Oncofusion

start / graph / ana-comparative-interface-druggability-scan

Analysismoderate confidence 3 of 5: One strong study, or several smaller consistent ones.Q11-comparative

Analysis (planned) — comparative druggability scan of every obligate fusion-partner interface in the atlas

also called interface triage scan · cross-family druggability ranking

Before doing any laboratory work, we can rank every known fusion-partner contact point by how drug-friendly it looks and how safely the partner could be blocked. The point is to find the next menin on a computer, cheaply, rather than by luck.

The same thing, for a scientist

A reproducible computational triage that takes every fusion-partner interaction catalogued in this graph, and for each scores: (1) structural evidence and buried surface area from PDB or AlphaFold-Multimer models; (2) pocket or hot-spot character at the interface using standard cavity-detection and fragment hot-spot mapping; (3) fusion-dependency versus normal essentiality using DepMap Chronos gene effect stratified by fusion genotype, GTEx/HPA normal-tissue expression and gnomAD LOEUF constraint; (4) existing chemical matter from ChEMBL. Output is a ranked table with an explicit dependency-to-essentiality ratio, plus a flag for cross-genotype sharing. It is designed to reproduce the menin result as a positive control: if the pipeline does not put menin-KMT2A near the top, it is wrong.

Go deeper
Design and positive controls1 deeper

Inputs: the partner list from the atlas (menin/MEN1, LEDGF/PSIP1, DOT1L, AFF1/AFF4, ELL, CDK9, BRD4, WDR5, KAT7, CBX8, PAF1c, RUNX1, CBFB, NCoR/SMRT, BAF/SMARCA4, NSD1, KDM5A), each paired with the fusion that recruits it.

Positive controls the pipeline must recover: (i) menin-KMT2A, which yielded revumenib 1; (ii) the CBFB-SMMHC:RUNX1 interface, which yielded AI-10-49 2. If both do not rank in the top decile the scoring is miscalibrated.

Negative control it must recover: the EWSR1 prion-like-domain to BAF interaction 3, which is biologically obligate but diffuse and should score poorly on pocket character. A pipeline that ranks a multivalent IDR contact alongside menin is not measuring druggability.

Key derived metric: dependency-to-essentiality ratio = (mean Chronos gene effect in fusion-positive lines) minus (mean Chronos in fusion-negative lines), divided by a normal-tissue tolerance term built from LOEUF and breadth of GTEx expression. The menin case is instructive because it does not score perfectly on tolerance, which is why the metric must be reported rather than thresholded.

Why this is worth doing before any wet-lab spend

Every data source is public and free, the whole scan is a few hundred API calls plus structure prediction, and the output directly prioritises exp-cross-family-interface-disruption-panel. It also produces a falsifiable prediction for hyp-obligate-partner-interface-generalises: if the ranked list shows that only one or two interfaces in the entire atlas have genuine pocket character, the hypothesis reduces from a principle to a lucky accident, and the lab should know that before committing a grant cycle to it.

