Beyond one drug–one target: AI-guided discovery of multi-target antiviral therapeutics

1 Introduction

Viruses have repeatedly demonstrated their ability to challenge modern medicine, overwhelm healthcare systems, and reshape global society. Despite remarkable progress in vaccination and antiviral therapy, viral diseases continue to remain a major cause of mortality and morbidity across the world. Chronic viral infections such as HIV and hepatitis B and C continue to affect millions of people, while acute infections caused by influenza, dengue, respiratory syncytial virus, Ebola, Zika, and coronaviruses remain a persistent threat (1). The devastating impact of the COVID-19 pandemic further highlighted how rapidly an emerging virus can spread across populations, disrupt economies, and expose the limitations of existing therapeutic strategies. Beyond their direct impact on human health, viral infections also contribute to long-term complications, immunological disorders, organ damage, and substantial financial burden on healthcare systems. One of the major challenges in antiviral therapy is the extraordinary ability of viruses to mutate and adapt. Many viruses, especially RNA viruses, possess high mutation rates because of the lack of efficient proofreading mechanisms during replication (2). This constant genetic variation allows viruses to develop resistance against drugs that target only a single viral protein or pathway. Traditionally, antiviral drug discovery has followed the “one drug-one target” approach, in which a molecule is designed to inhibit a specific viral enzyme, receptor, or replication mechanism. Examples include reverse transcriptase inhibitors for HIV, neuraminidase inhibitors for influenza, and protease inhibitors for hepatitis C virus (3). While these drugs have transformed the management of several viral diseases, their long-term effectiveness is often limited by the rapid emergence of resistant viral strains. The limitations of single-target therapy become even more evident in chronic viral infections and in rapidly evolving outbreaks. Monotherapy frequently loses effectiveness over time, requiring the use of combination therapy to suppress viral replication and delay resistance. Although combination regimens have been successful in diseases such as HIV, they are often associated with high treatment costs, complex dosing schedules, drug-drug interactions, cumulative toxicity, and poor patient adherence (4). In addition, many viruses rely on multiple viral and host-cell pathways to complete their life cycle. Targeting only one of these pathways may not be sufficient to completely block infection, allowing the virus to persist or escape through alternative mechanisms. These challenges have led to increasing interest in MTAs, which are designed to simultaneously inhibit two or more viral or host-associated targets within a single molecule (5). Unlike conventional drugs that focus on one biological pathway, MTAs aim to interfere with several stages of the viral life cycle at the same time, including viral entry, replication, protein processing, and assembly. This strategy offers several important advantages, such as broader antiviral spectrum, improved therapeutic efficacy, lower probability of resistance development, and reduced need for multiple-drug regimens. Multi-target approaches are particularly valuable for treating rapidly mutating viruses, mixed infections, and newly emerging pathogens for which no highly specific therapy is available (6).

However, designing a molecule capable of acting on multiple targets while maintaining acceptable selectivity, potency, safety, and pharmacokinetic behavior is an extremely complex process. Conventional experimental approaches alone are often too slow, expensive, and resource-intensive to efficiently identify such compounds. In recent years, AI, ML, and computational drug design have emerged as powerful tools for overcoming these challenges (7). These technologies can rapidly analyze large biological datasets, predict drug-target interactions, identify hidden molecular patterns, optimize lead compounds, and generate novel antiviral scaffolds with multi-target potential. Techniques such as molecular docking, virtual screening, pharmacophore modeling, quantitative structure-activity relationship analysis, deep learning, and network pharmacology are increasingly being integrated into antiviral research to accelerate the discovery process (8). The application of AI in antiviral drug discovery has become especially important after the COVID-19 pandemic, where computational methods played a major role in identifying repurposed drugs, predicting viral protein structures, and screening large chemical libraries in a short period of time. As viral threats continue to evolve, the integration of multi-target drug design with AI-guided approaches is expected to play a central role in the development of next-generation antiviral therapeutics (9). This review article includes the concept and rationale of multi-target antiviral agents, computational and AI-based approaches used in their design, medicinal chemistry strategies, natural product-derived scaffolds, host-targeted therapies, major case studies involving HIV, HCV, SARS-CoV-2, and influenza, as well as challenges related to resistance, pharmacokinetics, and future clinical development.

1.1 Scope, novelty and distinguishing features of the present review

Over the last few years, there have been several review articles that explored the use of AI, ML, and computational drug discovery strategies in antiviral studies. The studies concentrated mainly on AI-assisted virtual screening, drug repurposing, molecular docking, DL models and computational approaches in the COVID-19 pandemic (10–13). Overall, although these reviews have shed light on individual computational approaches, they have focused mainly on certain technologies, specific viral diseases, or specific steps in the antiviral drug discovery process.

The presented review is unique in offering a multivariate and multi-disciplinary point of view to MTAs development. This review will not only solely on AI methodologies or on the more classical medicinal chemistry strategies, but will emphasis on integrating the advancements in computational drug discovery, AI, medicinal chemistry, host-targeted therapies, natural product-derived antivirals and translational development strategies into one holistic structure. The new idea of multi-target antiviral therapy which can be used to overcome viral resistance, to increase the duration of antiviral therapy and the antiviral spectrum, is of particular importance.

One of the other key differences of this review is that it focuses on the evaluation of novel technologies recently introduced in the field of AI, such as protein structure prediction using AlphaFold, graph neural networks, generative AI models, multimodal learning and advanced ML frameworks that are now playing an important role in the discovery of antiviral drugs. Furthermore, the review critically evaluates major obstacles in the implementation of AI, such as data bias, model overfitting, lack of reproducibility, false predictions, ethical issues, and regulatory considerations, which are not extensively discussed elsewhere.

Moreover, this review goes beyond computational prediction, and covers experimental validation strategies, barriers to translation, clinical development considerations and future possibilities in the field of precision antiviral medicine. The current work is designed to offer a thorough guide for the development of future anti-viral therapeutic drugs which will be directed by AI, and to close the divide between the innovation of the computational approach and the translation of the technology into clinical practice was discussed.

2 Concept and rationale of multi-target antiviral therapeutics

MTAs are therapeutic molecules designed to simultaneously interact with two or more biological targets involved in the viral life cycle or host response as shown in Figure 1. These targets may include viral enzymes such as proteases, polymerases, integrases, neuraminidases, and helicases, as well as host factors such as kinases, receptors, transport proteins, and molecular chaperones (14). Unlike conventional antiviral drugs that act on a single protein or pathway, MTAs are intended to block multiple stages of viral infection within a single therapeutic framework. This can be achieved either through a single molecule possessing multiple pharmacological activities or through hybrid structures created by combining two pharmacophores with distinct mechanisms of action. The rationale for developing MTAs arises from the highly complex nature of viral replication. Viruses do not rely on only one target to survive and propagate; instead, they exploit a network of interconnected viral and host-cell pathways (6). Viral entry, genome replication, protein processing, immune evasion, and virion assembly are all regulated by multiple molecular events. Therefore, inhibiting a single viral target may not be sufficient to completely suppress infection. Even when a single-target drug initially shows strong antiviral activity, the virus may adapt by mutating the targeted protein or by activating alternative pathways that compensate for the inhibited function. MTAs address this problem by producing a broader inhibitory effect across multiple biological mechanisms (15). One of the most important advantages of MTAs over single-target drugs is their ability to reduce the development of antiviral resistance. Viral resistance occurs when mutations arise in the viral genome that decrease the ability of a drug to bind its intended target. This is particularly common in RNA viruses because of their high mutation rates and rapid replication cycles. In the case of single-target therapies, a single mutation in the drug-binding region may be enough to render treatment ineffective. However, when a molecule acts on multiple targets simultaneously, the virus must accumulate several mutations at the same time in order to fully escape therapeutic pressure (3). This substantially lowers the probability of resistance and can prolong the clinical usefulness of the drug. Another major advantage of MTAs is their potential to provide broader antiviral coverage. Since many viruses share common replication pathways or depend on similar host-cell factors, a single multi-target molecule may demonstrate activity against more than one virus. Such broad-spectrum activity is especially valuable during emerging outbreaks when specific antiviral agents are not yet available. In addition, MTAs may reduce the need for complex multidrug regimens, thereby improving patient adherence, minimizing drug-drug interactions, and lowering cumulative toxicity. The importance of viral escape mechanisms further strengthens the case for multi-target antiviral design (3). Viruses possess remarkable genetic flexibility and can evade therapy through several mechanisms, including point mutations, recombination, altered protein expression, and changes in host-cell dependency. Some viruses may also establish latent reservoirs or hide within immune-privileged tissues, making them difficult to eradicate with conventional therapies. In many cases, resistance does not arise from changes in a single viral protein alone but from coordinated adaptations across multiple pathways. Consequently, targeting several viral and host processes simultaneously offers a more durable and comprehensive therapeutic strategy (16). Host-virus interactions have also become increasingly important in the development of MTAs. Because viruses rely heavily on host-cell machinery for replication, targeting selected host factors can provide an effective means of inhibiting infection while reducing the likelihood of viral mutation-driven resistance. Host-targeted antivirals may interfere with cellular kinases, endosomal pathways, chaperone proteins, lipid metabolism, immune signaling pathways, or membrane fusion processes required for viral entry and replication (17). Combining host-targeted and virus-targeted mechanisms within a single molecule may further enhance antiviral efficacy and create synergistic effects. Nevertheless, the development of MTAs is not without challenges. A molecule capable of interacting with multiple targets must maintain an optimal balance between potency, selectivity, safety, and pharmacokinetic properties. Excessive activity toward host proteins may increase toxicity, while inadequate target selectivity may reduce therapeutic effectiveness. Therefore, rational design strategies supported by medicinal chemistry, computational modeling, and AI are essential for identifying successful multi-target antiviral candidates (18). The concept and rationale of multi-target antiviral agents is represented in Figure 1.

Table 1. Major viral targets and host targets used in multi-target antiviral design.

3 Computational approaches in multi-target antiviral drug discovery

The multi-target antiviral discovery is a very complicated process since multi-target antiviral agents are anticipated to engage on various targets of virus or the host and retain appropriate potency, selectivity, safety, and pharmacokinetic characteristics. Conventional experimental screening techniques are usually lengthy, costly and ineffective in detecting such molecules. Consequently, computational methods have emerged as a significant part of the contemporary antiviral drug discovery. Such methods enable the rapid analysis of large chemical and biological data, prediction of molecular interactions, optimization of lead compounds, and minimize the list of compounds which should be experimentally tested. Over the last several years, the field of multi-target antiviral therapy has gained momentum in terms of development due to the integration of computational chemistry, structural biology, bioinformatics and systems pharmacology (25, 26).

3.1 Structure-based drug design

One of the most commonly used computational strategies in the field of antiviral drug discovery is structure-based drug design (SBDD). This method depends on the accessibility of three-dimensional structure of viral proteins or host-cell proteins related to infection and replication. Structural data captured by methods like X-ray crystallography, cryo-electron microscopy and nuclear magnetic resonance spectroscopy gives a detailed description of active sites, catalytic residues, binding pockets and conformational changes within the target protein (27).

SBDD has been widely applied in antiviral studies to identify viral protease, polymerase, integrase, helicase, and envelope protein inhibitors. As an illustration, structural studies have been of significant use in the development of protease inhibitors against HIV and SARS-CoV-2 due to the disclosure of the shape and chemical context of the active sites of the viruses. SBDD is even more useful in the situation with multi-target antiviral agents since the researcher can easily compare the structural properties of many proteins and generate compounds that can bind more than one target (28).

This approach is also used to optimize leads rationally because it determines what structural changes can enhance the binding affinity, selectivity, and stability. Moreover, SBDD is able to unveil conserved amino acid residues across various viral proteins and as such, it is possible to come up with broad-spectrum antiviral agents. Since multi-target drug discovery demands a delicate equilibrium between activity and safety, structure-based methods offer a valuable base in the way molecules with optimized interaction patterns are designed (29).

3.2 Ligand-based drug design

Ligand-based drug design (LBDD) is especially applicable in cases where the detailed structure of the target protein is absent or challenging to get. LBDD uses the chemical and biological properties of active compounds. The comparison of the molecular properties of antivirally active compounds allows identifying patterns linked to potency, selectivity, and multi-target behavior (30).

Computational approaches that are part of the ligand-based design are quantitative structure-activity relationship (QSAR) analysis, molecular similarity studies, machine learning-based prediction models, and chemical fingerprint analysis. As an illustration, antiviral agents can be similar in their hydrophobic groups, hydrogen-bonding formats, aromatic systems or charge distributions, which are vital to activity (31, 32).

Ligand-based methods are particularly beneficial with respect to hybrid molecule development, in which two pharmacologically active entities are fused into a single molecule. Medicinal chemists can develop new analogues which have the positive properties of more than one drug class by examining the available antiviral agents (33). In addition, it is possible to use LBDD to facilitate scaffold hopping, where scientists substitute the core skeleton of a known compound with a novel scaffold and retain biological activity. It is also very effective in terms of overcoming toxicity, enhancing pharmacokinetics and preventing resistance (34).

3.3 Virtual screening and molecular docking

One of the most crucial computational tools, which are utilized in the modern process of drug discovery, is the virtual screening. It entails in silico screening of high-throughput chemical libraries to discover compounds that have biological activity. Since screening thousands or millions of compounds is costly and time-consuming in experiments, virtual screening can be used as an efficient initial process (35).

Virtual screening can be performed using either structure-based or ligand-based methods. In structure-based screening, the compounds are screened based on their predicted binding capacity to a target protein. In ligand-based screening, the compounds are chosen depending on their resemblance to known active molecules. The methods are especially convenient when it comes to multi-target antiviral studies, since they can screen the same compound library on multiple targets (36).

