TL;DR
SHAP zero is a breakthrough algorithm that slashes the computational cost of explaining biological sequence models by up to 1,000x. By leveraging a hidden mathematical link between the Sparse Fourier Transform and Shapley values, it allows researchers to uncover complex genetic and protein interactions across entire datasets with near-zero marginal cost per query.
The Scalability Wall in Biological AI
In the world of "black-box" biology—where AI models predict everything from gene-editing success to protein fitness—we desperately need to know why a model makes a specific prediction. The gold standard for this is SHAP (SHapley Additive exPlanations), which assigns importance scores to individual features (like a specific nucleotide in a DNA strand).
However, biology is rarely about single features; it’s about complex interactions (epistasis). Calculating these interactions using traditional SHAP methods is computationally ruinous. For a protein sequence, the number of possible combinations is astronomical. Current tools like SHAP-IQ are so slow that explaining a single dataset of gene-editing sequences could take over 80 days on a high-end GPU. This "scalability wall" prevents scientists from gaining global insights into how biological systems work.
Methods: The SHAP zero Pipeline
The authors introduce SHAP zero, which shifts the paradigm from "explaining one sequence at a time" to "sketching the model once and querying it forever."
- Fourier Sketching: The algorithm performs a one-time "global sketch" of the black-box model. It uses techniques from signal processing to identify the most important Fourier coefficients of the model. This is efficient because biological models are typically "sparse"—only a few high-order interactions actually drive the predictions.
- The Möbius Bridge: The core mathematical innovation is a formal pipeline connecting three distinct concepts: the global Fourier Transform, the localized Möbius Transform, and the final Shapley Values.
- Amortized Computation: Once the global Fourier sketch is complete, calculating the SHAP values for any new sequence becomes a simple mathematical mapping. This results in "near-zero marginal cost"—the more sequences you explain, the cheaper it gets per sequence.
Key Insights and Biological Discoveries
The success of SHAP zero relies on the fact that biological functions are "Fourier sparse." While a DNA sequence has millions of potential interactions, only a small, manageable number of high-order "motifs" actually dictate the biological outcome. Because it is so fast, SHAP zero can look for 3rd, 4th, or 5th-order interactions that were previously "invisible" due to compute constraints.
The authors evaluated SHAP zero on three distinct biological frontiers:
- TIGER: A deep learning model for CRISPR-Cas13d guide RNA binding efficacy.
- inDelphi: A model predicting DNA repair outcomes after CRISPR-Cas9 cuts.
- Tranception: A large-scale protein language model used to predict the fitness of protein variants.
Massive Speedups and High Fidelity
SHAP zero estimated feature interactions 1,000x faster than SHAP-IQ. In the TIGER model, it computed explanations in 6.5 seconds that would have taken SHAP-IQ two hours. Despite the speed, the results are highly accurate. SHAP zero’s estimates showed a high correlation ($\rho = 0.83$ to $0.97$) with the much slower, sampling-based KernelSHAP.
In the Tranception protein model, SHAP zero analyzed a sequence space of $\sim10^{13}$ possibilities, identifying specific amino acid pairs (like Proline acting as a "secondary structure breaker") that contribute to protein fitness. It successfully recovered known biological phenomena, like the importance of GC content in RNA "seed" regions and microhomology patterns in DNA repair.
Limitations and Future Work
- The Sparsity Assumption: If a model is not sparse in the Fourier domain (i.e., if every single nucleotide interacts with every other nucleotide in a complex way), the "sketch" will fail to capture the model's behavior accurately.
- Uniform Marginalization: Currently, SHAP zero assumes all nucleotides/amino acids are equally likely (a uniform distribution). In reality, biological sequences follow specific evolutionary distributions. Future versions could incorporate "training data distributions" into the Fourier sampling process.
- Black-Box Only: While designed for black-box models, the authors note that for specific "white-box" architectures (like some Transformers), there might be even faster ways to extract these values, though SHAP zero remains the most versatile across different model types.
Summary updated