Research map / Architectures and efficiency / Attention and transformers
Adaptive DNA Sequence Modeling via Synergistic Plasticity Units
Binghao Liu, Wenzheng Zhao, Zhijie Zheng, Fei Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 LPL (ablation) | SPU-DNA · Genomic Benchmark: Human vs. Worm | accuracy: 94.67 -> 97.19 | Improved |
| add DPL (ablation) | SPU-DNA · Genomic Benchmark: Human vs. Worm | accuracy: 90.01 -> 97.19 | Improved |
| add SPL (ablation) | SPU-DNA · Genomic Benchmark: Human vs. Worm | accuracy: 95.22 -> 97.19 | Improved |
| add DPL (ablation) | SPU-DNA · Genomic Benchmark: Human Enhancers Ensembl | accuracy: 85.91 -> 91.08 | Improved |
| add LPL (ablation) | SPU-DNA · Genomic Benchmark: Human Nontata Promoters | accuracy: 91.05 -> 94.78 | Improved |
| add SPL (ablation) | SPU-DNA · Genomic Benchmark: Mouse Enhancers | accuracy: 81.02 -> 82.38 | Improved |
| SPU-DNA vs strongest baselines on Genomic Benchmark | SPU-DNA · Genomic Benchmark (8 datasets) | accuracy: best on 7 of 8; Nontata Promoters 94.78 vs ConvNova 95.21 | Mixed |
| SPU-DNA vs baselines (<5M params) | SPU-DNA · NT Benchmark (18 datasets) | MCC/F1/accuracy: best on 16 of 18 | Mixed |
| scale SPU-DNA to 500M params | SPU-DNA · NT Benchmark | MCC/F1/accuracy: leads on all datasets vs 1.9M model, e.g. Enhancer types 51.17 -> 62.98 | Improved |
| SPU-DNA (6.5M) vs ConvNova (27.4M) | SPU-DNA · Chromatin Profile Prediction | AUROC: DHS 93.71 -> 94.01 | Improved |
| replace SENet and Hyena gating with SPL (ablation) | SPU-DNA · Information Bottleneck analysis | predictive power and compression | Improved |
| SPU vs Transformer + FlashAttn at 100k length | SPU · Efficiency, training | VRAM and wall clock: wall clock 4.87s -> 0.93s; VRAM 7.8 GB -> 6.8 GB | Improved |
Closest papers by meaning
- Revisiting Convolution Architecture in the Realm of DNA Foundation Models (ICLR 2025 · similarity 0.88)
- Hyperbolic Genome Embeddings (ICLR 2025 · similarity 0.84)
- JanusDNA: A Powerful Bi-directional Hybrid DNA Foundation Model (NeurIPS 2025 · similarity 0.83)
- Model Decides How to Tokenize: Adaptive DNA Sequence Tokenization with MxDNA (NeurIPS 2024 · similarity 0.83)
- From Mechanistic Interpretability to Mechanistic Biology: Training, Evaluating, and Interpreting Sparse Autoencoders on Protein Language Models (ICML 2025 · similarity 0.83)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 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.
