The review, authored by researchers working at the interface of computational chemistry and virology, takes aim at a quiet but consequential problem in the application of machine learning to antiviral discovery. Performance claims for AI models in drug design are frequently inferred from general protein-ligand benchmarks, datasets dominated by well-studied human targets such as kinases and proteases. Viral targets, and respiratory RNA virus polymerases in particular, are underrepresented in these benchmarks. The consequence is that a model may appear excellent on paper while failing to capture the specific structural and mechanistic complexities that govern whether a molecule actually inhibits a viral polymerase in an infected cell.

RNA-dependent RNA polymerases occupy a special position among antiviral targets. They are among the most conserved components of RNA viruses, which means that a compound active against the polymerase of one strain has a reasonable chance of retaining activity against related strains, and the enzyme has no direct human homolog, reducing the risk of host toxicity. These features make RdRps pharmacologically attractive, and the clinical success of nucleoside analogs such as remdesivir and the influenza polymerase inhibitor baloxavir marboxil confirms that the target class is druggable. Yet the same enzymes are notoriously difficult computational subjects. Their ligandable states depend on the presence of RNA templates and products, on divalent metal ions such as magnesium that coordinate the catalytic chemistry, on cofactors and accessory proteins that remodel the polymerase architecture, and on conformational transitions that shuttle the enzyme between initiation, elongation, and proofreading states. A static crystal structure captures only one frame of a dynamic molecular machine.

The review synthesizes the current landscape of computer-aided drug design as applied to these targets, covering classical molecular docking, machine learning-assisted virtual screening, modern molecular representations, pose-generation models, generative approaches to molecular design, and the benchmarking frameworks used to assess protein-ligand affinity prediction. Geometric deep learning, a branch of machine learning that respects the three-dimensional symmetry and structure of molecular data, features prominently as a technology that has advanced pose prediction and affinity estimation. Generative models that propose entirely novel chemical structures have likewise expanded the design space accessible to computational screening campaigns. The authors acknowledge this progress while insisting that progress and validation are not the same thing.

At the heart of the critique is a set of metrics that the authors consider insufficient on their own. Docking scores, predicted binding affinities, and root-mean-square error values on affinity benchmarks are the currencies in which computational models are typically evaluated. The review contends that these numbers become meaningful only when they are linked to a chain of downstream evidence: demonstration that the compound engages the intended target, evidence of enzymatic inhibition in biochemical assays, measurable antiviral activity in cell-based infection models, acceptable cytotoxicity profiles, characterization of resistance mutations that arise under drug pressure, and pharmacological feasibility including absorption, distribution, metabolism, excretion, and toxicity considerations. A model that ranks compounds well by docking score but produces candidates that fail at the first cell-based hurdle has not solved the problem it claims to address.

Part of the difficulty, the authors suggest, lies in biases embedded in molecular docking itself. Docking algorithms and the scoring functions they employ were largely developed and calibrated on datasets enriched for non-viral targets with different binding-site characteristics. Viral polymerase active sites present unusual challenges: they are often larger and more flexible than typical enzyme pockets, they accommodate nucleic acid substrates alongside small molecules, and their catalytic chemistry depends on metal coordination that many scoring functions do not model explicitly. When these biases propagate into AI systems trained on docking outputs or on benchmarks derived from non-viral proteins, the resulting models inherit systematic blind spots that are invisible to standard evaluation metrics.

To address these shortcomings, the review proposes an RdRp-centered AI pipeline that reorganizes the discovery workflow around virological reality. The pipeline begins with curated structural ensembles of polymerase complexes that capture the conformational diversity of the target rather than a single static structure. Benchmarking is designed to be leakage-resistant, meaning that the evaluation sets are constructed to prevent models from succeeding through memorization of training data rather than genuine generalization. Evaluation tasks are antiviral-specific rather than borrowed wholesale from general protein-ligand prediction. Computational prioritization is then filtered through ADME and toxicity screens before candidates proceed to biochemical validation against the polymerase, cell-based infection assays, and resistance profiling that reveals whether a compound’s activity is durable or easily circumvented by viral mutation. The results of each experimental round feed back into iterative model refinement, closing the loop between prediction and measurement.

The emphasis on leakage-resistant benchmarking deserves particular attention because it addresses a problem that extends well beyond virology. Machine learning models in drug discovery can achieve impressive benchmark results by exploiting subtle statistical regularities in how datasets are constructed, for example when highly similar compounds appear in both training and test sets, or when assay artifacts correlate with activity labels. In the antiviral context, where the ultimate question is whether a computational pipeline enriches for compounds with experimentally actionable antiviral activity, such artifacts can produce a misleading picture of readiness. The review’s proposal to validate whether computational prioritization actually enriches active compounds, rather than merely whether it scores well on held-out affinity data, reframes the evaluation question in terms that matter to medicinal chemists and virologists.

The polymerase-centered framing also carries strategic significance for pandemic preparedness. Because RdRps are conserved across many RNA viruses, advances in computational methods validated on one polymerase may transfer more readily to related enzymes than structure-specific approaches would. Benchmarking AI models against a panel of respiratory virus polymerases, rather than against generic protein-ligand datasets, would produce evaluation infrastructure that is directly relevant to the pathogens most likely to drive the next therapeutic emergency. The review positions respiratory RNA virus polymerases not only as high-value drug targets in their own right but as a complementary test case for the broader discipline of AI-guided drug discovery, one whose peculiarities expose weaknesses that conventional benchmarks conceal.

The broader message is a call for epistemic discipline in a field moving at remarkable speed. Generative AI and deep learning have genuinely expanded what computational drug discovery can attempt, and the urgency of respiratory viral threats gives that capability real public health weight. But the review insists that the measure of success is not the elegance of the model or the magnitude of the docking score. It is whether a computationally prioritized compound, when placed in a biochemical assay, an infected cell culture, and eventually a pharmacologically informed development pathway, demonstrates target engagement, antiviral activity, an acceptable safety margin, and a resistance profile that permits durable use. Moving beyond docking scores, the authors conclude, requires benchmarking models under virologically realistic conditions, and the respiratory RNA virus polymerase is where that discipline can be built.

Subject of Research: Artificial intelligence methods for antiviral drug discovery targeting respiratory RNA virus RNA-dependent RNA polymerases

Article Title: Virology-aware artificial intelligence for respiratory antiviral discovery: from molecular docking biases to RdRp-centered benchmarking

Article References: Ramos-Cortes, L. R., Santos-Mena, A. O., Flores-Padilla, L. E., Miranda-Lizárraga, K. P., Cañedo-Figueroa, D. M., Ontiveros-Zúñiga, L. F., Lira-Hernández, F. I., Osuna-Ramos, J. F., Noriega, S., Cardoso-Ortiz, J., Alcalá-Ramírez, V., & De Jesús-González, L. A. (2026). Virology-aware artificial intelligence for respiratory antiviral discovery: from molecular docking biases to RdRp-centered benchmarking. Frontiers in Pharmacology, 17, Article 1940098. https://doi.org/10.3389/fphar.2026.1940098

Image Credits: AI Generated

DOI: 10.3389/fphar.2026.1940098

Keywords: artificial intelligence, antiviral drug discovery, RNA-dependent RNA polymerase, respiratory RNA viruses, molecular docking, machine learning, geometric deep learning, virtual screening, benchmarking, SARS-CoV-2, influenza, drug resistance