Pre-trained model weights for PepForge, a hierarchical deep learning framework for generating peptides with special connections using HELM notation.
Architecture
PepForge uses a three-stage cascade (Layout → Content → Connection) for generation and a 4-model MCC-weighted ensemble for AMP activity prediction. The prediction ensemble was retrained 2026-04-28/29 on CLSI MIC-only DBAASP data with members selected by validation MCC (test set never consulted at the selection step).
Generation Models
Stage
File
Architecture
Test PPL / Metric
Layout
Generation/Layout/260210_GPT.pt
GPT (d=64, L=1)
PPL = 2.24
Content (autoregressive, default)
Generation/Content/GPT_L_260226.pt
GPT (d=768, L=12)
PPL = 6.61
Content (masked, infilling)
Generation/Content/BERT_L_260301.pt
BERT (d=768, L=12)
PPL = 9.15
Connection
Generation/Connection/GAT_L_260226.pt
GAT (d=768, L=6)
Exist F1 = 0.971, Type Macro-F1 = 0.912
Prediction Models — AMP Ensemble (260428/260429)
Each member is the best of its (encoding, model-type) quadrant by validation MCC.
File
Type
Encoding
Test Acc
Test Macro-F1
Test MCC
Weight (val MCC)
Prediction/AMP/LSTM_L_260428_SMILES.pt
LLM
SMILES
0.7167
0.5663
0.5871
0.6121
Prediction/AMP/LSTM_M_260429_HELM.pt
LLM
HELM
0.7058
0.5811
0.5717
0.6021
Prediction/AMP/GCN_L_260429_HELM.pt
GNN
HELM
0.6355
0.5047
0.4844
0.5136
Prediction/AMP/GCN_L_260428_SMILES.pt
GNN
SMILES
0.6165
0.4478
0.4630
0.4791
Held-out ensemble performance (test split, 2,206 samples; full report in ensemble_test_eval.json):
Strategy
Acc
Macro-F1
Weighted-F1
MCC
soft_vote (uniform 0.25 each)
0.7393
0.6049
0.7377
0.6175
weighted_vote (val-MCC weights, default)
0.7421
0.6092
0.7403
0.6216
The weighted ensemble exceeds the best single member (LSTM/L SMILES, MCC 0.5871) by +0.0345.
Quick Start
bash
1git clone https://github.com/wqx1999/PepForge.git
2cd PepForge
3python install.py # Installs env + downloads all models & data
bash
1# Generation + AMP prediction in one cascade call2python Pipelines/Inference.py --num_samples 100 --predict amp
