qingxin1999/pepforge-training-data

PepForge — Training Data

Training data for PepForge, a hierarchical deep learning framework for generating peptides with special connections using HELM notation.

Overview

Category	Description	Size
Raw HELM data	Generation corpus (383,817 molecules) + AMP corpus (11,026 active peptides)	~80 MB
Generation splits	Train/val/test + per-stage hierarchical (layout/content)	~395 MB
Prediction splits	AMP 5-class classification (train/val/test/train_active/unlabeled_pool)	~43 MB
Tokenizers	Layout + content monomer vocabularies	<1 MB
Monomer library	HELM monomer definitions and R-group rules	<1 MB
Embeddings	Pre-computed ChemBERTa monomer embeddings (384-dim)	<1 MB

Total size: ~520 MB

Data sources: PubChem, CycPeptMPDB, ChEMBL, DBAASP, UniProt, MacrocycleDB (all publicly available). The AMP corpus is restricted to CLSI MIC-only DBAASP measurements (2026-04-28 update).

Generation Corpus

Data/all_peptides/HELM.csv — 383,817 unique molecules in HELM notation, deduplicated by InChIKey.

Split	Records	File	Size
train	307,055	Generation/processed/train.jsonl	210 MB
val	38,381	Generation/processed/val.jsonl	26 MB
test	38,381	Generation/processed/test.jsonl	26 MB

Per-stage hierarchical splits in Generation/processed/hierarchical/:

Stage	Train	Val	Test
Layout (block-type sequences)	307,055	38,381	38,381
Content (monomer sequences per block)	387,948	48,294	48,529

Average blocks per HELM (train split): 1.26.

AMP Prediction Corpus

Data/amp_peptides/HELM.csv — 11,026 antimicrobial peptides with curated MIC values, restricted to CLSI MIC measurements only (DBAASP, MIC-only filter).

5-class label scheme (activity bins [8, 32, 128] μg/mL):

Class	Definition
background	not classified active
class_1	MIC ≥ 128 μg/mL
class_2	32 ≤ MIC < 128 μg/mL
class_3	8 ≤ MIC < 32 μg/mL
class_4	MIC < 8 μg/mL

Split	Records	background	class_1	class_2	class_3	class_4
train	17,640	8,820	1,858	2,495	2,847	1,620
val	2,206	1,103	250	310	364	179
test	2,206	1,103	232	312	360	199
train_active	8,820	–	1,858	2,495	2,847	1,620
unlabeled_pool	47,467	47,467	–	–	–	–

train_active.jsonl excludes the background class for active-only training experiments. unlabeled_pool.jsonl contains additional background peptides used for negative-mining experiments.

Quick Start

git clone https://github.com/wqx1999/PepForge.git
cd PepForge
python install.py          # Installs env + downloads all models & data

For details, see the GitHub repository.

File Structure

pepforge-training-data/
├── Data/
│   ├── all_peptides/HELM.csv            (75 MB, 383,817 molecules)
│   └── amp_peptides/HELM.csv            (4.9 MB, 11,026 active peptides)
├── Generation/processed/
│   ├── {train,val,test}.jsonl
│   ├── stats.json
│   └── hierarchical/
│       ├── {train,val,test}_layout.jsonl
│       ├── {train,val,test}_content.jsonl
│       └── {train,val,test}_hierarchical_stats.json
├── Prediction/processed/
│   ├── {train,val,test}.jsonl
│   ├── train_active.jsonl
│   └── unlabeled_pool.jsonl
├── Tokenizers/
│   ├── content_vocab.json
│   └── layout_vocab.json
├── Monomer/HELMLibrary.json
└── Embeddings/
    ├── monomer_embeddings.pt           (384-dim, ChemBERTa)
    └── monomer_embeddings.manifest.json

Related Resources

• Code: wqx1999/PepForge
• Models: pepforge-model
• Generated library: pepforge-generated-data
• Figure data: pepforge-fig-data

Citation

@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