Research map / Probabilistic and scientific modeling / Molecules and proteins
Fourier-transform-based attribution priors improve the interpretability and stability of deep learning models for genomics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Fourier-based attribution prior | binary CNN, SPI1 · SPI1 binary test | test accuracy: 87.996% -> 88.347% | Helped |
| add Fourier-based attribution prior | binary CNN, GATA2 · GATA2 binary test | test auPRC: 0.733 -> 0.745 | Helped |
| add Fourier-based attribution prior | binary CNN, K562 · K562 binary test | test accuracy: 90.049% -> 90.362% | Helped |
| add Fourier-based attribution prior | binary models · SPI1/GATA2/Nanog-Oct4-Sox2/K562 | attribution stability across random initializations: 95.6 -> 134.3 (SPI1 binary) | Helped |
| add Fourier-based prior when comparing models trained on 1% vs all data | profile models K562 · K562 profile | Jaccard similarity all vs 1% data: 31.4 -> 121.3 | Helped |
| add Fourier-based prior | binary and profile models · Nanog/Oct4/Sox2 profile ChIP-nexus peaks | fraction of importance in peak: 0.689 -> 0.787 | Helped |
| add Fourier-based prior | binary models · Nanog/Oct4/Sox2 binary ChIP-nexus peaks | auPRC of overlap with peaks: 0.644 -> 0.647 | No clear effect |
| add Fourier-based prior | binary K562 model · K562 binary DNase-seq footprints | auPRC of overlap with footprints: 0.235 -> 0.229 | Hurt |
| add Fourier-based prior | binary SPI1 model on synthetic sequences · simulated SPI1 | fraction of importance in motif instances: 0.218 -> 0.659 | Helped |
| add Fourier-based prior | binary SPI1 model on synthetic sequences · simulated SPI1 | auPRC of importance overlap with motif instances: 0.422 -> 0.607 | Helped |
| replace Fourier-based prior with L2-regularization (ablation) | binary SPI1 model · SPI1 binary | peak-overlap auPRC: 0.616 -> 0.585 | Hurt |
| replace Fourier-based prior with L2-regularization (ablation) | binary SPI1 model · SPI1 binary | Shannon entropy of attributions (lower is better): 7.891 -> 8.908 | Hurt |
Closest papers by meaning
- What DNA Foundation Models Learn Beyond Sequence Composition (NeurIPS 2026 · similarity 0.85)
- Hyperbolic Genome Embeddings (ICLR 2025 · similarity 0.84)
- Prospector Heads: Generalized Feature Attribution for Large Models & Data (ICML 2024 · similarity 0.83)
- Curriculum Learning for Biological Sequence Prediction: The Case of De Novo Peptide Sequencing (ICML 2025 · similarity 0.83)
- Context-Aware Regularization with Markovian Integration for Attention-Based Nucleotide Analysis (NeurIPS 2025 · similarity 0.82)
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation (NeurIPS 2026 · similarity 0.82)
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.