Boltz co-folding plan: six complexes worth folding, three that must not be, and the positive controls that decide whether to believe any of it
also called Boltz plan · co-folding readiness · structure prediction plan
The lab has credits for Boltz, a tool that predicts how two molecules fit together and how tightly they stick. We list exactly which pairs are worth predicting, which would produce confident-looking nonsense, and the known-answer cases we should run first so we can tell the difference.
The same thing, for a scientist
Boltz-2 (current public release v2.2.1) is the right tool for three phase-2 questions: menin partner selectivity, YEATS and MYST paralogue selectivity, and triage of designed binders against the LEDGF IBD. It is the wrong tool for the fusion junction, for the CXXC-DNA interface and for ternary degrader complexes, and the reasons are stated by its own authors. The plan below is therefore organised around a rule: **every campaign begins with a positive control whose answer is already known from a crystal structure, and a negative control that should fail.** If Boltz cannot reproduce the 1.55 A menin-MLL peptide complex or the 2MSR LEDGF IBD-KMT2A peptide complex, nothing downstream of it is believable. This discipline matters more here than usual because a predicted structure renders in the project dashboard looking exactly as authoritative as an experimental one.
The hard structural constraint every plan inherits
There is no AlphaFold model of full-length KMT2A - the prediction API returns HTTP 404 for Q03164 - and no experimental structure of an intact fusion oncoprotein. AFF1's model has a global pLDDT of 52.34 over 1,210 residues. Every Boltz plan in this project must therefore work on fragments, with boundaries chosen on domain annotation, and must state which fragment was folded whenever a result is reported.
The six complexes worth folding
1. Positive control: menin + KMT2A MBM1 peptide (4-15). The answer is known at 1.55 A (4GQ6). If the predicted pose does not reproduce the RWRFP insertion, stop and recalibrate. Cost: negligible.
2. Positive control: LEDGF IBD (347-435) + KMT2A 140-160. Known from 2MSR, including the F148/F151 anchors. Second calibration point, and the one closest to the programme this review recommends starting.
3. The selectivity question: menin + KMT2A peptide versus menin + JUND peptide, each against each clinical menin ligand. A useful result is a consistent rank ordering of compounds by predicted KMT2A-over-JUND differential, not an absolute number. This directly serves hyp-partner-selective-menin-ligand-separates-mandates and ana-menin-partner-selectivity-profiling.
4. Resistance recapitulation: menin S160T, M327I, M327V, G331D, G331R, T349M, each with revumenib. Six known-answer cases. A useful result is that the six resistant variants score worse than six matched non-resistant surface substitutions. This is the only legitimate quantitative validation available, because these six are the entire patient-observed label set.
5. Paralogue selectivity: ENL YEATS versus AF9 YEATS versus GAS41 YEATS, same ligand. Cheap, and it addresses the mandatory selectivity gate in the ENL dossier.
6. Designed binders against the LEDGF IBD. Only after gate 2 passes. Macrocycles and miniproteins reproducing the F148/F151 anchor geometry. Triage only - the output is a shortlist to synthesise, never a claim of affinity.
The three that must not be folded, and why
The fusion junction. Co-folding two disordered chains returns coordinates with a pose. The pose will look like a pocket. It will not be one. This is the single most likely way for this project to waste a year on a computational artefact.
The CXXC-DNA interface as a ligand-design target. Boltz-2's authors state the affinity module does not explicitly handle ions, water or multimeric partners; a zinc-coordinating domain reading DNA is exactly that regime.
Ternary degrader complexes, as a decision input. Same limitation, same reason. Ternary triage is permitted as a filter and forbidden as evidence, and the degrader dossier says so.
What Boltz-2 actually claims, and what it admits
Claims: an average Pearson correlation of 0.66 on a four-target free-energy-perturbation subset, approaching physics-based methods while running more than 1,000 times faster, and out-of-the-box performance exceeding all submitted entries on the CASP16 affinity track 1.
