DREAM x CACHE Target 2035 Drug Discovery Challenge — A DNA-Encoded Library Machine-Learning Challenge Targeting PGK2 Organizers: Structural Genomics Consortium (SGC), DREAM Challenges (IBM), Sage Bionetworks, Conscience, BEACON, MAINFRAME, and the labs of Damian Young (Baylor College of Medicine) and Neelagandan Kamariah (BRIC-inStem) · Open science (no patents) · Public access Challenge page: cache-challenge.org — A DEL-ML Challenge Targeting PGK2 · Data & submissions via the Synapse platform.
TL;DR This is the largest open DNA-encoded library (DEL) machine-learning benchmark released to date: a DEL affinity-selection screen of close to a billion (~890 million) compounds against phosphoglycerate kinase 2 (PGK2) — an enzyme essential for sperm motility and a target for non-hormonal contraception. The task is to train models on the DEL screen and retrospectively retrieve confirmed PGK2 binders from a hidden affinity-selection mass-spectrometry (ASMS) hit set, then prospectively nominate novel compounds from Enamine's catalog for experimental testing. This dataset's
input/folder holds the enumerated OpenDEL structures, the PGK2 selection read-count table, a curated ML training table, and the retrospective validation/test compound splits.
0. Summary
Biological premise. Phosphoglycerate kinase 2 (PGK2) is a testis-specific glycolytic enzyme essential for sperm motility and is a promising target for non-hormonal male contraception. It is the same target as CACHE Challenge #7. Finding chemical starting points (ligands/binders) for PGK2 is hard: it lacks a classical deep druggable pocket, so this challenge tests whether modern machine learning (ML) can extract binders from noisy, large-scale screening data.
DNA-encoded libraries (DELs). A DNA-encoded library is a mixture of many small molecules, each covalently tagged with a unique DNA barcode that records the building blocks used to synthesize it. Incubating the library with an immobilized protein target and sequencing the DNA of the retained molecules yields a read count per compound that acts as a noisy proxy for binding/enrichment. As is typical for DEL data, the signal-to-noise ratio is low and read counts must be interpreted accordingly.
The challenge task. Participants (1) train ML models on the DEL screen (SMILES structures, read counts, and building-block compositions of ~890 million compounds); (2) retrospectively predict true PGK2 binders that were independently confirmed by an affinity-selection mass-spectrometry (ASMS) screen of ~400,000 compounds followed by a kinase assay — an out-of-distribution retrieval task with a small number of hidden true positives; and (3) top teams then prospectively predict novel hits from Enamine's catalog, which are procured and tested experimentally. The retrospective phase deadline is September 15, 2026.
Assay and processing. DNA-encoded library selection (read counts for PGK2, a PGK2-with-inhibitor competition arm, and a no-target control) plus affinity-selection mass spectrometry (ASMS) for orthogonal hit confirmation. The curated training table additionally ships nine precomputed molecular-fingerprint feature blocks per molecule — Extended-Connectivity Fingerprints (ECFP4, ECFP6), Functional-Class Fingerprints (FCFP4, FCFP6), MACCS keys, RDKit fingerprint (RDK), Avalon, Atom-Pair (AtomPair), and Topological-Torsion (TopTor) — alongside molecular weight (MW) and Wildman–Crippen logP (ALOGP).
Sources. All files were obtained from the challenge's Synapse distribution for the DREAM x CACHE Target 2035 DEL-ML Challenge. This is a community challenge / benchmark dataset, not a single peer-reviewed paper; the inaugural Target 2035 DEL-ML challenge overview manuscript has been submitted (preprint pending). No analysis code is bundled here.
| Domain | Small-molecule drug discovery / DNA-encoded library (DEL) machine learning |
| Target protein | Phosphoglycerate kinase 2 (PGK2), Homo sapiens |
| Assay | DNA-encoded library (DEL) affinity selection + affinity-selection mass spectrometry (ASMS) confirmation |
| Molecular representation | SMILES strings; nine fingerprint blocks (ECFP4, ECFP6, FCFP4, FCFP6, MACCS, RDK, Avalon, AtomPair, TopTor); MW; ALOGP |
| Organisms | Homo sapiens (target PGK2) |
| Repository | Synapse (Sage Bionetworks) |
| Accession | DREAM x CACHE Target 2035 Challenge (Synapse) |
| Access | Public (open science) |
| Number of studies | 1 (challenge with multiple provided files) |
| Number of files | 5 |
| Total size | 6,264,772,911 bytes (≈ 6.3 GB) |
| Distinct data units | OpenDEL ≈ 890 million enumerated compounds; PGK2 selection table = 7,703,070 rows; training table = 339,258 molecules; validation split = 244,328 compounds; test split = 184,632 compounds; challenge test file = 233 compounds |
| Instrument / method | DEL selection sequencing + affinity-selection mass spectrometry (ASMS) |
| Code license | Not bundled (open-science challenge; no patents) |
A note on the training table's target. In
DREAM_Challenge_1_TrainSet.parquet, theTARGET_IDcolumn is uniformlyWDR91(339,258 rows; 138 positive labels), i.e. this shipped table is the worked example / template carried over from the inaugural Target 2035 challenge showing the expected feature layout (SMILES + fingerprints + LABEL), not the PGK2 labels themselves. The PGK2 experimental signal lives inPGK2_selection.parquet(read counts / z-scores) and the hidden ASMS hit set. Treat the training table as the feature-schema exemplar and derive PGK2 training labels from the selection table.