Molecular docking closely relates to virtual screening and is applied to predict the behavior of a compound at the molecular level in its interaction with a protein. Docking methods determine the orientation, conformation and binding affinity of a ligand in the active site of the target protein. They also give details on hydrogen bonding, electrostatic interactions, van der Waals forces, hydrophobic contacts and steric compatibility (35).

In multi-target antiviral discovery, docking can be done on multiple proteins, including viral proteases, polymerases, and host-cell receptors. Those compounds that show positive affinity to bind to more than one target are deemed as good candidates to be pursued. Docking experiments can also be used in the study of the mechanism of action of antiviral compounds, as well as in the rationalization of structural changes that can enhance biological activity (25).

3.4 Pharmacophore modeling

Pharmacophore modeling is a computerized method of determining the necessary structural characteristics needed to achieve biological activity. A pharmacophore is the spatial structure of functional groups that facilitate a molecule to bind with a target protein as shown in Figure 2. Hydrogen bond donors, hydrogen bond acceptors, aromatic rings, hydrophobic regions, ionizable groups and charged centers are also common pharmacophoric features (37).

Pharmacophore models are produced by examining similar properties of active compounds in antiviral drug discovery. It is then possible to use these models to search large databases to find new molecules that have similar properties. Pharmacophore modeling is of particular use in the context of multi-target antiviral agents since it aids in determining common binding features that can be applied to multiple targets (38).

By combining the pharmacophoric characteristics that are needed in the various viral proteins or host targets, researchers can create merged or hybrid pharmacophore models (Figure 2). These models can be used to design compounds that can meet several different pathways at the same time. Pharmacophore analysis can also be useful in discovering lead compounds containing new scaffolds and thus increasing the chemical diversity of antiviral drug candidates (39).

3.5 Molecular dynamics simulations

The behavior and stability of drug-target complexes when subjected to dynamic and near-physiological systems are determined using molecular dynamics (MD) simulations. In contrast to molecular docking, which is only a static representation of molecular binding, MD simulations permit the researcher to study the movement and interaction of proteins and ligands over time. Such simulations can be used to determine the stability of a molecule to stay bound to its target, detect conformational transitions in proteins, and assess the flexibility of the binding site. MD simulations are especially useful in the context of multi-target antiviral drug discovery to ensure that the compounds which are designed to bind multiple proteins are stable (40, 41).

3.6 Network pharmacology approaches

Network pharmacology has become an effective systemic level method of investigating the multifaceted interactions between viruses, host cells, signaling pathways and therapeutic agents. Network pharmacology does not target any specific drug, but rather looks at the whole biological network of viral infection (42). Such an approach aids in revealing interdependent routes, dependency variables in the host, and molecular crossroads that can be used as potential therapeutic targets. Network pharmacology particularly finds application in the design of multi-target antiviral drugs since it can reveal synergistic interactions among targets, and is able to predict how a compound might affect multiple biological pathways concurrently (43).

3.7 Integration of computational approaches

Multi-target antiviral agents have been discovered with much faster rates by the integration of structure-based design, ligand-based methods, virtual screening, docking, pharmacophore modeling, molecular dynamics and network pharmacology (Table 2). These computational methods minimize the experimental screening process which involves large experimental screens and enhances the accuracy of the predictions and enable the researchers to design compounds in a more efficient manner. They can be used in conjunction with AI and ML to offer a highly efficient platform to find safer, more potent, and broad-spectrum antiviral therapy (51).

Table 2. Computational and AI-based approaches used in multi-target antiviral discovery.

4 AI and ML in MTA design

ML and AI are considered as the revolution in contemporary drug discovery. These tools are becoming common these days in antiviral research to hasten the discovery, optimization, and forecasting of MTAs. Conventional methods used in drug discovery are usually characterized by high costs, long development cycles and low success rates. MTAs are even more difficult to design since a single molecule should be active on many targets and has to have good pharmacokinetic and toxicity profiles. The AI-based approaches are capable of handling vast quantities of biological, chemical, genomic, and pharmacological data within a relatively brief period, and thus, they are of great significance when it comes to antiviral research (52).

4.1 ML algorithms in drug discovery

ML is a set of computation approaches that allow computers to recognize a pattern in the data and make some predictions without writing a program to do all the tasks. ML algorithms are also used in drug discovery by training on large datasets of data on chemical structures, biological activity, toxicity, target interactions, and pharmacokinetic properties. The most popular ML algorithms are a random forest, support vectors machine, decision tree, k-nearest neighbors, gradient boosting algorithms and artificial neural networks. Antiviral activity, classification of compounds as active or inactive, prediction of toxicity, and multi-target potential molecules can be predicted by these models (53). ML algorithms can also be particularly helpful in the context of MTA design since they can examine several target interactions concurrently and determine compounds with balanced activity patterns. This saves time and resources since such methods minimize the number of compounds to be synthesized and tested experimentally (54).

4.2 Public databases and datasets used for AI-based antiviral drug discovery

High-quality datasets are crucial to the success of AI and ML models in antiviral drug discovery. There are a number of publicly available databases available containing chemical, biological, genomic, pharmacological and clinical data, which are commonly employed for training, validation and benchmarking anti-viral prediction models. These resources enable identification of antiviral compounds, prediction of target interactions, toxicity assessment, drug repurposing and multi-target drug discovery (11, 55).

ChEMBL is one of the most widely used databases and provides a large amount of bioactivity data for small molecules against a wide variety of biological targets, including viral proteins. ChEMBL includes experimentally determined activity measurements like IC50, EC50, Ki and Kd values, which are vital for applications like QSAR modeling and DL (56). Another important resource is PubChem, which provides biological assay data, molecular descriptors, screening data from high throughput assays, and chemical structures. Due to its size, it is useful for training ML algorithms which are data-intensive (57).

DrugBank integrates chemical, pharmacological, and pharmaceutical data, while also providing comprehensive data about drug-target interactions. It is a database that is often used in the study of AI-assisted drug repurposing, as it contains approved drugs, investigational compounds, and their corresponding biological targets (58). The other widely used database is Binding DB, which includes experimentally determined binding affinity data for small molecules and proteins, aiding the development of predictive models for target identification and lead optimization (59).

The GenBank database, the Global Initiative on Sharing All Influenza Data (GISAID), and the UniProt database offer huge collections of genomic and protein sequences from viruses, along with functional annotations, for viral sequence analysis and analysis of proteins. These resources have been especially instrumental in developing AI models that focus on new viral pathogens such as SARS-CoV-2, influenza viruses and hepatitis viruses (60, 61). The Protein Data Bank (PDB) also provides experimentally determined three-dimensional structures of viral and host proteins which are crucial in structure-based AI approaches, molecular docking, and DL models based on structure information (62, 63).

These databases have a number of drawbacks. The quality of data may also vary depending upon the experimental protocols, the assay conditions and reporting practices. The presence of missing data, duplicate records, class imbalance, and inconsistent annotations can lead to bias in AI models and affect how well they predict outcomes. Also, many antiviral datasets are dominated by well-studied viruses, like HIV, influenza, hepatitis C virus, SARS-CoV-2 etc. and are lacking in data for emerging or neglected viral pathogens (64, 65). In addition, these repositories are generally much larger than active molecules, which poses difficulties in training and validating the models. The other major disadvantage is the lack of comprehensive data that incorporates all the important attributes of successful development of multi-target antiviral agents, such as antiviral activity, toxicity, pharmacokinetic parameters and multi-target interactions (66).

However, public databases are only the starting point for antiviral discovery using AI, and the curation, standardization and integration of multiple datasets, as well as rigorous external testing and validation of AI models, are still required to ensure the production of robust and clinically relevant predictive models.

4.3 DL for antiviral activity prediction

DL is a particular field of machine learning, which employs neural networks with multiple hidden layers to process very complex data. In contrast to conventional ML algorithms, DL algorithms are able to automatically extract pertinent features out of raw data without having to engage in a lot of manual feature extraction. DL has been very promising in the discovery of antiviral drugs since it is capable of processing large chemical data, protein structures, viral genomes, and molecular interaction networks (67). Convolutional neural networks, recurrent neural networks, graph neural networks and transformer-based models are also being applied to predict antiviral activity and to identify multi-target behavior compounds (68).

These models are capable of examining complicated associations between molecular structure and biological action more precisely than numerous conventional computational methods. Moreover, deep learning-based techniques are able to give insight into the concealed patterns that might otherwise be obscured by conventional medicinal chemistry analysis. This renders them very applicable in the prediction of compounds that may at the same time inhibit two or more viral or host targets (69).

4.4 Critical comparison of traditional computational methods, machine learning, and DL in antiviral drug discovery

The conventional computational methods in antiviral drug discovery have been molecular docking, virtual screening, pharmacophore modelling, QSAR analysis, and molecular dynamics simulations. These approaches offer mechanistic understanding of drug–target interactions and are especially useful in cases where there is a good amount of structural information available (70, 71). Molecular docking and pharmacophore-based approaches can be used to identify the rational leads, and molecular dynamics simulations can help to gain detailed knowledge of protein–ligand stability and conformational behavior. These methods, however, typically involve assumptions about the physicochemical properties of the molecule and may fail to account for biological complexity sufficiently, and can be challenging to compute for large chemical libraries or multiple targets, respectively (72, 73).

One way to address some of these limitations is through ML techniques that learn patterns directly from the experimental data. Random Forest, Support Vector Machines, Gradient Boosting and k-nearest neighbors are among the algorithms that can accurately predict antiviral activity, toxicity, pharmacokinetics, and target interactions (74). The advantage of ML models over traditional computational models is that they often have a quicker prediction time and can model data from various sources such as chemical descriptors, omics data, and biological activity profiles. However, they rely significantly on the size, quality and variety of training data. There are still problems of limited interpretability and bias in the dataset (75).

A subset of ML, DL is an approach that can be used to extract complex features from raw molecular, structural, and genomic data automatically. Convolutional neural networks (CNNs), graph neural networks, recurrent neural networks, and transformer-based models have been shown to be better at predicting antiviral activity, generating new molecules from scratch, and optimizing multi-target molecules (76). DL models are able to model nonlinear relationships which are not possible with traditional QSAR or docking approaches. But they generally require a much larger amount of information, high computing resources, and careful tuning to prevent overfitting. They may also be difficult to understand mechanistically and to be acceptable for regulatory purposes because of their “black-box” nature (75, 77).

Each method has its own advantages for multi-target antiviral discovery. Mechanistic understanding and structural validation are the domain of traditional computational techniques, while ML can be used to rapidly prioritize candidate compounds and DL can be used to discover hidden molecular patterns and generate novel antiviral scaffolds. Therefore, the integrated workflows combining structure-based modeling, machine learning, and DL with experimental validation have become the main approach for antiviral discovery today. Combination approaches boost the potential for predicting drug activity, shorten drug development timelines and increase the chance of finding clinically relevant multi-target antiviral drugs.

4.5 QSAR and multi-target prediction models

One of the oldest and most popular computational methods in medicinal chemistry is the quantitative structure-activity relationship (QSAR) modeling. QSAR models are mathematical correlations of the structure of molecules and their biological behavior. Commonly used molecular descriptors to create QSAR models are lipophilicity, molecular weight, the number of hydrogen bond donors, polar surface area, aromaticity, and electronic properties. It is then possible to predict using these models whether a new compound will have any antiviral activity or not (31).

QSAR models find application in the multi-target antiviral drug discovery process in that they can be built on multiple targets at once. Multi-target QSAR models are useful in the discovery of compounds that are balanced in action among various viral enzymes or host proteins. They also help in maximizing selectivity, lowering toxicity, and enhance pharmacokinetic characteristics (78).

4.6 AI-assisted de novo drug design

The de novo drug design is the production of an entirely new structure of a drug, not a modification of an existing one. AI-assisted de novo design is a computational algorithm applied in the design of new molecules with desirable biological and physicochemical characteristics. This strategy frequently starts with a list of predetermined goals, e.g. activity against various viral proteins, low toxicity, good oral bioavailability, and reasonable metabolic stability. The AI systems then create and assess thousands of potential molecular structures to produce the most promising structures (79).

Antiviral research specifically is of great need of de novo drug design since new viruses could lack drugs that can be repurposed. The AI-assisted approaches offer a chance to quickly develop a new set of antiviral molecules and aim at different stages of the viral life cycle (25).

4.7 Generative models for novel antiviral scaffolds

Generative AI models are now one of the most promising fields of computational drug discovery. Such models can form entirely novel molecular scaffolds having the desired properties. This area usually uses variational autoencoders, generative adversarial networks, reinforcement learning models and transformer-based systems. This is possible because the databases of known antiviral compounds can be trained on generative models that are used to design new structures with enhanced potency, selectivity, and multi-target activity. Other properties can also be optimized by such models including solubility, permeability, metabolic stability and synthetic accessibility (80).

Generative AI can be useful in multi-target antiviral drug discovery, where it can be necessary to find molecules that merge the structural properties of multiple antiviral classes. This can be specifically applicable in designing hybrid pharmacophores that can be used to deliver multiple viral proteins in an identical molecular construct (10).

4.8 AI-driven protein structure prediction and antiviral target identification

Identification of appropriate antiviral targets is a key point in drug discovery. Historically, protein structure determination by X-ray crystallography, nuclear magnetic resonance (NMR) spectroscopy and cryogenic electron microscopy (cryo-EM) has been key to the understanding of viral proteins and the design of antiviral drugs. These methods, however, can be time consuming, costly, and technically difficult especially in the event of an outbreak of an emerging viral disease (10, 81). However, with the recent developments in AI-based protein structure prediction, this process has been greatly expedited, facilitating rapid and accurate prediction of protein three-dimensional structures (82).

