Table 3 Ablation study of the denoising network modules in MapDiff
From: Mask-prior-guided denoising diffusion improves inverse protein folding
Module | Component | MapDiff | Variant 1 | Variant 2 | Variant 3 | Variant 4 | Variant 5 |
|---|---|---|---|---|---|---|---|
G-EGNN | EdgeUpdate | ✓ | ✓ | ✓ | ✓ | ✓ | |
CoordinateUpdate | ✓ | ✓ | ✓ | ✓ | |||
GlobalContext | ✓ | ✓ | ✓ | ✓ | ✓ | ||
Refinement | MaskAdaptor | ✓ | ✓ | ✓ | ✓ | ||
IPA network | ✓ | ✓ | ✓ | ✓ | ✓ | ||
Sequence | Recovery (↑, %) | 60.93 | 58.64 | 59.76 | 58.38 | 60.16 | 56.46 |
NSSR62 (↑, %) | 78.57 | 76.73 | 77.32 | 77.04 | 77.80 | 75.69 | |
NSSR90 (↑, %) | 75.66 | 73.52 | 74.82 | 74.24 | 75.02 | 72.58 | |
Foldability | pLDDT (↑) | 88.63 | 88.08 | 88.24 | 87.95 | 88.30 | 86.95 |
PTM (↑) | 79.00 | 78.42 | 78.61 | 78.24 | 78.74 | 77.20 | |
PAE (↓) | 5.42 | 5.57 | 5.54 | 5.62 | 5.49 | 5.86 | |
TM-Score (↑, %) | 88.77 | 88.25 | 88.47 | 88.16 | 88.58 | 87.50 | |
GDT-TS (↑, %) | 87.75 | 87.12 | 87.40 | 86.96 | 87.53 | 85.72 | |
RMSD (↓) | 2.57 | 2.67 | 2.65 | 2.65 | 2.53 | 2.76 | |
Summary | Change | - | ↓↓ | ↓ | ↓↓ | ↓ | ↓↓↓ |
