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

Respiratory RNA viruses remain a persistent threat because they combine high transmissibility, rapid evolution, immune escape, zoonotic spillover, and incomplete therapeutic coverage. RNA-dependent RNA polymerases (RdRps) and polymerase complexes are among the most conserved and pharmacologically actionable viral targets, yet they are also challenging computational targets because their ligandable states depend on RNA templates, cofactors, divalent metals, multidomain architecture, and conformational transitions. Artificial intelligence (AI), machine learning (ML), deep learning (DL), geometric deep learning (GDL), and generative molecular design are increasingly used to accelerate antiviral discovery, yet their performance is often inferred from general protein-ligand benchmarks dominated by non-viral targets. This review argues that respiratory RNA virus polymerases are a high-value but comparatively underrepresented class of antiviral targets and a complementary test case for evaluating AI-guided drug discovery. We synthesize current approaches in computer-aided drug design (CADD), molecular docking, ML-assisted virtual screening, molecular representations, pose-generation models, generative AI, and protein-ligand affinity benchmarking. We emphasize that docking scores, predicted affinity, and root-mean-square error (RMSE) values are insufficient unless linked to target engagement, enzymatic inhibition, cell-based antiviral activity, cytotoxicity, resistance, and pharmacological feasibility. We propose an RdRp-centered AI pipeline that integrates curated polymerase structural ensembles, leakage-resistant benchmarking, antiviral-specific evaluation tasks, ADME/Tox filtering, biochemical validation, cell-based infection assays, resistance profiling, and iterative model refinement. Moving beyond docking scores requires benchmarking models under virologically realistic conditions and validating whether computational prioritization enriches compounds with experimentally actionable antiviral activity.