Protein Generation Evaluation Set 30K samples
1.28sc-RMSDPhysio-DPO
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Physio-DPOBackbone=ProGen2-XL 6.4B + LoRA, Evaluation Protocol=Alignment (Fine-tuning)2026.01 | 1.28 | 87.5 | 92.8 | -3.05 | 8.2 | 33.8 | |
| KTOBackbone=ProGen2-XL 6.4B + LoRA, Evaluation Protocol=Alignment (Fine-tuning)2026.01 | 1.79 | 82.1 | 84.1 | -2.71 | 8.1 | 36.1 | |
| DPO (Standard)Backbone=ProGen2-XL 6.4B + LoRA, Evaluation Protocol=Alignment (Fine-tuning)2026.01 | 1.82 | 81.3 | 83.6 | -2.65 | 8.6 | 36.5 | |
| IPOBackbone=ProGen2-XL 6.4B + LoRA, Evaluation Protocol=Alignment (Fine-tuning)2026.01 | 1.88 | 80.8 | 82.9 | -2.58 | 7.8 | 35.2 | |
| PPO (RLHF)Backbone=ProGen2-XL 6.4B + LoRA, Evaluation Protocol=Alignment (Fine-tuning)2026.01 | 2.15 | 78.5 | 79.2 | -2.48 | 12.5 | 39.8 | |
| SFTBackbone=ProGen2-XL 6.4B + LoRA, Evaluation Protocol=Supervised Fine-tuning2026.01 | 2.35 | 75.2 | 71.5 | -2.25 | 7.4 | 34.1 | |
| ProGen2-XLNumber of Parameters=6.4B, Evaluation Protocol=Zero-shot Generation2026.01 | 3.25 | 67.8 | 52.4 | -1.65 | 6.1 | — | |
| ESM-3Number of Parameters=1.4B, Evaluation Protocol=Zero-shot Generation2026.01 | 3.95 | 62.5 | 45.8 | -1.31 | 6.5 | — | |
| ESM-2Number of Parameters=3B, Evaluation Protocol=Zero-shot Generation, Sampling method=Gibbs sampling2026.01 | 4.55 | 56.2 | 35.4 | -1.05 | 6.8 | — | |
| ProtGPT2Number of Parameters=762M, Evaluation Protocol=Zero-shot Generation2026.01 | 5.12 | 48.5 | 22.1 | -0.85 | 8.2 | — |