The most impactful among these has been the AlphaFold, an AI-driven protein structure prediction program by DeepMind. AlphaFold showed the highest level of accuracy in predicting protein structure directly from their amino acid sequences, matching experimental methods for many proteins (83). The Protein Structure Database of AlphaFold has made available millions of protein structures predicted by the AlphaFold models, greatly augmenting the structural information available for drug discovery applications (84).

AlphaFold has been instrumental in the structural characterization of viral proteins where experimental structures were either missing or incomplete in antiviral research. In the midst of the COVID-19 pandemic, AI-generated structures of multiple SARS-CoV-2 proteins facilitated a rapid investigation of viral cycle, protein–protein interactions and possible drug binding sites (85, 86). Influenza viruses, flaviviruses, hepatitis viruses and other emerging pathogens have been reported with similar applications. AI-driven prediction tools can help speed up target prioritization and hypothesis generation by giving structure during the early stages of the outbreak response (87).

AI-based protein structure generation also improves the drug design process based on structure. The predicted protein models can be used in combination with molecular docking, molecular dynamics simulation, virtual screening and ML platforms to explore possible antiviral compounds (88). Structural predictions also contribute to the identification of conserved binding pockets, allosteric sites, protein interaction interfaces, which are potential attractive targets for multi-target antiviral agents. This is especially useful when experiential structural information is lacking or unavailable (25).

However, there are some shortcomings in these advances. Structures predicted may not be as accurate for very flexible proteins, intrinsically disordered regions, multi-protein complexes, or proteins that undergo large conformational shifts. Moreover, it is important to note that static structural prediction may not completely reflect dynamic biological processes such as viral replication and host–virus interactions. Thus, for an AI-generated structure to be biologically relevant, experimental validation, molecular dynamics, and biochemical studies should be performed on the structure.

4.9 Emerging AI paradigms in antiviral drug discovery

The recent development of AI, especially beyond the traditional ML and DL methods, has broadened the field of antiviral drug discovery. New techniques like foundation models, graph neural networks (GNNs), large language models (LLMs), multimodal AI and reinforcement learning (RL) are being used to enhance target identification, molecular design, drug repurposing and multi-target optimization (89).

4.9.1 Graph neural networks

The molecule graph representation lends itself naturally to graph neural networks, where atoms are modeled as nodes and chemical bonds as edges, making the networks extremely useful for predicting molecular properties. GNNs learn the molecular representations directly from the structural information allowing for more accurate prediction of antiviral activity, toxicity, drug–target interactions, and pharmacokinetic properties, unlike traditional descriptor-based ML approaches (90). The ability of GNN-based models to accurately describe complex structural relationships within molecules has been shown to be effective in virtual screening and lead optimization applications, outperforming traditional ML algorithms (91).

4.9.2 Foundation models

The foundation models are a new class of AI models that is trained on very large datasets and then fine-tuned for specialized biomedical tasks. Just as they have revolutionized NLP in the natural world, foundation models in chemistry and biology are capable of learning general representations of molecular structures, protein sequences, biological networks and pharmacological data (92). They can be readily adapted for prediction of antiviral activity, identification of antiviral targets, drug repurposing, and multi-target optimization, and require significantly fewer task-specific training data than conventional DL methods (93).

4.9.3 Large language models

Recently, large language models (LLMs) have shown remarkable capabilities in biomedical knowledge extraction and drug discovery. LLMs can also analyze large volumes of scientific publications, patent data, clinical trial information, and biomedical research databases to help researchers discover new drug targets, create biological hypotheses, summarize emerging evidence, and explore the potential for drug repurposing (94). Antiviral research can benefit from automated literature mining, knowledge graph construction, target prioritization, and the incorporation of multidisciplinary information that aids in therapeutic development by leveraging LLMs. Nevertheless, issues of inaccuracies, hallucinations, and limited domain-specific validation are important (95).

4.9.4 Multimodal AI platforms

Multimodal AI systems combine many types of data, such as molecular structure, protein sequence, transcriptomic profile, proteomic data, imaging data, clinical data and scientific data. Multimodal models can provide a more holistic view of host–virus interactions and therapeutic response by integrating multiple biological data sources (96). These approaches are especially appealing for multi-target antiviral drug discovery as they represent the complex interplay of biology occurring in viral infection, immune response, and drug action. The combination of omics data and structural and pharmacological data has the potential to significantly enhance the predictive accuracy and translational relevance (97).

4.9.5 Reinforcement learning for molecular design

Reinforcement learning is an emerging method exploring the problem of de novo molecular generation and optimization. For RL-based drug discovery, the AI agents produce and tweak molecular structures, and the rewards are computed by specific criteria including antiviral activity, target specificity, toxicity, pharmacokinetics and synthetic feasibility. This structure can be used to access huge chemical spaces and to perform multi-objective optimization in the process of development of multi-target antiviral agents. Reinforcement learning is often used in conjunction with generative models to create new compounds with more desired therapeutic properties (98).

These new AI paradigms are a step beyond the classic ML workflows. Their capacity to bring together disparate biological data, create new molecular structures, and provide for complex decision-making will be expected to speed the development of next-generation antiviral therapeutics. However, issues of data quality, data interpretability, computational demands and experimental validation are important factors in their successful implementation.

4.10 Drug repurposing using AI

Drug repurposing is the process of finding new therapeutic uses of existing drugs. This approach has received considerable interest in antiviral drug development since repurposed drugs already have established safety, pharmacokinetic and manufacturing profiles, so they do not need as much time and cost to develop. AI has made the process of repurposing drugs highly effective, as it utilizes massive amounts of data representing approved drugs, disease mechanisms, target proteins, gene expression, and clinical outcomes (99). Unexpected correlations between known drugs and viral diseases can be detected by ML and network-based approaches. In the context of the COVID-19 pandemic, AI-based drug repurposing methods were popularly applied to discover possible antiviral therapies, including Remdesivir, Favipiravir, and Molnupiravir. The same strategies are being used on other viral infections, to determine multi-target drugs that have multi-target activity on viral and host pathways (10).

4.11 Limitations and false-positive outcomes in AI-guided antiviral drug discovery

Although AI has made significant strides in antiviral drug discovery, a number of studies have noted some critical flaws in AI predictions. Among the most common issues is that of false-positive candidates, which show high computational scores but are not successful when tested in the experiments. Some of these failures can be caused by bias in the datasets, lack of biological complexity, lack of generalizability of the model, and poor quality of training data (100).

In the pandemic, many ML and DL platforms have been used to quickly target candidate antivirals in virtual screening and drug repurposing. However, several compounds with high docking scores and rankings using AI algorithms that were set or predicted to have strong SARS-CoV-2 antiviral activity did not show strong antiviral activity in cell-based or animal studies. Predicted binding affinities often did not predict the actual biological activity, as models cannot capture all the factors involved in uptake, metabolic stability, protein dynamics, off-target effects, and pharmacokinetic behavior of the compounds (101).

Drug repurposing using AI also yielded several candidates with initially encouraging computational results, but failed to show any clear or beneficial effects in clinical trials. A number of compounds found by network-based and machine-learning methods have shown poor activity in randomized clinical trials even though they have been predicted to be active in silico (102). These results highlight that computational predictions are hypothesis-generating devices, and not evidence of therapeutic potential.

A further major drawback is the fact that “activity cliffs” exist within data sets of structurally related molecules, which show markedly different biological activity. For such cases, AI models may not be good at predicting them, especially if the training data is not extensive enough or chemically balanced (103). Additionally, there are many examples of DL models that function as black-box models, where there is a lack of understanding of the mechanistic basis of the predictions, which leads to the possibility of choosing compounds that are not biologically relevant, but rather have spurious correlations (104).

Thus, to achieve successful antiviral drug discovery, the use of AI methods needs to be closely coupled with experimental validation, such as biochemical assays, cell-based screening, pharmacokinetic studies, toxicity assessment, and animal models. By integrating AI predictions with detailed biological understanding, quality data sets and meticulous experimental validation, these programs can significantly lower false positives and enhance the likelihood of successful AI-driven antiviral drug development that translates to human use.

4.12 Addressing dataset bias, overfitting, and reproducibility in AI-based antiviral drug discovery

The increasing use of ML and DL in antiviral drug discovery has raised several important issues of bias in the datasets, model overfitting, and the lack of reproducibility. Such hurdles can have a profound impact on the accuracy, applicability, and transferability of AI-driven forecasts. Therefore, significant work has been undertaken on model best practices to enhance model robustness and scientific reproducibility (105).

The most frequent cause of error in antiviral AI models is the presence of dataset biases. There could be bias due to unequal representation of viral species, overrepresentation of common chemical scaffolds, class imbalance between active and inactive compounds, or differences in experimental assay conditions (106). There are many public datasets that are quite rich in compounds acting against known viruses like HIV, influenza virus, hepatitis C virus and SARS-CoV-2, but there are less compounds acting against emerging or neglected viruses. To overcome these drawbacks, researchers have been able to rely on data curation, standardization processes, balanced sampling methods, and data augmentation methods to enhance data diversity and minimize systematic bias in their datasets (107).

Over fitting takes place when a model learns patterns in the training set instead of relationships. Overfitted models can perform very well on the training set, but usually fail to predict well on external sets. To prevent overfitting, various strategies that have been employed include cross validation, independent test sets, external validation sets, regularization techniques, dropout techniques, early stopping techniques, and ensemble-learning approaches. These approaches allow for greater generalization of the models and increased confidence in their predictive ability (108).

Reproducibility is another major hurdle in the quest of AI-aided drug discovery. The results reported may vary widely due to variations in the preprocessing of the data sets, the methods used for feature selection, the architectures of the models used, the process of hyperparameter optimization, or the methods used to evaluate the models (109). In response to these concerns, researchers are increasingly advocating open-science approaches, such as open access to datasets, source code, model parameters, and/or protocol for benchmarking. Standardized reporting guidelines and independent validation studies also help to provide a transparent comparison among predictive models (110).

For antiviral drug discovery applications, it is especially significant to have external validation. A model developed using data from one viral family/chemical space may not have a high accuracy for new viruses/pathogens or for structurally dissimilar compounds. As a result, it is increasingly recognized that thorough validation with independent data, predictions for the future, and experimental testing is critical to ensure model reliability (111). The use of techniques in explainable AI (XAI) can further enhance confidence in model predictions by revealing molecular characteristics and biological processes that help guide decisions (78).

4.13 Correlation between in silico predictions and experimental antiviral efficacy

These computational methods, such as molecular docking, QSAR modeling, ML, DL, molecular dynamics simulations, and AI-assisted virtual screening, have become essential tools in the antiviral drug discovery process. These methods can provide quick identification of the candidates from huge chemical library and prioritize them with a lower cost and time required for experimental screening. The correlation of the in-silico prediction with in vitro or in vivo antiviral activity, however, is not perfect and is a major challenge in computational drug discovery (112).

In general, the computational predictions are moderately successful in predicting compounds with potential biological activity, especially if good structural and experimental data is available. Molecular docking scores are often correlated with the target binding potential and ML models can be useful to predict antiviral activity, toxicity and pharmacokinetic properties in the chemical space defined by their training sets (113). However, computational predictions don’t always lead to experimental effectiveness as biological systems are much more complex than a computer model can accurately represent.

There are various reasons for discrepancies between in silico and experimental results. The most common molecular docking methodologies simulate binding affinities by considering simplified scoring functions that might not capture protein conformational flexibility, solvent effects, induced-fit conformational changes and dynamic molecular interactions (114). Likewise, ML and DL models could detect statistical trends within a training dataset but would not be able to capture the important biological features, like uptake into cells, distribution within cells, metabolic stability, immune modulation, tissue penetration, and host–virus interactions (115).

Additional complexity is added by transitioning from in vitro to in vivo efficacy. Potent antiviral activity in cell-based assays does not necessarily mean that the compound has good pharmacokinetic characteristics, minimal metabolism, good bioavailability, or minimal toxicity in animals or humans (115). Many compounds with successful pre-clinical or clinical development fail in the in-vivo phase due to these reasons, resulting in a failure to reach the market. These findings highlight the importance of computational predictions being used mainly to prioritize rather than prediction of therapeutic success (116).

To increase the accuracy of predictions, researchers are increasingly looking to using integrated workflows that involve molecular docking, MD simulation, ML models, pharmacokinetic prediction, toxicity analysis and experimental testing. Another method used to reduce false positives and to increase the chance of obtaining biologically relevant antiviral candidates is consensus scoring approaches, ensemble ML methods, and multi-parameter optimization strategies (117). In addition, future work is needed to validate the computational models with independent data and experimental tests to investigate their usefulness.

In general, in silico methods significantly speed up the discovery of antiviral drugs and enhance the prioritization of candidates, but their predictive power relies on the quality of the available data, the design of the model and the biological complexity. The combination of computational predictions with meticulous in vitro and in vivo validations is typically the most effective approach, offering a complementary mix of speed and reliability, with the advantage of computational methods and the precision of experimental science.

4.14 Experimental validation and translational challenges of AI-predicted multi-target antivirals

Over the past few years, AI-based methodologies have facilitated a dramatic improvement in the speed of antiviral drug discovery, with the ability to identify targets quickly, perform virtual screening, and identify potential antiviral drugs from repurposing or de novo molecular design (118). But calculation predictions don’t provide enough information to determine therapeutic efficacy. Experimental validation is a crucial process to ensure biological relevance of AI-generated candidates and their potential for clinical development (104, 119).

