daylight-00/prepibind-demo

PREpiBind — demo checkpoints (float16)

The same four models as daylight-00/prepibind, cast to float16. 106.5 MiB each instead of 213.0 MiB. This is what the notebooks and the Colab demo download, and what configs/predict/*.py load by default.

Use these to try the model. Use the float32 ones for anything you intend to report.

• Code: https://github.com/daylight-00/PREpiBind
• Colab: demo/run_colab.ipynb in that repository
• Backbone: daylight-00/esmc-300m-2024-12

The four files

file	arm	seed	fold	val_loss	predict config
prepibind_qualitative_s100_f0_fp16.pt	qualitative binding assays	100	0	0.35582	configs/predict/config_demo.py
prepibind_ms_s128_f3_fp16.pt	mass-spectrometry eluted ligands	128	3	0.12274	configs/predict/config_ms.py
prepibind_ic50_500_s128_f2_fp16.pt	IC50, binder threshold < 500 nM	128	2	0.46631	configs/predict/config_ic50_500.py
prepibind_ic50_1000_s42_f1_fp16.pt	IC50, binder threshold < 1000 nM	42	1	0.49456	configs/predict/config_ic50_1000.py

Each is 106.5 MiB, a one-key {"model_state_dict": ...} dict of 64 float16 tensors, 55,820,161 parameters. Nothing is quantised or pruned: this is tensor.half() of the float32 release, produced by demo/build_demo_assets.py checkpoint, which is a pure function of the training checkpoint and can be re-derived at any time.

All four rows are the val_loss argmin of their arm. Not yet published: the files exist locally and nothing has been uploaded to HuggingFace; the sequence for doing so is tools/release/upload_plan.md.

How these four were chosen

Each arm was trained as 15 runs: three seeds (42, 100, 128) by five cross-validation folds. Every number reported in the paper is a mean over those 15 runs. A released checkpoint is one run out of the fifteen, the one with the lowest validation loss in its arm — the value in the table. Validation loss is the selection criterion; the Val ROC-AUC in the training logs is invalid — train.py calls its own roc_auc_score(outputs, labels) with the arguments swapped, so 21.7 % of the logged values fall outside [0, 1] — and it was never used. No test-set quantity took part in the choice either. The val_loss column was re-derived on 2026-09-10 from analysis/val_metrics.csv; the numbers published before that date were minima over a filtered subset of the epoch lines and were too high.

That correction moved one selection. The ms row was seed 100, fold 1 (true val_loss 0.13128) and is now seed 128, fold 3 (0.12274) — a different run of the same fifteen the paper averages for that arm, trained from the same 77,954-row data/dataset/ms_ql/train.csv under a byte-identical config. The research card sets out why the two are comparable. The other three arms kept the run they always had; only the number printed beside them changed.

A single checkpoint therefore does not reproduce a paper number, and this float16 copy of it reproduces one even less exactly. The full comparison, mean over 15 runs against this run alone, is on the research card.

This does not reproduce the research path bit for bit

Two separate reasons, and both of them are accepted rather than fixed:

1. Half precision throughout. The head runs in float16 and so does ESMC, against float32 head and bfloat16 ESMC on the research path.
2. The demo encodes epitopes with ESMC at run time instead of reading the precomputed store the training and the paper's evaluation used. That alone makes bit-exactness impossible whatever precision it runs in. On Colab's free tier the GPU is a T4, which is pre-Ampere, so flash-attn is off and the runtime says so when it drops it.

Measured, on 2,000 demo rows, this float16 path against the as-trained path:

mean absolute difference in score	0.0031
worst case	0.0619
Spearman correlation	0.99985
peak GPU memory	1.95 GiB, against 2.95 GiB

Ranking is essentially preserved; individual scores move in the third decimal, and one row in two thousand moved by 0.06. If a decision turns on that, use the float32 checkpoints and the as-trained precision.

Input format

Two columns. The MHC field is the two chain names joined by an underscore, and both must be keys in the mapping table the config points at. The datasets and the demo write beta first, then alpha; the order does not change the score, the head having no positional encoding, but it keeps inputs comparable with the published ones.

MHC,Epitope
HLA-DRB1*01:01_HLA-DRA*01:01,PKYVKQNTLKLATA
HLA-DPB1*03:01_HLA-DPA1*01:03,AAAARYPNVTIAAAA
H2-IAbB_H2-IAbA,AAAAPAAAATTAAPA

Output is the input with Logits and Score (sigmoid) appended. Every training epitope is a 15-mer; other lengths run but are outside the training distribution.

The demo ships its own HLA store, demo/data/emb_hla_esmc_small_demo_fp16.h5 (16 MB, float16, 116 alleles, cut to the peptide-binding window) - every allele the four dataset arms use, which is 98 human HLA class II chains and 18 mouse H2 chains. That file is in the GitHub repository, not here, and is a verified derivation of the full-length research store: python demo/build_demo_assets.py hla-store --check.

Usage

git clone https://github.com/daylight-00/PREpiBind && cd PREpiBind
pip install --extra-index-url https://download.pytorch.org/whl/cu126 -e .

hf download daylight-00/prepibind-demo      prepibind_qualitative_s100_f0_fp16.pt --local-dir models
hf download daylight-00/esmc-300m-2024-12   esmc_300m_2024_12_v0_fp16.pth         --local-dir models

python -m prepibind.inference configs/predict/config_demo.py --plot

Predictions land in outputs/prediction.csv; --plot also writes a KDE of the scores. The config already points at models/prepibind_qualitative_s100_f0_fp16.pt; switch models by switching config, or override any single key:

python -m prepibind.inference configs/predict/config_ms.py --test_path my_input.csv --out_path out

The same thing from Python, which is what the notebooks do:

from prepibind.inference import load_config, main

cfg = load_config("configs/predict/config_demo.py", test_path="my_input.csv", out_path="outputs")
df = main(cfg)

The ESMC encoder is vendored into prepibind/esmc/, so the demo needs only torch and huggingface_hub, both already present on Colab. Do not pip install esm.

No GPU is required. Without CUDA the runtime falls back to float32 throughout and turns off flash-attn; on a pre-Ampere card it turns off flash-attn and keeps float16. Either way it prints what it changed, because a silently downgraded run is one whose numbers cannot be compared with anything.

Limitations

Everything on the research card applies — one run rather than the reported mean, class II only, the disclosed H2 chain correction in the embedding store, binding rather than immunogenicity — plus the precision deviation measured above.

Licence

MIT for these weights, same as the repository. The ESMC 300M backbone is not ours but is also MIT: it moved to biohub/esmc-300m-2024-12 (ungated, tagged mit + other), and the vendored encoder source is copied from esm 3.4.0, whose licence is plain MIT, "Copyright 2026 Chan Zuckerberg Biohub, Inc.". Checked 2026-09-10; see THIRD_PARTY_NOTICES.md in the repository.

Citation

@article{jang2026prepibind,
  title   = {PREpiBind: Protein Representation-integrated Epitope-MHC Class II Binding Prediction},
  author  = {Jang, David Hyunyoo and Kim, Dongwoo and Park, Byungho and Hwang, Untaek and Choi, Yoonjoo and Lee, Juyong},
  journal = {bioRxiv},
  year    = {2026}
}