mechanistic-ddi-pipeline

Mechanistic DDI Pipeline (cross-stack orchestrator)

Overview

This skill is the GLUE between two servers. It fuses ToolUniverse knowledge (CYP/UGT/transporter roles, DDI records, FDA labels, literature) with Schrodinger structure-level computation (3D generation, drug-likeness descriptors, QikProp ADMET, optional Glide docking) into one perpetrator-victim DDI verdict. It does not replace the deep tooluniverse-drug-drug-interaction skill - it adds the structure layer.

Stacks used:

• tooluniverse MCP (knowledge) - see the ddi-admet-research and tooluniverse-drug-drug-interaction skills for routing and tool details.
• schrodinger MCP (structure) - see the schrodinger-maestro-driver skill for the 67 tools, the license-token split, and the full prep -> grid -> dock -> MM-GBSA pipeline (glide_grid builds the grid in-server).

When to Use

• A DDI question where mechanism must be backed by both pathway knowledge AND molecular evidence (descriptors, ADMET, or a docking pose into a CYP/target structure).
• A novel or under-documented compound (limited DDI records) where structure-derived properties inform the prediction.
• A case where you must reconcile a knowledge-base claim with computed physchem/ADMET (e.g., does the structure support the proposed CYP-substrate role).

If the question is purely knowledge-based, use ddi-admet-research instead - do not spin up structure tools unnecessarily.

Orchestrated Workflow

1. Frame the pair with the perpetrator-victim model (who changes whom). Apply the CYP/UGT reasoning from tooluniverse-drug-drug-interaction BEFORE any tool call.
2. Knowledge pass (ToolUniverse): resolve identifiers, then query CYP/transporter roles (ChEMBL_get_drug_mechanisms), known interactions (drugbank_get_drug_interactions_by_drug_name_or_id), FDA label evidence (DailyMed_get_spl_by_setid), and literature (PubMed_search_articles).
3. Structure pass (Schrodinger): for each drug, compute the drug-likeness panel (molecule_properties(smiles)); generate 3D where needed (smiles_to_3d); run qikprop_admet on a prepared structure for ADMET/physchem that supports or challenges the predicted pathway.
4. Optional docking corroboration: prep the receptor (prepwizard) and build a grid (glide_grid) for a CYP isoform/target, then dock the perpetrator/victim with glide_dock to test the proposed binding/inhibition hypothesis, optionally rescoring with prime_mmgbsa. No receptor structure on hand = skip docking (see structure-based-screening).
5. Synthesize the verdict: combine mechanism (PK via CYP/UGT/transporter, PD overlap), severity (Contraindicated/Major/Moderate/Minor), direction (inhibitor up / inducer down), and structure-derived support into a single ranked conclusion with management + monitoring.
6. Grade every claim: FDA label (★★★) > clinical study (★★☆) > theoretical/structure-inferred (★☆☆). Mark structure-derived inferences as supporting, not definitive.

Conflict Resolution

• Knowledge says interaction, structure says implausible (e.g. descriptors/ADMET inconsistent with the claimed pathway): report both, flag the conflict, prefer the FDA-label/clinical-study tier and note the structural caveat.
• Structure suggests a pathway with no DDI record: label it theoretical (★☆☆); recommend monitoring, not a hard verdict.

Common Mistakes

• Letting this skill duplicate tooluniverse-drug-drug-interaction - delegate the knowledge reasoning, add only the structure layer.
• Calling glide_dock without first running glide_grid - it needs the grid .zip (and glide_grid needs a box center: coords or ligand_molnum).
• Treating computed ADMET/descriptors as clinical evidence - they are supporting (★☆☆), never override FDA labels.
• Burning license tokens on Schrodinger drivers (ligprep/qikprop_admet/glide_dock, long -WAIT timeouts) when the knowledge pass already answers the question.

Related Skills

• ddi-admet-research - knowledge-only ToolUniverse front-end / router.
• schrodinger-maestro-driver - the structure MCP details and the full docking pipeline.
• tooluniverse-drug-drug-interaction - the deep DDI reasoning framework this orchestrator delegates to.