity, and define constraints for generative models.

Interpretability should be integrated throughout the design workflow as a source of actionable biological insight. For example, attribution maps can be used to propose targeted residue substitutions, while counterfactual analysis can identify minimal sequence changes that alter predicted activity (Chou et al., 2022). This enables explanatory information to guide sequence optimisation, candidate refinement, and hypothesis generation.

Uncertainty awareness constitutes a complementary requirement. Antimicrobial peptide datasets are often sparse and biased, particularly for resistant phenotypes (Medina-Ortiz et al., 2026), making point predictions unreliable under distributional shift (Wan et al., 2024). Methods such as deep ensembles, Bayesian neural networks, and conformal prediction provide complementary mechanisms for quantifying predictive uncertainty and supporting risk-informed decision-making (Majlatow et al., 2025). While conformal prediction offers distribution-free confidence guarantees that are particularly valuable in low-data settings (Vellore and Jha, 2026), ensemble-based approaches can identify regions of elevated uncertainty and support candidate ranking under distributional shift (Abdullah et al., 2024). Integrated within active learning frameworks, these estimates can further prioritise sequences expected to maximise experimental information gain (Barrett and White, 2020).

Explainability and uncertainty can transform predictive models from static scoring functions into adaptive decision-support components capable of guiding design, prioritisation, and iterative refinement under resistance-driven conditions.

3.2 Conditional and rule-aware generative peptide design

Generative models enable exploration of peptide sequence space beyond known examples (Torres et al., 2025). However, unconstrained generation is insufficient in resistance-driven discovery, as it frequently produces sequences that violate biological or physicochemical constraints (Chang et al., 2022).

Conditional generative design provides partial control by directing generation toward predefined objectives such as MIC or activity spectrum (Zhao et al., 2025b). However, conditioning alone does not guarantee biological plausibility or resistance robustness. Clinically relevant peptide optimisation requires balancing multiple therapeutic objectives simultaneously, including antimicrobial potency, spectrum of activity, toxicity, stability, pharmacokinetic behaviour, manufacturability, and robustness against resistance emergence (Wan et al., 2024; Brizuela et al., 2025). Consequently, successful peptide design requires a multi-objective optimisation framework capable of balancing antimicrobial efficacy, toxicity, stability, pharmacological properties, and resistance-related constraints.

Rule-aware generative frameworks address this limitation by incorporating constraints directly into the generative process (Cheng et al., 2026). These constraints can be operationalised at different levels depending on their role in the design objective. In many cases, physicochemical properties such as sequence length, charge distribution, or compositional balance are treated as soft constraints, implemented through multi-objective optimisation or penalty functions that guide exploration without strictly restricting it (Zhao et al., 2025b). In contrast, biologically critical criteria, such as toxicity thresholds, structural validity, or minimum activity requirements, can be enforced as hard constraints that define admissible regions of sequence space (Cheng et al., 2026).

From an implementation perspective, these constraints can be integrated during generation through conditional modelling or constrained decoding, or applied post hoc through filtering and re-ranking strategies. In reinforcement learning settings, both hard and soft constraints can be encoded within reward functions, enabling adaptive optimisation across competing objectives (Yang et al., 2025).

In antimicrobial resistance contexts, such constraints could explicitly encode resistance-relevant properties, including membrane disruption mechanisms, stability against proteolysis, reduced susceptibility to efflux, or activity against specific resistant phenotypes. Embedding these considerations within the generative process reframes peptide design as a multi-objective optimisation problem that simultaneously balances antimicrobial efficacy, toxicity, stability, pharmacological properties, and resistance robustness. Such an approach guides sequence exploration toward biologically meaningful objectives, promoting candidate designs that are both biologically plausible and better aligned with resistance-aware discovery goals.

3.3 Orchestrated modular systems and design–test–learn principles

The complexity of antimicrobial peptide discovery under resistance pressure exposes the limitations of monolithic computational models (Brizuela et al., 2025). Predictive evaluation, generative exploration, and interpretability impose distinct requirements, motivating a shift toward modular system architectures (Lavecchia, 2025; Soto-Garcia et al., 2026).

In this framework, peptide discovery is structured as the interaction of predictive, generative, and interpretability-oriented modules. Predictive modules evaluate candidates under uncertainty, generative modules propose new sequences within constrained design spaces, and interpretability modules contextualise outputs in biological terms (Szymczak et al., 2025).

