Dataset: ChEMBL chemical matter - open, quantitative, and easy to misread as a druggability score
also called ChEMBL access plan · chemical matter census · pChEMBL pull
ChEMBL is a free catalogue of every published measurement of a drug-like molecule binding a protein. It tells us instantly whether anyone has ever made chemistry against a target. The catch is that a big number can mean the chemistry was aimed at a different part of the protein entirely.
The same thing, for a scientist
ChEMBL supplies the chemical-matter axis of every dossier in this beat, and the census was run this cycle for all candidates. The counts, at pChEMBL 6 or above, span three orders of magnitude and they order the portfolio: menin 499 (best Ki 0.104 nM), menin/MLL PPI target entity 1,014, KMT2A 577, DOT1L 193, ENL 144, KAT7 98, CBX8 15, LEDGF/PSIP1 3 (best IC50 470 nM). The resource is open, key-free and trivially automatable. Its danger is interpretive rather than technical, and this beat found a live example: KMT2A's 577 potent rows would rate it a well-drugged protein, but none of them engage the CXXC domain that a phase-2 CXXC programme would target - they belong to the WDR5-MLL and menin-MLL partner programmes. Any ML model or heuristic that uses target-level ChEMBL counts as a druggability feature will inherit this error systematically.
What question it answers
Has anyone made a potent molecule against this protein? How potent, in what assay, and is there a curated mechanism-of-action drug? For a phase-2 plan this determines whether a programme starts from an existing series, from a probe, or from nothing.
And, read carefully, it answers a second question: has anyone tried and failed? Three rows against LEDGF after a decade of published interest is either open ground or a graveyard, and ChEMBL cannot tell you which - but it can tell you the question exists.
How to reach it
Open REST, no key 1:
- Resolve target:
https://www.ebi.ac.uk/chembl/api/data/target/search?q=menin&format=json - Count potent rows:
https://www.ebi.ac.uk/chembl/api/data/activity.json?target_chembl_id=CHEMBL1615381&pchembl_value__gte=6&limit=1and readpage_meta.total_count - Best affinity: add
&order_by=-pchembl_value
The pchembl_value field is ChEMBL's harmonised -log10 potency across IC50, Ki, Kd and EC50, which is what makes cross-target comparison possible at all.
The census, and what it means for the portfolio
| Target | ChEMBL id | Rows at pChEMBL >=6 | Best reported |
|---|---|---|---|
| Menin/MLL PPI | CHEMBL2093861 | 1,014 | - |
| KMT2A | CHEMBL1293299 | 577 | - |
| Menin | CHEMBL1615381 | 499 | Ki 0.104 nM |
| DOT1L | CHEMBL1795117 | 193 | Kd 0.01 nM |
| ENL | CHEMBL4295801 | 144 | Ki 1.43 nM |
| KAT7 | CHEMBL3774299 | 98 | Ki 8 nM |
| CBX8 | CHEMBL3232684 | 15 | Kd 120 nM |
| PSIP1/LEDGF | CHEMBL3988590 | 3 | IC50 470 nM |
The portfolio reads off this table directly: menin and DOT1L are solved chemistry problems whose questions are elsewhere; ENL and KAT7 have real starting points; CBX8 and LEDGF have essentially nothing, and those two sit at opposite ends of the dependency ranking, which is why one is a no-go and the other is the recommended first programme.
The access barrier, and the interpretive trap
Access barrier: none. Open, key-free, safe to automate, CC BY-SA licensed.
The trap, stated so a model cannot inherit it: a ChEMBL target-level count is a count against a protein, not against a site. KMT2A's 577 rows engage WDR5 and menin interfaces, not the CXXC domain. Before any count is used as evidence for a specific pocket, the rows must be audited for which assay they came from. That audit costs a day and is a G1 item in the CXXC dossier.
