First clear positive result in the project. Ran Phase 1 on Modal L4
(~$0.70). Boltz-2 P(binder), cofactors co-folded:
- HDAC2 (+Zn): vorinostat 0.9994 vs negatives ~0.1 -> PASS, decisive
- hemoglobin (+heme): voxelotor 0.46 -> PASS (weak; covalent/tetramer)
- PKR (+FBP/Mg): mitapivat 0.32 < hydroxyurea 0.40 -> FAIL (allosteric)
HDAC2/Zn is the exact case classical Vina failed (no metal term, 7.9A
redock). Co-folding handles the Zn-chelation chemistry -> the structure-
binding modality pivot (PLAN §12) is validated on its decisive test.
Engineering fixes that got it running: image needs cuequivariance kernels;
max_containers=1 so weights download once (parallel corrupted the shared-
Volume checkpoint); rank by P(binder) not affinity_pred_value (sign).
Adds docs/results/phase1_affinity.csv (committed; raw under data/ gitignored).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add metal/cofactor handling to the Boltz-2 YAML as CCD ligand entries -
the modes classical docking couldn't model:
- HDAC2 + catalytic Zn (vorinostat chelates it)
- PKR + FBP + Mg (allosteric activator + metal)
- hemoglobin + heme
Same cofactors present when co-folding negatives into a target (fair test).
build_boltz_yaml() gains a cofactor_ccds arg (emits `ligand: {ccd: ...}`
entries); TARGETS carries per-target cofactors; cofold()/main() thread them
through. Verified locally: YAML builds correctly with Zn / FBP+Mg.
Honest limitation noted: Hb's voxelotor site is at the tetramer centre and
covalent (Schiff base), so single-chain+heme only approximates it - HDAC2
(Zn) and PKR (cofactor) are the real co-folding tests. Ready for
`modal run gpu/modal_app.py`.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Flesh out the Modal app into a runnable Phase-1 positive-control test and
reconcile it with the plan:
- cofold() GPU fn: build Boltz-2 YAML (protein+ligand+affinity), run
`boltz predict --use_msa_server --cache /weights/boltz`, parse affinity
JSON + predicted pose; weights persist via Volume.
- Local helpers (CPU, import-tested against our PDBs): binding_chain_sequence
(gemmi -- correctly picks the binding chain, e.g. alpha-globin for 5E83),
pubchem_smiles, build_boltz_yaml, fetch_pdb (RCSB).
- main(): fan out cofold.starmap over 3 targets x (known binder + 2
negatives); tabulate; PASS if known binder has top P(binder) for its target.
Alignment fixes:
- Rank by P(binder) (higher=better), NOT raw affinity_pred_value whose sign
(~log IC50) is version-dependent -- avoids a backwards positive-control test.
- gpu_plan.md Phase 1 updated to affinity/P(binder) ranking; pose-RMSD noted
as a later refinement (needs receptor superposition).
Local half verified (sequence/SMILES/YAML); cofold() needs a live `modal run`
(account + `modal token new`) to validate end-to-end.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Document and wire the weight-caching mechanism:
- modal.Volume is a cloud-backed FS independent of the GPU/container;
run 1 downloads weights into /weights, run 2+ reuses them (no GPU time
wasted re-downloading).
- Point downloaders at the mount: HF_HOME/TORCH_HOME/boltz --cache; persist
via weights.commit(), see updates via weights.reload().
- Volume storage costs pennies, separate from GPU = near-free caching.
modal_app.py cofold(): set cache env vars to /weights, reload()/commit()
around the (stubbed) boltz call.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
docs/gpu_plan.md: cost-efficient plan for running AF3-class co-folding
(Boltz-2/DiffDock) on a GPU then paying nothing when idle.
- Key insight: structure-track data is tiny (MB of PDBs/SMILES); only the
GPU + model weights are heavy -> serverless is ideal.
- Recommend Modal (per-second billing, scales to zero = nothing to kill);
RunPod as the SSH-box alternative with idle auto-terminate.
- Lifecycle: image -> weights Volume (cache, don't re-download) -> run ->
git push small results -> teardown automatic.
- Phase 1 validate on 3 known binders (~$1) before paying for a screen;
Boltz-2 (affinity) on an L4/A10 (24-48GB); est total ~$5-15.
gpu/modal_app.py: Modal app skeleton (image, weights volume, GPU cofold()
function, local entrypoint); boltz invocation stubbed with TODOs.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
§12.4 pushed to its limit. Meeko ligand prep + in-place symmetry RMSD
(spyrmsd, not obrms) on clean HDAC2/vorinostat: 7.9A -> 4.76A. Prep and
metric mattered, but still FAIL.
Residual cause is fundamental: vorinostat binds via hydroxamate-Zn
chelation and Vina has no metal-coordination term. Real finding: sickle's
druggable targets bind via non-classical chemistry classical docking
handles poorly -- Hb (covalent), PKR (allosteric+cofactor), HDAC (Zn
chelation). Vina is the wrong tool for this target landscape.
Redirect: data-driven AF3-class co-folding (Boltz-2/Chai-1/DiffDock)
handles these modes -- the indicated next tool, gated by the 24GB local
memory ceiling (cloud GPU needed). The "GPU breaks all-local" §12.6
prediction is now the binding constraint of the track.
Adds: scripts/dock_production.py; deps meeko, spyrmsd, gemmi.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
§12.4 de-biased validation (scripts/dock_validate.py).