Operationally, orchestration can be implemented through iterative loops in which predictive models score candidate sequences, generative models update proposals based on constraints and feedback, and selection strategies prioritise candidates according to uncertainty, diversity, novelty, and predicted performance. These loops can be formalised using active learning strategies, reinforcement learning policies, or multi-objective optimisation frameworks (Castle et al., 2024; Das et al., 2021).

Emerging systems illustrate partial implementations of this paradigm. M3-CAD integrates multimodal representations and multi-objective prediction (Li et al., 2026), BioAMPify employs hierarchical predictive modules for refined prioritisation (Rashid et al., 2026), and EvoGradient demonstrates iterative optimisation guided by interpretability signals (Wang B. et al., 2025). These studies demonstrate that elements of integrated discovery are already emerging within the field. Nevertheless, resistance dynamics, uncertainty-aware acquisition, adaptive retraining, and systematic feedback integration are not yet consistently incorporated across all stages of the discovery process.

Design–test–learn principles provide the governing logic for such systems (Castle et al., 2024). Computational design generates candidate peptides, experimental testing evaluates their performance under selective pressure, and learning mechanisms update models based on observed outcomes (Das et al., 2021). Under antimicrobial resistance, this loop must explicitly account for evolving pathogen responses, shifting activity landscapes, and uncertainty in experimental measurements.

In practice, design–test–learn cycles can be implemented through workflows in which candidate selection is guided by uncertainty-aware acquisition functions, experimental results are incorporated through adaptive retraining, and generative models are updated using reinforcement learning or feedback-conditioned optimisation (Mishra et al., 2025). This enables discovery systems to continuously incorporate new biological evidence and refine future prediction, prioritisation, and design decisions, aligning computational discovery with the dynamic nature of antimicrobial resistance (Castle et al., 2024). More broadly, these workflows provide a practical mechanism for transforming experimental observations into actionable design knowledge, allowing discovery systems to progressively improve through iterative interaction with evolving biological environments (Das et al., 2021).

4 Toward AI-native peptide discovery ecosystems

Building on the methodological pillars outlined above, this section examines how predictive, generative, and interpretability-oriented components can be coordinated within adaptive antimicrobial peptide discovery workflows. Here, AI-native discovery denotes the integration of computational modelling, experimental feedback, and iterative learning into a unified design–test–learn process capable of responding to evolving resistance pressures.

4.1 Process-oriented organisation for integrated discovery

In resistance-aware discovery settings, biological targets, evaluation criteria, and experimental constraints evolve continuously, rendering static workflows increasingly ineffective (Cunha-Ferreira et al., 2025). Predictive performance achieved under fixed conditions provides limited value if computational systems cannot revise their reasoning as resistance mechanisms emerge or experimental contexts shift (Ghislat et al., 2024).

AI-native discovery ecosystems address this limitation by organising prediction, generation, evaluation, and learning as interdependent functions operating within a shared process. In these systems, behaviour emerges from the coordinated exchange of information between modules, the propagation of uncertainty across decision points, and the iterative refinement of design hypotheses as new evidence becomes available (Wu et al., 2025; Chang et al., 2022).

This process-oriented perspective enables methodological coherence across heterogeneous discovery scenarios without requiring commitment to specific algorithms or software implementations. Instead, it establishes a set of operational invariants, including traceability of decisions, explicit treatment of uncertainty, and continuous integration of new information into the modelling process.

In practice, achieving these objectives requires infrastructure capable of supporting reproducible and auditable workflows. Deterministic data processing, versioned datasets, model lineage tracking, and reproducible execution environments constitute core elements of the discovery ecosystem (Herrera-Rocha et al., 2025; Wilkinson et al., 2016; Heil et al., 2021). These components ensure that predictions, design decisions, and experimental outcomes can be systematically traced, reproduced, and re-evaluated as biological knowledge evolves.

Such organisation is particularly important in antimicrobial peptide discovery campaigns spanning multiple pathogens, experimental conditions, and resistance scenarios (Brizuela et al., 2025). By explicitly defining how decisions are generated, evaluated, and updated, AI-native systems reduce the risk of optimisation being driven by artefacts of static datasets or benchmark-specific biases and instead promote sustained, evidence-driven refinement of candidate peptides (Ghislat et al., 2024; Herrera-Rocha et al., 2025).

4.2 Functional roles within AI-native discovery ecosystems

Within AI-native discovery ecosystems, computational components are most effectively characterised by the functions they fulfil within the discovery process, independent of specific implementation choices (Gridach et al., 2025). This abstraction clarifies how different forms of computational reasoning contribute to antimicrobial peptide discovery under resistance pressure while remaining agnostic to particular implementation choices.

