We develop SHAP zero into a Python package, which opens the door for efficient, principled, and scalable interpretability of biological sequence models!
⭐ Paper: arxiv.org/abs/2410.19236
⭐ Code: github.com/amirgroup-co...

We then moved to apply SHAP zero to extract epistatic interactions in protein language models. Despite the total feature space being larger than a trillion, SHAP zero ran up to 7x faster in amortized time and uncovered interactions associated with structural stability.

We demonstrate the power of SHAP zero by applying it to guide RNA and DNA repair models. SHAP zero uncovered high-order interactions at scale corresponding to known biological motifs, a task previously inaccessible due to the space of feature interactions.

Our secret? We connect the sparse Fourier transform of a model with Shapley explanations. If the model is "compressible" (which many biological sequence models are!), SHAP zero amortizes the computation of feature interactions up to 1000x faster than current methods.

Our core idea: instead of recomputing explanations for every new sequence from scratch, we pay a one-time cost to create a global sketch of the model. This enables SHAP zero to explain biological sequences from this sketch with near-zero marginal cost.

The success of biological sequence models has created an urgent need to explain their predictions. However, computing Shapley values, often the gold standard of explanations, over thousands of sequences to extract biological insight remains computationally prohibitive.