dn-gh/dti-merged-preprocessed

MERGED (SPRINT) — preprocessed for DTI model training

Preprocessed, ready-to-train version of the MERGED drug-target interaction dataset introduced by SPRINT:

McNutt, A. T., Adduri, A., Ellington, C. N., et al. (2024). Sprint: Vector-based screening of protein-ligand interactions. arXiv:2411.15418 Code: abhinadduri/panspecies-dti (MIT)

MERGED compiles interactions from PubChem BioAssay, BindingDB, and ChEMBL into a single binary (binder / non-binder) benchmark with MMSeqs2-clustered train/val/test splits, covering ~9,067 targets (paper-reported) and 3.5M ligands.

This repo is a derived, filtered, feature-extracted version of that data, produced for pretraining drug-target interaction models with a pre-computed ligand fingerprint matrix and cleaned, split-ready interaction tables.

What was done to the raw data

1. Target filtering: of 12,065 raw UniProt target sequences, 2,678 (22%) were junk (< 20 residues) and dropped, leaving 9,387 usable targets — matches the paper's reported ~9,067 (small gap likely a stricter/different filter threshold).
2. Row filtering: all six raw interaction tables ({pos,neg} x {train,val,test}) filtered to rows referencing only non-junk targets.
3. Fixed eval sets: val/test are each pre-sampled once (seed 42) at a 1:1 positive:negative ratio and shuffled, so checkpoint selection is comparable across training epochs. train keeps the full filtered positive set and the full filtered negative pool on disk — a common recipe for this kind of heavily imbalanced (94:1) binary DTI data is to resample a fresh negative subset from the pool every training epoch rather than training on all ~64M negatives at once, so that resampling is left to train time, not baked into this preprocessing.
4. Ligand features: 2048-bit ECFP4 (Morgan, radius 2) fingerprints computed for every one of the 2,005,744 ligands actually referenced by a kept interaction row (out of 3.5M total in the raw dataset) via RDKit.

Verified against the paper's reported totals before processing (854,118 positive / 80,681,825 negative raw interactions, 3,529,822 drugs) — exact match, i.e. this is the complete, untampered dataset upstream of the filtering above.

Files

File	Rows / shape	Description
merged_train_pos.parquet	574,754	Full filtered positive training interactions (Drug_ID, Target_ID)
merged_train_neg_pool.parquet	64,361,902	Full filtered negative training pool — resample from this per epoch
merged_val_eval.parquet	215,950	Fixed val set, 1:1 pos:neg, pre-shuffled, Y label included
merged_test_eval.parquet	342,776	Fixed test set, 1:1 pos:neg, pre-shuffled, Y label included
merged_targets.parquet	5,488	Target_ID (UniProt accession) → Target (AA sequence), restricted to targets actually referenced by kept rows
merged_drugs_order.csv	2,005,744	Row order for merged_ligand_features.npy (Drug_ID per row index)
merged_ligand_features.npy	(2,005,744, 2048) float32	ECFP4 fingerprints, one row per merged_drugs_order.csv entry

Drug_ID is an integer ligand ID (not a raw SMILES string — join against merged_drugs_order.csv's row index to recover the fingerprint). Target_ID is a UniProt accession string despite the raw column being named aa_seq upstream.

License

MIT, inherited from the source panspecies-dti repository. Please cite the SPRINT paper (above) if you use this data.