Predictive components function as evaluators, contextualising candidate peptides in terms of antimicrobial efficacy, toxicity, uncertainty, and resistance-related properties (Zhou et al., 2025). Their role extends beyond sequence ranking to supporting risk-aware prioritisation under incomplete and evolving knowledge. In resistance-aware settings, predictive models must account for extrapolation, distributional shift, pathogen-specific variability, and uncertainty associated with sparse experimental observations, reflecting the challenges encountered in real-world antimicrobial peptide discovery (Cunha-Ferreira et al., 2025).

Generative components act as hypothesis generators, proposing candidate peptides within constrained regions of sequence space defined by biological priors, physicochemical requirements, and design objectives (Perron et al., 2022). Their effectiveness depends on continuous interaction with predictive evaluation, uncertainty estimates, and interpretability signals, transforming generation from unconstrained novelty exploration into guided design informed by accumulated evidence.

Interpretability-oriented components provide the connective layer linking computational outputs to biological reasoning and experimental decision-making (Alizadehsani et al., 2024). By exposing explanatory signals, uncertainty structure, and decision rationale, they facilitate hypothesis refinement, error diagnosis, and calibration of trust throughout iterative discovery cycles. Treating interpretability as an integral component of the discovery workflow reinforces its importance for reproducibility, transparency, and translational relevance (Medina-Ortiz et al., 2025).

These roles acquire meaning through their interactions within the broader discovery system. System-level behaviour emerges from the coordinated exchange between evaluation, hypothesis generation, interpretation, and learning. This interaction becomes particularly important under antimicrobial resistance, where predictive performance, design objectives, and biological constraints evolve simultaneously.

A central feature of this framework is the explicit incorporation of resistance into the modelling process. Predictive components may incorporate pathogen-specific susceptibility patterns, minimum inhibitory concentration (MIC) estimation, and resistance-associated descriptors, while generative models can be conditioned on biologically relevant objectives such as proteolytic stability, membrane disruption mechanisms, and resilience to adaptive resistance responses (Li et al., 2024; Sharma et al., 2023). Integrating these factors throughout the discovery process enables the design and prioritisation of candidates with a greater likelihood of maintaining activity across evolving resistance landscapes, supporting a transition from retrospective validation to proactive resistance-aware optimisation.

4.3 Design–test–learn cycles as a governing principle

Design–test–learn cycles provide the organising principle that binds AI-native discovery ecosystems. In antimicrobial peptide discovery, this principle reflects both the iterative nature of experimental science and the dynamic behaviour of resistance-driven biological systems (Nedyalkova et al., 2024).

Computational design generates hypotheses in the form of candidate peptides, experimental testing evaluates their behaviour under biologically relevant conditions, and learning mechanisms update predictive and generative models based on observed outcomes (Szymczak et al., 2025). When implemented as a continuous process, this cycle transforms experimentation from a terminal validation step into an active source of information for model refinement.

Under antimicrobial resistance, this loop must explicitly account for evolving pathogen responses, shifting activity landscapes, and uncertainty associated with experimental measurements (Castle et al., 2024). Predictive models must adapt to new data distributions, generative models must refine exploration strategies, and selection mechanisms must continuously balance exploitation of known active regions with exploration of underrepresented areas of sequence space (Ghislat et al., 2024).

From an operational perspective, design–test–learn cycles can be implemented through workflows in which candidate selection is guided by uncertainty-aware acquisition strategies, experimental outcomes are incorporated through adaptive retraining, and generative models are updated via reinforcement learning or feedback-conditioned optimisation (Castle et al., 2024; Mishra et al., 2025). In this manner, discovery systems can continuously incorporate new biological evidence and refine future prediction and design decisions.

As illustrated in Figure 3, successive design cycles progressively reshape the explored peptide space, reducing uncertainty and promoting convergence toward candidates that remain robust under evolving resistance pressures. Such dynamic refinement is essential for maintaining relevance in non-stationary biological environments.

4.4 From methodological principles to operational workflows

The principles described above define a conceptual framework for resistance-aware antimicrobial peptide discovery, but their value ultimately depends on how they are instantiated in practice. Translating explainability, uncertainty estimation, constrained generation, and iterative learning into operational workflows requires explicit coordination across multiple decision layers (Castle et al., 2024). Predictive models must support candidate prioritisation beyond predictive probability alone, interpretability signals must inform sequence modification and constraint definition, uncertainty estimates must guide exploration and experimental allocation, and newly acquired evidence must be systematically incorporated into model refinement (Li et al., 2026).