1. Datasets — files in this dataset's input/ folder
Landed sources
| Source | Access | Files |
|---|---|---|
| Synapse — DREAM x CACHE Target 2035 Challenge | Public | 5 |
The provided files
| File | What it is | Rows / scale | Size |
|---|---|---|---|
OpenDEL_libraries.zip | Enumerated OpenDEL structures: 21 qDOS*.enumeration.parquet files (full SMILES of the ~890 M-compound library) plus per-library building-block SMILES CSVs under BBids_SMILES/ | ≈ 890 million compounds | 5,435,654,144 bytes (≈ 5.1 GB) |
PGK2_selection.parquet | PGK2 DEL affinity-selection results: per-compound read counts and z-scores for PGK2, PGK2-with-inhibitor (competition), and no-target control (NTC), plus a historic_hits flag | 7,703,070 rows | 193,260,896 bytes (≈ 184 MB) |
DREAM_Challenge_1_TrainSet.parquet | Curated ML training table (feature-schema exemplar): SMILES, RandomID, TARGET_ID (WDR91), binary LABEL, nine fingerprint list-columns (ECFP4/6, FCFP4/6, MACCS, RDK, Avalon, AtomPair, TopTor), MW, ALOGP | 339,258 molecules (138 positives) | 628,270,836 bytes (≈ 599 MB) |
Val-Test-set.zip | Retrospective evaluation SMILES splits: PGK2_Validation_split.csv (244,328 compounds) and PGK2_Test_split.csv (184,632 compounds); columns CatalogID,SMILES | 428,960 compounds total | 7,573,328 bytes (≈ 7.2 MB) |
DREAM_Target2035_Challenge_test_data.csv | Small challenge test manifest; columns Ccompound_ID,SMILES | 233 compounds | 13,707 bytes (≈ 13 KB) |
1.1 Column reference
PGK2_selection.parquet—compound(library-barcode ID such asqDOS11-33-191-22),SMILES,count_PGK2,count_PGK2_with_inhibitor,count_NTC(integer read counts),zscore_PGK2,zscore_PGK2_with_inhibitor,zscore_NTC(enrichment z-scores),historic_hits(integer flag). The inhibitor-competition and no-target-control (NTC) arms let you distinguish target-specific enrichment from non-specific DNA/matrix binding.DREAM_Challenge_1_TrainSet.parquet—SMILES,RandomID,TARGET_ID,LABEL(0/1), plus nine fingerprint columns stored as integer lists and two scalar descriptors (MW,ALOGP). Fingerprints are precomputed so models can train without an RDKit featurization step.Val-Test-set.zipsplits —CatalogID(Enamine catalog ID, e.g.Z317469132) andSMILES. These are the compounds over which retrospective PGK2-binder predictions are scored.OpenDEL_libraries.zip— eachqDOS<n>.enumeration.parquetenumerates the full molecular structures of one DEL sub-library;BBids_SMILES/qDOS<n>.csvmaps building-block IDs to SMILES so tri-synthon compounds can be decomposed into their building blocks.
1.2 How this became a dataset
Files were downloaded from the challenge's Synapse distribution and copied verbatim (original filenames, byte-for-byte) into this dataset's input/ folder. Archive integrity was verified with unzip -t (both zips OK) and both Parquet files re-opened and row-counted after the copy. No files were renamed, unpacked, or modified; output/ is intentionally empty and reserved for downstream processing.
1.3 Completeness
All five distributed files are present and size-verified against the source (6,264,772,911 bytes total). The hidden ASMS true-positive hit set used for final scoring is deliberately not distributed (that is the prediction target held by the organizers), and no analysis code accompanies the release. The retrospective phase is scored on Synapse against the withheld labels.
2. Analyses this dataset supports
- DEL-to-binder ML models. Train binary or ranking models on
PGK2_selection.parquet(usingcount_*/zscore_*and the PGK2-vs-inhibitor-vs-NTC contrast to define enrichment labels) with SMILES/fingerprint features, then score theVal-Test-set.zipsplits — the core retrospective challenge task. - Fingerprint / representation benchmarking. Use the nine precomputed fingerprint blocks plus
MW/ALOGPinDREAM_Challenge_1_TrainSet.parquetto compare classical descriptors against learned representations (graph neural networks, chemical foundation models) on the same feature schema. - Out-of-distribution generalization. The training DEL chemistry (tri-synthon
qDOSlibraries) differs structurally from the Enamine catalog validation/test compounds — a realistic domain-shift study for hit-finding. - Building-block SAR. Decompose enriched compounds into building blocks via
OpenDEL_libraries.zip/BBids_SMILES/*.csvto study per-synthon structure–activity relationships and enrichment additivity. - Denoising DEL read counts. Model non-specific binding using the NTC and inhibitor-competition arms to separate true PGK2 affinity from matrix/DNA artifacts.
- Enter the challenge: register and submit retrospective predictions via Synapse; see the challenge page for scoring and the prospective phase.
Citation
DREAM x CACHE Target 2035 Drug Discovery Challenge — A DEL-ML Challenge Targeting PGK2. Structural Genomics Consortium (SGC), DREAM Challenges, Sage Bionetworks, Conscience, and collaborators, 2026. Challenge page: https://cache-challenge.org/challenges/a-del-ml-challenge-targeting-pgk2 · Data and submissions: https://www.synapse.org. (Inaugural Target 2035 DEL-ML challenge overview manuscript submitted; preprint pending.)
About this dataset
This DNA-encoded library (DEL) affinity-selection screen + ML feature tables (SMILES, read counts, molecular fingerprints) dataset provides single-cell gene expression from homo sapiens samples.
Analyze it in Drylab
Open this dataset in Drylab and ask a question in plain language. The AI agent loads the files, runs the analysis in a ready sandbox and explains each result. A typical first pass:
- Load the filtered feature-barcode matrix (HDF5) and run quality control on counts, genes and mitochondrial reads per cell.
- Normalize, select variable genes, then cluster the cells and draw a UMAP.
- Annotate clusters with marker genes and compare the cell-type mix between conditions.