Medicinal Chemistry Deterministic: ADMET prediction, structural alerts, and selectivity scoring Medicinal chemists and DMPK/tox scientists de-risking a lead series before candidate selection
Catch developability and off-target liabilities early — flag ADMET red flags and predicted selectivity issues while chemistry can still fix them.When it applies: a potent lead series needs a fast, structured liability read before expensive in vivo studies commit the program to a scaffold.

Scenario

A potent lead series looks great on target but may hide developability liabilities — poor absorption, rapid clearance, hERG risk, or off-target activity. Discovering these in vivo is slow and expensive. The team wants a structured early read so chemistry can design the liabilities out.

Worked example. A CNS series with strong potency must be checked for hERG and permeability risk before committing to an analog campaign.

Agent workflow

  1. Read the series (SMILES).
  2. ADMET liability triage — predict key ADMET endpoints and flag structural alerts / reactive groups.
  3. Selectivity & off-target screen — profile predicted off-target activity against a relevant panel.
  4. Aggregate into a per-compound liability scorecard.

Demo output

A liability scorecard per compound with green/amber/red flags across ADMET and selectivity, and the structural feature driving each flag. The magic moment: a top-potency compound is flagged for a reactive motif and hERG risk before it ever reaches an animal study.

Deterministic vs LLM

  • Deterministic — the ADMET predictions, structural-alert matching, and selectivity scoring.
  • LLM — the plain-language liability summary and the mitigation strategy.

Every flag traces to a validated model or rule. The agent reads and routes; the engine decides.

Limits

Predictive de-risking, not a replacement for measured DMPK and safety pharmacology. It focuses the assay panel; the lab confirms the profile.