49 of 91 proteins in this graph have never had a drug-like molecule bound to them
also called ligandability census · structural tractability dossier
For every protein in the graph we asked two questions that need no new experiment: how much of it has anyone ever seen at atomic resolution, and has anyone ever made a small molecule that sticks to it and solved the picture. Slightly more than half have never had such a molecule, and the ones that have are mostly not the ones the fusion keeps.
The same thing, for a scientist
Across 91 UniProt accessions, distinct non-polymer chemical components >300 Da co-crystallised with each protein were counted from the RCSB search API after removing an audited blocklist of cofactors, lipids, detergents, cryoprotectants, buffers and sugars. The median is zero. 49/91 have no drug-like ligand at all, and 12 of those are covered end-to-end by deposited constructs — structure is not the limiting factor, chemical matter is. MEN1 carries 45 distinct ligands; among proteins retained in KMT2A fusions it is effectively the only one with a real medicinal-chemistry campaign behind it.
What was measured, and what the numbers do not mean
Two quantities per protein, both recoverable from public structural data:
Construct coverage. The union of UniProt spans over every PDB entry containing the accession (PDBe SIFTS best_structures). Read this as the fraction of the chain that appears in at least one deposited construct — not the fraction resolved in density. It systematically overstates how much is actually ordered, because a construct that nominally spans residues 1–628 may have half of them invisible. Treated as an upper bound throughout.
Demonstrated ligandability. The count of distinct non-polymer chemical components over 300 Da ever co-crystallised with the protein. This is deliberately not a pocket prediction: it is a count of compounds somebody actually got to bind and then solved.
The filter matters more than the query. A naive count over 300 Da gives KMT2A 14 and looks authoritative. Inspected: DA/DC/DG/DT are DNA from the CXXC–DNA complex, SAH is the methyl-donor product, 12P and PE8 are polyethylene glycol from the cryoprotectant, CHD is a detergent. The real count is 4 — the MLL1 SET-domain inhibitor series. So the query filters server-side to chem_comp.type = non-polymer (removing nucleotides, amino acids and saccharides) and then against an explicit blocklist. Every excluded identifier is recorded per protein in the output, so the filter is auditable rather than trusted.
The finding: being seen and being druggable are different problems
The median protein in this graph has zero drug-like ligands. 49 of 91 have never had one. What makes that interesting is which ones.
Twelve proteins are covered end-to-end by deposited constructs and still have zero ligands. They are, almost exactly, the scaffold of the complexes this project cares about:
| protein | construct coverage | PDB entries | drug-like ligands |
|---|---|---|---|
| ASH2L | 100.0% | 32 | 0 |
| RBBP5 | 100.0% | 32 | 0 |
| DPY30 | 100.0% | 12 | 0 |
| CDC73 | 100.0% | 21 | 0 |
| CTR9 | 100.0% | 21 | 0 |
| LEO1 | 100.0% | 20 | 0 |
| PAF1 | 100.0% | 21 | 0 |
| RTF1 | 100.0% | 18 | 0 |
| SKIC8 | 100.0% | 28 | 0 |
| MED26 | 100.0% | 8 | 0 |
| ELL2 | 100.0% | 4 | 0 |
RBBP5 and ASH2L have been solved 32 times each. SKIC8 28 times. Nobody has ever put a molecule on any of them. Structure is not the bottleneck for this class of target; chemical matter is, and no amount of further structural work will change that on its own. This is the quantitative form of an objection the graph already records qualitatively about scaffold and complex-subunit targets.
Menin is the only retained handle with real chemistry behind it
| protein | retained in the fusion? | construct coverage | drug-like ligands |
|---|---|---|---|
| MEN1 | yes | 100.0% | 45 |
| PSIP1 (LEDGF) | yes | 100.0% | 4 |
| WDR5 | no — lost | 100.0% | 63 |
| DOT1L | not in the fusion | 37.3% | 23 |
| BRD4 | not in the complex | 54.2% | 421 |
| AFF1 (the 3′ partner) | it IS the fusion | 3.5% | 0 |
Two things fall out, and they point in opposite directions.
