Research map / Large language models / Language, safety and interpretability
Predicting evolutionary rate as a pretraining task improves genome language model representations
Micaela Elisa Consens, Kevin K Yang, James Brian Hall, Ashley Mae Conard, BO WANG, Lorin Crawford, Alan M Moses, Alex Xijie Lu
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 PhyloP prediction (CEP) to autoregressive sequence pretraining (ablation) | ArGamba (66M) · Functional ROIs vs random regions | balanced accuracy (%): 71.54 -> 76.25 | Improved |
| add PhyloP prediction (MEM) to masked LM pretraining (ablation) | Bi-Gamba (4M) · Functional ROIs vs random regions | balanced accuracy (%): 71.39 -> 76.80 | Improved |
| add evolutionary-rate prediction to pretraining (ablation) | ArGamba and Bi-Gamba · Functional ROIs vs upstream regions | balanced accuracy (%): 66.99 -> 71.40 (ArGamba); 65.92 -> 72.65 (Bi-Gamba) | Improved |
| add evolutionary-rate prediction to pretraining (ablation) | ArGamba · Multiclass functional region classification | balanced accuracy (%): 53.97 -> 53.23 | No clear effect |
| add MEM to masked LM pretraining (ablation) | Bi-Gamba · Multiclass functional region classification | balanced accuracy (%): 34.53 -> 42.33 | Improved |
| add evolutionary-rate prediction to pretraining (ablation) | ArGamba and Bi-Gamba · ATG codon context classification | balanced accuracy (%): 44.59 -> 53.50 (ArGamba); 46.94 -> 56.16 (Bi-Gamba) | Improved |
| rank variants by predicted evolutionary rate instead of sequence likelihood (ablation) | Bi-Gamba MEM-only vs MLM-only · ClinVar missense | AUROC: 0.476 -> 0.641 | Improved |
| rank variants by predicted evolutionary rate instead of sequence likelihood (ablation) | ArGamba · gnomAD rare vs common missense | AUROC: 0.454 -> 0.589 | Improved |
| add NTP+CEP joint training vs CEP-only (ablation) | ArGamba · VEP non-coding benchmarks | AUPRC: 0.158 (CEP-only) vs 0.156 (NTP+CEP) on OMIM | No clear effect |
| add MEM to MLM training, scoring by sequence likelihood (ablation) | Bi-Gamba · COSMIC missense | AUPRC: 0.022 -> 0.013 | Hurt |
| add evolutionary rate-only pretraining vs sequence-only (ablation) | ArGamba · Multiclass functional region classification | balanced accuracy (%): 53.97 -> 50.31 | Hurt |
| use alternative loss instead of Gaussian NLL (ablation) | Gamba · downstream evaluation | performance | No clear effect |
Closest papers by meaning
- Training Compute-Optimal Protein Language Models (NeurIPS 2024 · similarity 0.87)
- PhyloLM: Inferring the Phylogeny of Large Language Models and Predicting their Performances in Benchmarks (ICLR 2025 · similarity 0.86)
- Ultrafast classical phylogenetic method beats large protein language models on variant effect prediction (NeurIPS 2024 · similarity 0.86)
- Hyperbolic Genome Embeddings (ICLR 2025 · similarity 0.86)
- HELM: Hierarchical Encoding for mRNA Language Modeling (ICLR 2025 · similarity 0.86)
- DART-Eval: A Comprehensive DNA Language Model Evaluation Benchmark on Regulatory DNA (NeurIPS 2024 · similarity 0.85)
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.