The validation process usually starts with biochemical assays to verify the binding of the compound to the target protein and to verify the inhibitory activity against viral or host proteins. These assays provide direct evidence of interaction of predicted compounds with their intended molecular target (119). Cell-based antiviral assay is conducted after biochemical evaluation to evaluate the efficacy of the candidate molecules to inhibit the viral replication under physiologically relevant conditions. These experiments often show a mismatch between the in-silico results and biological activity because of membrane permeation, accumulation in cell, metabolic instability, and off-target effects (120).

Experimental validation is especially challenging for multi-target antiviral agents which require multiple virus or host pathways to be modulated for therapeutic efficacy. Therefore, the assessment of target inhibition should be assessed in each biological system as well as synergistic, additive and antagonistic effects at multiple biological levels. These mechanisms of action are increasingly studied with systems biology approaches, transcriptomic analyses, proteomic profiling and network pharmacology studies (25).

While some AI-designed compounds show great promise in the lab, others face significant hurdles in the real world. There are factors such as poor bioavailability, rapid metabolism, poor tissue distribution, toxicity, immune responses, and suboptimal free drug levels, etc., that can affect the effectiveness of the therapy. As such, the full spectrum of assessments for absorption, distribution, metabolism, excretion, and toxicity (ADMET) are an essential part of candidate optimization.

There are further translational challenges associated with transitioning from the preclinical to the clinical stage. Before clinical trials, there needs to be a lot of evidence to support efficacy, safety, reproducibility and manufacturing feasibility from regulatory agencies. Additionally, computationally derived evidence may have issues of explainability, model transparency, IP issues, and regulatory acceptance. These factors emphasize the need for incorporating AI predictions into robust experimental and clinical validation processes (121).

Modern drug discovery for antiviral agents increasingly relies on computational-experimental drug discovery processes for improving the chances of translational success. In these methods the output of the experiments is continually fed into ML models, which are then used to make better predictions and boost the model’s performance. Strategies that have emerged to lower false-positive rates and to speed up the lead optimization process are active learning, closed-loop optimization, and human-in-the-loop AI systems (10).

4.15 Ethical and regulatory considerations in AI-guided antiviral drug discovery

AI is increasingly becoming a vital tool in the discovery of new antiviral drugs, presenting opportunities for swift progress in therapeutic development but also posing potential ethical, regulatory, and governance issues. With AI-generated predictions becoming more influential in target selection, lead optimization, and clinical decision-making, it is crucial to ensure transparency, reliability, and accountability in order to successfully translate predictions into clinical practice (122).

Data quality and representation is one of the significant ethical issues. The datasets used for training and validation are of significant importance for the ML models. Unfortunately, the datasets used to make predictions might be biased, incomplete, or unrepresentative, and so the performance of models may be different for different viral pathogens, patient populations, or geographic regions. These biases can inadvertently reduce the usefulness of AI-generated therapeutics and help to overcome emerging or under-researched viral diseases (123).

There are more challenges to transparency and explainability. For many of the more sophisticated DL systems, the inner workings of the system are not easily explainable, and the process used to make a decision could be considered a “black-box” system. However, limited explainability can result in decreased confidence among researchers, clinicians, and regulatory bodies in assessing AI predictions. As a result, explainable AI techniques are being created and developed to enhance the interpretability of models and to aid in their scientific and regulatory review.

Reproducibility and data sharing are other important factors. The results of studies may differ due to variability in datasets, preprocessing techniques, model structures, and evaluation techniques. Implementing open-science practices, common measurement tools and reporting standards can enhance reproducibility and increase trust in the results of AI in the drug discovery process (124).

Regulators are actively considering guidelines for the use of AI in the pharmaceutical R&D process. Agencies are exploring regulatory frameworks concerning the application of AI in pharmaceutical R&D. Model validation, data integrity, traceability, transparency and continuous performance monitoring have become key focuses of regulatory bodies. While AI tools are more about assisting decision making, it is essential that the predictions made by AI tools are scientifically sound, reproducible and backed up by relevant experimental evidence before they can be translated to the clinic (125).

The increasing volume of genomic, clinical, and pharmacological data incorporated into AI systems emphasizes the critical need for privacy and cybersecurity measures. Secure data handling, patient confidentiality and adherence to relevant data protection regulations should be ensured through appropriate safeguards. Additionally, the legal challenges and uncertainties surrounding the ownership of AI-generated molecular structures and algorithm-driven discovery efforts are ongoing and may need to be addressed further in the regulations (125). The schematic representation of the AI-guided multi-target antiviral discovery is illustrated in Figure 3.

4.16 Critical evaluation of computational and AI approaches in multi-target antiviral drug discovery

In recent years, the advancement of computational and AI based methodologies has revolutionized the field of antiviral drug discovery by greatly increasing the number of novel therapeutic candidates identified in a quicker and more efficient way than traditional experimental strategies. Each computational strategy, however, has its own merits and drawbacks which can affect its viability for different phases of multi-target antiviral development (93).

While molecular docking and molecular dynamics (MD) simulations are widely used techniques for structure-based methods, these approaches continue to be essential tools for understanding protein–ligand interactions. Molecular docking is a quick screening of large compound libraries and gives insights into the mechanism of potential binding modes. Docking predictions are, however, often based on simplified scoring functions and may not include the flexibility of the protein, solvent effects, or dynamic conformational changes. These are partially overcome by molecular dynamics simulations, where molecular behavior is modeled through time, making the predictions more reliable. However, MD simulations are highly computationally demanding tasks, and are not ideal for large-scale screening applications (126, 127).

However, ML techniques provide significant benefits in terms of the predictive efficiency and scalability. Random forests, support vector machines and gradient boosting models can quickly process huge chemical data sets and find compounds that have good antiviral activity. ML can combine various biological, chemical and pharmacological characteristics, which is not possible with the traditional approach to docking structures. Prediction accuracy is still highly sensitive to the quality of training data, feature engineering, and the diversity of the datasets (74, 128).

This is also extended by DL techniques, which automatically learn complex molecular representations directly from the raw data. The performance of these tasks in activity prediction, target identification, and molecular generation tasks has been shown to be better than that of CNN, GNN, RNN, and transformer architectures (117). However, DL models, in general, demand huge amounts of data, significant computing power, and a lot of hyperparameter tuning. They might also have a black-box status that could restrict mechanistic understanding and regulatory acceptance (129).

Advances in AI-led protein structure prediction, such as AlphaFold, have greatly accelerated the antiviral target characterization process by providing models with unprecedented accuracy of the protein structure. These tools have helped to increase the structural information of viral proteins and made it possible to structure-based drug discovery for a number of viral targets that could not be solved before. For conformationally dynamic proteins, transient protein complexes and flexible regions the structures predicted might not be as reliable and it is advisable to validate them experimentally and perform complementary computational analyses (130).

Generative AI approaches are a significant leap forward in de novo antiviral design, allowing the generation of novel molecular scaffolds with multiple goals optimized for them. These models can be used to traverse enormous chemical space that is not present in any compound libraries. Many molecules generated, however, may have difficulties with synthesis, metabolic stability, toxicity, or experimental validation, so there is a need for an integrated medicinal chemistry assessment (131).

However, none of the computational methodologies alone can be used to solve all the problems involved in antiviral drug discovery. Mechanistic understanding is gained using traditional structure-based methods, rapid prioritization is achieved with machine learning, complex nonlinear relationships are captured with deep learning, and chemical exploration is extended with the help of generative AI for structural characterization via AlphaFold (132). In this context, the most successful antiviral discovery approaches today are more and more hybrid computational approaches, involving the integration of various complementary methods and systematic experimental validation. Integrated approaches are especially useful in the discovery of antivirals against multiple targets, in which greater biological complexity may hinder the use of any single computational method (133).

5 Medicinal chemistry strategies for multi-target antivirals

Medicinal chemistry provides the structural foundation for the successful development of MTAs. In contrast to single-target drugs, which are constructed to selectively interact with a single viral protein, the design strategy required by MTA is more advanced with one molecule having to interact well with a variety of viral or host targets without sacrificing the desirable pharmacokinetic, pharmacodynamic, and safety properties. Such molecules are extremely difficult to design due to the fact that activity to one target can occasionally decrease activity to another target, whereas overly complicated design might diminish oral bioavailability, membrane permeability, metabolic stability or toxicity (134).

Recently, medicinal chemists started paying more attention to the development of molecules that could have a multifaceted effect on various steps of the viral life cycle, such as viral attachment, membrane fusion, viral genome replication, viral protein processing, virion assembly, and host immune response. These compounds are especially useful in the treatment of highly mutable viruses since it minimizes the likelihood of a single mutation that can cause a therapeutic failure. Multiple medicinal chemistry approaches have thus been attained to realize multi-target antiviral activity (135).

5.1 Hybrid pharmacophore design

One of the most popular medicinal chemistry strategies that have been used extensively in the development of MTAs is hybrid pharmacophore design. In this method, pharmacologically active fragments of two or more drugs or lead compounds are coupled to form one molecular part. The resulting hybrid molecule is predicted to have the biological properties of both parent pharmacophores and antiviral activities via several mechanisms (135).

As an example, a hybrid antiviral molecule can be a nucleoside analogue pharmacophore in combination with a viral protease inhibitor or an entry blocker. In this type of design, a component of a molecule could prevent viral replication of the genome, another could disrupt viral protein production, or host-cell entry. Hybrid molecules thus offer a chance to inhibit several phases of the viral life cycle at the same time (136).

The hybrid pharmacophore design relies on the compatibility of the pharmacophores of choice. Medicinal chemists have to pay attention to steric hindrance, electronic distribution, lipophilicity, potential of hydrogen-bonding, molecular flexibility, and conformational stability. The linker connecting the two pharmacophores is also extremely important because its length, rigidity, polarity, and orientation can strongly influence biological activity (137).

Short rigid linkers can also in certain instances maintain the orientation of both pharmacophores and enhance binding efficiency, but the highly flexible ones can enable the molecule to conform to the various target environments. But too much flexibility may cause the entropy loss during binding and decrease potency. On the same note, large hybrid molecules can be highly antiviral but low membrane permeable or low oral bioavailability due to their higher molecular weight and polarity (138).

Hybrid pharmacophore design has extensively been used in the discovery of HIV, HCV, and SARS-CoV-2 drugs. A number of investigational compounds have been found to have combined reverse transcriptase inhibitory effect with integrase inhibitory effect or a combination of viral protease inhibitory effect and host-targeted effects. These types of designs are particularly handy in overcoming resistance since the virus will need to mutate through several different pathways to avoid being inhibited (139).

5.2 Molecular conjugation and fusion strategies

Molecular conjugation and fusion strategies are closely related approaches that are frequently used to develop MTAs. Molecular conjugation is the chemical conjugation of two active molecules or pharmacophoric fragments using a covalent linker. Fusion strategies on the other hand tend to combine the necessary attributes of two molecules into one compact scaffold without the need to have a lengthy connecting linker (140). There are a number of linkers that can be employed in conjugation strategies, such as ester, amide, ether, carbamate, hydrazone, triazole, peptide, and disulfide linkers. The linker type selected depends on the chemical stability required, metabolic behavior and release characteristics of the drug. As an example, ester linkers are commonly employed when controlled hydrolysis is required and amide linkers are preferred when more stability is needed (141). The conjugation methods can generate bifunctional molecules that can target viral and host proteins at the same time. As an example, a conjugate can include a viral polymerase enzyme with a kinase enzyme that inhibits host-cell signaling pathways needed to carry out replication. These compounds can have synergistic antiviral properties and will decrease the necessity of combination therapy with several distinct drugs (142).

Fusion-strategies are commonly used in cases where a medicinal chemist wants to lower molecular weight and enhance pharmacokinetic properties. Fusion strategies can allow the formation of smaller, metabolically stable and orally bioavailable molecules by incorporating overlapping structural motifs of multiple pharmacophores into a single scaffold compared to the simple conjugates (143).

Nevertheless, conjugation and fusion strategies are both quite challenging. The higher the molecular complexity, the worse the aqueous solubility, the low absorption in the intestine, the rapid metabolism, or the tissue penetration. Thus, these strategies need to be optimized on a large scale to establish a balance between the biological activity and drug-like characteristics (142).

5.3 Design of dual inhibitors targeting viral enzymes

The viruses require a number of enzymes to replicate and survive, among them proteases, polymerases, integrases, helicases, neuraminidases, and methyltransferases. Consequently, the development of dual inhibitors that address two viral enzymes has emerged as a significant field of activity in the antiviral medicinal chemistry (144). Dual inhibitors are structures that are meant to disrupt two important enzymatic processes at the same time. As an example, a single compound can be both a viral protease and polymerase inhibitor and, therefore, stop the maturation of proteins and the replication of the genome simultaneously. Molecules that inhibit neuraminidase as well as viral entry have been investigated in the context of influenza research, whereas dual reverse transcriptase/integrase inhibitors have been examined in the context of HIV therapy (145).

The design of dual inhibitors involves intensive understanding of structural and mechanistic properties of the two antagonists. The active sites of two enzymes in certain situations can have some common hydrophobic pockets, hydrogen-bonding patterns or catalytic residues and it is therefore possible to design a single molecule capable of binding the two protein enzymes. In other instances, each target may be optimized on different regions of the same molecule (145).

One of the biggest problems in designing dual inhibitors is the problem of a balance of potency between the two targets. Any structural changes which enhance affinity with one enzyme can decrease the affinity with the second target. Consequently, medicinal chemists tend to use the structure-activity relationship studies, the molecular docking and the molecular dynamics simulations to improve the activity profile (146). The use of dual inhibitors is especially useful in resistance prevention since the virus would have to mutate various proteins to avoid the inhibitor. This contributes to a great deal of reducing the chances of treatment failure and, possibly, increases the clinical lifespan of antiviral treatment (147).

