Analyze the SAR of a measured compound series

You have a table of analogs with measured potencies. Drive the SAR Analysis skill to find the shared scaffold, decompose each molecule into core + R-groups, map substituent to activity, and flag the activity cliffs that should steer the next synthesis round.

   
Problem class Data analysis
Subject areas Chemistry, Drug Repurposing and Discovery
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

A medicinal-chemistry project accumulates a series of close analogs with measured IC50 / Ki / EC50 values, and the recurring question is which structural change moved potency, and in which direction? Answering it by eye across dozens of SMILES is slow and inconsistent: the shared scaffold has to be identified, each molecule split into a common core plus its variable substituents, the substituents lined up at matching R-positions, and the potency deltas read off — the same maximum-common-substructure and R-group-decomposition logic every time. The high-value output is the activity cliff: a pair of near-identical structures with a large potency gap, which pinpoints the substituent position worth exploiting. Solved looks like: a CSV of analogs in, a scaffold + R-group-vs-activity table out, activity cliffs flagged, in a few minutes on a laptop — reproducibly, so two chemists analyzing the same series get the same decomposition.

This is the retrospective, analyze-what-you-measured counterpart to the forward enumerate-analogs recipe, which proposes new structures to make.

  1. Install the SAR Analysis skill (SciAgent-Skills; runs its Python locally via RDKit). Clone the collection and load it as a plugin:

    git clone https://github.com/jaechang-hits/SciAgent-Skills
/plugin install sciagent-skills

    It declares its own dependencies (RDKit, pandas) in SKILL.md; install them when prompted, and pin them (rdkit, pandas) in a requirements.txt you commit.

  2. Prepare an input CSV with one row per compound — a smiles column, a compound id, and a numeric activity column (IC50/Ki in nM, or pIC50). Standardize the structures first if they came from mixed sources (the datamol skill strips salts and neutralizes charges); a clean, deduplicated table makes the scaffold detection robust.

  3. Have the assistant author a versioned analysis script, not an interactive chat. A prompt that lands a durable artifact:

    Use the sar-analysis skill on series.csv (columns: id, smiles, ic50_nM).
Write the analysis to a committed script sar_analysis.py that:
  1. Converts ic50_nM to pIC50 for the activity scale.
  2. Finds the maximum common substructure (rdFMCS, threshold 0.8)
     and reports the shared scaffold as SMARTS.
  3. Runs R-group decomposition against that core; emits a table of
     core + R1/R2/... substituents per compound.
  4. Joins substituents to pIC50 and, for each R-position, tabulates
     substituent -> mean pIC50 so the SAR trend at each position is
     explicit.
  5. Flags activity cliffs: pairs differing at a single R-position
     with |delta pIC50| >= 1.0 (>=10x potency change).
Save sar_table.csv, activity_cliffs.csv, and the skill's HTML report
to results/. Do not summarize beyond what those files contain.

  4. Run the script and read the outputs. sar_table.csv is the core-plus-R-group decomposition; activity_cliffs.csv is the shortlist of single-change pairs with large potency swings — those are the transformations to prioritize or avoid next. The HTML report gives aligned 2D structures with the activity heatmap for a figure. Any written SAR narrative must cite only rows in those tables.

  5. Record provenance. Emit a provenance.json capturing the RDKit and pandas versions, the SAR-skill commit, the MCS threshold and cliff cutoff you used, the input CSV’s sha256, the run date, and the model id — so the decomposition can be re-run and audited. Commit sar_analysis.py, requirements.txt, results/, and provenance.json. See the reproducibility guide for the artifact pattern.

Why this assembly

Rung 2 of the simplicity ladder. The whole task — MCS detection, R-group decomposition, activity alignment, cliff flagging — is exactly what the one SAR Analysis skill wraps over RDKit, so a single skill solves it. Rung 1 (Claude Code alone) loses RDKit’s validated rdFMCS/rdRGroupDecomposition and risks a scaffold the model eyeballs incorrectly, which corrupts every downstream R-group column. Rung 3+ (adding a QSAR model or an autonomous system) answers a different question — predict the potency of an unmade analog — and is overkill for reading the SAR of a series you already measured. Start here; escalate only when you want quantitative prediction of new compounds.

Availability

Fully open. The SAR Analysis skill ships in the CC-BY-4.0 SciAgent-Skills collection and runs entirely locally over RDKit and pandas — your structures and assay data never leave the machine, so there is no subscription or data-residency gate. Any current Claude plan suffices.

Compute requirements

Laptop. MCS and R-group decomposition over a series of tens to a few hundred analogs is CPU-only and completes in seconds to a minute; runtime is dominated by the skill’s reasoning, not the RDKit calls. No GPU. Very large series (thousands of compounds) with an expensive MCS threshold can take longer — subset to one chemotype at a time if so.

Evidence

Reported. No published paper documents this exact Claude + SAR-Analysis-skill assembly, but its two layers are both grounded. (1) The underlying methods are field-standard for lead optimization: R-group decomposition and matched-pair-style single-substituent comparison are the canonical way to read SAR and detect activity cliffs, established across the medicinal-chemistry literature and reaffirmed in current work — context-based matched-pair analysis of a ChEMBL CYP1A2 set to guide lead optimization (Raut & Dixit, RSC Med. Chem. 2025), matched-pair/matched-series bioactivity models reaching R² ≈ 0.83 for potency change (Ding et al., Curr. Med. Chem. 2020), and the “retro-optimization” matched-pair-network framing of optimization logic (Kombo & LaMarche, J. Med. Chem. 2025). (2) The skill is part of the SciAgent-Skills collection evaluated on BixBench, and it calls RDKit’s validated rdFMCS/rdRGroupDecomposition — so the assembly inherits those routines’ correctness. What is not separately benchmarked is the Claude-driven path versus a hand-coded RDKit script; they call the same functions.

Alternatives considered

  • Hand-coded RDKit script. The same rdFMCS + rdRGroupDecomposition in a notebook works for a one-off. Reach for it when you are scripting a batch pipeline across many series. The skill’s value is doing the decomposition, alignment, and cliff-flagging conversationally and emitting the aligned-structure HTML report without boilerplate.
  • Enumerate new analogs instead of analyzing measured ones. If the question is what should we make next rather than what did our measurements tell us, use the enumerate-analogs recipe (forward, hypothesis-generation) — these are siblings.
  • Property-focused medicinal-chemistry filtering. To flag structural-alert / drug-likeness liabilities across the series rather than potency SAR, the medchem skill and the filter-virtual-screening-hits recipe are the better fit.
  • True matched-molecular-pair transformation networks with search-time rules. The skill does single-position R-group comparison, not a full MMP transformation database (e.g., mmpdb). For systematic transformation-rule mining across a large corpus, a dedicated MMP tool is the right escalation — none is catalogued as a Claude-installable component today.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.