An international AI competition has tested machine learning models on real antiviral data to assess their use in developing coronavirus treatments.
The ASAP-Polaris-OpenADMET Challenge involved 66 teams from across the world, each tasked with predicting how well potential drugs could block coronavirus enzymes, as well as assessing their safety and how they bind to viral targets.
Participants worked with previously unseen data focused on the main proteases of SARS-CoV-2 and MERS-CoV – enzymes that are essential for coronavirus replication.
The challenge was jointly organised by the Open Molecular Software Foundation, the AI-driven Structure-enabled Antiviral Platform, the Polaris benchmarking platform and the OpenADMET project, with backing from the NIH’s Antiviral Drug Discovery programme.
Top-performing models predicted molecular potency – how well a compound blocks a virus enzyme – with accuracy approaching laboratory precision.
Errors averaged around half a log unit, falling within typical experimental variability.
Models were also effective in predicting drug-like characteristics such as lipophilicity and cell permeability, both important for oral treatments.
However, solubility and liver clearance were harder to model accurately, likely due to sparse or noisy training data.
When asked to predict how drug molecules bind inside viral enzymes, some AI-powered “co-folding” models succeeded in identifying the correct binding pose in more than 80 per cent of cases.
Still, accuracy varied across different targets and compound types, with some deep learning and physics-based methods underperforming.
The challenge reflected three key tasks in real-world antiviral development: predicting antiviral potency, modelling safety and metabolic profiles, and identifying the 3D binding poses of molecules in viral enzyme active sites.
All model outputs were tested against a held-out, blinded dataset from ASAP’s actual drug discovery campaign to ensure unbiased evaluation.
The project supports the wider OpenADMET initiative, funded by ARPA-H, which aims to build open-access tools and datasets to predict how drug compounds are absorbed, distributed, metabolised and excreted in the body.
Approaches that performed best included pretraining on large chemical datasets, multi-task learning and the use of ensemble models.
The challenge also exposed ongoing limitations, particularly in modelling complex traits like solubility and metabolic clearance.
AI tools that can speed up candidate evaluation may help cut drug discovery timelines from years to months, supporting more agile responses to future pandemics.
Final datasets and evaluation metrics have been made publicly available, with many participants publishing their methods.
Further challenges are planned as part of future phases of the OpenADMET initiative, as the ASAP Discovery Consortium progresses toward clinical trials for its lead pan-coronavirus candidate.