A second, milder trap: ChEMBL under-represents unpublished industry chemistry, so a low count is evidence of a gap in the public record only.
| resource | ChEMBL bioactivity database |
|---|---|
| question answered | Does potent, published chemical matter exist against this target, and how potent? |
| access barrier | None. The barrier is interpretive: target-level counts are not site-level counts. |
| mandatory audit rule | Before a ChEMBL count is used as evidence for a specific pocket, audit the rows for assay target site. A day of work; a G1 item wherever a count carries argumentative weight. |
| status | run |
| result summary | Census completed for all nine dossiers. The chemical-matter axis orders the portfolio cleanly and identifies the two extremes: LEDGF has the best dependency profile and almost no chemistry (the opportunity), CBX8 has almost no chemistry and no dependency (the no-go). |
| compute | laptop; no GPU |
access
{
"auth": "none",
"endpoint": "https://www.ebi.ac.uk/chembl/api/data",
"key_field": "pchembl_value (harmonised -log10 potency across IC50/Ki/Kd/EC50)",
"licence": "CC BY-SA"
}census this cycle — the numbers behind the chart
[
{
"target": "menin/MLL PPI",
"chembl_id": "CHEMBL2093861",
"rows_pchembl_ge_6": 1014
},
{
"target": "KMT2A",
"chembl_id": "CHEMBL1293299",
"rows_pchembl_ge_6": 577,
"warning": "none of these engage the CXXC domain"
},
{
"target": "menin",
"chembl_id": "CHEMBL1615381",
"rows_pchembl_ge_6": 499,
"best": "Ki 0.104 nM (CHEMBL5267644)"
},
{
"target": "DOT1L",
"chembl_id": "CHEMBL1795117",
"rows_pchembl_ge_6": 193,
"best": "Kd 0.01 nM (CHEMBL5764114)"
},
{
"target": "ENL",
"chembl_id": "CHEMBL4295801",
"rows_pchembl_ge_6": 144,
"best": "Ki 1.43 nM (CHEMBL4210431)"
},
{
"target": "KAT7",
"chembl_id": "CHEMBL3774299",
"rows_pchembl_ge_6": 98,
"best": "Ki 8 nM (CHEMBL5928174)"
},
{
"target": "CBX8",
"chembl_id": "CHEMBL3232684",
"rows_pchembl_ge_6": 15,
"best": "Kd 120 nM (CHEMBL4635096)"
},
{
"target": "PSIP1/LEDGF",
"chembl_id": "CHEMBL3988590",
"rows_pchembl_ge_6": 3,
"best": "IC50 470 nM (CHEMBL3981070)",
"note": "all antiviral LEDGIN chemistry, no oncology series"
}
]- Any ML druggability model that uses ChEMBL target-level counts as a feature will systematically rate KMT2A as tractable and LEDGF as intractable, which is exactly backwards for this project's purposes.
- A low ChEMBL count is evidence about the public record, not about the molecule space. Concluding 'nobody tried' from three rows may mean several groups tried and published nothing.
- Do any of KMT2A's 577 potent rows actually assay the CXXC domain, and does an audit change the CXXC feasibility verdict?
- Is the LEDGF gap an untried opportunity or an unpublished graveyard, and can that be distinguished by looking at patent literature rather than ChEMBL?
- Should a site-level rather than target-level chemical-matter feature be constructed for this project, and what would it cost to build?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
The ChEMBL REST API is open and key-free, and a census run this cycle returned the following counts of activity rows at pChEMBL 6 or above: menin (CHEMBL1615381) 499 with best Ki 0.104 nM; DOT1L (CHEMBL1795117) 193 with best Kd 0.01 nM; ENL (CHEMBL4295801) 144 with best Ki 1.43 nM; KAT7 (CHEMBL3774299) 98 with best Ki 8 nM; CBX8 (CHEMBL3232684) 15 with best Kd 120 nM; PSIP1 (CHEMBL3988590) 3 with best IC50 470 nM; KMT2A (CHEMBL1293299) 577; menin/MLL PPI (CHEMBL2093861) 1,014.
ChEMBL REST API, target search and activity endpoints with pchembl_value >= 6 · EMBL-EBI ChEMBL 2026What this evidence cannot show
All counts fetched directly this cycle. Counts mix biochemical, cellular and functional assays and include patent-derived series; a best single affinity value is an extreme of a distribution and is not representative. ChEMBL v34 or later; the exact version was not recorded from the API response.
Evidence that argues against this (1)
- The KMT2A row count is the counter-example to its own use: 577 potent rows against a protein whose CXXC domain has no known chemical series at all. A count against a protein is not a count against a pocket. source 2026
3 potent rows against LEDGF - the chemistry gap that is the opportunity
KMT2A's 577 potent rows engage partner interfaces, not the CXXC domain - the target-versus-site trap
chemical-matter census orders the portfolio the gates are written against