comp-pharm-cascade

comp-pharm-cascade

Orchestration skill: how a target + ligand series flows through the available compute stack to a defensible affinity/mechanism result. This is a routing law over executor skills, not a new engine. Reference implementation: the phenibut α2δ-1 / GABA_B campaign (worked numbers at the bottom).

Core principle — a cascade with two guarded gates

Escalate cost only when the cheaper tier can't answer. But a naive cascade is brittle where campaigns actually break: it trusts each tier and escalates upward, so a wrong pose makes MM-GBSA, FEP, and QM all rank garbage equally. So spend redundancy surgically at the two checkpoints where being wrong is catastrophic and silent, single-track everywhere else.

• Gate 1 — Pose. A pose is trusted only when independent methods agree (Boltz-2 co-fold ipTM/PAE ∩ Glide/IFD ∩ BPMD stability). Cheap ensemble at the known failure point. Never rescore an unvalidated pose.
• Gate 2 — The published number. The one headline ΔΔG gets FEP plus one orthogonal corroborator (Boltz affinity or QM decomposition) plus the experimental value. Everything else in the SAR runs single-method, flagged lower-confidence.

Tier map (cheap → expensive)

Tier	Capability	Executor	Note
0 — Anchor ⭐	Experimental Ki/IC50/EC50, the y-axis	ChEMBL MCP / GtoPdb / tooluniverse pharmacology	Run FIRST. No ground truth → no rankable claim (see Gate 0).
1 — Cheminfo	descriptors, pKa/protonation, filters	cheminformatics-rdkit, Inductive Bio, Epik/jaguar_pka	Zwitterion state decides GB desolvation — get it right.
2 — ML structure+affinity	co-fold + affinity, license-free	Boltz-2 MCP (boltz_*) / qdx-rush boltz_fold	Breadth ranking + pose gate. Minutes, no license.
3 — Physics endpoint	MM-GBSA, BPMD pose-stability	schrodinger-maestro-driver (prime_mmgbsa, thermal_mmgbsa), mcp__schrodinger__*	Fast, but endpoint scoring often fails on flexible/charged ligands.
4 — Physics free-energy	alchemical FEP+, MD, WaterMap	mcp__schrodinger__fep_plus/desmond_md/watermap	Gold standard for the headline. Cloud GPU.
5 — QM	interaction-energy / QM-MM mechanism	qdx-rush (exess_interaction_energy, exess_qmmm)	Captures salt-bridge/polarization MM underestimates.
6 — Deliverables	figures, methods, unit/error math	BioRender, Mermaid, protocols.io, pymol_render, Wolfram	Automate the write-up.

Canonical phase pipeline

• Gate 0 — Verify the anchor (do this before any compute). Pull the full experimental affinity ladder (ChEMBL target ID → all Ki/IC50/Kd with pchembl_value; dedup to one row/compound by max pChEMBL). Measure its dynamic range. If the congeners span <~2 log or the subjects are unmeasured, the maximum achievable rank correlation is ~0 — reframe the claim (binding-mode/mechanism) instead of burning compute on an unwinnable metric. This gate is cheap and repeatedly pays for itself.
• Phase 1 — Pose (Gate 1). Boltz-2 co-fold target+ligands (prefer a holo template when a co-crystal exists; else sequence-only with the known anchor residue as pocket condition / reference ligand). Confirm the conserved anchor contact. Keep only poses where Boltz ∩ BPMD agree.
• Phase 2 — ML breadth rank. Boltz-2 affinity over the whole ladder → Spearman(score, experimental pChEMBL). Method is validated for the series iff the correlation holds on the real dynamic range.
• Phase 3 — Physics headline (Gate 2). FEP+ on 3–5 ladder-spanning edges + the subject edge; QM decomposition of the key contact; require FEP ∩ Boltz ∩ QM agreement before publishing the number.
• Phase 4 — Predict. Apply the validated method to any unmeasured subject; label it a computational hypothesis, not an experimental value.
• Phase 5 — Deliverables. Mechanism panel, pipeline DAG, methods, error math.

Substrate & ops (hard-won)

• Local vs cloud. Schrödinger prep/dock/MM-GBSA/QM-setup run locally if a Windows license is valid; FEP/Desmond MD need a real GPU (≥16–40 GB) → cloud (Lightning/Modal A100) or license-free AToM-OpenMM. A laptop GPU (e.g. 6 GB) cannot do FEP. Never bypass a license verifier.
• Boltz-2 free tier. Orgs hold ~$5, run ~1 job concurrent (excess startup-times-out → stopped at 0, $0 cost), OAuth token expires on long runs, and switching orgs orphans in-flight jobs (workspace-scoped → 404 from the new org). Some orgs are pathologically slow (progress bursts near the end, not incrementally). Pattern: one small solo job → collect immediately (results paginate 100/page via after_id) → then move. Screen caps 500 mol/call; the default SMARTS filter silently drops ~15–20%. ~$0.025/molecule.
• Verify the account before a paid call (boltz_get_account_context) and re-estimate cost (boltz_estimate_*) — never assume credits.
• Honesty gates. State whether a contrast is matched (same ligands, same assay) or qualitative (broad set vs narrow). Distinguish equilibrium score from kinetic effects (a co-fold score won't capture off-rate selectivity). Log every silent cap (SMARTS-filtered, truncated, unmeasured).

Worked reference — phenibut (2026-07-04)

• Gate 0 caught a dead metric: α2δ-1 (ChEMBL CHEMBL1919) — only gabapentin (27 nM) and pregabalin (19 nM) measured, ~0.15 log apart; phenibut/baclofen unmeasured in ChEMBL. Real ground truth = Zvejniece 2015 (PMID 26234470, [³H]gabapentin rat-brain, subtype-unresolved): gabapentin 0.05 µM, R-phenibut 23 µM, S-phenibut 39 µM, baclofen 156 µM.
• Tier 2 win: Boltz-2 co-fold vs α2δ-1 (P54289, pocket Arg241/Asp452/Asp491, mirogabalin ref) → Spearman(binding_confidence, pChEMBL) = +0.48, n=486 (stable: +0.54@212 → +0.48@369 → +0.48@486). Recovers the breadth ranking the MM-GBSA/FEP endpoint could not (≈0 on the narrow β-aryl-GABA set). Core-6: potent gaba/prega separate from weak phenibut/baclofen; R-phe > S-phe correct (matches 23 < 39 µM) — independent of the ladder join.
• Tier 4 headline: GABA_B FEP ΔΔG(baclofen→phenibut) = +1.74 vs exp +1.62 (0.12 kcal/mol); ChEMBL confirms the baclofen 6 µM leg.
• Honest framing: the +0.48 (broad) vs MM-GBSA ≈0 (narrow) is a qualitative contrast, not a matched head-to-head; α2δ endpoint failure is MM-specific, not "computation can't rank α2δ."

Full audit trail: D:/Research/projects/phenibut/docs/COMPUTE_STACK_GRILL.md.