The objective of such workflows extends beyond predictive performance to the establishment of adaptive discovery systems capable of continuously updating their internal representations as biological knowledge accumulates (Szymczak et al., 2025). Different modelling choices, data modalities, and experimental infrastructures may be appropriate depending on the specific discovery context (Wan et al., 2024). Nevertheless, the underlying principles of integration, transparency, uncertainty awareness, and feedback-driven adaptation remain consistent.

To illustrate the practical implementation of the proposed framework, the following worked example integrates explainable prediction, uncertainty-aware prioritisation, interpretable rule extraction, and adaptive model updating within a simplified antimicrobial peptide discovery workflow. The objective is not to introduce a new predictive model, but to demonstrate how these complementary capabilities can support iterative, transparent, and evidence-driven discovery.

4.5 A worked example of explainable and adaptive AMP discovery

Figure 4 summarises a demonstrative antimicrobial peptide discovery workflow combining predictive modelling, explainability, uncertainty estimation, interpretable rule extraction, candidate prioritisation, and iterative model updating. The objective of this example is not to establish a new state-of-the-art predictive model, but to illustrate how complementary methodologies can be integrated into a coherent decision-support framework that supports candidate evaluation, experimental prioritisation, and progressive refinement of predictive knowledge throughout successive discovery cycles.

The workflow begins with the construction of a curated peptide dataset derived from experimentally annotated sequences obtained from Peptipedia v2.0 (Cabas-Mora et al., 2024). The initial collection comprised approximately 86,000 peptides associated with antimicrobial and non-antimicrobial annotations. To improve consistency and reduce biases associated with redundancy and atypical peptide compositions, non-canonical sequences were removed, peptides outside the range of 5–60 amino acids were excluded, and sequence redundancy was reduced through MMseqs2 clustering at 80% sequence identity (Steinegger and Söding, 2017). Following preprocessing, the final dataset contained approximately 47,000 non-redundant peptides. For each sequence, a panel of 41 descriptors was calculated, encompassing amino acid composition, physicochemical properties, sequence complexity, entropy-related measures, and structural propensity estimates. Detailed preprocessing criteria, descriptor definitions, implementation settings, and reproducible workflows are available through the accompanying repository.

As illustrated in Figure 4A, exploratory analysis revealed systematic differences between antimicrobial and non-antimicrobial peptides across multiple descriptor families, including net charge, positively charged residue content, hydrophobicity-related properties, and sequence entropy. These trends are consistent with established mechanistic characteristics of membrane-active antimicrobial peptides (Kumar et al., 2025; Bin et al., 2021). Beyond demonstrating class separability, these patterns establish an interpretable descriptor space that provides the foundation for predictive modelling, uncertainty estimation, and rule extraction.

Predictive modelling was performed using an ensemble strategy trained under distance-aware data partitioning to better approximate prospective screening scenarios (Xiao et al., 2026). The training and evaluation datasets comprised approximately 33,000 and 14,000 peptides, respectively. To preserve geometric diversity across partitions and reduce biases associated with highly similar samples, peptides were grouped into up to 80 clusters within a z-score standardised descriptor space using MiniBatchKMeans, and training, validation, and test subsets were subsequently generated through cluster-aware stratified sampling. The final ensemble consisted of 10 predictive models and achieved an overall MCC of 0.65 on the held-out evaluation set. Although predictive performance was not the primary objective of this demonstrative application, these results confirm that the selected representation space captures sufficient biological signal to support downstream interpretation and decision-support analyses.

To facilitate model interpretation, global SHAP analysis was applied to quantify the contribution of individual descriptors to antimicrobial activity predictions (Figure 4B). Descriptors associated with amino acid composition, charge distribution, physicochemical balance, and sequence complexity emerged among the most influential variables. These explanatory signals provide mechanistically interpretable insights into the sequence characteristics driving antimicrobial predictions and offer a rational basis for candidate selection and optimisation.

Prediction uncertainty was estimated through ensemble variability and incorporated directly into the prioritisation process (Figure 4C). While highly confident predictions accumulated near the extremes of the probability distribution, regions of elevated uncertainty emerged for peptides occupying intermediate areas of the learned descriptor space. Such information becomes particularly valuable under limited-data conditions, where predictive confidence may exceed actual model reliability (Vellore and Jha, 2026). By explicitly accounting for uncertainty, candidate selection can distinguish between sequences suitable for immediate experimental validation and candidates that may require additional evidence before prioritisation (Abdullah et al., 2024). More broadly, uncertainty estimates provide a principled mechanism for balancing exploitation of promising candidates with exploration of underrepresented regions of sequence space.

