ESM3di/config_esm2_3B_balanced_ADD_PLDDT.txt at master · DessimozLab/ESM3di · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
================================================================================
TO ADD pLDDT WEIGHTING TO config_esm2_3B_balanced.json
================================================================================

CURRENT STATUS:
  ✓ Cyclical Focal Loss enabled
  ✗ pLDDT weighting NOT enabled (no pLDDT bins files available)

The balanced dataset at /mnt/data1/bfvd/balanced/ does not include pLDDT bins.

================================================================================
OPTION 1: Generate pLDDT Bins from PDB Files
================================================================================

If you have the original PDB files that were used to create the balanced
dataset, you can generate pLDDT bins:

1. Find the source PDB directory for balanced dataset

2. Run extract_plddt_bins.py:
   python -m esm3di.extract_plddt_bins \
     --pdb-dir /path/to/balanced/pdbs \
     --output /mnt/data1/bfvd/balanced/phylum_balanced_plddt_bins.fasta

3. Split the pLDDT bins to match train/val:
   python -m esm3di.split_data \
     --aa-fasta /mnt/data1/bfvd/balanced/phylum_balanced_aa.fasta \
     --three-di-fasta /mnt/data1/bfvd/balanced/phylum_balanced_3di.fasta \
     --plddt-bins-fasta /mnt/data1/bfvd/balanced/phylum_balanced_plddt_bins.fasta \
     --train-ratio 0.8 \
     --val-ratio 0.1 \
     --test-ratio 0.1 \
     --output-prefix /mnt/data1/bfvd/balanced/phylum_balanced

4. Add to config_esm2_3B_balanced.json:
   {
     ...
     "plddt_bins_fasta": "/mnt/data1/bfvd/balanced/phylum_balanced_train_plddt_bins.fasta",
     "val_plddt_bins_fasta": "/mnt/data1/bfvd/balanced/phylum_balanced_val_plddt_bins.fasta",
     "plddt_min_bin": 5,
     "plddt_weight_exponent": 1.5,
     ...
   }

================================================================================
OPTION 2: Use Without pLDDT Weighting (CURRENT CONFIG)
================================================================================

The current config uses CyclicalFocalLoss without pLDDT weighting.
This is still an improvement over regular Focal Loss and will work fine!

Advantages:
  ✓ Can use immediately - no additional data needed
  ✓ Cyclical learning strategy still provides benefits
  ✓ Asymmetric weighting (gamma_pos/gamma_neg) handles class imbalance

Disadvantages:
  ✗ No confidence-based position weighting
  ✗ Cannot down-weight low-confidence predictions

================================================================================
RECOMMENDATION
================================================================================

If pLDDT bins are not easily available:
  → Use current config as-is, it's already improved with cyclical loss

If you can generate pLDDT bins:
  → Follow Option 1 to enable PLDDTWeightedCyclicalFocalLoss
  → Expected accuracy improvement: 1-3% on high-confidence regions

================================================================================