Mechanistic interpretability of biological foundation models

AIs trained on millions of cells are increasingly used to predict how biology works, but nobody looks inside them to understand how they make their predictions. We build tools that open these models up, showing what knowledge and algorithms the AI has learned and whether it matches biology. Then AIs can be used more efficiently and reliably. Today, scientists working with biological foundation models face a binary choice: either trust the model’s predictions blindly and risk wasting resources on artifacts, or fall back on traditional analysis tools that ignore what the model has learned entirely. The few explainability methods that exist (SHAP, attention maps) only approximate which input genes influenced a prediction from the outside, without revealing what internal biological concepts the model has constructed or how it uses them to reason. Our approach is fundamentally different: we go inside the model, using sparse autoencoders to decompose its representations into individually meaningful biological features, causal circuit tracing to map the computational pathways it uses to make decisions, and manifold extraction to recover continuous biological trajectories directly from learned representations. This enables not just transparency but entirely new capabilities: extracting compact, efficient algorithms from bloated models by identifying which circuits actually matter for specific biological tasks, optimizing model performance by pinpointing and correcting internal representations that don’t correspond to real biology, and using the decomposed internal knowledge as a hypothesis generation engine that surfaces biological relationships the model has learned from data but that no human has yet described. Success means that biological foundation models become not just prediction tools but transparent engines of biological discovery, where every prediction comes with a mechanistic explanation, every model ships with an interpretability layer, and the extracted knowledge feeds back into both better models and new science.