Files changed (1) hide show
  1. README.md +0 -38
README.md CHANGED
@@ -13,44 +13,6 @@ tags:
13
 
14
  Multi-Task Learning peptide classifier covering 22 binary peptide-activity tasks. Built on a frozen ESM-2 (650M) backbone with a parallel Transformer + CNN feature extractor and per-task heads, following a PDeepPP-inspired design.
15
 
16
- ## Held-out Test Set Performance (Averaged across 22 tasks)
17
-
18
- | Metric | Value |
19
- |---|---|
20
- | Accuracy | 86.63% |
21
- | F1 | 84.99% |
22
- | AUC | 93.47% |
23
- | MCC | 72.67% |
24
-
25
- Best Val Avg F1 (used for checkpoint selection): 85.56%
26
-
27
- ## Per-Task Test Metrics
28
-
29
- | Task | ACC | F1 | AUC | MCC |
30
- |---|---|---|---|---|
31
- | AntiMRSA | 0.9899 | 0.9667 | 0.9970 | 0.9607 |
32
- | Anticancer | 0.7035 | 0.7344 | 0.8007 | 0.4184 |
33
- | ACE_inhibitory | 0.6801 | 0.7392 | 0.8355 | 0.4040 |
34
- | Antioxidant | 0.7117 | 0.7309 | 0.8173 | 0.4382 |
35
- | Bitter | 0.8203 | 0.8456 | 0.9591 | 0.6782 |
36
- | Antimalarial | 0.9736 | 0.7692 | 0.9177 | 0.7579 |
37
- | Anti_inflammatory | 0.9886 | 0.9887 | 0.9979 | 0.9773 |
38
- | Antimicrobial | 0.9746 | 0.9552 | 0.9915 | 0.9379 |
39
- | Signal_peptide | 0.9927 | 0.9927 | 0.9997 | 0.9854 |
40
- | Antifungal | 0.9465 | 0.9456 | 0.9863 | 0.8935 |
41
- | Antimalarial_alt | 0.9877 | 0.9630 | 0.9942 | 0.9566 |
42
- | Anticancer_alt | 0.9330 | 0.9316 | 0.9784 | 0.8667 |
43
- | Anti_parasitic | 0.7826 | 0.7500 | 0.9216 | 0.5855 |
44
- | Umami | 0.8427 | 0.7308 | 0.9297 | 0.6243 |
45
- | Quorum_sensing | 0.9250 | 0.9231 | 0.9850 | 0.8511 |
46
- | Antibacterial | 0.9431 | 0.9424 | 0.9789 | 0.8863 |
47
- | NeuroPred | 0.8660 | 0.8543 | 0.9444 | 0.7416 |
48
- | Toxicity | 0.9086 | 0.8971 | 0.9699 | 0.8178 |
49
- | Antiviral | 0.8307 | 0.8319 | 0.9098 | 0.6614 |
50
- | DPPIV_inhibitory | 0.8647 | 0.8732 | 0.9478 | 0.7361 |
51
- | BBP | 0.6579 | 0.5185 | 0.9141 | 0.3873 |
52
- | TTCA | 0.7360 | 0.8129 | 0.7858 | 0.4221 |
53
-
54
  ## Architecture
55
 
56
  - **Shared encoder**: frozen ESM-2 (`facebook/esm2_t33_650M_UR50D`, 650M params) + learnable base embedding, mixed at `esm_ratio=0.9`
 
13
 
14
  Multi-Task Learning peptide classifier covering 22 binary peptide-activity tasks. Built on a frozen ESM-2 (650M) backbone with a parallel Transformer + CNN feature extractor and per-task heads, following a PDeepPP-inspired design.
15
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
16
  ## Architecture
17
 
18
  - **Shared encoder**: frozen ESM-2 (`facebook/esm2_t33_650M_UR50D`, 650M params) + learnable base embedding, mixed at `esm_ratio=0.9`