For the strategy. Among the fragment the fusion actually retains, MEN1 is essentially the only protein with a prosecuted medicinal-chemistry campaign — 45 distinct compounds, and the chain fully covered. This analysis reaches that from structural and chemical data alone, with no reference to the biology, which makes it independent corroboration of the target choice rather than a restatement of it.
Against it. The most ligandable proteins here are the ones the fusion throws away. WDR5 has 63 distinct ligands, more than menin, and is lost at the breakpoint. BRD4 has 421 and is not in the complex at all. And AFF1, the partner that actually drives the disease, is 3.5% covered with zero ligands — the business end of the oncoprotein is both structurally dark and chemically untouched.
The downstream programme is not druggable by this measure either
HOXA9 (0.0% coverage, 0 ligands, 0 PDB entries), MEF2C (0.0%, 0), MEIS1 (14.9%, 0). Twelve proteins in the graph have no PDB entry whatsoever.
These are transcription factors, so the result is unsurprising — but it is worth having as a number rather than an intuition, because it sets the realistic options for the downstream programme to degradation, expression-level control or interface disruption, not orthosteric inhibition.
Why a ligand count is a biased instrument
It measures attention as much as tractability. BRD4's 421 compounds reflect a decade of bromodomain fashion. WDR5's reflect one group prosecuting the WIN site hard. A protein nobody has tried is indistinguishable here from a protein many people have tried and failed on — and those are very different prospects. Nothing in this analysis distinguishes them, which is its single biggest weakness.
A co-crystallised compound is not a drug. No potency, selectivity, cell activity or developability is captured. Fragment screens deposit weak binders; those count here identically to a clinical candidate.
Cofactor-site binders flatter enzymes. DOT1L's count survives partly because SAM-competitive chemistry is easy to crystallise. The blocklist removes SAM and SAH themselves but not close analogues designed to occupy that site.
| script | analyses/structural-tractability-ligandability/run.py |
|---|---|
| status | executed |
| run date | 2026-09-19 |
| n accessions | 91 |
| min ligand mw | 300 |
| blocklist size | 144 |
| n never liganded | 49 |
| n fully covered but zero ligands | 12 |
| n no PDB entry at all | 12 |
| median druglike ligands | 0 |
| headline | Median drug-like ligand count across the graph is zero. 49/91 proteins have never had one. Twelve are covered end-to-end by deposited constructs and still have none — the entire WRAD/PAF1C/SEC scaffold set. MEN1 (45) is the only retained handle with real chemistry. |
| coverage metric warning | pct_solved is the union of SIFTS construct spans and is an UPPER BOUND on resolved residues, not a measure of them. |
| caveat | A co-crystallised compound is not a drug and a count is not potency. Ligand counts track how hard a protein has been prosecuted as much as how tractable it is. Cofactor-site binders that survive the blocklist flatter enzymes over scaffolds. Structural coverage counts residues present in a deposited model, which is not the same as residues that are ordered and well resolved. |
ligandability — the numbers behind the chart
{
"MEN1": {
"pct_construct_covered": 100,
"pdb_entries": 69,
"druglike_ligands": 45
},
"WDR5": {
"pct_construct_covered": 100,
"pdb_entries": 198,
"druglike_ligands": 63
},
"BRD4": {
"pct_construct_covered": 54.2,
"pdb_entries": 636,
"druglike_ligands": 421
},
"DOT1L": {
"pct_construct_covered": 37.3,
"pdb_entries": 48,
"druglike_ligands": 23
},
"BCL2": {
"pct_construct_covered": 91.2,
"pdb_entries": 63,
"druglike_ligands": 31
},
"FLT3": {
"pct_construct_covered": 94.3,
"pdb_entries": 11,
"druglike_ligands": 5
},
"PSIP1": {
"pct_construct_covered": 100,
"pdb_entries": 40,
"druglike_ligands": 4
},
"KAT7": {
"pct_construct_covered": 45.2,
"pdb_entries": 8,
"druglike_ligands": 5
},
"AFF1": {
"pct_construct_covered": 3.5,
"pdb_entries": 1,
"druglike_ligands": 0
},
"MEIS1": {
"pct_construct_covered": 14.9,
"pdb_entries": 2,
"druglike_ligands": 0
},
"HOXA9": {
"pct_construct_covered": 0,
"pdb_entries": 0,
"druglike_ligands": 0
},
"MEF2C": {
"pct_construct_covered": 0,
"pdb_entries": 0,
"druglike_ligands": 0
},
"ASH2L": {
"pct_construct_covered": 100,
"pdb_entries": 32,
"druglike_ligands": 0
},
"RBBP5": {
"pct_construct_covered": 100,
