Table 1 Performance comparison on the CATH 4.2 and CATH 4.3 datasets with topology classification split
From: Mask-prior-guided denoising diffusion improves inverse protein folding
Models | External | Model | Perplexity (↓) | Median recovery rate (%, ↑) | |||||
|---|---|---|---|---|---|---|---|---|---|
knowledge | parameters | Short | Single-chain | Full | Short | Single-chain | Full | ||
CATH 4.2 | aStructGNN26 | ✗ | 1.4M | 8.29 | 8.74 | 6.40 | 29.44 | 28.26 | 35.91 |
aGraphTrans26 | ✗ | 1.5M | 8.39 | 8.83 | 6.63 | 28.14 | 28.46 | 35.82 | |
aGVP43 | ✗ | 2.0M | 7.09 | 7.49 | 6.05 | 32.62 | 31.10 | 37.64 | |
aAlphaDesign44 | ✓ | 6.6M | 7.32 | 7.63 | 6.30 | 34.16 | 32.66 | 41.31 | |
ProteinMPNN1 | ✗ | 1.9M | 6.90 | 7.03 | 4.70 | 36.45 | 35.29 | 48.63 | |
PiFold13 | ✗ | 6.6M | 5.97 | 6.13 | 4.61 | 39.17 | 42.43 | 51.40 | |
LM-Design45 | ✓ | 659M | 6.86 | 6.82 | 4.55 | 37.66 | 38.94 | 53.19 | |
GRADE-IF38 | ✗ | 7.0M | 5.65 | 6.46 | 4.40 | 45.84 | 42.73 | 52.63 | |
MapDiff (uniform prior) | ✗ | 14.7M | 3.99 | 4.43 | 3.46 | 52.85 | 50.00 | 61.03 | |
MapDiff (marginal prior) | ✗ | 14.7M | 3.96 | 4.41 | 3.43 | 54.04 | 49.34 | 60.93 | |
CATH 4.3 | aGVP-GNN-Large27 | ✗ | 21M | 7.68 | 6.12 | 6.17 | 32.60 | 39.40 | 39.20 |
a+ AF2 predicted data | ✓ | 142M | 6.11 | 4.09 | 4.08 | 38.30 | 50.08 | 50.08 | |
aGVP-Transformer27 | ✗ | 21M | 8.18 | 6.33 | 6.44 | 31.30 | 38.50 | 38.30 | |
a+ AF2 predicted data | ✓ | 142M | 6.05 | 4.00 | 4.01 | 38.10 | 51.50 | 51.60 | |
ProteinMPNN1 | ✗ | 1.9M | 6.12 | 6.18 | 4.63 | 40.00 | 39.13 | 47.66 | |
PiFold13 | ✗ | 6.6M | 5.52 | 5.00 | 4.38 | 43.06 | 45.54 | 51.45 | |
LM-Design45 | ✓ | 659M | 6.01 | 5.73 | 4.47 | 44.44 | 45.31 | 53.66 | |
GRADE-IF38 | ✗ | 7.0M | 5.30 | 6.05 | 4.58 | 48.21 | 45.94 | 52.24 | |
MapDiff (uniform prior) | ✗ | 14.7M | 3.88 | 3.85 | 3.48 | 55.95 | 54.65 | 60.86 | |
MapDiff (marginal prior) | ✗ | 14.7M | 3.90 | 3.83 | 3.52 | 55.56 | 54.99 | 60.68 | |
