Protein Representation-integrated Epitope–MHC Class II Binding Prediction
PREpiBind Webserver Jang DH, Kim D, Choi Y, Lee J. TBD (2026)
A Streamlit-based web server for predicting peptide binding to human and mouse MHC class II molecules using ESM C 300M protein language model embeddings. The server implements four task-specific models trained on qualitative, IC50, and mass spectrometry ligandomics data from IEDB, and provides pre-computed embeddings for 7,282 alleles (838 alpha, 6,444 beta chains) from IPD-IMGT/HLA.
Live server: https://lcdd.snu.ac.kr/prepibind
PREpiBind-web/
├── app.py # Entry point: model loading, session state, navigation
├── security_config.py # HTTP security headers via Tornado monkey-patching
├── config_demo.py # Inference config — Standard Mode
├── config_demo_custom_hla.py # Inference config — Custom HLA Mode
├── robots.txt
├── .streamlit/config.toml # Streamlit server settings
│
├── pages/
│ ├── 0_home.py # Home / overview
│ ├── 1_prediction.py # Standard prediction interface
│ ├── 2_evaluation.py # Model benchmarking with labeled data
│ ├── 3_instructions.py # Full user guide
│ ├── 4_about.py # Authors, citation, licensing
│ └── 5_custom.py # Custom HLA sequence mode
│
├── code/
│ ├── model.py # Neural network architecture (plm_cat_mean_inf)
│ ├── inference.py # Inference pipeline and DataLoader logic
│ ├── encoder.py # Dataset classes for standard/custom HLA modes
│ ├── dataprovider.py # Data loading and MHC sequence mapping
│ └── collate.py # Batch collation with padding and masking
│
├── data/
│ ├── mhc_mapping.csv # 7,282 allele name → domain-trimmed sequence
│ ├── dataset_demo.csv # Demo input for smoke-testing
│ ├── test.csv # Built-in test set — Qualitative model
│ ├── test_ms.csv # Built-in test set — MS model
│ ├── test_ic50_500.csv # Built-in test set — IC50 <500 nM model
│ └── test_ic50_1000.csv # Built-in test set — IC50 <1000 nM model
│
├── models/ # Model checkpoints (download required; see below)
└── outputs/ # Prediction output directory (runtime)
- Python ≥ 3.10
- CUDA-enabled GPU (tested on CUDA 12.8)
| Package | Version tested |
|---|---|
| streamlit | 1.55.0 |
| torch | 2.7.1+cu128 |
| esm | 3.2.0 |
| flash-attn | 2.8.0.post2 |
| numpy | 2.3.0 |
| pandas | 2.3.0 |
| plotly | 6.1.2 |
| scikit-learn | 1.7.0 |
| h5py | 3.14.0 |
Install ESM (EvolutionaryScale):
pip install esmInstall Flash Attention (requires CUDA toolkit and NVCC):
pip install flash-attn --no-build-isolationgit clone https://github.com/daylight-00/PREpiBind-web.git
cd PREpiBind-webDownload from HuggingFace and place in models/:
# PREpiBind task-specific checkpoints
huggingface-cli download daylight-00/prepibind-esmc-300m --local-dir models/
# ESM C 300M base model (FP16)
huggingface-cli download daylight-00/esmc-300m-2024-12 --local-dir models/Expected files after download:
models/
├── esmc_300m_2024_12_v0_fp16.pth
├── prepi_esmc_small_e5_s128_f4_fp16.pth # Qualitative
├── prepi_esmc_small_ms_e5_s100_f0_fp16.pth # Mass Spectrometry
├── prepi_esmc_small_ic50_500_e5_s128_f4_fp16.pth # IC50 <500 nM
└── prepi_esmc_small_ic50_1000_e5_s128_f1_fp16.pth # IC50 <1000 nM
Pre-computed ESM C 300M embeddings for all 7,282 IPD-IMGT/HLA alleles are required for Standard Mode:
huggingface-cli download daylight-00/emb_hla_esmc_small_0601_fp16 \
--repo-type dataset \
--local-dir data/emb_hla_esmc_small_0601_fp16The embedding directory should resolve to data/emb_hla_esmc_small_0601_fp16/.
The server defaults to GPU device 1 (CUDA_VISIBLE_DEVICES=1 in app.py line 18). Change this to match your system before running.
streamlit run app.py \
--server.port 8501 \
--server.baseUrlPath /prepibindThe server will be available at http://localhost:8501/prepibind.
For production deployment behind a reverse proxy, set --server.baseUrlPath to match your proxy path prefix. Once TLS is configured, uncomment the HSTS header and secure cookie flag in security_config.py.
The prediction model (code/model.py) takes two inputs — a pre-computed HLA embedding (concatenated alpha + beta chain from ESM C 300M) and a tokenized peptide sequence — and processes them through:
- Separate two-block self-attention encoders for HLA and epitope representations
- A single joint interaction block (concatenation + self-attention)
- Masked mean pooling → two-layer MLP → scalar logit
The output score is sigmoid-transformed; scores ≥ 0.5 indicate predicted binding.
Standard Mode uses pre-computed embeddings stored in data/emb_hla_esmc_small_0601_fp16/.
Custom HLA Mode computes ESM C embeddings on-the-fly for arbitrary MHC sequences not in the database (including non-human alleles such as Mamu, SLA, and BoLA).
Prediction (CSV upload):
MHC_alpha,MHC_beta,Epitope
HLA-DRA*01:01,HLA-DRB1*15:01,GELIGILNAAKVPAD
HLA-DQA1*05:01,HLA-DQB1*02:01,PKYVKQNTLKLATAAEvaluation (CSV upload with binary labels):
MHC_alpha,MHC_beta,Epitope,Target
HLA-DRA*01:01,HLA-DRB1*15:01,GELIGILNAAKVPAD,1
HLA-DQA1*05:01,HLA-DQB1*02:01,PKYVKQNTLKLATAA,0Column headers must match exactly. Allele names must follow WHO HLA nomenclature and be present in data/mhc_mapping.csv. The model was trained on 15-mer peptides; predictions for other lengths are accepted but not validated.
The server provides pre-computed ESM C embeddings for 7,282 alleles (838 alpha, 6,444 beta chains) from IPD-IMGT/HLA. However, the underlying models were trained on 151 unique HLA dimers (27 alpha × 89 beta allele combinations) from IEDB. Predictions for alleles outside the training set rely on PLM-based zero-shot generalization.
Non-human primate (Mamu), porcine (SLA), and bovine (BoLA) MHC alleles evaluated in the original study are not available in Standard Mode; use Custom HLA Mode with explicit sequence input for these species.
If you use PREpiBind or this web server, please cite:
@article{Jang2026PREpiBind,
author = {Jang, David Hyunyoo and Kim, Dongwoo and Park, Byungho
and Hwang, Untaek and Choi, Yoonjoo and Lee, Juyong},
title = {{PREpiBind}: Protein Representation-integrated
Epitope--{MHC} Class {II} Binding Prediction},
journal = {TBD},
year = {2026},
}The ESM C 300M model is developed by EvolutionaryScale and is subject to the Cambrian Open License Agreement.
The PREpiBind web server source code is released under the MIT License. ESM C 300M model weights are subject to the Cambrian Open License (EvolutionaryScale).