5.4 Incorporation of host-targeting components

Viruses have a high dependency on host-cell machines to enter, replicate, translate, fold proteins, and release viruses. The host proteins mutate much slower than those of the virus, so integrating host-targeting systems into the MTA can be used to offer a more sustainable antiviral approach. The most common targets of host-targeted antiviral design include cellular kinases, heat shock proteins, cyclophilins, endosomal traffic, enzymes of lipid metabolism, membrane fusion, and immune signaling. Examples include viruses that have a need to use host kinases to regulate viral transcription and replication, and others that need chaperone proteins to allow viral proteins to fold correctly (148, 149).

By using viral enzyme inhibitors together with fragments that disrupt host-cell processes, medicinal chemists can add host-targeting elements to MTAs. These compounds can inhibit the viral replication as well as the cellular support systems that are needed to infect cells.

Host-targeted strategies also can be particularly effective in the development of broad-spectrum antivirals since several viruses tend to use the same host pathways. As an example, endosomal acidification/host protease inhibitors can be active against influenza, coronaviruses, and filoviruses (150).

Although these benefits exist, host-targeting strategies present a higher toxicity risk since host proteins are not only engaged in normal physiology, but also management. Thus, medicinal chemists should be cautious in the identification of host factors which are significant to viral replication but not as important to normal cell processes. To minimize the adverse effects, selective modulation instead of complete inhibition is usually desirable (151).

5.5 Optimization of selectivity and toxicity

Among the most challenging things in multi-target antiviral design are the high selectivity and low toxicity. Since the purpose of MTAs is to engage multiple targets at the same time, they have very broad biological activity in many cases compared to traditional single-target drugs. Although this general activity can enhance the antiviral activity, it can also pose the risk of off-target action and toxicity. Medicinal chemists adopt various methods of enhancing selectivity. The size of the molecule, lipophilicity, polarity, stereochemistry, aromaticity, and bonding properties are highly optimized to ensure that the molecule feels inclined to interact with the target of interest and to reduce non-specific binding (6). Bioisosteric replacement, conformational restriction, prodrug formation and fluorination are structural modifications that are commonly employed to enhance selectivity and metabolic stability. Premature testing of absorption, distribution, metabolism, excretion and toxicity characteristics have proven to be a significant aspect of MTA development. Excessively lipophilic compounds can lead to deposition in the tissues and result in toxicity, and the highly polar molecules can lack membrane permeability and oral absorption. Equally, molecules with a strong ability to inhibit cytochrome P450 enzymes can give rise to any meaningful drug-drug interactions (152). In vitro cytotoxicity testing, microsomal stability testing, plasma protein binding testing, ERG channel screening, mitochondrial toxicity testing, and computational toxicity prediction are thus gaining more and more importance in modern medicinal chemistry during the early drug development. These techniques can be used to determine the presence of problematic compounds prior to preclinical or clinical research incursion. Finally, the effective development of MTAs should be based on a fine line between antiviral activity, multi-target activity, safety, selectivity, and behavior as a drug. Medicinal chemistry coupled with computational modeling, structure-based design, and AI are likely to have a significant contribution to producing safer and more efficient antiviral therapeutics in the future.

6 Host-targeted multi-target antiviral approaches

Host-targeted antiviral therapy has emerged as a highly promising strategy in the development of multi-target antiviral agents because viruses are heavily dependent on host-cell machinery for successful infection, replication, assembly, and release. In contrast to direct-acting antivirals that target viral enzymes (e.g. proteases, polymerases or neuraminidases), host-targeted antiviral therapy blocks host proteins and pathways that are exploited by the virus. These range from protein kinases and chaperone proteins to immune signaling, host proteases, endosomal trafficking, lipid metabolism and receptor-mediated entry (97).

An important benefit of host-targeted drugs is that host proteins are more genetically stable than viral proteins. Viral proteins, especially for RNA viruses such as HIV, influenza, hepatitis C virus and SARS-CoV-2, rapidly mutate, giving rise to resistance. However, host-cell proteins do not mutate easily and host-targeted drugs may have a longer lifespan with reduced chance of resistance. Moreover, many viruses use the same host protein pathways, making single host-targeted antivirals potentially active against a range of viruses (153).

However, host-targeted therapy also presents major challenges because host proteins play essential roles in normal physiology. Thus, blocking these pathways may be toxic, immunosuppressive or disrupt metabolism. As a result, efficacy of host-targeted MTAs hinges on the ability to target host pathways that are essential for viral survival but not for host cell survival (154).

6.1 Kinase inhibitors

Protein kinases play a critical role in regulating cell growth, death, metabolism, endocytosis, trafficking and immune responses. Viral infection and replication often hijack these kinase-dependent pathways to facilitate viral entry, replication, protein translation and evasion of immune responses. Consequently, host kinase inhibitors have emerged as promising broad-spectrum antiviral agents. The phosphoinositide 3-kinase (PI3K)/Akt/mammalian target of rapamycin (mTOR) pathway is one of the most well-studied kinase pathways exploited by viruses. It plays a key role in protein synthesis and cell survival, two processes critical for successful viral replication (155). The activation of PI3K/Akt by influenza viruses, hepatitis viruses, and coronaviruses protects cells from death and extends the life of the virus. So, blocking this pathway has been found to decrease viral replication and promote cell death (156). Cyclin-dependent kinases (CDKs) are also important because many viruses rely on them for regulating transcription. CDK9, for instance, is required by HIV for phosphorylation of RNA polymerase II, leading to the activation of virally encoded genes. Roscovitine, a CDK inhibitor, has been shown to inhibit HIV, herpes simplex virus and cytomegalovirus. It has been shown that Roscovitine inhibits HIV transcription by preventing CDK9-mediated activation of viral genes, and therefore inhibiting viral RNA synthesis and replication (157).

Abelson tyrosine kinases play a role in cytoskeletal organization, endocytosis and membrane fusion, which are all processes relevant to viral entry. Two FDA-approved Abl kinase inhibitors, Imatinib and Dasatinib, have demonstrated antiviral activity against poxviruses, Ebola virus, and coronavirus. Imatinib was found to suppress viral fusion and syncytium formation in SARS-CoV and SARS-CoV-2 infection models by targeting Abl2 kinase. Moreover, reports during the COVID-19 pandemic have indicated that Imatinib may prevent vascular leakage and inflammation in severe COVID-19 (158).

Janus kinase (JAK) inhibitors have received major attention because of their dual antiviral and anti-inflammatory effects. A prominent example is Baricitinib. Baricitinib is a Janus kinase 1 and 2 inhibitor, and also blocks AP2-associated protein kinase 1, which is required for clathrin-mediated endocytosis of SARS-CoV-2. In clinical trials, Baricitinib combined with Remdesivir shortened recovery time and improved outcomes in COVID-19 patients admitted to hospital when compared with Remdesivir alone (159).

6.2 Chaperone-targeting agents

Molecular chaperones are host proteins that assist in the folding, stabilization, trafficking, and degradation of cellular and viral proteins. Host chaperones are required for the folding and assembly of many viral proteins; therefore, these proteins are potential targets for host-directed antiviral strategies.

Heat shock protein 90 (Hsp90) is one of the most important chaperones involved in viral replication. Hsp90 is involved in the stabilization of viral polymerases, capsid proteins, reverse transcriptase and regulatory proteins. Hsp90 is required for replication of influenza virus, HIV, hepatitis C virus, dengue virus and coronaviruses (160).

The Hsp90 inhibitors most widely studied are geldanamycin and 17-AAG. These drugs compete with ATP binding to Hsp90, thereby inhibiting viral protein maturation. In influenza virus, Hsp90 inhibition prevents the formation of the viral RNA polymerase complex, which is required for viral RNA synthesis. Similar effects have been reported in hepatitis C virus and dengue virus model (161).

Another molecular chaperone that plays a role in viral infection is Hsp70. It aids in the folding of newly translated viral proteins and in the assembly of replication complexes. Studies have shown that inhibition of Hsp70 suppresses dengue virus, Zika virus, and Japanese encephalitis virus replication (162).

Cyclophilins are another family of host chaperones that are critical for replication. Cyclophilin A is required for the replication of hepatitis C virus and HIV. Cyclosporine A has also been shown to have broad antiviral effects by inhibiting cyclophilins’ involvement in viral protein folding and viral replication complex assembly. But the immunosuppressive activity of Cyclosporine A precludes its use as an antiviral agent. To overcome this issue, non-immunosuppressive derivatives such as Alisporivir were developed. Alisporivir showed potent anti-HCV activity and potential to be used in combination with direct-acting antivirals (163).

6.3 Immune-modulating pathways

Host immune responses are essential for control of viral infections and clearance of infected cells. However, many viruses suppress innate immune responses during early infection and trigger excessive inflammatory responses during advanced disease stages. Therefore, modulation of immune signaling pathways has become an important strategy in host-targeted antiviral therapy (164).

Interferons are among the most important endogenous antiviral mediators. Type I interferons activate the transcription of interferon-stimulated genes that suppress viral replication, enhance antigen presentation and stimulate immune cells like natural killer cells and macrophages. Interferon alfa was extensively used for treatment of chronic hepatitis B and hepatitis C before the advent of direct-acting antivirals. Interferon-based treatments have been shown to decrease viral load and improve long-term liver disease outcomes in clinical trials (165).

Toll-like receptors (TLRs) are another important component of innate antiviral immunity. Activation of TLR3, TLR7, and TLR8 stimulates interferon production and antiviral cytokine release. As a result, TLR agonists have been explored as immune adjuvants in chronic viral infections, including hepatitis B and HIV (166).

In severe viral infections, activated immunity can be detrimental. This was the case during the COVID-19 pandemic, where an elevated inflammatory response (cytokine storm) was responsible for acute respiratory distress syndrome (ARDS) and death. Dexamethasone is one of the most commonly used drugs as it reduces hyper-inflammatory responses via the glucocorticoid receptor. The trial demonstrated that Dexamethasone significantly reduced mortality in patients requiring oxygen therapy or mechanical ventilation (167).

Interleukin-6 inhibitors such as Tocilizumab were also used to reduce inflammation in severe COVID-19 patients. Tocilizumab inhibits IL-6 receptor, which decreases cytokine secretion, blood vessel damage and inflammation. Likewise, Baricitinib had both antiviral and anti-inflammatory effects through the inhibition of Janus kinases (JAKs) and their downstream effects on cytokine signaling (167).

6.4 Host proteases and entry factors

Host proteases and receptors are essential for viral entry into host cells. Many enveloped viruses require cleavage of their surface glycoproteins by host proteases before membrane fusion can occur. Therefore, inhibition of host proteases can block an early and essential step in viral infection.

Transmembrane serine protease 2 (TMPRSS2) is one of the most important host proteases involved in viral entry. TMPRSS2 cleaves the spike protein of SARS-CoV-2, enabling fusion of the viral envelope with the host-cell membrane. Camostat mesylate and Nafamostat mesylate are serine protease inhibitors that block TMPRSS2 activity. In vitro studies showed that these drugs significantly reduce SARS-CoV-2 entry and replication (168).

Furin is another important host protease because it activates viral glycoproteins in influenza virus, HIV, Ebola virus, and coronaviruses. SARS-CoV-2 contains a furin cleavage site within its spike protein that enhances infectivity. Therefore, furin inhibitors are being explored as broad-spectrum antiviral agents. Cathepsins are lysosomal proteases involved in viral entry through the endosomal pathway. Viruses such as Ebola virus and SARS-CoV-2 can use cathepsin-mediated activation after endocytosis. Inhibition of cathepsins may therefore provide an additional strategy to block viral entry (169).

Host receptors are equally important targets for antiviral intervention. HIV requires CD4 and CCR5 receptors for attachment and entry into host cells. Maraviroc is a CCR5 antagonist that prevents HIV from interacting with host cells and has been successfully used in HIV therapy. Similarly, SARS-CoV-2 relies on ACE2 for attachment, while influenza viruses bind to sialic acid receptors present on respiratory epithelial cells (17).

6.5 Benefits and risks of host-targeted therapy

Host proteins are less likely to mutate than virus proteins, so the possibility of antiviral resistance is less likely. In addition, many viruses depend on similar host pathways, making host-targeted agents potentially useful as broad-spectrum antivirals (170). However, host-targeted therapy has potential drawbacks. Host proteins play an important role in normal cell function, and may cause toxicity, immune suppression or organ dysfunction. Kinase inhibitors can cause blood and heart defects, while immunomodulators can promote secondary infections (150).

So, the future of host-targeted MTAs lies in identifying host processes that play a critical role in the viral replicative cycle but are not needed for normal cellular processes. Advances in medicinal chemistry, systems biology, and AI are expected to improve the safety and selectivity of these therapies (171).

7 Natural product-derived multi-target antiviral scaffolds

Nature has long been the source of many pharmacologically active molecules used in drug discovery as shown in Figure 4. The complexity, diversity and pleiotropic bioactivity of natural products make them especially valuable for the discovery of poly-pharmacological antiviral drugs. Natural products frequently interact with multiple targets, whereas many synthetic small molecules are designed to target a single protein. This is a desirable feature in antiviral drug discovery as viruses exploit multiple mechanisms for entry, replication, assembly and evasion of the immune system (172). Antivirals derived from natural products may inhibit viral enzymes, block viral entry, disrupt replication complexes, interfere with host pathways, or modulate host immune cells. Furthermore, many natural products have antioxidant and anti-inflammatory effects, which may also contribute to their value in treating viral infections with an inflammatory component (173).