Admits, in its own Limitations section 1: predictions are unreliable if the pocket or interface is mis-built; the affinity module does not explicitly handle cofactors, ions, water or multimeric binding partners; the model often fails to capture large binding-induced conformational changes; and performance varies strongly between assays for reasons the authors cannot yet attribute. The repository also states that updated evaluation code for Boltz-2 is not yet released, so the headline numbers cannot currently be reproduced independently.
Four targets is a small benchmark. None of them resemble a peptide-in-groove pocket like menin's or a reader domain like ENL's. Treat 0.66 as an existence proof that the method can work, not as an expected performance on our systems.
| tool | Boltz-2, current public release v2.2.1 |
|---|---|
| governing rule | Every campaign opens with a positive control whose answer is known from an experimental structure and a negative control that should fail. No downstream result is reported without them. |
| structural constraint | no full-length KMT2A model (Q03164 returns 404); no intact fusion structure; all work is fragment-wise with boundaries recorded |
| status | planned |
| compute | Boltz credits, via the boltz-api CLI. Campaigns 1-5 are small. Campaign 6 is the only one that could consume meaningful credit and is gated behind an experimental result. |
campaigns — the numbers behind the chart
[
{
"n": 1,
"complex": "menin + KMT2A MBM1 peptide 4-15",
"type": "positive control",
"known_answer": "PDB 4GQ6, 1.55 A",
"useful_result": "reproduces RWRFP insertion into the TPR cavity",
"cost": "negligible"
},
{
"n": 2,
"complex": "LEDGF IBD 347-435 + KMT2A 140-160",
"type": "positive control",
"known_answer": "PDB 2MSR, NMR",
"useful_result": "reproduces the F148/F151 aromatic anchors",
"cost": "negligible"
},
{
"n": 3,
"complex": "menin + KMT2A peptide vs menin + JUND peptide, against each clinical menin ligand",
"type": "selectivity",
"useful_result": "a consistent RANK ORDERING of compounds by KMT2A-over-JUND differential, not an absolute number",
"serves": [
"hyp-partner-selective-menin-ligand-separates-mandates",
"ana-menin-partner-selectivity-profiling"
]
},
{
"n": 4,
"complex": "menin S160T / M327I / M327V / G331D / G331R / T349M + revumenib",
"type": "quantitative validation",
"useful_result": "the six resistant variants score worse than six matched non-resistant surface substitutions",
"note": "the only legitimate quantitative validation available, because these six are the entire patient-observed label set"
},
{
"n": 5,
"complex": "ENL YEATS vs AF9 YEATS vs GAS41 YEATS with the same ligand",
"type": "paralogue selectivity",
"serves": "the mandatory selectivity gate in the ENL dossier"
},
{
"n": 6,
"complex": "designed macrocycles and miniproteins vs LEDGF IBD",
"type": "design triage",
"precondition": "only after LEDGF gate G2 passes",
"useful_result": "a shortlist to synthesise; never an affinity claim"
}
]forbidden campaigns
[
{
"complex": "fusion junction co-fold",
"reason": "two disordered chains; co-folding returns a confident-looking pose that is not a pocket. The single most likely way to waste a year on an artefact."
},
{
"complex": "CXXC-DNA as a ligand-design target",
"reason": "zinc-coordinating domain on nucleic acid; the affinity module does not explicitly handle ions or nucleic-acid partners"
},
{
"complex": "ternary degrader complexes as a decision input",
"reason": "multimeric binding partners are outside the affinity module's declared domain; permitted as a filter, forbidden as evidence"
}
]reported accuracy
{
"pearson_on_4_target_fep_subset": 0.66,
"speed": ">1000x faster than FEP",
"caveat": "four targets, none resembling a peptide-in-groove pocket or a reader domain; evaluation code for Boltz-2 not yet released"
}- A predicted structure rendered in this project's 3D viewer is visually indistinguishable from a crystal structure. Every predicted model that reaches the dashboard must carry its pLDDT or confidence and the word 'predicted' in the caption, or the graph will mislead its own readers.
