Research map / Probabilistic and scientific modeling / Molecules and proteins
PoET: A generative model of protein families as sequences-of-sequences
Timothy Fei Truong Jr, Tristan Bepler
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 |
|---|---|---|---|
| Use PoET for conditional variant scoring instead of TranceptEVE L | Protein variant fitness prediction · ProteinGym substitutions | Average Spearman correlation: Δρ̄=0.013 | Improved |
| Ensemble PoET with TranceptEVE L | Protein variant fitness prediction · ProteinGym substitutions | Average Spearman correlation: Δρ̄=0.021 | Improved |
| Use PoET or PoET plus TranceptEVE L instead of TranceptEVE L | Protein variant fitness prediction · ProteinGym indels | Average Spearman correlation: Δρ̄∈[0.044,0.055] | Improved |
| Replace TranceptEVE M with PoET or PoET plus TranceptEVE L | Protein variant fitness prediction · ProteinGym indels | Average Spearman correlation: Δρ̄∈[-0.006,0.005] | No clear effect |
| Increase training context length from 4K to 8K (ablation) | PoET variant fitness prediction · ProteinGym validation | Variant effect prediction performance | Improved |
| Increase training context length from 8K to 16K (ablation) | PoET variant fitness prediction · ProteinGym validation | Variant effect prediction performance | Mixed |
| Use inverse-count sequence-set sampling instead of uniform sampling (ablation) | PoET · ProteinGym validation | Variant fitness prediction performance | Improved |
| Increase model size from 57M to 201M parameters (ablation) | PoET · ProteinGym validation | Variant fitness prediction performance | Improved |
| Increase model size from 201M to 604M parameters (ablation) | PoET · ProteinGym validation | Variant fitness prediction performance | No clear effect |
| Use tiered Transformer decoder layers instead of regular RoPE Transformer layers (ablation) | PoET · Protein sequence generation at long context lengths | Perplexity | Improved |
| Use shortest-to-longest rather than random sequence ordering (ablation) | PoET · Heldout UniRef50 next-sequence prediction | Perplexity: 7.44075 -> 7.44828 | No clear effect |
| Use longest-to-shortest rather than random sequence ordering (ablation) | PoET · Heldout UniRef50 joint likelihood | Mean joint log likelihood: -8099.74973 -> -8091.89462 | No clear effect |
Closest papers by meaning
- Understanding protein function with a multimodal retrieval-augmented foundation model (NeurIPS 2025 · similarity 0.94)
- Diffusion on Language Model Encodings for Protein Sequence Generation (ICML 2025 · similarity 0.90)
- LEMON-ZEST: Evolution-Informed Tokenization for Efficient Protein Language Modeling (NeurIPS 2026 · similarity 0.89)
- Steering Protein Language Models (ICML 2025 · similarity 0.89)
- Training Compute-Optimal Protein Language Models (NeurIPS 2024 · similarity 0.89)
- Diffusion Language Models Are Versatile Protein Learners (ICML 2024 · similarity 0.88)
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.
