Paramus — Selectivity & Off-Target Screen

Overview

Predict anti-target activity and profile similarity to known liabilities, then flag selectivity concerns and a safety tier. The value is the fused off-target risk view, not a single prediction.

Guidance vs execution: anti-target predictions come from tested BRAIN models with provenance. Do not assert off-target risk from model memory.

When to Use

• Assessing selectivity / safety-pharmacology risk of hits or leads
• Flagging hERG / CYP / promiscuity concerns before advancing Do not use for: definitive tox prediction (hand off to the regulatory tox workflow).

Workflow

Candidate set + (optional) primary target
  ↓ 1. Parse & standardize        → Paramus cheminformatics tool
  ↓ 2. Anti-target predictions    → BRAIN models (hERG, CYP isoforms, key receptors)
  ↓ 3. Liability similarity       → similarity to known off-target/toxicophore sets
  ↓ 4. Selectivity assessment     → primary vs anti-target margin (if primary given)
  ↓ 5. Report                     → per-compound risk flags + tier + provenance

Procedure

1. Discover tools via search/get_schema ("hERG prediction", "CYP inhibition", "similarity search").
2. Standardize inputs; quarantine failures.
3. Predict anti-target activities; capture confidence + model versions.
4. Profile similarity to known liability sets; flag close matches.
5. If a primary target activity is provided, compute the selectivity margin.
6. Report risk flags, selectivity margin, and a safety tier with rationale + versions.

Domain Judgment

• A predicted off-target hit is a flag to test, not a verdict — state confidence.
• Selectivity is a margin, not a binary; report the ratio, not just pass/fail.
• Structural similarity to a liability is a warning even absent a model hit.

Fallbacks

• An anti-target model missing → screen on available anti-targets; mark gaps.
• Endpoint unreachable → stop and report; no local off-target heuristics.