kfold-multimer-lmdb

Training data for the multimer / protein-ligand fine-tune in hsjang0/k-fold-multimer-tuning. Pre-tokenized complexes as LMDB, plus the eval splits and label tables the configs read.

Do not clone the whole thing to run one experiment. It is 201 GB and a single arm reads a handful of its sources. The repo's scripts/fetch_data.py takes a config name, works out which sources that config actually needs, and pulls only those.

python scripts/fetch_data.py --config CB_v7_bias4bin_v2mix_100k --dest /data/kfold

Layout

lmdb/<source>/              CA-coordinate records, 40 sources
lmdb_cb/<source>/           CB-coordinate sidecars for the sources that have one
eval_split_global/*.txt     held-out keys per source. Every config references these
stability_pairs/*.jsonl     cDNA dG pairs for the stability head

The configs expect lmdb_root to point at the lmdb/ tree and cb_root at lmdb_cb/.

The larger sources

source	size
lmdb/afdb_homodimer_holo_pairs	43.6 G	AFDB homodimers, holo pair schema
lmdb/teddymer_all_holo_pairs	37.4 G
lmdb/sair_ligand_pairs	27.9 G	protein-ligand, SAIR
lmdb/afdb_synth_ligand	16.4 G	synthetic protein-ligand
lmdb/AFDB_Homodimer_apo_token	16.3 G	apo tokens
lmdb_cb/afdb_homodimer_holo_pairs	16.5 G	CB sidecar
lmdb/gator_affinity_pairs	9.5 G	BindingDB Kd/Ki, cofolded
lmdb/stability_monomer	3.4 G	cDNA dG
lmdb/rcsb_ligand	3.2 G	experimental RCSB protein-ligand

What is NOT here

monomer_80M (607 GB) and its cluster index (3.5 GB). Deliberately excluded for size. Several configs carry nonzero weight on it, so those cannot be reproduced exactly from this repo alone. fetch_data.py names the shortfall rather than failing silently.

homodimer_apo_holo (52.8 GB). One data.mdb over HuggingFace's 50 GB per-file limit.

Checkpoints. Encoder weights are separate. The released TriProRep encoders are at k-fold-structure/triprorep-{35M,150M,650M,3B}.

Reading a record

import lmdb, pickle
env = lmdb.open("lmdb/rcsb_holo_pairs", readonly=True, lock=False)
with env.begin() as txn:
    rec = pickle.loads(txn.next())

lock.mdb is a runtime lock and is deliberately not included; lmdb.open(..., lock=False) does not need it.

Provenance

Every source was built by a script in the GitHub repo, listed in its CODEMAP.md section 4, and described in DATASET.md. Eval keys are held out by sequence identity at the AlphaFold-Multimer interface-level convention, not by random assignment.