7.1 Flavonoids

Flavonoids are polyphenolic compounds found in fruits, vegetables, tea, herbs and medicinal plants. These compounds have diverse biological activities such as anti-viral, anti-inflammatory, anti-oxidant and immuno-modulatory. Flavonoids have several hydroxyl groups that allow them to bind to viral proteins and host enzymes (174). Quercetin is a well-studied anti-viral flavonoid. It is effective against influenza virus, dengue virus, hepatitis viruses, Zika virus and SARS-CoV-2. Quercetin targets viral proteases, polymerases and helicases, and reduces the secretion of inflammatory cytokines. In recent SARS-CoV-2 research, quercetin was found to block the main protease (Mpro) of SARS-CoV-2 and has been shown to block viral entry by modulating the ACE2 receptor (175).

Another significant antiviral flavonoid is luteolin. It has been shown to inhibit Japanese encephalitis virus, dengue virus, HIV, and SARS-CoV. Studies suggest that luteolin inhibits viral replication by blocking RNA synthesis and interfering with host-cell signaling pathways required for viral protein production. Baicalein and baicalin, isolated from Scutellaria baicalensis, have shown potent antiviral effects against influenza virus, hepatitis B virus, HIV, and SARS-CoV-2. Baicalein blocks the SARS-CoV-2 main protease and has viral replication inhibitory effects while baicalin also has anti-inflammatory effects, which may help to neutralize cytokine-induced tissue damage (176).

Epigallocatechin gallate (EGCG), the most abundant catechin found in green tea, has antiviral effects against HIV, hepatitis C virus, influenza virus and coronaviruses. EGCG may prevent attachment of virus to target cells, inhibit viral entry, and inhibit membrane fusion. In HIV infection, EGCG has been shown to bind to the CD4 receptor and reduce viral attachment (177).

7.2 Alkaloids

Alkaloids are nitrogenous natural products with various therapeutic effects. Many alkaloids have demonstrated their antiviral activity by interacting with viral enzymes, viral nucleic acid synthesis and host cell signaling. Berberine is a potent antiviral alkaloid. It is active against influenza virus, hepatitis C virus (HCV), human immunodeficiency virus (HIV), chikungunya virus and SARS-CoV-2. Berberine has several modes of action including inhibition of viral replication, down-regulation of MAPK, suppression of viral protein synthesis and down-regulation of pro-inflammatory cytokines (178).

Lycorine, an alkaloid found in Amaryllidaceae species, is a potent antiviral agent against coronaviruses, poliovirus, Zika virus, and SARS-CoV. Lycorine has been shown to suppress viral RNA synthesis and interfere with protein elongation, resulting in the inhibition of viral replication. The biscoclaurine alkaloid cepharanthine was of particular interest during the COVID-19 pandemic due to its reported in vitro activity against SARS-CoV-2 (179). It is thought to block viral entry, membrane fusion, and inflammation-associated gene expressions, including NF-kB. In vitro experiments demonstrated that cepharanthine decreased replication of SARS-CoV-2. Emetine is another alkaloid with broad antiviral activity against Zika virus, cytomegalovirus, Ebola virus and coronaviruses. It inhibits protein synthesis and interferes with viral RNA replication. However, the clinical application of emetine is limited by cardiotoxicity (180).

7.3 Terpenoids

Terpenoids are a group of naturally occurring compounds, derived from isoprene. A large number of terpenoids have antiviral activity by blocking viral entry, membrane fusion, protease enzyme activity and host inflammatory responses. One of the most studied antiviral terpenoids is the triterpenoid saponin glycyrrhizin from licorice root. It is active against hepatitis viruses, HIV, influenza virus, SARS-CoV, and SARS-CoV-2. Glycyrrhizin blocks adsorption and penetration, inhibits replication and regulates immune activity. During the SARS outbreak, glycyrrhizin demonstrated greater antiviral activity against SARS-CoV than several conventional antiviral agents (181, 182).

Andrographolide, a diterpenoid lactone from Andrographis paniculata, exhibits activity against dengue virus, influenza virus, hepatitis C virus, and SARS-CoV-2. It has been shown to inhibit proteases and polymerases of viruses and also inhibits inflammatory cytokines and oxidative stress. Artemisinin and its derivatives (first developed as antimalarial drugs) are also antiviral against hepatitis viruses, cytomegalovirus, Epstein-Barr virus and SARS-CoV-2. They have been shown to suppress NF-kB, oxidative stress, and protein synthesis to exert antiviral activity (183). Another triterpenoid with antiviral properties is betulinic acid, which has activity against HIV and influenza virus. It has been used as a scaffold for the development of HIV maturation inhibitors such as bevirimat, which blocks viral particle maturation (184).

7.4 Polyphenols

Polyphenols are a family of plant-derived compounds that have antioxidant, anti-inflammatory, and antiviral properties. These compounds contain several hydroxyl groups that interact with viral and host proteins and signaling pathways. The stilbene polyphenol resveratrol, which is present in grapes and berries, has been found to have antiviral effects against influenza virus, respiratory syncytial virus, herpes simplex virus, Middle East Respiratory Syndrome Coronavirus (MERS-CoV), and SARS-CoV-2. Resveratrol inhibits the nuclear factor-kappa B (NF-kB) pathway, viral gene expression and oxidative stress (185).

Turmeric’s active ingredient, curcumin, has also been shown to have broad-spectrum antiviral actions against HIV, hepatitis viruses, influenza virus, Zika virus, and coronaviruses. Curcumin can bind to viral proteins to inhibit viral entry, suppress viral proteases, and prevent inflammatory cytokines from being produced. It also modulates host cell processes such as NF-kB, PI3K/Akt and MAPK (186). Another class of antiviral polyphenols are tannins. Pomegranate punicalagin has inhibited influenza virus and herpes simplex virus by interfering with viral attachment and membrane fusion. Similarly, ellagic acid has demonstrated activity against hepatitis viruses and HIV (187).

7.5 Marine-derived natural products

Marine organisms are a valuable source of structurally novel antiviral compounds. Organisms such as sponges, algae, tunicates, fungi and bacteria, produce secondary metabolites that are often structurally unique and biologically active. Red algal sulfated polysaccharides (carrageenans) have been reported to have antiviral activity against HPV, herpes simplex virus, influenza virus, and coronaviruses. Their antiviral activity is associated with interference in viral adsorption and cell entry (188).

Fucoidan, a sulfated polysaccharide from brown algae, has been shown to inhibit influenza virus, HIV, and SARS-CoV-2. It exerts antiviral activity by blocking adsorption and modulating immunity. Griffithsin, a lectin from red algae, is a powerful antiviral agent. It binds to glycoproteins present on the viral envelope and prevents viral attachment and fusion. Griffithsin has potent activity against HIV, SARS-CoV, MERS-CoV and SARS-CoV-2. Marine sponges have also provided potent antiviral compounds. Manoalide and avarol from marine sponges have been shown to have antiviral effects against HIV by inhibiting reverse transcriptase and replication (189).

Overall, natural products offer considerable promise for the discovery of multi-target antiviral agents due to their diversity and ability to regulate multiple pathways within the virus and/or host. Moreover, research is needed to enhance their pharmacokinetic and pharmacodynamic properties, bioavailability, selectivity and clinical potential through medicinal chemistry modifications and drug delivery technologies.

7.6 Comparative evaluation of promising natural product-derived multi-target antiviral scaffolds

The structural diversity, biological activity, and multi-target nature of natural products make them one of the most promising sources of antiviral drug leads (Figure 4). But, not every class of natural products has equal translatability. A thorough analysis of their antiviral activity, pharmacology, toxicity, and progress in development is essential to determine the most promising scaffolds to pursue further development for therapeutic purposes (190).

Flavonoids, one of the different natural product classes, have been the focus of great interest in view of their wide spectrum of antiviral activity against various influenza viruses, coronaviruses, dengue virus, hepatitis viruses, and human immunodeficiency virus (HIV). Quercetin, baicalein, luteolin and myricetin potently disrupt viral entry, replication, viral proteases and host inflammatory pathways (191). The relatively good safety profiles and the amount of preclinical data that has been performed make flavonoids potentially good candidates for further development. However, there are still challenges pertaining to poor oral bioavailability, high rate of metabolism and limited systemic exposure (192).

Another promising group of antiviral agents are the alkaloids. Multiple alkaloids have been shown to exhibit strong inhibitory properties on viral polymerases, proteases and host-cell signaling pathways. Some alkaloids are found to be more potent and target specific than many polyphenolic compounds. However, toxicity, narrow therapeutic range and pharmacokinetic properties could restrict the clinical progress of certain members of this class (178).

Terpenoids and triterpenoids have proven to be highly promising antiviral scaffolds due to the potential to target multiple viral and host targets. Derivatives of betulinic acid and related terpenoids exhibit antiviral, anti-inflammatory and immunomodulatory activity, while glycyrrhizin has been shown to have antiviral activity. They have poly-functional mechanisms which are very suitable for multi-target antiviral strategies. Still, optimization of solubility, bioavailability and formulation is required to make the product available to wider clinical use (193, 194).

Polyphenols and phenolics related to them also have wide antiviral activity by modulating oxidative stress, inflammatory responses, viral enzyme activity and host-cell pathways. These compounds can have very good safety parameters, but the pharmacokinetics factors and bioavailability are often sub-optimal therefore, limiting clinical translation (195).

Considering their developmental stage, flavonoids and terpenoid scaffolds currently seem to have the best combination of antiviral breadth, multi-target activity, safety and translation (Table 3), and natural product-derived multi-target antiviral compounds are represented in Table 4. Alkaloids can be more potent, but typically, they need to be fully optimized for safety. Thus, for future development, it is essential to optimize the pharmacokinetic properties, target selectivity, as well as to combine medicinal chemistry optimization and AI-guided lead design to accelerate the advancement of the natural-product-based antivirals toward clinical evaluation.

Table 3. Comparative evaluation of major natural product-derived antiviral scaffolds.

Table 4. Natural product-derived multi-target antiviral compounds.

8 Resistance studies and strategies to overcome viral escape

Resistance to antiviral drugs is a major challenge for the sustained treatment of viral infections. Viral resistance is driven by genetic mutations that confer reduced susceptibility to antiviral drugs while maintaining the viruses’ ability to reproduce. Resistance is more likely to occur with RNA viruses such as HIV, influenza, hepatitis C, and SARS-CoV-2 due to high viral mutation rates, short generation times and error-prone viral polymerases (Table 5). Resistant viruses can lead to treatment failure, persistent infection, disease progression and spread of resistant viruses (3). During infection, viral replication is a dynamic process involving the generation of viral progeny. With each viral replication, the mistake-prone viral polymerases introduce mutations, resulting in a population of viral variants (quasi species). Some of these variants may possess mutations that make them partially or fully resistant to a specific drug (5).

Table 5. Examples of multi-target antiviral agents and their mechanisms.

8.1 Mechanisms of antiviral resistance

Antiviral resistance can arise through several molecular mechanisms, including target-site mutations, altered drug activation, reduced intracellular drug accumulation, compensatory mutations, and changes in host-cell pathways (3).

One of the most common mechanisms is mutation of the viral target protein. In HIV infection, resistance to reverse transcriptase inhibitors develops through amino acid substitutions in the reverse transcriptase enzyme that reduce drug binding. For example, the M184V mutation in HIV reverse transcriptase causes resistance to lamivudine and emtricitabine by altering the geometry of the nucleoside-binding pocket. Similarly, the K103N mutation reduces susceptibility to non-nucleoside reverse transcriptase inhibitors such as efavirenz and nevirapine (215).

Resistance to HIV protease inhibitors can also occur through mutations within the protease active site. Mutations such as D30N, I50V, and V82A reduce inhibitor binding while preserving enzymatic cleavage of viral polyproteins. However, primary resistance mutations often reduce viral fitness. To compensate for this loss, viruses frequently acquire secondary or compensatory mutations that restore replication efficiency (216).

In influenza virus, resistance to neuraminidase inhibitors such as oseltamivir is commonly associated with the H275Y mutation in neuraminidase. This mutation reduces drug binding but preserves enzymatic activity required for viral release. Similarly, resistance to M2 ion channel inhibitors such as amantadine occurs through mutations in the transmembrane domain of the M2 protein, particularly the S31N substitution (217).

Hepatitis C virus demonstrates resistance through mutations in NS3 protease, NS5A, and NS5B polymerase proteins. NS5A inhibitors are particularly vulnerable to resistance because single amino acid substitutions can markedly reduce drug efficacy. For example, the Y93H mutation in NS5A is associated with high-level resistance to daclatasvir and ledipasvir (218).

In SARS-CoV-2, mutations in the spike protein can reduce susceptibility to neutralizing antibodies and entry inhibitors, whereas mutations in the RNA-dependent RNA polymerase and main protease may affect susceptibility to direct-acting antivirals. The rapid emergence of variants such as Alpha, Delta, and Omicron demonstrated the ability of SARS-CoV-2 to evolve under immune and therapeutic pressure (219).

In addition to viral target mutations, resistance may also occur through altered intracellular drug activation. Nucleoside analogues require phosphorylation by host kinases to become active. Reduced expression or activity of these kinases may decrease intracellular concentrations of active metabolites, thereby reducing antiviral efficacy. Another important mechanism involves changes in cellular drug transporters. Overexpression of efflux pumps such as P-glycoprotein can reduce intracellular drug accumulation and contribute to multidrug resistance. This mechanism is particularly relevant in chronic viral infections requiring prolonged therapy (3).

