An AI system that tells you why it thinks a drug works — not just that it does — has cleared a meaningful benchmark in computational drug discovery.
Researchers built MARS, short for MoA Retrieval System, to tackle a task they call mechanism-of-action deconvolution: given a drug, figure out which biological pathways it actually targets. The system combines logical rules with neural networks — a hybrid approach known as neurosymbolic AI — and runs those rules over a custom knowledge graph called MoA-net. In head-to-head tests, MARS matched the predictive accuracy of current top models. More importantly, its explanations lined up with mechanisms scientists already know to be real.
That last point is the one that matters. Most high-performing drug discovery models are neural networks that produce predictions with no auditable reasoning — a problem when a wrong answer could shape clinical decisions. MARS surfaces the rules it used to reach a conclusion, giving researchers something to interrogate rather than just trust. The team also caught and fixed a subtle failure mode they call "degree-bias," where the model was gaming predictions by exploiting how connected a node is in the graph rather than following domain logic — a reminder that interpretability tools can themselves mislead if no one checks their work.
Neurosymbolic AI has been a niche research direction for years, mostly overshadowed by the raw benchmark gains of pure deep learning. MARS does not claim to beat neural models outright — it claims to match them while being auditable, which is a different and arguably more useful bar for regulated, high-stakes fields like drug development.