README.md

---
license: cc-by-4.0
task_categories:
  - text-generation
language:
  - en
tags:
  - peptide
  - HELM
  - chemistry
  - drug-discovery
  - antimicrobial-peptide
  - evaluation
  - figure-data
size_categories:
  - 1K<n<10K
---

# PepForge — Figure Reproduction Data

Evaluation data and pre-computed metrics for reproducing all figures in the PepForge manuscript. Used by the visualization notebooks in [PepForge](https://github.com/wqx1999/PepForge).

## Overview

| Category | Description | Files | Size |
|----------|-------------|------:|------|
| **End-to-end evaluation** | Cascade vs Flat (5+5 runs × 100K), GPT vs BERT infilling (3 ratios × 5+5 runs), constrained generation (9 scenarios) | 219 | 1.4 GB |
| **Per-stage evaluation** | Layout (3 archs), Content (4 archs × 3 sizes), Connection (5 archs × 3 sizes) | 75 | 92 MB |
| **Prediction evaluation** | LLM + GNN (4 archs × 3 sizes × 2 encodings), ensemble comparison | 97 | 1.2 MB |
| **Candidate analysis** | Screening results from 4.78M novel generated molecules | 4 | 19 MB |

Total size: ~1.5 GB

## Data Format

Each experiment directory contains a subset of:

| File | Format | Description |
|------|--------|-------------|
| `eval_report.json` | JSON | Pre-computed evaluation metrics (validity, uniqueness, novelty, internal diversity, NN distance, per-type statistics). **Primary data source for visualization.** |
| `raw.jsonl` | JSONL | Per-molecule generation results (HELM, SMILES, connection types, roundtrip validation). |
| `candidates.csv` | CSV | Valid molecules with SMILES (subset of `raw.jsonl`). |
| `eval_summary.txt` | Text | Human-readable evaluation summary. |
| `eval_test.json` | JSON | Per-stage model test metrics (loss, accuracy, confusion matrix). |
| `inference.json` | JSON | Per-stage inference outputs for distribution analysis. |
| `metrics.json` | JSON | Per-stage training metrics (per-epoch loss / accuracy curves). |

## File Structure

```
pepforge-fig-data/
└── Data/
    ├── Generation_Model/
    │   ├── Layout/{GPT,GRU,LSTM}/                 # 3 archs × eval_test + inference + metrics
    │   ├── Content/{GPT,GRU,LSTM,BERT}/{SMALL,MEDIUM,LARGE}/
    │   ├── Connection/{GAT,GCN,GIN,MPNN,GRAPH_TRANSFORMER}/{SMALL,MEDIUM,LARGE}/
    │   └── End2end/
    │       ├── Cascade_vs_Flat/                   # Main comparison (Fig 3 / S5)
    │       │   ├── Cascade/run{1..5}/             # eval_report.json + raw.jsonl + ...
    │       │   └── Flat/run{1..5}/
    │       ├── GPT_vs_BERT_Infilling/             # Infilling comparison (Fig 3 / S6)
    │       │   ├── constraints/                   # Pre-sampled constraint files
    │       │   └── ratio_{0.3,0.5,0.7}/{GPT,BERT}/run{1..5}/
    │       └── Constrained_Generation/{CG1_L,CG2_LC_Cys,CG3_LCC_SS,CG4_LCC_Lanthi,
    │                                  CG5_L_multi,CG6_LCC_multi_Sulfa,CG7_L_filter_SS,
    │                                  CG8_LayoutRange,CG9_PEPonly}/
    ├── Prediction_Model/
    │   ├── HELM/{LLM,GNN}/{arch}/{SMALL,MEDIUM,LARGE}/    # 48 evals
    │   ├── SMILES/{LLM,GNN}/{arch}/{SMALL,MEDIUM,LARGE}/  # 48 evals
    │   └── Inference/ensemble_comparison.json
    └── Candidate_Analysis/
        ├── eval_report.json                       # Screening evaluation (NN distance, FP stats)
        ├── AMP_active.csv                         # AMP-active candidates (39,891 rows, class 3/4)
        └── AMP_hit.csv                            # Hit set: Active ∧ non_hem ∧ non_tox (799 rows)
```

The Candidate_Analysis split reflects the rescored library after retraining the AMP ensemble on MIC-only DBAASP data: 4.78M novel → 39,891 active → 799 hit (triple filter). The legacy `AMP_active_druglike.csv` from the previous release has been replaced by `AMP_hit.csv` (same column schema, current numbers).

## Corresponding Figures

| Figure | Data Source | Notebook |
|--------|-----------|----------|
| Fig 3a–f | `Generation_Model/End2end/Cascade_vs_Flat/{Cascade,Flat}/run{1..5}/eval_report.json` | `Generation_Model/End2end/Cascade_vs_Flat/flat_cascade_multirun.ipynb` |
| Fig 3g–i | `Generation_Model/End2end/GPT_vs_BERT_Infilling/ratio_*/eval_report.json` | `Generation_Model/End2end/GPT_vs_BERT_Infilling/Infilling_Generation.ipynb` |
| Fig 4a–b | `Prediction_Model/{HELM,SMILES}/{LLM,GNN}/*/*/eval_test.json` + `Prediction_Model/Inference/ensemble_comparison.json` | `Prediction_Model/prediction.ipynb` |
| Fig 4c–e | `Candidate_Analysis/eval_report.json` + `Candidate_Analysis/AMP_active.csv` + `AMP_hit.csv` | `Candidate_Analysis/candidate_analysis.ipynb` |
| SI: Layout / Content / Connection | `Generation_Model/{Layout,Content,Connection}/...` | per-stage notebooks |
| SI: Constrained generation | `Generation_Model/End2end/Constrained_Generation/CG*/` | `constrained_generation.ipynb` |

## Quick Start

```bash
# Clone PepForge and download figure data
git clone https://github.com/wqx1999/PepForge.git
cd PepForge
python install.py --download paper    # Downloads this dataset
python install.py --download data     # Downloads training data (some notebooks need it)

# Run a visualization notebook
cd Paper/Figures/Generation_Model/End2end/Cascade_vs_Flat/
jupyter notebook flat_cascade_multirun.ipynb
```

## Related Resources

- **Code + Notebooks**: [wqx1999/PepForge](https://github.com/wqx1999/PepForge)
- **Models**: [pepforge-model](https://huggingface.co/qingxin1999/pepforge-model)
- **Training data**: [pepforge-training-data](https://huggingface.co/datasets/qingxin1999/pepforge-training-data)
- **Generated library**: [pepforge-generated-data](https://huggingface.co/datasets/qingxin1999/pepforge-generated-data)

## Citation

```bibtex
@article{wang2026pepforge,
  title={PepForge: Hierarchical HELM-Based Peptide Generation},
  author={Wang, Qingxin and Süssmuth, Roderich D.},
  journal={bioRxiv},
  year={2026},
  doi={10.64898/2026.05.29.728379},
  url={https://www.biorxiv.org/content/10.64898/2026.05.29.728379v1}
}
```

## License

CC-BY-4.0