omtx/om-del-binders

Om DEL Binders

Prospective small-molecule binding labels from Om's DNA-encoded library screening, released one protein target per configuration. Each row is a protein–molecule–label triple: the target sequence travels with every row, so configurations concatenate directly for multi-target training.

from datasets import load_dataset
ds = load_dataset("omtx/om-del-binders", "tbxt")

Targets

Config	Target	Sequence	Rows	Positives	Negatives
tbxt	TBXT (brachyury), UniProt O15178	435 aa, full length	1,359,232	339,808	1,019,424
brd4_412	BRD4 tandem BD1–BD2 bromodomain, UniProt O60885	412 aa fragment	7,722,864	1,930,716	5,792,148
brd4_1363	BRD4 full length, UniProt O60885	1363 aa	12,207,744	3,051,936	9,155,808

Every configuration is balanced at exactly three negatives per positive, so the three pool cleanly if you concatenate them.

The two BRD4 configurations are different constructs of the same protein, not duplicates. brd4_412 is the tandem bromodomain module, the region every BET inhibitor engages.

Schema

Column	Type	Meaning
smiles	string	molecule structure
protein_sequence	string	the target sequence, identical on every row of a config
target	string	target label, e.g. TBXT
label	int8	1 = binder, 0 = not a binder of this target
binding_rank	int32	1 = most target-specific; null for rows with no target signal

What the labels mean

A positive is a molecule that showed more signal against the target than against a no-protein control run on the same library. Positives are ordered by that margin in binding_rank, so rank 1 is the most target-specific molecule rather than simply the most abundant one.

A negative is not an inert compound. Negatives are molecules that enriched in the absence of the target protein, meaning they bind something nonspecifically but not this target. They are hard negatives, and a model trained on them is learning selectivity rather than general stickiness. Do not describe them as inactive.

Negatives come from two distinct places, and you can tell them apart. binding_rank is populated for every molecule that came off the target's own screen and null for everything else, which separates them without an extra column:

• Tested losers — label = 0 with a rank. These appeared against the target and still enriched harder on the control. They are the hardest negatives here.
• Untested decoys — label = 0 with a null rank. Drawn from the control, these never registered against the target at all, so they carry no target-side measurement.

If you are building a benchmark you will probably want to weight those two groups differently, or hold one of them out entirely.

Nothing was deleted to express a judgement: molecules that appeared against the target and failed the comparison are kept as negatives rather than discarded.

Intended use and limits

These are DEL screening labels, not measured affinities. There is no Kd, no IC50, and no dose response. A positive means enrichment consistent with binding under the screen; it is not a validated hit.

No holdout split is provided. Splitting is left to you deliberately, so that no published split becomes an answer key for Om's Open Pipeline Challenges, where the same targets are validated experimentally by affinity-selection mass spectrometry.

Two further caveats. Read counts are normalized for sequencing depth but not for input library representation, so abundance is ordinal rather than quantitative. And DEL chemical space is not the same as purchasable chemical space: these molecules train a model to extrapolate onto compounds that can actually be bought, which is what the challenges require.

Provenance

Single screening vintage, v20260608, one internally consistent experiment per target. Structures were identity-checked: each configuration's sequence hashes to the protein UUID that keys it in Om's catalogue.

Generated and quality-checked with a fixed random seed. Every build asserts that positives and negatives are disjoint, that ranks are contiguous, that labels agree with the ranking, and that every SMILES parses under RDKit.

Citation

If you use this dataset, please cite Om and link back to this repository.