"pdb_entries": 32,
"druglike_ligands": 0
},
"DPY30": {
"pct_construct_covered": 100,
"pdb_entries": 12,
"druglike_ligands": 0
},
"CDC73": {
"pct_construct_covered": 100,
"pdb_entries": 21,
"druglike_ligands": 0
},
"CTR9": {
"pct_construct_covered": 100,
"pdb_entries": 21,
"druglike_ligands": 0
},
"LEO1": {
"pct_construct_covered": 100,
"pdb_entries": 20,
"druglike_ligands": 0
},
"PAF1": {
"pct_construct_covered": 100,
"pdb_entries": 21,
"druglike_ligands": 0
},
"RTF1": {
"pct_construct_covered": 100,
"pdb_entries": 18,
"druglike_ligands": 0
},
"SKIC8": {
"pct_construct_covered": 100,
"pdb_entries": 28,
"druglike_ligands": 0
},
"MED26": {
"pct_construct_covered": 100,
"pdb_entries": 8,
"druglike_ligands": 0
},
"ELL2": {
"pct_construct_covered": 100,
"pdb_entries": 4,
"druglike_ligands": 0
}
}- Using a ligand count to rank targets rewards fashion. BRD4 tops this table with 421 compounds and is not part of the complex; the protein everyone has already worked on will always look the most tractable, which is exactly how a field stays where it is.
- The comfortable reading — menin is the only retained handle with chemistry, therefore menin is the right target — is also the reading that would follow if menin were simply the only thing anyone had tried on the retained fragment. This analysis cannot tell those apart.
- Declaring the scaffold proteins chemically intractable risks foreclosing the one class of target that a proximity-based modality is best suited to. The number says no orthosteric ligand exists, not that no molecule can act there.
- For the twelve fully-covered, never-liganded scaffold proteins, is the absence of ligands a record of failure or of nobody having tried? A patent and screening-deck search would separate those, and they imply opposite decisions.
- Does a degrader or molecular-glue modality change the picture for ASH2L, RBBP5 and DPY30, none of which would leave a co-structure against the target alone?
- AFF1 is 3.5% covered with no chemical matter. Is any part of it foldable enough to be a structure-based target, or is the only route to it through its partners?
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
Across 91 UniProt accessions in this graph, the median number of distinct drug-like (>300 Da, non-polymer, non-cofactor) co-crystallised ligands is zero; 49 of 91 have none. Twelve proteins are covered end-to-end by deposited constructs and still have zero ligands, including ASH2L, RBBP5, DPY30, CDC73, CTR9, LEO1, PAF1, RTF1, SKIC8, MED26 and ELL2. MEN1 has 45 distinct ligands with full construct coverage; WDR5 has 63; AFF1 has zero at 3.5% coverage.
RCSB PDB search API, chemical-component query over UniProt-mapped entries · RCSB Protein Data Bank 2026What this evidence cannot show
Run by the coordinator on 2026-09-19; script, full per-protein output and the complete list of excluded chemical components are in analyses/structural-tractability-ligandability/. Construct coverage is the union of SIFTS unp spans and therefore an UPPER BOUND on what is resolved in density, not a measure of it. Ligand counts measure how hard a protein has been prosecuted as much as how tractable it is, and cannot distinguish 'never attempted' from 'attempted and failed'. A co-crystallised compound is not a drug: no potency, selectivity or cell activity is captured, and weak fragment hits count the same as clinical candidates.
Evidence that argues against this (1)
- Zero co-crystallised ligands is not proof that a protein is undruggable, and the scaffold list is where that matters most. Absence of deposited ligand structures can reflect absence of effort, failure to crystallise a complex rather than failure to bind, or a modality that leaves no co-structure at all — molecular glues and degraders act through induced interfaces that are rarely deposited against the target alone. WDR5 is the cautionary case in the other direction: it has more ligands than menin and is lost from the fusion, and the WIN-site inhibitors built on that chemistry turn out to act in KMT2A-rearranged cells through ribosome protein gene suppression and p53 rather than through MLL1 biology. A large ligand count is therefore not evidence that the right thing is being drugged. Howard GC et al. 2024
Selectivity bound (Seeliger 2007, PMID 17355866): c-Src binds imatinib in the same inactive Abl/Kit-like conformation, so imatinib's specificity is a distributed free-energy penalty between inactive states, not a unique ABL1 conformation.
