Scenario

A screening campaign returns dozens of hits against a target. Progressing all of them is impossible, and eyeballing 2-D structures misses the real binding and potency signal. The team needs a reproducible triage that ranks hits by predicted binding quality and groups them into chemically coherent series before any medicinal chemistry starts.

Worked example. A kinase program has 60 fragment hits; the goal is to pick the two or three scaffolds worth optimizing.

Agent workflow

1. Read the hit list (SMILES + assay readouts).
2. Docking pose screen — dock each hit, score the pose, and flag poses that make the key hinge interactions.
3. SAR series analysis — cluster hits by scaffold, compute descriptors, and surface activity cliffs and the emerging SAR trend within each cluster.
4. Rank series by a combined pose-quality and SAR-consistency signal.

Demo output

A ranked short-list of lead series (not just single molecules), each with a representative pose, the shared scaffold, and the within-series SAR trend. The magic moment: two scaffolds with weak individual hits reveal a strong, consistent SAR once clustered — a signal invisible in the raw hit list.

Deterministic vs LLM

• Deterministic — docking pose scores, RDKit descriptors, and the SAR clustering / activity-cliff math.
• LLM — the series-progression narrative and the recommendation of which scaffolds to advance.

Every score and cluster is a validated calculation. The agent reads and routes; the engine decides.

Limits

Screening-level prioritization, not a substitute for confirmatory assays, free-energy perturbation, or crystallography. It narrows the series; the bench confirms the lead.