Facts and figures
scriptanalyses/comparative-interface-druggability/run.py
statusplanned
cost tierlow
computelaptop-scale except AlphaFold-Multimer for unresolved pairs, which should run on the Explorer cluster or Modal
reproducibility notePin the DepMap release, the ChEMBL version and the AlphaFold model versions in the output manifest; all sources are public and free.
inputs
[
  "data/graph/nodes/comparative/*.json (fusion-partner pairs)",
  "PDB / RCSB REST API for experimental interface structures",
  "AlphaFold DB and AlphaFold-Multimer for unresolved pairs",
  "DepMap Chronos gene effect via Open Targets GraphQL, stratified by fusion genotype",
  "GTEx and Human Protein Atlas normal-tissue expression",
  "gnomAD LOEUF and pLI constraint",
  "ChEMBL bioactivity per partner target"
]
outputs
[
  "ranked interface table (csv) with buried surface area, pocket/hot-spot score, dependency-to-essentiality ratio, chemical-matter count, cross-genotype flag",
  "calibration report against the two positive and one negative control"
]
tests hypotheses
[
  "hyp-obligate-partner-interface-generalises",
  "hyp-oligomerisation-interfaces-as-mutation-resistant-targets",
  "hyp-nup98-partner-obligate-interface"
]
required additional columns
[
  "does an interface POINT MUTANT (not a knockout) phenocopy the small molecule? -- the WDR5 WIN-site case (PMIDs 30865883, 38682900) shows a liganded interface can give an off-interface phenotype",
  "is the dependency paralogue-buffered? -- RUNX2 compensates for RUNX1 (PMID 23979164), so single-gene DepMap scores understate it",
  "has any molecule against this interface reached a registered trial? -- CBFB-SMMHC has not, eleven years on (ClinicalTrials.gov, 2026-09-19)"
]
How this could bite us back
  • A scan like this will rank partners that are broadly essential highly on dependency and then be over-trusted; the tolerance term is doing the important work and is the least reliable part.
  • AlphaFold-Multimer confidence on IDR-containing interfaces is poor, so the pipeline will systematically under-rank exactly the condensate-class interfaces it is meant to triage away — which looks like success but may be circular.
  • Cell-line dependency data for paediatric fusion genotypes is thin, so an apparently quantitative ranking may rest on two or three lines per genotype.
Still open
  1. Are there enough fusion-positive cell lines per genotype in DepMap for the dependency contrast to be meaningful, particularly for infant KMT2A-r ALL and NUP98-r AML?
  2. Can AlphaFold-Multimer be trusted for interfaces involving long intrinsically disordered regions, or should those pairs be reported as unscoreable rather than scored badly?
  3. What is the right normal-tissue tolerance term, given that menin — the one clinically validated target — would itself score poorly on naive constraint metrics?
  4. Can pocket character be scored jointly with a requirement that the drug phenotype phenocopy an interface point mutant? Without that column the scan measures ligandability, which the WDR5 case shows is not the same as therapeutic relevance.
  5. The scan scores pocket character and ligand availability, which are properties relevant to occupancy. Adding a modality column - is this partner degradable, is there a ligandable surface anywhere on it that a degrader could exploit even if it is not the functional interface - would have reordered the WDR5 row entirely. What else does it reorder?
Every source behind this entry (5)
  1. E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The menin-KMT2A interface is a validated positive control for interface druggability, having produced a clinically active drug.

    The menin inhibitor revumenib in KMT2A-rearranged or NPM1-mutant leukaemia · Issa GC, Aldoss I, DiPersio J, et al. · Nature 2023
    PMID 36922593DOI
    What this evidence cannot show

    Clinical validation of a target does not validate any particular computational score, so the control demonstrates calibration only, not correctness.

    Evidence that argues against this (1)
    • The scan's design assumes that chemical matter against a partner interface indexes the interface's therapeutic relevance. The WDR5 WIN site breaks that assumption: it is a structurally ideal, picomolar-liganded, chromatin-displacing partner interface whose inhibitors act in MLL-rearranged cells through ribosome protein gene suppression, nucleolar stress and p53/MDM4 activation -- not through the partner's chromatin biology. A scan scoring pocket character plus available ligands would rank this interface highly and would be scoring the wrong mechanism. Howard GC et al. 2024
  2. E2strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The CBFB-SMMHC:RUNX1 interface is a second validated positive control, drugged by AI-10-49 with restoration of RUNX1 activity and delayed leukaemia in mice.

    PMID 25678665DOI
    What this evidence cannot show

    The compound never reached the clinic, so it validates the interface as chemically engageable but not as a clinical target.

