Paramus — ADMET Liability Triage

Overview

Turn "which of these compounds are worth keeping?" into a reproducible parse → descriptors → flags → ML ADMET → rank pipeline. The value is the fused triage judgment, not any single prediction.

Guidance vs execution: every number comes from a tested Paramus tool. Do not hand-roll RDKit or a property model — route to the tools.

When to Use

• A library/set of molecules to prioritize, filter, or triage
• Early hit-to-lead developability screening Do not use for: single-molecule deep study, or final regulatory ADMET reporting.

Workflow

Candidate set (SMILES / SDF / IDs)
  ↓ 1. Parse & standardize        → Paramus cheminformatics tool
  ↓ 2. Physicochemical descriptors → MW, logP, TPSA, HBD/HBA, rotatable bonds
  ↓ 3. Rule-based flags            → Lipinski / Veber / PAINS / reactive groups
  ↓ 4. ML ADMET predictions        → BRAIN models (solubility, hERG, CYP, clearance…)
  ↓ 5. Score & tier                → Keep / Review / Drop
  ↓ 6. Report with provenance      → per-compound rationale + model versions

Procedure

1. Discover tools via search/get_schema ("molecular descriptors", "ADMET prediction", "standardization").
2. Parse & standardize every input; quarantine parse failures — never silently drop.
3. Compute descriptors (batched if supported).
4. Apply rule-based flags (cheap deterministic gates).
5. Run ML ADMET; capture each prediction's confidence/units + model version.
6. Tier with an explicit, echoed rule; report a table with rationale + versions.

Domain Judgment

• Flags are gates, predictions are signals: a PAINS/reactive hit drops regardless of a good score.
• Low-confidence ML prediction → route to Review, not Drop.
• Batch, don't loop — consistent model versions across the set.

Fallbacks

• A model unavailable → triage on available signals; mark the missing dimension. No surrogate models.
• Endpoint unreachable → stop and report; no local RDKit/model fallback.