Abstract: The growing adoption of machine learning models for biological sequences has intensified the need for interpretable predictions, with Shapley values emerging as a theoretically grounded standard for model explanation. While effective for local explanations of individual input sequences, scaling Shapley-based interpretability to extract global biological insights requires evaluating thousands of sequences—incurring exponential computational cost per query. We introduce SHAP zero, a novel algorithm that amortizes the cost of Shapley value computation across large-scale biological datasets. After a one-time model sketching step, SHAP zero enables near-zero marginal cost for future queries by uncovering an underexplored connection between Shapley values, high-order feature interactions, and the sparse Fourier transform of the model. Applied to models of guide RNA efficacy, DNA repair outcomes, and protein fitness, SHAP zero explains predictions orders of magnitude faster than existing methods, recovering rich combinatorial interactions previously inaccessible at scale. This work opens the door to principled, efficient, and scalable interpretability for black-box sequence models in biology.
The Scalability Bottleneck in Genomic Interpretability
Machine learning models are now essential for predicting everything from CRISPR guide RNA efficiency to protein fitness. However, these models remain "black boxes." To understand why a model makes a specific prediction, researchers rely on SHAP (SHapley Additive exPlanations), the gold standard for feature importance.
The primary challenge is scale. Calculating exact SHAP values is exponentially expensive relative to sequence length. While approximations like KernelSHAP exist, they still require thousands of model evaluations per query. Explaining a dataset of 100,000 sequences to find global biological motifs can take weeks of GPU time. This "scalability bottleneck" prevents the discovery of complex, high-order interactions—such as epistasis in proteins—that drive biological function.
Amortized Interpretability via Fourier Sketching
SHAP zero introduces a paradigm shift: amortized interpretability. Instead of explaining each sequence from scratch, the algorithm uses a two-step process to "map the terrain" once and then navigate it instantly.
- Global Fourier Sketching (One-time cost): The algorithm treats the biological model as a function over a discrete alphabet ($q$-ary). It leverages the fact that most biological models are "sparse" in the Fourier domain—meaning only a few feature combinations (motifs) truly matter. Using a sample-efficient sparse Fourier transform (q-SFT), it identifies the most significant global coefficients.
- Möbius Localization (Near-zero marginal cost): To explain a specific sequence, SHAP zero "localizes" this global sketch. The authors establish a novel mathematical connection between Fourier coefficients, the Möbius transform (a tool from combinatorial theory), and Shapley values. By mapping the global sketch to a local Möbius transform, the marginal cost of explaining a new sequence drops to nearly zero.
Results: Speed, Fidelity, and Biological Discovery
The authors validated SHAP zero across three distinct biological domains: CRISPR-Cas13d guide RNA efficacy (TIGER), DNA repair outcomes (inDelphi), and protein fitness (Tranception).
Massive Computational Speedup
SHAP zero is 1,000x+ faster than existing state-of-the-art methods like SHAP-IQ for calculating high-order interactions. In the TIGER model, explaining 1,000 sequences took a full day using standard methods; SHAP zero completed the task in 6.5 seconds after the initial sketch. This shift from "per-query cost" to "one-time cost" enables interpretability at the scale of entire genomes.
High Fidelity and Biological Insights
Despite being an approximation, SHAP zero estimates show very high correlation with standard KernelSHAP (Pearson $\rho$ up to 0.97). This fidelity allowed for significant biological discoveries:
- TIGER: It correctly identified the importance of the "seed region" in guide RNAs.
- inDelphi: It uncovered microhomology motifs (repeating DNA patterns) that drive frameshift mutations, which were previously difficult to isolate at scale.
- Tranception: It successfully extracted complex epistatic interactions, aligning with known protein secondary structures.
Discussion and Future Directions
The success of SHAP zero hinges on the observation that biological functions are often driven by a limited number of motifs and epistatic interactions. This sparsity allows a "sketch" to capture the model's essence efficiently.
While currently assuming uniform marginalization (treating all nucleotides as equally likely), future work could incorporate empirical training distributions. Additionally, while highly efficient for DNA ($q=4$), scaling to the 20-letter amino acid alphabet for long protein sequences remains a frontier for further optimization. SHAP zero represents a critical step toward principled, efficient, and scalable interpretability for black-box sequence models in biology.
Summary updated