Hybrid AI model aims to make drug discovery more explainable and trustworthy

Scientists have introduced MARS (MoA Retrieval System), a hybrid artificial intelligence approach designed to tackle one of drug discovery's key challenges: understanding precisely how candidate compounds affect biological systems in ways that humans can verify and trust.

Why interpretability matters in drug development#

Traditional deep learning models excel at pattern recognition but often function as "black boxes"—making predictions without showing their reasoning. For pharmaceutical research, this opacity creates risk: decisions affecting human health need explanations that biologists and chemists can validate against known biology. MARS addresses this by combining neurosymbolic AI (which fuses logical rules with neural networks) to generate predictions that align with established biological knowledge.

The approach: MARS and mechanism-of-action prediction#

The team developed a specialized task called drug mechanism-of-action (MoA) deconvolution—essentially training models to predict and explain how drugs produce their effects. They built MoA-net, a tailored knowledge graph encoding known drug-target relationships, then trained MARS to leverage weighted logical rules within this structured domain knowledge.

Crucially, MARS achieved performance matching current state-of-the-art models while producing interpretations that reflected genuine biological mechanisms rather than statistical artifacts.

A critical discovery: reasoning shortcuts#

The researchers uncovered a subtle but important flaw in such systems: degree bias, where models exploit simple network patterns (like the frequency of highly connected nodes) rather than learning domain-appropriate rules. The model would predict correctly but for the wrong reasons—a "shortcut" that could fail when applied to novel molecules or transferred between datasets. The team demonstrated methods to identify and correct this bias, improving the reliability of predictions for downstream research.

Implications for drug discovery#

By deploying interpretable models that can be validated against known biology and protected from reasoning shortcuts, MARS exemplifies how AI can become more trustworthy for high-stakes applications. The approach underscores why transparency and domain alignment matter as much as raw accuracy in pharmaceutical innovation.