Research map / Learning theory and algorithms / Prediction and decision losses
SHAP zero Explains Biological Sequence Models with Near-zero Marginal Cost for Future Queries
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SHAP zero, which sketches a black-box biological sequence model's sparse Fourier transform once, then maps it via a q-ary Möbius transform to get SHAP values and Faith-Shap interactions for new queries at near-zero marginal cost.
Problem
Computing Shapley values and high-order Shapley interactions for black-box biological sequence models requires exponentially many model evaluations per query, which becomes prohibitive when explaining thousands of sequences for global insight.
Approach
Connects Shapley values and Faith-Shap interactions to the q-ary Möbius transform, which in turn can be computed from the model's sparse Fourier transform; a one-time sparse Fourier sketch (via subsampling, aliasing, and peeling) is reused across all future queries.
Key results
- On TIGER, SHAP zero's SHAP values agree with KernelSHAP (rho=0.83) and amortized Faith-Shap runtime is 6.5 s versus 2 hours for SHAP-IQ (>1000x faster).
- On inDelphi, SHAP zero's recovered Fourier coefficients predict frameshift frequencies with rho=0.74, versus 0.37 for linear and 0.47 for pairwise models.
- On Tranception GFP, SHAP zero runs 7x faster than SHAP-IQ in amortized time (5.2 vs 38.1 minutes) and 3x faster than KernelSHAP (5.2 vs 17.7 minutes).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SHAP zero Fourier coefficients instead of the linear model | TIGER guide-score prediction · TIGER held-out guide scores (1038 sequences) | Pearson correlation: 0.44 -> 0.58 | Helped |
| use SHAP zero Fourier coefficients instead of the pairwise model | TIGER guide-score prediction · TIGER held-out guide scores (1038 sequences) | Pearson correlation: 0.48 -> 0.58 | Helped |
| replace KernelSHAP with SHAP zero for amortized explanations | TIGER · TIGER query explanations | amortized runtime: 109 seconds -> 6.5 seconds | Helped |
| replace SHAP-IQ with SHAP zero for Faith-Shap interactions | TIGER · TIGER Faith-Shap interactions | amortized runtime: 2 hours -> 6.5 seconds | Helped |
| use SHAP zero Fourier coefficients instead of the linear model | inDelphi frameshift prediction · inDelphi held-out frameshift frequencies (84 sequences) | Pearson correlation: 0.37 -> 0.74 | Helped |
| use SHAP zero Fourier coefficients instead of the pairwise model | inDelphi frameshift prediction · inDelphi held-out frameshift frequencies (84 sequences) | Pearson correlation: 0.47 -> 0.74 | Helped |
| replace KernelSHAP with SHAP zero for amortized explanations | inDelphi · inDelphi query explanations | amortized runtime: 55 minutes -> 40 minutes | Helped |
| replace SHAP-IQ with SHAP zero for Faith-Shap interactions | inDelphi · inDelphi Faith-Shap interactions | amortized runtime: 37 hours -> 40 minutes | Helped |
| replace KernelSHAP with SHAP zero for amortized explanations | Tranception · Tranception GFP query explanations | amortized runtime: 17.7 minutes -> 5.2 minutes | Helped |
| replace SHAP-IQ with SHAP zero for Faith-Shap interactions | Tranception · Tranception GFP Faith-Shap interactions | amortized runtime: 38.1 minutes -> 5.2 minutes | Helped |
| use SHAP zero rather than FastSHAP for amortized SHAP estimates | TIGER · TIGER SHAP values | correlation with KernelSHAP estimates: -0.39 -> 0.83 | Helped |
| use SHAP zero rather than FastSHAP for amortized SHAP estimates | inDelphi · inDelphi SHAP values | correlation with KernelSHAP estimates: 0.08 -> 0.84 | Helped |
Surprising findings
- FastSHAP shows poor empirical correlation with KernelSHAP on TIGER (rho=-0.39), inDelphi (rho=0.08), and Tranception (rho=0.03).
- On inDelphi, SHAP zero's speedup over KernelSHAP is small (40 vs 55 minutes) due to the smaller query set.
Limitations
- Relies on the assumption that the model's Fourier transform is sparse; denser models need more samples (e.g., 4-6 million samples).
- Uses uniform marginalization over the alphabet, which differs from the empirical training distribution used by KernelSHAP, so the two estimate slightly different quantities.
- Interaction order is bounded (typically up to 5) and sketching cost grows with the sparse-recovery requirements.
Experimental setup
Datasets: TIGER guide RNA scores (1038 held-out targets) inDelphi repair outcomes (84 held-out targets) GFP epistatic sites from prior study
Benchmarks: TIGER guide RNA efficacy inDelphi DNA repair outcomes Tranception GFP fitness
Metrics: Pearson correlation amortized runtime R^2
Baselines: KernelSHAP SHAP-IQ FastSHAP linear model pairwise model DeepSHAP DeepLIFT
Builds on: KernelSHAP Faith-Shap interaction index SHAP-IQ sparse Fourier transform (sparse-graph codes / peeling) FastSHAP
Closest papers by meaning
- Kernel-Based Evaluation of Conditional Biological Sequence Models (ICML 2024 · similarity 0.84)
- SHAP values via sparse Fourier representation (NeurIPS 2025 · similarity 0.84)
- Prediction via Shapley Value Regression (ICML 2025 · similarity 0.84)
- Beyond TreeSHAP: Efficient Computation of Any-Order Shapley Interactions for Tree Ensembles (AAAI 2024 · similarity 0.84)
- On the Tractability of SHAP Explanations under Markovian Distributions (ICML 2024 · similarity 0.83)
- Provably Accurate Shapley Value Estimation via Leverage Score Sampling (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.