daylight-00/prepibind-embeddings

PREpiBind HLA embeddings — emb_hla_esmc_small_0430.h5

The full-length, float32 ESMC 300M embeddings of the MHC class II chains PREpiBind was trained and evaluated on. One HDF5 file, 142.9 MiB, 154 datasets.

This is the store the research path reads. It is not the demo's store: the demo ships a smaller float16 file cut to the peptide-binding window, inside the GitHub repository.

• Code and the pipeline that produced this: https://github.com/daylight-00/PREpiBind
• Models that read it: daylight-00/prepibind
• Backbone that produced it: daylight-00/esmc-300m-2024-12

Contents

file	emb_hla_esmc_small_0430.h5
bytes	149,865,560 (142.9 MiB)
datasets	154, one per chain, keyed by bare allele name
dtype	float32 throughout
shape	(L, 960) — one 960-dimensional vector per residue
lengths	21 distinct L, from 81 to 266
attributes	none, on the file or on any dataset

Keys are allele names exactly as data/mhc_mapping/ spells them, alpha and beta chains as separate entries:

group	count	examples
HLA-DRB1	53	HLA-DRB1*01:01
HLA-DPB1 / HLA-DQB1 / HLA-DQA1 / HLA-DPA1	18 / 17 / 13 / 4	HLA-DQA1*01:01
HLA-DRA / HLA-DRB3 / HLA-DRB4 / HLA-DRB5	1 / 4 / 3 / 3	HLA-DRA*01:01
non-human: H2 / BoLA / Mamu / SLA	18 / 8 / 6 / 6	H2-IAbA, BoLA-DRB3*001:01

116 human HLA class II chains and 38 non-human chains curated alongside them.

The four released models were trained on 116 of these 154 keys, and it is not the same 116: the arms use 98 of the human alleles and all 18 mouse H2 chains. The remaining 38, which are 18 further human alleles plus the 20 BoLA, Mamu and SLA chains, are in the store because they were part of the curation and are used by analyses of molecule coverage, not by any released checkpoint.

How it was produced

ESMC 300M forward pass over each chain's full amino-acid sequence, per-residue hidden states kept, [CLS] and [EOS] stripped, written as float32. The code is pipeline/embeddings/esmc/ in the repository; the sequences come from data/mhc_mapping/, which derives from IPD-IMGT/HLA and UniProt. Nothing here is a measurement — it is a deterministic function of the sequences and the backbone weights.

The peptide-binding window is not applied in this file. It is carried separately, in data/mhc_mapping/HLA2_IMGT_MSA_idx_edit.csv, as sequence|start|end, and applied at load time. Use that mapping table with this store. The demo's store is already cut and needs the mapping table without windows; applying a window twice raises rather than silently returning the wrong residues.

Disclosure: three keys changed after the released models were trained

On 2026-09-08 an H2 chain-swap error was corrected in the upstream sequences, and this file was rebuilt for the affected chains. Exactly 3 of the 154 keys differ from the arrays the four released checkpoints were trained on:

key	before	after
H2-IAdA	(265, 960)	(256, 960)
H2-IAdB	(256, 960)	(265, 960)
H2-IAg7A	(256, 960)	same shape, different values

H2-IAdA and H2-IAdB had been written from each other's sequences. The governing decision is disclose, do not retro-apply: re-running finished training on account of them would buy nothing at the scale involved. The May 2025 training runs behind the released checkpoints read the pre-fix arrays.

The scale, measured on the arms rather than asserted — rows whose alpha or beta chain is one of the three changed keys:

arm	training rows affected	test rows affected
qualitative	547 of 112,871 (0.48 %)	265 of 48,352 (0.55 %)
ms	450 of 77,954 (0.58 %)	215 of 33,490 (0.64 %)
ic50	218 of 33,004 (0.66 %)	85 of 14,150 (0.60 %)

Three consequences worth stating plainly:

• This file carries no HDF5 attributes recording the fix, so the difference is not detectable from the file itself. That is what this section is for. (The sibling …_0329.h5 store does carry h2_chain_fix* attributes; this one does not.)
• Every human HLA embedding is unchanged. Anything that does not touch mouse H2 is unaffected.
• The 18 mouse H2 chains, these three included, are part of the four training arms. The affected rows are a real, if small, part of what the released checkpoints were fit on, which is why this is disclosed rather than dismissed.

If you re-train on this store you will get slightly different H2 behaviour from the released checkpoints. If you run inference on human alleles, you will not.

Usage

hf download daylight-00/prepibind-embeddings emb_hla_esmc_small_0430.h5 \
    --repo-type dataset --local-dir emb

Read it directly:

import h5py
with h5py.File("emb/emb_hla_esmc_small_0430.h5", "r") as f:
    print(len(f))                       # 154
    emb = f["HLA-DRB1*01:01"][()]       # (L, 960) float32, full length

Or hand it to the model, which is what it is for:

from prepibind.inference import load_config, main

cfg = load_config(
    "configs/predict/config_demo.py",
    chkp_path="models/prepibind_qualitative_s100_f0.pt",     # the float32 research checkpoint
    hla_emb_path="emb/emb_hla_esmc_small_0430.h5",           # this store
    hla_path="data/mhc_mapping/HLA2_IMGT_MSA_idx_edit.csv",  # the mapping WITH the windows
    test_path="my_input.csv",
    out_path="outputs",
)
cfg["Test"]["precision"] = "as-trained"
df = main(cfg)

The demo's 116-allele float16 store is derived from this file, and the derivation is checkable:

PREPIBIND_EMB_ROOT=$(pwd)/emb python demo/build_demo_assets.py hla-store --check

PREPIBIND_EMB_ROOT is the directory you downloaded this file into. The script's built-in default path predates a repository move and no longer exists, so set the variable.

Licence

MIT. These arrays are outputs of ESMC 300M. That model moved to Chan Zuckerberg Biohub — biohub/esmc-300m-2024-12, ungated, card tagged mit + other — so the Cambrian Open License Agreement this card previously cited no longer governs it, and the same MIT grant that covers the PREpiBind code and weights applies here. Checked 2026-09-10; the repository's THIRD_PARTY_NOTICES.md records the evidence.

The underlying sequences come from IPD-IMGT/HLA and UniProt, both free to use with attribution; cite them as those databases ask.

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}
}

Cite ESMC as EvolutionaryScale asks, as well: these embeddings are its output.