8.2 Multi-target design to reduce resistance

Multi-target antiviral drugs represent a promising approach against drug resistance because they simultaneously target multiple viral or host factors. In order for a virus to develop resistance to a multi-target drug, it would require multiple simultaneous mutations. At the molecular level, simultaneous blockage of multiple steps of the viral life cycle can greatly limit the virus’s ability to adapt. For instance, a single antiviral agent that simultaneously targets viral proteases and host infection factors can have two effects on viral replication. Even if a mutation develops in the viral protease, the virus may still remain susceptible because entry inhibition continues to block infection (220).

Dual inhibitors of viral polymerase and protease enzymes are of particular interest because these enzymes are critical for replication and maturation. Polymerases replicate genomes, while proteases help to cleave viral polyproteins to produce mature proteins. Both enzymes can be targeted simultaneously to have a greater antiviral effect and prevent mutations (221).

Multitargeting the host as well as the virus can increase the resistance threshold. Host-targeted aspects are unaffected by viral mutations. For instance, the combination of a viral protease inhibitor and a host kinase inhibitor can reduce viral replication even if the virus has mutations resistant to the protease inhibitor. Multi-target design may also reduce the probability of compensatory mutations. Viruses can compensate for resistance mutations that reduce fitness by acquiring additional mutations. But if multiple pathways are shut down, the virus may not be able to compensate without seriously disrupting other functions (222).

8.3 Combination therapy versus single-molecule MTAs

The historical approach towards reducing antiviral resistance has been combination therapy. Highly active antiretroviral therapy in the treatment of HIV is a combination of drugs against reverse transcriptase, protease, integrase, and viral entry. This strategy significantly lowered the chances of viral resistance and made HIV a chronic disease. In the same way, combination regimens are used in the treatment of hepatitis C virus, in which protease, NS5A, and polymerase inhibitors are administered concomitantly to achieve sustained virological response and reduce resistance (223).

Despite the high effectiveness of combination therapy, there are a number of limitations associated with the method. The pill burden, the adherence levels, the drug-drug interactions, the overlapping toxicities, and the unequal pharmacokinetics in various drugs may occur. Lack of compliance may result in subtherapeutic levels of drugs and this provides a perfect ground where the development of resistant types may be developed (224). To address such shortcomings, single-molecule multi-target antivirals seek to combine multiple drug actions into one compound. These agents provide easier dosing, enhanced predictability of pharmacokinetics, better patient adherence and even reduced chances of resistance. As an illustration, hybrid molecules that integrate protease inhibitory activity and host-targeting functions can deliver a broad-spectrum antiviral response using a single chemical scaffold (225).

Nevertheless, single-molecule MTA development is scientifically difficult since the substance needs to have balanced activity across all desired sites and not be very toxic. Moreover, physicochemical characteristics like solubility, membrane permeability, metabolic stability, and oral bioavailability are harder to optimize with increasing molecular complexity (226).

8.4 Predictive computational models for resistance

There is an increasing use of computational methods to forecast antiviral resistance prior to the rise of clinically relevant resistant variants. These methods combine systems biology, molecular dynamics, machine learning, structural biology, and genomic surveillance. The atomic level of change in drug binding that results due to certain mutations can be studied using molecular docking and molecular dynamics simulations. As an example, computer models of HIV protease mutants have shown how resistance related mutations distort the active site and decrease the affinity of inhibitors. Other similar studies have been conducted on influenza neuraminidase, HCV NS5A, and SARS-CoV-2 main protease (227).

Mutations that cause decreased drug susceptibility may be detected with the use of ML models that are trained on viral sequences data and phenotypic resistance profiles. Predictive algorithms are routinely applied in HIV to estimate viral genotype resistance to reverse transcriptase, protease, and integrase reverse transcriptase, protease, and integrase reverse transcriptase inhibitors. DL methods can also advance the predictive performance by combining various factors like the genome sequence of the virus, protein structure, host-cell pathways, and medication features. Such models could assist in determining areas of resistance in viral proteins and inform the development of drugs that will be effective against future strains (228, 229).

Critical host-virus interaction networks which are less prone to resistance may also be identified by means of network pharmacology and systems biology approaches. By focusing on conserved host pathways instead of highly mutated viral proteins, researchers can create antivirals that are more long-lasting. Finally, to develop effective antiviral therapies in the future, it is necessary to understand the mechanisms of resistance on the molecular, cellular and systems level. The multi-target design, host-targeted strategies, rational combination therapy, predictive computational models are likely to dominate and seek significant roles in overcoming viral escape and prolonging clinical lifespan of antiviral agents (97).

9 Pharmacokinetic and pharmacodynamic considerations

The pharmacokinetic and pharmacodynamic are the important preconditions of the clinical effectiveness of multi-target antiviral agents. Even extremely active antiviral agents cannot pass the development stage since they could have poor absorption, poor tissue penetration, high metabolism, poor elimination, and undesirable toxicity. Multi-target antivirals are aimed to engage with more than one target viral or host at a time and therefore, it is even more difficult to balance efficacy, selectivity, and safety (230).

In the case of multi-target agents, pharmacokinetic optimization is especially crucial since the compound should be able to reach sufficient concentrations on all target sites of action. In most viral infections, it may include the drug penetration in tissues like lungs, liver, lymphoid tissues, central nervous system or intracellular compartments where the viruses are able to replicate. Similarly, the pharmacodynamic analysis must be done to ascertain the ability to achieve the right drug levels over a reasonable period of time to inhibit viral replication and avoid resistance (225).

9.1 Absorption, distribution, metabolism, and excretion

The therapeutic potential of a drug is highly dependent on the ADME profile. Absorption is the measurement of the drug effectiveness in entering the systemic circulation after being administered. In the case of antivirals orally, absorption is significantly influenced by gastrointestinal solubility, membrane permeability, gastric stability and first-pass metabolism. Most antiviral agents lack oral bioavailability due to low aqueous solubility or intestinal permeability (231). As an example, a number of protease inhibitors have limited absorption and necessitate booster enzymes like ritonavir to obstruct metabolism and raise plasma concentrations. Prodrug strategies are frequently employed in the case of nucleoside analogues to enhance membrane permeability and intracellular delivery. An example is Remdesivir, which was developed as a phosphoramidate prodrug to increase intracellular conversion to its active triphosphate metabolite (232).

Distribution is the circulation of a drug to tissues and organs. In viral infections, tissue penetration is especially crucial since the virus usually replicates in particular compartments. HIV may remain in lymphoid tissues and the central nervous system, hepatitis viruses may infect the liver, and SARS-CoV-2 may infect lung tissue primarily. Antiviral agents should therefore reach sufficient levels in these locations (233). In liver, the primary drug metabolism is by cytochrome P450 enzymes, esterases, and conjugation. The reason why metabolic stability is important is that the rapid degradation may cause a decrease in the half-life and efficacy of the drug. Nonetheless, metabolic stability can be too high, leading to the risk of accumulation and toxicity. CYP3A4 is a substrate to many antivirals, which is susceptible to metabolic interactions. It is excreted either in kidneys or the bile. Drugs excreted mainly by the kidney might need to be given in a higher or lower dose in patients who have renal problems and hepatically excreted drugs might build up in patients with hepatopathy. The knowledge of excretion pathways is of particular significance in chronic antiviral therapy where accumulation of antivirals on a long-term basis can lead to toxicity (234).

9.2 Drug-drug interactions

Drug-drug interactions represent a significant issue in antiviral treatment due to the fact that patients often take many drugs at the same time. Patients with HIV, hepatitis, and influenza complications, as well as those with severe COVID-19, often need antivirals, anti-inflammatory agents, antibiotics, anticoagulants, or comorbidity drugs. Most of the interactions take place by modulation of cytochrome P450 enzymes or membrane transporters like P-glycoprotein. A classic example is ritonavir as it is a strong inhibitor of CYP3A4 and is usually applied as a pharmacokinetic boost of protease inhibitor plasma levels. Nevertheless, ritonavir can also raise the concentration of other medicines which can lead to toxicity (235).

On the same note, antiviral drugs which prevent or promote P-glycoprotein may result in a change in the intestinal absorption, tissue distribution and excretion of co-administered drugs. In particular, certain kinase and protease inhibitors can enhance the occurrence of cardiovascular toxicity in case of their interactions with other drugs that are metabolized via identical pathways (236).

Multi-target antiviral drugs can be especially challenging, as these substances usually have more complicated chemical structures and can act on multiple enzymes or transport systems. Thus, screening of metabolic interactions is vital in early drug development (25).

9.3 Bioavailability and tissue distribution

Bioavailability can be described as the percentage of an administered dose that is active and reaches systemic circulation. In the case of antiviral drugs, bioavailability is a key factor because it determines sufficient plasma and tissue concentrations. Ineffective bioavailability may result in under therapeutic exposure, insufficient viral inhibition, and resistance. Some of the natural products that can be used as antivirals, including curcumin, quercetin, and resveratrol, have poor bioavailability due to rapid metabolism and poor absorption. To eliminate these restrictions, nanoparticles, liposomes, phospholipid complexes and prodrugs have been considered (231).

The distribution of tissues is also significant since the virus may frequently replicate in shielded areas. Indicatively, the HIV reservoirs found in the brain and lymphoid tissues can hardly be eradicated since most of the drugs are not well penetrative across the blood-brain barrier. In a similar manner, antiviral agents that target the respiratory viruses should preferably reach peak concentrations in the lung tissue (237).

Tissue distribution is determined by lipophilicity, plasma protein binding, molecular size, and transporter affinity. Compounds that are very lipophilic can easily enter membranes but are also concentrated in fatty tissues thereby becoming very toxic. Alternatively, polar compounds can be poorly penetrated in the tissues even in favorable plasma concentrations (233).

9.4 Toxicity and safety profiling

Multi-target antiviral agents are of particular concern with respect to safety evaluation since these agents are frequently active against viral and host targets. The side effects of these drugs include off-target effects that can cause hepatotoxicity, nephrotoxicity, cardiotoxicity, neurotoxicity, or hematological toxicity (238).

Indicatively, cardiovascular complications and myelosuppression have been linked to some of the kinase inhibitors that have been used as host-targeted antiviral therapy. Nucleoside analogues can lead to mitochondrial toxicity as they disrupt host mitochondrial polymerases. Protease inhibitors may cause metabolic disorders like dyslipidemia and insulin resistance (239). Reactive metabolites produced in the process of hepatic metabolism may also cause toxicity. Thus, to investigate the risks of early toxicity, in vitro experiments with the models of hepatocyte, cardiomyocytes, and renal cells are frequently employed. Animal trials also assist in the determination of dose-limiting toxicities, organ toxicities, and long-term safety issues (240).

Another significant factor is therapeutic index. Ideally, multi-target antiviral agents are meant to exhibit excellent antiviral properties at very low concentrations, which do not cause toxicity. The biggest challenge in MTA development is hence maintaining equilibrium between potency and safety (6).

9.5 PK/PD modeling in MTA development

Pharmacodynamics/pharmacokinetic modeling contributes significantly to the optimization of antiviral drug development. PK/PD models combine data on the concentration of drugs, the target effect, viral suppression, and toxicity with time. These models are especially applicable in the identification of the minimum effective concentration, the optimal dosing interval, and the exposure-response relationship. In the case of antivirals, it is crucial to keep concentrations of drugs above the effective inhibitory level to avoid the occurrence of viral rebound and resistance (241). Combination regimens and multi-target agents can also be assessed through PK/PD modeling. E.g. simulation studies can be used to predict synergistic effects of two antiviral mechanisms in a single molecule or to predict whether one mechanism will dominate pharmacological responses. More complex physiologically based pharmacokinetic models are also finding application in predicting tissue-specific drug distribution, blood-brain barrier penetration, intracellular drug accumulation and possible drug-drug interactions (242). Such techniques are particularly useful in the early-phase antiviral development since they limit the use of extensive animal testing and enhance the process of promising candidate selection. In general, pharmacodynamic and pharmacokinetic optimization is the key to the translation of multi-target antiviral drugs in the laboratory to clinical practice. Next-generation antiviral agents are likely to be more effective and safer in the future due to advances in the formulation science, computational modeling, biomarker-based dosing, and precision medicine (243).

10 Clinical development, translational challenges, and regulatory considerations for multi-target antiviral agents

While significant strides have been made in the discovery of multi-target antiviral agents through medicinal chemistry, computational modelling and AI, the successful translation into the clinic is continuing to be a challenge. Many antiviral compounds with good computational prediction and in vivo activity, but with suboptimal efficacy, safety, PK, manufacturing or regulatory considerations, fail during clinical development (244).

10.1 Ongoing clinical development of multi-target antiviral strategies

Multiple antiviral agents in clinical use have shown the benefit of attacking more than one aspect of viral life cycle. When considering multi-target antiviral interventions, one of the most effective is the use of combination regimens in the treatment of HIV, where a combination of drugs that inhibit the reverse transcriptase, integrase, protease, and HIV entry into cells are used to interfere with HIV replication and minimize development of resistance. Multi-mechanistic studies have been conducted for hepatitis C virus and SARS-CoV-2 infections (245, 246).

In the last few years, clinical studies have increasingly tested antiviral agents with dual action, both against the virus and another target of the cellular host. These include attempts to increase therapeutic durability, expand the range of antiviral activity and minimize the development of resistance. Furthermore, during the outbreak of a new virus, AI-driven drug repurposing has been a way to speed up the process of identifying antiviral candidates suitable for clinical testing (25, 247).

10.2 Translational barriers

Although the preclinical findings have been favorable, several challenges still exist with regard to successful clinical translation. The structural complexity of multi-target compounds can also have a negative impact on absorption, distribution, metabolism, excretion, and toxicity profiles. The challenge for medicinal chemistry is to achieve good PK properties and acceptable balance of multiple targets (248).

