English | 简体中文

AMP Forge is a de novo antimicrobial peptide (AMP) design platform built on a joint Transformer-based VAE + Latent Diffusion Model architecture. The system leverages pre-trained protein language models (ESM-2 / ProtT5 / Ankh) to extract deep sequence-level representations, compresses them into a low-dimensional latent space via a BiGRU encoder, and employs a latent diffusion process coupled with a non-autoregressive Transformer decoder for parallel sequence generation. Six conditional generation modes — mixed, c_sub, c_ext, c_trunc, tag, and latent — enable precise and controllable AMP variant design.

• Cross-database AMP corpus — consolidated 6 major sources into a curated 25,622-sequence dataset covering nearly all publicly accessible AMP collections we could obtain.
• Multi-PLM backbone — unified interface over ESM-2, ProtT5, and Ankh; pre-computed embeddings avoid training-time bottleneck.
• Latent diffusion generation — 50-step Gaussian diffusion in a 64-dim latent space with classifier-free guidance (CFG), balancing sample diversity and quality.
• Non-autoregressive decoding — parallel prediction of all residue positions eliminates exposure bias and error accumulation.
• 6 conditional variant modes — C-terminal substitution / extension / truncation-rebuild, tag appending, latent perturbation, and mixed stochastic sampling.
• 3-phase training pipeline — VAE MLE pre-training → RL adversarial fine-tuning → latent diffusion training, with cyclical KL annealing + free-bits to prevent posterior collapse.
• MIC prediction (ESM-MIC) — gated multi-branch regression model predicts Minimum Inhibitory Concentration from pre-computed ESM-2 embeddings. Dual-branch architecture (multi-head attention pooling + multi-scale CNN) with gated fusion, OOF data cleaning, and 21-model multi-seed snapshot ensemble achieves PCC = 0.90, R² = 0.81 on held-out test set, enabling in-silico candidate ranking before wet-lab synthesis.
• End-to-end reproducibility — data crawling, embedding computation, training, generation, and evaluation all scripted with a single YAML config and fixed random seeds.

Joint architecture: PLM representation -> VAE latent compression -> latent diffusion -> non-autoregressive Transformer decoding.

.
├── esm_diffvae/               # Core model — data, training, generation, evaluation
│   ├── models/                #   Neural network components
│   ├── training/              #   3-phase training scripts
│   ├── generation/            #   Unconditional, variant, interpolation
│   ├── evaluation/            #   Metrics, physicochemical, visualization
│   ├── data/                  #   Crawling, cleaning, embedding computation
│   ├── mic_prediction/        #   ESM-MIC: MIC value prediction module
│   │   ├── model.py           #     Gated multi-branch architecture
│   │   ├── train.py           #     Training with multi-seed snapshot ensemble
│   │   ├── dataset.py         #     Data loading & OOF filtering
│   │   ├── features.py        #     Physicochemical feature extraction
│   │   ├── precompute_embeddings.py  # ESM-2 embedding pre-computation
│   │   └── config.yaml        #     Hyperparameter configuration
│   └── configs/default.yaml   #   Global configuration
├── frontend/                  # Interactive web UI (React + Three.js)
├── docs/                      # Bilingual documentation (EN + ZH)
├── PROJECT_SUMMARY.md         # Detailed technical summary
└── DATA_COLLECTION_REPORT.md  # Data sources & pipeline report

cd esm_diffvae
pip install -r requirements.txt

cd esm_diffvae
python data/crawl/parse_local_sources.py
python data/crawl/crawl_dramp.py
python data/crawl/crawl_uniprot.py
python data/crawl/merge_and_clean.py
python data/compute_embeddings.py --backend prot_t5 --model prot_t5_xl_half

cd esm_diffvae
python training/train_vae.py --config configs/default.yaml
python training/train_vae_rl.py --config configs/default.yaml --vae-checkpoint checkpoints/vae_best.pt
python training/train_diffusion.py --config configs/default.yaml --vae-checkpoint checkpoints/vae_best_recon.pt

Unconditional generation:

cd esm_diffvae
python generation/unconditional.py \
  --config configs/default.yaml \
  --checkpoint checkpoints/esm_diffvae_full.pt \
  --n-samples 100 \
  --top-p 0.9

Variant generation:

cd esm_diffvae
python generation/variant.py \
  --config configs/default.yaml \
  --checkpoint checkpoints/esm_diffvae_full.pt \
  --input-sequence "GIGKFLHSAKKFGKAFVGEIMNS" \
  --mode mixed \
  --n-variants 50

Latent interpolation:

cd esm_diffvae
python generation/interpolation.py \
  --config configs/default.yaml \
  --checkpoint checkpoints/esm_diffvae_full.pt \
  --seq-a "GIGKFLHSAKKFGKAFVGEIMNS" \
  --seq-b "ILPWKWPWWPWRR" \
  --n-steps 10

cd esm_diffvae
python evaluation/run_evaluation.py \
  --config configs/default.yaml \
  --checkpoint checkpoints/esm_diffvae_full.pt

Pre-compute ESM-2 embeddings, then train the MIC regression model

cd esm_diffvae

# Step 1: Pre-compute ESM-2 embeddings (CPU recommended, ~5 min)
python -m mic_prediction.precompute_embeddings --device cpu

# Step 2: Train single model (with OOF filtering + snapshot ensemble)
python -m mic_prediction.train

# Step 3: Train multi-seed ensemble for best results (3 seeds x 7 snapshots = 21 models)
python -m mic_prediction.train --multi-seed

cd frontend
pnpm install
pnpm dev

• EN: docs/en/quickstart.md
• EN: docs/en/training.md
• EN: docs/en/generation.md
• EN: docs/en/evaluation.md
• EN: docs/en/data-pipeline.md