- ← related toCBFB::MYH11 (CBFbeta-SMMHC), inv(16)(p13;q22)
Absence bound (Blankenfeldt 2006, PMID 17095604): MYH11 (P35749) maps to exactly one PDB entry, 9FU2, but human cardiac beta-myosin II S2 coiled-coil structures exist under other accessions, so the retained SMMHC rod is unsolved yet modellable.
- ← related toCBFA2T3::GLIS2, inv(16)(p13.3q24.3)
Homology bound (Pavletich 1993, PMID 8378770): the five-finger GLI1-DNA crystal structure at 2.6 A is a real template for the GLIS2 zinc-finger module, so the structural blindness here is protein-specific, not fold-specific.
- ← related toSix countermeasures, no head-to-head — so rank their targets: CDK6 is 13th and CDK4 is 16,052nd
The other cheap target-triage axis: this asks whether a target is selectively required, that one whether anyone has ever made a molecule that binds it.
MEN1 carries 45 distinct drug-like co-crystallised ligands with full construct coverage - among proteins retained in the fusion it is effectively the only one with a prosecuted medicinal-chemistry campaign.
- refutes →WDR5 — the partner that is not a partner: it binds the half of KMT2A the fusion throws away
WDR5 has 63 distinct ligands, MORE than menin, and is lost from the fusion at the breakpoint - ligandability and relevance come apart, so a ligand count cannot be used to choose a target on its own.
- related to →ASH2L — the WRAD subunit that has DPY30's essentiality profile and none of its co-dependency signal
Covered end-to-end by deposited constructs across 32 PDB entries and never liganded. Structure is not the bottleneck for this target; chemical matter is.
32 PDB entries, full construct coverage, zero drug-like ligands ever co-crystallised.
- related to →DPY30 — the smallest COMPASS subunit, and the only one that behaves like a selective dependency
Full construct coverage across 12 entries and zero ligands - independent of the DepMap designs on which DPY30 already fails.
- related to →KMT2A / MLL1 — the gene that breaks
The naive ligand count for KMT2A is 14 and is wrong: DNA from the CXXC complex, the SAH cofactor, PEG and a detergent. The real count is 4, the MLL1 SET-domain inhibitor series.
AFF1, the partner that drives the disease, is 3.5% construct-covered with zero drug-like ligands - the business end of the oncoprotein is both structurally dark and chemically untouched.
23 drug-like ligands at 37% coverage, but the count is flattered by SAM-competitive chemistry being easy to crystallise.
- related to →HOXA9 — required for survival of established MLL-rearranged leukemia, and dispensable for starting it
Zero PDB entries and zero ligands - the downstream programme is not addressable by orthosteric inhibition, which restricts realistic options to degradation or expression-level control.
- supports →Dataset: PDB and AlphaFold - the structural census, including the two absences that shape every plan
Supplies the quantitative dossier that entry describes: per-protein construct coverage and demonstrated ligandability for all 91 accessions in the graph.
Adds a chemistry column to the lost-module ledger: the most ligandable proteins in the graph are largely the ones the fusion discards.
Complicates it. Menin IS the only retained handle with real chemistry (45 ligands), which the geometry argument predicts - but this analysis cannot distinguish 'the only one that is tractable' from 'the only one anyone has tried', so it corroborates the conclusion without confirming the mechanism.
Supplies the tractability side of that generalisation: across 91 proteins the median drug-like ligand count is zero, so an obligate interface being real is not evidence it can be drugged.
- ← related toAQR (Aquarius) - a spliceosomal RNA helicase whose loss impairs MLL-rearranged leukaemia stem cells
A new dependency whose tractability is untested: AQR is core spliceosomal machinery, the class where the therapeutic window usually fails.
- ← related toSix of 85 proteins are both selectively required and drugged — and MEN1 is not close to the others
One of the two axes joined here.