In addition, biological effects seen in cell-based assays do not necessarily translate to therapeutic efficacy in humans. Clinical outcomes can vary greatly with host immune responses, disease heterogeneity, viral evolution, inter-individual variability. High attrition rates are also experienced with antiviral drug development, in part due to manufacturing scalability, formulation development, and long-term safety evaluation (249).

10.3 Regulatory challenges

Multi-target antiviral agents pose special regulatory challenges. In general, evidence of mechanism of action, efficacy, safety, quality control and reproducibility is required by regulatory agencies. Determination of clear links between target engagement and clinical benefit may be more complicated for compounds with multiple targets compared to those with a single target (250).

AI becomes increasingly embedded in drug discovery, new regulatory challenges arise, such as the need for transparency and validation of the models, standards for model reproducibility, traceability of predictions, and documentation of the computational workflow. The governments are actively working on schemes to assess AI-powered pharmaceutical development, ensuring patient safety and scientific integrity (251).

10.4 Successful case studies

Multi-target therapeutics have been beneficial in several antiviral success stories. The use of combination antiretroviral therapy (cART) for HIV has changed the course of HIV from a deadly disease to a manageable chronic condition, because it targets multiple viral processes and reduces the risk of viral resistance (252). Likewise, the combined use of direct-acting antiviral drugs in hepatitis C virus (HCV) has resulted in an unprecedented success rate in achieving viral eradication (253). More recently, during the COVID-19 pandemic, antiviral drugs developed by the four disciplines of computational, structural biology, medicinal chemistry and rapid clinical evaluation proved the importance of this integration to speed up therapeutic development (254).

Overall, these examples illustrate the importance of proper integration of computational prediction, medicinal chemistry optimization, experimental validation, clinical investigation and regulatory oversight to achieve successful innovation in the field of antiviral drugs. Future multi-target antiviral drugs will continue to benefit from ongoing progress in AI-aided drug discovery drug discovery, systems biology, and precision medicine.

11 Current challenges

Though AI, computational chemistry, and systems pharmacology have continued to gain success in the discovery of antiviral drugs, various significant shortcomings remain to limit the creation of clinically effective multi-target antiviral drugs. In contrast to the traditional single-target drugs, the MTA demands a coordinated action of the potency, selectivity, pharmacokinetics, safety and target balance. The complexity of such requirements poses significant problems in drug design as well as in the translational development (255).

Besides, even though the computational tools are useful in the rapid identification of promising compounds, a large fraction of the predicted molecules is unable to pass through experimental validation or clinical testing due to the fact that biological systems are far more complicated than can be effectively represented by computational models. Thus, the further evolution of the given area will rely on the enhanced data sets, combination of the computational and experimental processes, predictive models, and the more adaptable regulatory policies (256).

11.1 Data limitations in AI models

The success of AI models depends on high-quality and diverse well-annotated datasets. However, with antiviral studies, most of the datasets are either not complete, not consistent, or biased against a small range of viral species and molecular targets. The majority of publicly available antiviral databases include much information on viruses like HIV, influenza virus, hepatitis C virus, and SARS-CoV-2, but less on neglected or emerging viruses (257).

The other significant problem is inconsistency of experimental conditions. The data that are produced on the antiviral activity is usually produced under varied conditions of host-cell lines, viral strains, assay conditions, and multiplicities of infection, incubation times and biological endpoints (258). Consequently, the same compound might be found to be very active in one study and weak or inconsistent in another. This leads to a decrease in the reliability of ML models and makes it challenging to compare results on different datasets (258).

Also, in many cases, the toxicity-related data are underrepresented. Most ML models focus on antiviral potency without giving a proper consideration to hepatotoxicity, cardiotoxicity, nephrotoxicity, mitochondrial toxicity, or long-term safety. It may lead to the production of compounds which seem promising in vitro but fail in animal experiments or in clinical trials (259).

11.2 Validation of computational predictions

Molecular docking, virtual screening, QSAR analysis, pharmacophore modeling, and ML are some of the computational methods that can be used to quickly create large numbers of potential antiviral hits. These predictions however have to be experimentally verified since computational predictions are not necessarily biologically true. A limitation that is shared by docking studies is that they tend to assume that proteins are rigid structures (32). As a matter of fact, viral proteins are also very dynamic, and can experience significant conformational alterations during binding, replication, or interaction with host factors. The mutation can also change the protein flexibility, the geometry of the binding pocket, and the allosteric control (260).

As an example, a compound might be predicted by molecular docking to bind strongly to the SARS-CoV-2 main protease, but fail experimentally due to an inability to penetrate cells, high metabolic clearance, or insufficient intracellular concentrations. In the same way, the HIV reverse transcriptase inhibitors that are predicted to cause strong inhibition of HIV may be rendered inactive in vivo due to low membrane permeability or due to active drug efflux by transport proteins (261).

Thus, computational predictions should be corroborated by enzyme inhibition, cell-based antiviral, cytotoxicity, animal infection and biomarker. Higher-order validation models (organoids, organ-on-chip, and three-dimensional cell cultures) could enhance the accuracy of predictions even further due to its closer resemblance to human physiology (261).

11.3 Regulatory challenges

In the case of single-target drugs, the target protein is typically easily defined and therapeutic effects are easy to measure. But again, there is a possibility of the interaction between MTA and various viral proteins, host enzymes, transporters, immune pathways, or receptors at once. This complicates the identification of which of the targets plays the greatest role in efficacy and what interactions might be playing a part in toxicity (119).

Host-targeted antivirals are even more susceptible to regulatory interference due to their effect on the normal cellular processes. As an illustration, kinase, immune, and host protease inhibitors can influence normal immune activities, cellular growth, or metabolic processes. Therefore, regulators might demand prolonged long-term data on safety before they approve (18).

The other challenge that is arising is the regulatory approval of AI-designed drugs. Despite the fact that AI models can be quickly used to produce new molecular structures, there are presently few recommendations on how such predictions are to be validated, recorded and included in regulatory submissions. The regulatory measures in the future might require models transparency, reproducibility, dataset quality, and algorithm validation (262).

11.4 Translational gaps from in silico to clinical studies

One of the significant problems in the discovery of antiviral drugs is that there is a high failure rate between the prediction phase and the success of their clinical trials. A significant number of the compounds that look very promising during virtual screening or cell-based assays fail in animals or human trials (25). This translational gap is because of a number of factors. First, computational models might be insufficient in the complexity of ADME properties. A compound can be found to have excellent predicted binding affinity, but fail due to poor oral absorption, low bioavailability, rapid metabolism, or poor tissue penetration (121).

Second, in vitro studies cannot be a complete replication of the complexity of the human viral infection. Cell culture viral replication can differ significantly to viral replication in human tissues, as immunity, inflammation, tissue structure, blood flow, and metabolic rate are all factors contributing to antiviral response (263). Indicatively, SARS-CoV-2 compounds that appear potent in Vero cells are ineffective in human lung tissue due to the variation in protease expression, receptor density and intracellular trafficking routes. Likewise, the same compounds which seem harmless in mono-culture cell lines can have hepatotoxicity, cardiotoxicity or neurotoxicity in the living organism. To overcome this gap, more experimental systems and biomarker-directed research, more realistic PK/PD models, and earlier use of human-relevant biological data in preclinical development will be needed (264).

11.5 Future directions in AI-guided antiviral discovery

AI, network pharmacology, systems biology and multi-omics technologies are likely to further influence the future of antiviral drugs discovery. The development of genomics, proteomics, metabolomics, and transcriptomics is producing vast biological datasets that can be combined to discover new anti-viral targets and assess the response of multi-target drugs (42). The next generation of AI will not merely focus on mere prediction of antiviral efficacy but will include other variables like resistance risk, tissue distribution, immune reactions of a host, toxicity, metabolic stability and genetic factors unique to an individual. This would help in the creation of specific antiviral therapy for an individual patient and viral strain (265).

This has also been extended to medicinal chemistry with generative AI models predicted to be useful in creating completely novel antiviral scaffolds with multi-target optimization of activity. Reinforcement learning methods can be used to repeatedly optimize molecules in terms of potency, selectivity, oral bioavailability and safety (266). The other avenue that is showing promise is the application of network pharmacology in order to find key networks of host-virus interactions. Instead of attacking a single viral enzyme, future therapeutic options could be used to simultaneously attack interacting signaling pathways, host proteins and viral processes. This systems-level strategy can lead to more permanent antiviral immunity and minimization of the chances of resistance (267). Lastly, virtual clinical trials, real-world patient data and digital twins can be used to streamline dosing, forecast toxicity and shorten clinical development. When combined, these new technologies can achieve a substantial positive effect on the rate, accuracy, and effectiveness of antiviral drugs discovery (268).

12 Future perspectives: precision antiviral medicine and next-generation AI-driven antiviral discovery

Advances in AI, systems biology, precision medicine and integrative omics technologies are likely to drive the future of antiviral drug discovery. Current antiviral development strategies are still based on broad-spectrum therapeutic targets, but new technologies are emerging that will allow for more individualized and adaptive therapeutic paradigms that will likely enhance therapeutic effectiveness with reduced toxicity and resistance development.

12.1 Precision antiviral medicine

Precision antiviral medicine is designed to personalize treatment based on each patient’s genetic makeup, immune system, viral strain, disease severity and history of antiviral treatment. New, genomics-based sequencing and molecular diagnostic technologies allow for the swift elucidation of host and viral factors that affect therapy. The ability to integrate these datasets with AI-assisted predictive models potentially could help with tailored antiviral selection, dose optimization and treatment outcome prediction. These are methods that may be useful in the case of chronic viral infections, immunocompromised individuals and rapidly changing viral pathogens (269–271).

12.2 AI-driven personalized therapeutics

AI has come to be used to assist with personalized therapeutic decision making based on large-scale clinical, genomic, pharmacological, and epidemiological data. ML algorithms can detect intricate connections among patient-specific characteristics and treatment results that may not be obvious using traditional statistics. AI systems could help future clinicians to choose the best antiviral treatment plan, predict adverse drug reactions, assess the likelihood of resistance and fine-tune treatment options as new clinical data emerges (272).

12.3 Digital twins in antiviral drug development

Digital twin is a newly emerging paradigm of precision medicine. A digital twin is a digital model of a single patient, combining clinical, molecular, physiological and pharmacological data to simulate disease progression and therapeutic effects. Digital twins can be used in antiviral research to virtually test treatment strategies, predict drug efficacy, optimize dosing regimens and assess resistance development prior to clinical implementation. While in an early stage of development, digital twins can significantly accelerate the development of anti-viral drugs and provide better personalized care for patients (273).

12.4 Integrated omics-guided antiviral discovery

Genomics, transcriptomics, proteomics, metabolomics, and epigenomics are being integrated to give unprecedented insights into host-virus interaction and disease pathogenesis. Multi-omics approaches help to identify novel antiviral targets, markers of therapeutic response and pathways involved in the development of resistant strains. This analysis of these high-dimensional datasets can be performed using AI, which could reveal intricate biological interactions that could aid in creating more efficient multi-targeted antiviral medicines. For the future, it is hoped that discovery platforms for antiviral therapeutics will integrate and use omics data to gain biological insights, structural biology, medicinal chemistry, and ML to create highly optimized therapeutic candidates (274, 275).

12.5 Foundation models and autonomous drug discovery systems

The evolution of autonomous drug discovery platforms is fueled by recent advancements in foundation models, large language models, multimodal AI systems and reinforcement learning. These technologies could, one day, facilitate the entire antiviral development process from target identification to molecular generation, virtual screening to optimization, experimental design to clinical trial planning. When combined with automated laboratory platforms and active learning systems, new AI-powered discovery strategies could significantly shorten and cut down the time and expense needed to develop new antiviral treatments (276, 277).

12.6 Toward an integrated antiviral discovery ecosystem

The convergence of computational modeling, AI, experimental biology, clinical data science and precision medicine are likely to be key aspects of future antiviral innovation. Combination of real-world clinical data, electronic health records, wearable monitoring devices, digital twins and multi-omics data could form continuously evolving antiviral discovery ecosystems that can rapidly respond to new threats from infectious disease (278). These systems could revolutionize the process of developing antiviral drugs from being mostly reactive to proactively and adaptively predicting future outbreaks and speeding up therapeutic introduction (276, 277).

Overall, advances in precision medicine, AI-driven personalized therapeutics, digital twin technologies, and integrated multi-omics platforms are expected to redefine the future of antiviral drug discovery. Interdisciplinary cooperation, greater exchange of information, comprehensive validation processes, and sound regulatory oversight will be crucial to bring these innovations to clinically viable antiviral therapies.

13 Conclusion

Multi-target antiviral therapy is an exciting approach towards overcoming the shortcomings of the traditional single-target therapy especially in areas of viral resistance, mutational escape, and new infections. MTAs have the potential of providing broader-spectrum antiviral activity and longer durability through concomitant regulation of several viral proteins and host pathways. Identification, optimization and repurposing of such compounds have been greatly accelerated with AI, machine learning, and computational drug design. Although there is still a set of challenges associated with the quality of data, experimental validation, toxicity, and regulatory approval, further development in systems biology and AI-guided modeling is likely to enhance the success of antiviral drug discovery. In general, the combined approach of multi-target pharmacology and contemporary computational technologies could be the key to the next generation of antiviral medicines against the contemporary and future viral diseases.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Artificial intelligence-based tools were used to assist in language editing and grammar correction.

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: David Jesse Sanchez, Western University of Health Sciences, United States

Reviewed by: Semanti Ghosh, Swami Vivekananda University, India

Sumanta Mondal, Gandhi Institute of Technology and Management (GITAM), India