← Back to Home

Do AI-Selected Antiviral Molecules Work? Review Calls for Dedicated Tests for Viral Polymerases

High scores on general protein datasets may not predict antiviral activity. A new review argues for incorporating the RNA, accessory proteins, and metal ions required for viral replication into evaluations, then using enzyme and cell experiments to test AI’s ability to select drug candidates.

By SURL BioNews

A molecule that appears to bind tightly to a viral protein in a computer simulation may not stop the virus from multiplying once it enters an infected cell. For scientists hoping to use artificial intelligence to accelerate antiviral drug development, this gap determines which predictions warrant experimental testing. A review published on October 15 in *Frontiers in Pharmacology* notes that commonly used tests for evaluating AI may not yet adequately capture the biological characteristics of respiratory virus polymerases.

Written by researchers from the Mexican Social Security Institute and several universities, the article focuses on RNA-dependent RNA polymerases (RdRp) and related complexes. These enzymes synthesize viral RNA and are important targets for antiviral drugs. The authors argue that current benchmark data for protein–small molecule binding consist mainly of nonviral targets; whether models that perform well on these datasets can extend that performance to viral polymerases requires dedicated testing.

The difficulty is that a polymerase is a constantly changing molecular machine. Its drug-bound state is influenced by the RNA template, accessory proteins, divalent metal ions, and conformational changes. A model that relies on a single static structure may miss the binding environment present during actual viral replication, causing seemingly plausible molecular docking results to diverge from actual activity.

The authors therefore argue that evaluation should begin by assembling multiple structural states of polymerases and preventing information leakage between training and test data. This bears on whether a model has learned to predict the activity of unfamiliar molecules or has achieved high scores because it has encountered similar data. Tests designed specifically for antiviral development should also examine whether models can prioritize compounds that are experimentally effective.

The next stages of evidence must connect step by step: whether a compound acts on its intended target, whether it inhibits the polymerase, and whether it reduces viral activity in cell-based infection experiments. Cytotoxicity must also be measured to avoid mistaking changes caused by damage to host cells for antiviral effects. Resistance analysis and assessments of absorption, distribution, metabolism, excretion, and toxicity are used to judge whether candidate molecules warrant further development.

The significance of this proposal lies in bringing AI performance closer to laboratory decision-making: with the same resources for validation, can a model improve the chances of finding effective molecules? Molecular docking scores, predicted binding affinity, or statistical error alone cannot answer this question. Feeding experimental results back into models offers a chance to gradually narrow the gap between computational predictions and viral biology.

However, this article is a review that proposes an evaluation and development workflow. It does not demonstrate through new prospective comparative experiments that this workflow can improve drug success rates, nor does it provide evidence of efficacy in humans for new drug candidates. It offers more specific criteria for evaluating AI-driven antiviral development; whether these tests can improve candidate selection remains a question for subsequent experiments.

References

  1. Bioengineer.org
  2. Frontiers in Pharmacology
  3. PubMed — National Library of Medicine