To transform explanatory information into operational design knowledge, the most influential descriptors identified through SHAP analysis were converted into explicit decision rules using surrogate modelling approaches (Nerini et al., 2025). As illustrated in Figure 4D, the resulting rules capture interpretable relationships between sequence properties and predicted antimicrobial activity, generating human-readable criteria that can support candidate selection, optimisation, and downstream design decisions. A total of 12 decision rules were extracted, with representative examples shown in Figure 4D.

Candidate prioritisation was subsequently reformulated as a multi-criteria decision problem integrating predicted antimicrobial probability, uncertainty estimates, and rule-support scores (Figure 4E). This framework combines predictive confidence, exploration potential, and interpretable evidence within a unified ranking strategy, enabling more informed candidate selection than probability-based prioritisation alone. Application of this strategy to an external screening pool containing approximately 16,000 candidate peptides resulted in the prioritisation of 606 sequences for further evaluation.

Finally, the prioritised candidate pool was incorporated into a simplified model update stage, generating the design–test–learn cycle illustrated in Figure 4F. Newly acquired observations, represented here through a pseudo-feedback strategy, were used to update model parameters and revise candidate rankings. Although intentionally simplified, this step demonstrates how new information can be integrated into iterative model refinement, enabling predictive behaviour, uncertainty estimates, and prioritisation criteria to evolve as additional biological evidence becomes available.

Several limitations should be considered when interpreting this worked example. The objective of this demonstrative application was to illustrate the practical integration of explainability, uncertainty estimation, interpretable rule extraction, and iterative updating within a unified discovery workflow. Consequently, extensive hyperparameter optimisation, large-scale benchmarking, and exhaustive comparative evaluation were intentionally omitted. In addition, the implementation does not include generative modelling components, explicit modelling of resistance evolution, active learning-based experimental design, or feedback derived from wet-laboratory validation. The resulting design–test–learn cycle therefore serves as an illustrative example of methodological integration, highlighting how complementary sources of information can be coordinated throughout the discovery process.

Despite these limitations, the example illustrates how complementary sources of information can be transformed into actionable decision support throughout the discovery process. Predictive probabilities, uncertainty estimates, and interpretable rules contribute distinct yet interconnected perspectives that can support candidate prioritisation, experimental planning, and iterative model refinement. More broadly, this demonstration highlights the value of coordinating existing methodologies within a common workflow, allowing computational predictions and newly acquired observations to inform one another through successive discovery cycles. Complete implementation details, source code, datasets, and reproducible workflows are available through the public repository accompanying this Perspective.

5 Conclusion and future perspectives

Artificial intelligence has transformed antimicrobial peptide discovery by enabling large-scale sequence exploration, predictive modelling, and generative design. However, future progress will depend less on the development of increasingly sophisticated algorithms and more on the integration of predictive, generative, and experimental components into adaptive and interpretable discovery systems.

In this Perspective, we proposed a conceptual framework built upon three complementary principles: explainable and uncertainty-aware prediction, biologically constrained and rule-guided generative design, and iterative design–test–learn cycles capable of continuously incorporating new evidence. These elements provide a foundation for organising existing methodologies into coherent workflows that support more informed and transparent decision-making throughout the discovery process.

The demonstrative application presented here illustrates how many of these concepts can already be implemented using currently available methodologies. By combining interpretable prediction, uncertainty-aware prioritisation, rule extraction, and iterative model updating within a unified workflow, the example highlights a practical path toward adaptive and evidence-driven antimicrobial peptide discovery.

Looking forward, advances in active learning, foundation models, autonomous experimentation, scientific agents, and hybrid computational–experimental infrastructures may further accelerate the development of closed-loop discovery ecosystems. Ultimately, the impact of artificial intelligence on antimicrobial peptide development will depend not only on predictive performance, but also on the ability to continuously integrate prediction, experimentation, interpretation, and learning within a unified and evolving discovery framework.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. As non-native English speakers, the authors used ChatGPT (OpenAI) to assist with language polishing, clarity, and grammar during the preparation of this manuscript. All content was subsequently reviewed and edited by the authors, who take full responsibility for the final version of the 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.

Footnotes

Edited by: Vinícius Gonçalves Maltarollo, Federal University of Minas Gerais, Brazil

Reviewed by: Rodolpho C. Braga, InsilicAll, Brazil

Mia Md Tofayel Gonee Manik, University of the Cumberlands, United States