- Co-folding the fusion junction would produce exactly the result someone hoping for a fusion-specific pocket wants to see. Recording the prohibition in advance is the only protection against that.
- Selectivity rankings from a model that has never seen a menin-JUND complex are extrapolation. If they agree with a wet-lab result, that is validation; if they disagree, the model is wrong, not the assay.
- Are 4GQ6 and 2MSR in Boltz-2's training data, and if so what untrained positive control could serve instead?
- Does Boltz-2 rank the six known resistance variants correctly - and if it does not, does that invalidate campaign 3's selectivity ordering as well?
- Can a co-fold distinguish the KMT2A and JUND peptides in the menin pocket at all, given that both are short peptides in the same cavity?
- E1moderate confidence 3 of 5: One strong study, or several smaller consistent ones.preprint
Boltz-2 reports an average Pearson correlation of 0.66 on a four-target FEP subset while running over 1,000 times faster than free-energy simulation, and its authors state that the affinity module does not explicitly handle cofactors, ions, water or multimeric binding partners, that the model often fails to capture large binding-induced conformational changes, and that performance varies strongly between assays for unattributed reasons.
Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction · Passaro S, Corso G, Wohlwend J, et al. · bioRxiv 2025What this evidence cannot show
Preprint, not peer reviewed; the repository states updated Boltz-2 evaluation code has not been released, so the reported accuracy cannot be independently reproduced at present. The four FEP targets bear no resemblance to a peptide-in-groove pocket or a YEATS reader domain.
Evidence that argues against this (1)
- The authors themselves report that Boltz-2 does not significantly improve on its predecessors for structure prediction, and does not significantly deviate from AlphaFlow or BioEmu for ensembles. The advance is in affinity, on a narrow benchmark. source 2025
- E2established confidence 5 of 5: Multiple independent lines of evidence; treated as settled in the field.database
The two positive controls have experimentally determined answers: menin with an MLL peptide at 1.55 A (4GQ6) and the LEDGF/p75 IBD with the MLL1 140-160 peptide by NMR (2MSR), both verified against the RCSB data API this cycle.
RCSB PDB data API entry records for 4GQ6 and 2MSR · RCSB PDB 2026What this evidence cannot show
Reproducing a deposited structure that is very likely in the model's training data is a weak test. It is a necessary condition, not a sufficient one, and a pass should not be over-interpreted.
Evidence that argues against this (1)
- Both structures predate the model and may be in its training set, so success on them measures memorisation as much as prediction. The resistance-variant recapitulation (campaign 4) is the more honest test because the specific variant complexes are not deposited. source 2026
- E3strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.database
No AlphaFold model of full-length KMT2A exists (Q03164 returns HTTP 404) and the AFF1 model has a global pLDDT of 52.34, so every co-folding plan in this project must operate on fragments.
What this evidence cannot show
Fragment boundaries are a modelling choice that can create or destroy an apparent pocket; boundaries must be recorded with every result.
Evidence that argues against this (1)
- No AlphaFold model of full-length KMT2A and a low-confidence AFF1 model is a statement about DISORDER, not about coverage. Both proteins are largely intrinsically disordered, so a co-folding plan should expect low confidence as the ground truth rather than as a gap to be closed by a better predictor. RCSB PDB 2026
- ← depends onSeparation-of-function test: does KMT2A::AFF1 need its C-terminal homology domain to dimerise?
cheap first step: co-fold the AFF1 CHD homodimer and the AFF1-AFF4 CHD heterodimer before committing to mutants
campaign 3 is the structural half of this planned analysis
- depends on →Dataset: PDB and AlphaFold - the structural census, including the two absences that shape every plan
the positive controls and the KMT2A absence come from this structural census
co-folding the junction is explicitly forbidden: two disordered chains give a confident-looking pose that is not a pocket
campaign 5 serves the mandatory YEATS paralogue selectivity gate
- ← supportsCompute: which backend for which piece of phase-2 work, and the one constraint that matters for each
matches the co-folding campaigns to Boltz credits rather than cluster time