    Evidence that argues against this (1)
    • Calling this a validated positive control overstates what validation means for a scan whose output is a druggability ranking. Eleven years after AI-10-49, no interventional study of a CBFbeta-SMMHC:RUNX1 interface inhibitor is registered (re-checked against the ClinicalTrials.gov v2 API on 2026-09-19: query.term=CBFB-SMMHC and query.term=AI-10-49 return no matching interventional study). What IS registered in core-binding-factor AML is a phase 2 of avapritinib against the acquired KIT co-lesion, running since 2022-06-01. A scan calibrated on a positive control that never entered the clinic, while the clinic pursued a different node in the same disease, is calibrated on preclinical tractability rather than on developability - and those are the two things the scan is being used to tell apart. The First Affiliated Hospital of Soochow University 2022
  3. E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    The EWSR1 prion-like domain to BAF interaction is an obligate but structurally diffuse fusion-partner interface, suitable as a negative control for pocket-character scoring.

    Cancer-specific retargeting of BAF complexes by a prion-like domain · Boulay G, Sandoval GJ, Riggi N, et al. · Cell 2017
    PMID 28844694DOI
    What this evidence cannot show

    No structure of the contact exists, so its classification as diffuse is itself an inference from the phase-transition dependence.

    Evidence that argues against this (1)
    • Using the EWSR1 prion-like domain as the negative control assumes the relationship between interaction strength and oncogenic output is monotonic, so that 'diffuse and unliganded' maps onto 'undruggable'. Quantitative single-cell and single-molecule imaging in patient-derived Ewing cells shows the relationship has an optimum instead: EWS::FLI1 requires a narrow window of low-complexity-domain interaction, increasing those interactions toward putative phase separation REPRESSES its GGAA-microsatellite targets, and forcing the fusion into the nucleolus inhibits both its program and transformation. A surface that can be disabled by pushing interaction UP is not a negative control for druggability; it is an unusual positive one, and the scan has no column for that. Chong S et al. 2022
  4. E4strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database

    DepMap CRISPR (Chronos) gene effect and gnomAD constraint are available programmatically and have already been used in this graph to characterise KMT2A and MEN1 essentiality.

    What this evidence cannot show

    Cell-line essentiality is a poor proxy for tissue tolerance, fusion-positive line counts are small for several genotypes (especially infant KMT2A-r ALL), and DepMap release versions must be pinned for reproducibility.

    Evidence that argues against this (1)
    • DepMap-plus-constraint scoring has a documented blind spot in exactly this domain: RUNX2 compensates for RUNX1 loss in AML1-ETO and MLL-AF9 cells, so a paralogue-buffered dependency reads as a non-dependency in a single-gene CRISPR screen. This graph already records the failure mode as `hyp-paralogue-buffering-hides-dependencies`; RUNX1/RUNX2 is a published instance of it inside the scan's own target set. Goyama S et al. 2013
  5. E5strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper

    Interface engagement does not index interface-mediated mechanism: WDR5 WIN-site inhibitors displace WDR5 from chromatin at picomolar affinity yet act in MLL-rearranged cells through ribosome protein gene suppression, nucleolar stress and p53/MDM4.

    PMID 38682900DOI
    What this evidence cannot show

    Abstract only for PMIDs 30865883 and 38682900; cell lines and one chemical series. The scan should therefore carry a mandatory separation-of-function column (does an interface point mutant phenocopy the drug?) rather than scoring pocket character and ligand availability alone.

    Evidence that argues against this (1)
    • The WDR5 row also carries the opposite lesson, and the scan needs both. Occupancy-based WDR5 protein-protein-interaction inhibitors gave modest cancer-cell killing and lacked in vivo efficacy, but a structure-designed VHL-recruiting WDR5 degrader (MS67) was more effective than those inhibitors at suppressing WDR5-regulated transcription, at reducing the chromatin-bound fraction of MLL complex components and c-MYC, and at inhibiting proliferation, and it suppressed MLL-rearranged AML patient cells in vitro and in vivo while being well tolerated. So the WDR5 WIN site is not an example of interface engagement failing to produce interface-mediated mechanism; it is an example of OCCUPANCY failing where DEGRADATION of the same protein succeeded. A druggability scan that scores pocket character and ligand availability is scoring for occupancy, and this row says the scan needs a modality axis. Yu X et al. 2021
Connected to