Redock each co-crystal ligand into its own structure, RMSD vs crystal:
- voxelotor->Hb: NA (covalent binder, out of scope §12.7)
- mitapivat->PKR: 8.2A (allosteric, cofactors stripped)
- vorinostat->HDAC2 (4LXZ, zinc kept): 7.9A -- a CLASSICAL target that
should have worked
The clean target also failing => systematic pipeline-quality problem,
not target choice. Cheap Vina + open-babel prep gives scores but doesn't
reproduce known geometry, so affinities aren't trustworthy. Ligand
efficiency over-corrects (ranks tiny hydroxyurea best). Fix needs
production prep (Meeko/AutoDockTools prepare_receptor + reduce) and an
in-place RMSD metric. Consistent with the project theme: the quick
version of every method runs but fails honest validation.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
§12.3-12.4 first binding result on ARM Mac.
- Toolchain SOLVED: AutoDock Vina 1.2.5 mac binary (Rosetta) + open-babel
(brew). No conda, no MLX. dock_positive_controls.py runs end-to-end.
- Cross-dock known binders + negatives into Hb (5E83) and PKR (8XFD),
box centered on co-crystal ligands (5L7=voxelotor, WV2=mitapivat).
Finding: raw Vina affinity ranks almost perfectly by MOLECULAR SIZE
(mitapivat > voxelotor > decitabine/caffeine > hydroxyurea) in both
pockets — mitapivat wins even on hemoglobin it doesn't target. Raw score
can't distinguish target-specific binding: the docking analog of the
connectivity specificity problem. Next: redocking-RMSD validation +
ligand-efficiency normalization.
Note: machine is 24GB (not 96GB per PLAN §2), capping local AF3-class
inference. tools/ gitignored (vina binary).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Start the structure-based binding branch (PLAN §12), baseline-first.
- src/binding.py: validated RDKit ligand retrieval (morgan_fp, tanimoto,
retrieve_nearest = the §12.9 engine) + dock() stub documenting the
blocked ARM-Mac toolchain
- scripts/binding_ligand_baseline.py: 300 drugs vs known binders
- docs/structure_binding_notes.md: status, toolchain blocker, next steps
- pyproject: [structure] extra (rdkit); data/raw/structures/ for PDBs
Step-0 finding: retrieval engine VALIDATED on in-set classes
(decitabine->azacitidine 0.62; vorinostat->scriptaid/belinostat) but the
distinctive binders voxelotor/mitapivat have no analog in our 300-drug
set (Tanimoto ~0.2). Needs (a) bigger library, (b) real docking (§12.3),
which is blocked on the ARM-Mac docking toolchain (§12.6 pitfall 4).
Structures 5E83 (Hb+voxelotor) and 8XFD (PKR+mitapivat) fetched.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Post-hoc improvement after the pre-registered v1 recovery test failed.
Two changes, diagnosing v1's failure:
- score on the full 12,328-gene LINCS space (week2_lincs_extract.py),
lifting signature overlap from 12% to 85% (brings erythroid markers in)
- src/scoring.py: KS connectivity + per-drug specificity z-score
(spec_z = SDs below a 1,000 random-query null). Primary ranking is
now spec_z. (Textbook tau saturated at +/-100 for a coherent query —
documented; needs a reference-signature library, a v2 item.)
- week3_scoring.py: spec_z primary + WTCS reference + prior-blended
- tests: tau/spec_z calibration test; 19 passing
- scripts/exp_genespace.py: the BING vs all-12,328 comparison
Result: hydroxyurea recovers (rank 40 -> 18, top 6%, passes top-10%),
confirming the v1 failure was the landmark bottleneck not the algorithm.
Overall STILL FAILS: L-glutamine does not reverse (rank 213, metabolite),
and negative controls (norethindrone, ciprofloxacin) rank top-3 —
connectivity != therapeutic relatedness. v1.1 is post-hoc/exploratory,
not a confirmatory test; reported as such in recovery_test_report.md.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Run the formal recovery test against the pre-registered criteria and
write the deliverable report (PLAN §6 Week 4):
- week4_recovery_test.py: evaluate hydroxyurea/L-glutamine + 5
pre-specified negative controls vs the committed criteria
- recovery_test_report.md: methodology, FAIL result with diagnosis,
top-10, lisinopril as the non-obvious candidate, limitations, v2
- known_limitations.md: L-glutamine coverage resolved, 12%-overlap
driver, recovery outcome table
Outcome: FAIL on all 3 criteria (hydroxyurea top 13%, L-glutamine
WTCS=0, 1/5 negative controls bottom-half). Root cause is signature/
assay data limitations (lost erythroid+HbF axis, 12% landmark overlap),
not the matching algorithm — reported straight per the project ethos.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Build the drug profile dataset (PLAN §6 Week 2):
- week2_curate_drugset.py: 300-drug set (2 ground-truth + 32 related-
mechanism + 26 negative-control + 240 random), restricted to
LINCS-scorable compounds, seed=42
- week2_chembl.py: InChIKey->ChEMBL match (145/300), MoA + targets
- week2_lincs_extract.py: cmapPy-slice both Level-5 GCTX phases to 978
landmark genes, mean-aggregate per drug to one consensus signature
- week2_assemble.py: join into drug_profiles_v1.parquet, Tier B (LINCS
single-source), scored flag per PLAN §6 Week 3 task 2
- docs/data_sources.md: drug set composition + LINCS/ChEMBL provenance
Results (all gitignored data): 300/300 drugs scored, both ground-truth
drugs present (hydroxyurea Phase II = CHEMBL467, L-glutamine Phase I).
Key caveat recorded: only 56/477 (12%) of the disease signature genes
are LINCS landmarks, so Week-3 scoring uses a 30-up/26-down query.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
