phenibut-compchem

Phenibut Computational Pharmacology — Tool Routing & Pipeline

Companion to proof_empty/RESEARCH_PLAN.md and proof_empty/SCHRODINGER_TOOLKIT.md (the authoritative capability map + gap analysis). Read those for the full module/feature detail.

Ground truth

• You own the full Schrodinger Drug Discovery Suite (2026-2, licensed to 2050): Glide, IFD/IFD-MD, FEP+, Desmond, Jaguar/QSite (QM), Prime+STA_GPCR, QikProp, Phase, SiteMap, WaterMap, Epik, MacroModel, Canvas, AutoQSAR, AutoDesigner. Buy nothing for the structural/chemistry layer.
• PyMOL 3.1.7 (Incentive) is the only figure tool. Headless: Scripts\pymol.exe -cq script.pml. License lives at %USERPROFILE%\.pymol\license.lic (must be the real entitlement, not the auto-generated eval token — else figures get a watermark).
• The mechanistic "why" questions (dopamine puzzle, withdrawal, prosocial effect) are neurobiology/systems pharmacology, not docking. Corpus synthesis + free tools (PK-Sim, STRING/Cytoscape, NEURON/Brian2) are the main engine; computation tests falsifiable sub-hypotheses.

MCP: schrodinger (server at C:\Users\Prekzursil\source\repos\schrodinger-mcp\server.py)

After editing the server, reconnect the MCP / restart Claude Code to reload it.

Tested, ready: schrodinger_version, backend_info, smiles_to_3d, molecule_properties, convert_structure, ligprep, qikprop_admet, glide_grid, glide_dock, prepwizard, prime_mmgbsa, run_schrodinger_python, pdb_fetch, qikprop_parse_results, pymol_render. Implemented-untested (verify first live run): sitemap, jaguar_pka, epik_enumerate. Roadmap (Phase 2, async/GPU — spec in schrod_cards/_mcp_expansion.md): job-queue + ifd_dock, fep_plus_rbfe, desmond_md_*, watermap_run, phase_pharmacophore, autoqsar_*, prime_homology, glide XP/constrained variants.

Standard structure-based pipeline (GABA-B + analogs)

1. pdb_fetch('7eb2'|'7c7q'|'6uo8', out_dir) — GABA-B cryo-EM structure.
2. prepwizard(pdb, prepped.mae, ph=7.0) — clean receptor → QC with pymol_render.
3. sitemap(prepped.mae, ...) — locate orthosteric + cryptic pockets (verify first run).
4. smiles_to_3d / ligprep (+ epik_enumerate for the zwitterion at pH 7.4) — prep ligands.
5. jaguar_pka(ligand.mae, ...) — confirm protonation state (load-bearing for binding mode).
6. glide_grid(prepped.mae, grid.zip, center=...) → glide_dock(grid.zip, ligands.mae, 'SP'|'XP').
7. prime_mmgbsa(pv.maegz) — rescore dG_bind; qikprop_admet + qikprop_parse_results — ADMET/BBB.
8. pymol_render([prepped.mae, pv.maegz], fig.png, highlight_selection='polymer within 5 of hetatm') — publication pose figure.
9. (Phase 2) IFD-MD for the flexible VFT; FEP+ for <2 kcal/mol analog ranking; Desmond for dynamics.

Biomedical & literature MCP layer — ToolUniverse + direct MCPs (PREFER over external installs)

Most "gap" capabilities are already callable in-session — no desktop install. Reach for an external tool only when no MCP can do it (the short list below).

ToolUniverse (mcp__tooluniverse__*, 2,206 tools/511 categories, globally registered; 5 compact tools). Workflow: find_tools/grep_tools → get_tool_info(detail_level='full') → execute_tool. Map each phenibut question to its tools:

Need	Q	ToolUniverse tools
Bioactivity / Ki / IC50 / targets (SAR table)	A,B,D	ChEMBL (29), bindingdb, gtopdb (IUPHAR — has GABA-B), drugcentral, pubchem_bioassay
GPCR target data (GABA-B = class-C GPCR)	A,B	gpcrdb, gtopdb, uniprot
ADMET / BBB / CNS	D	ADMETAI_predict_BBB_penetrance, SwissADME_calculate_adme (49 props), admetai (9) — ADMET-AI v1 vs v2 predictions differ (v2 = Chemprop-v2, retrained): re-run or version-label any existing phenibut ADMET number; pin the exact version in every receipt
Off-target panel (dopamine puzzle)	B	swiss_target (SwissTargetPrediction), stitch, dgidb, pharos
Pathway / network (GABA-B→VTA, neuroadaptation)	B,E	reactome, kegg, wikipathways, string_network, omnipath (ligand-receptor)
Drug-drug interactions (phenibut + amphetamine/alcohol/benzo)	C	DailyMed_parse_drug_interactions (CYP/P-gp), sider, DrugSynergy_calculate_loewe, synergxdb, dose_response
Structures (no install)	all	rcsb_pdb (39), pdbe, alphafold, swissmodel, esmfold
Literature beyond the corpus	all	EuropePMC, pubmed, semantic_scholar, OpenAlex, unpaywall, biorxiv/medrxiv
Free docking cross-check	B	swissdock
Enzyme/inhibition kinetics math	C,D	EnzymeKinetics_calculate, DrugSynergy_calculate_loewe

Direct MCPs in-session (often richer than ToolUniverse's wrapper): mcp__claude_ai_ChEMBL__*, mcp__claude_ai_PubMed__*, mcp__claude_ai_bioRxiv__*, mcp__claude_ai_Clinical_Trials__*, mcp__rdkit__* (74 descriptor tools), mcp__scholarly-access__* (OA PDF resolve/fetch — corpus expansion). See the tooluniverse and consensus-tooluniverse-research skills; the bionlp workspace (Downloads/bionlp/) has the full setup.

Still needs an external install (no MCP covers it)

• PBPK brain kinetics (phenibut vs baclofen CNS exposure) → PK-Sim + MoBi (free, OSP Suite).
• VTA disinhibition circuit model (the dopamine puzzle's circuit answer) → NEURON/Brian2 + ModelDB (free).
• All-atom MD as a free Desmond backup / cross-validation → GROMACS / OpenMM.
• Protein-ligand co-folding for novel analogs (no PDB template) → Boltz-2 (pip install boltz); structure-only is covered by ToolUniverse alphafold/esmfold.
• SAR-cliff / MMP visual analysis → DataWarrior (free).

Gotchas

• The MCP never imports schrodinger in-process; everything goes through $SCHRODINGER/run.exe. The .mcp.json env block pins SCHRODINGER=D:\Program Files\Schrodinger 2026-2 (authoritative).
• IFD-MD, FEP+, Desmond, WaterMap-MD, DeepAutoQSAR are GPU / Linux-host; don't expect them to run synchronously on the Windows box — they need the job-queue (async) pattern.
• Run PyMOL as a separate subprocess from the Schrodinger Python runtime (interpreter conflict).