Prepare the correct protonation state of a ligand before docking

Hand Claude Code a list of ligand SMILES; get back each compound in its dominant protonation (and, where relevant, tautomeric) state at physiological pH, with the macro-pKa values that justify the assignment — so the structures you feed into docking or MD carry the right formal charges instead of whatever state the drawn SMILES happened to encode.

   
Problem class Experimental design
Subject areas Chemistry, Drug Repurposing and Discovery, Integrative Structural and Computational Biology
Evidence level Proposed
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

A SMILES string carries whatever protonation state the chemist happened to draw — a carboxylic acid keyed as neutral C(=O)O, a basic amine keyed as neutral N, a zwitterion drawn as either extreme. Docking and molecular dynamics do not fix this for you: they place the atoms and charges they are given. Feed an unionized carboxylate into a pose where the biological reality is a −1 anion salt-bridging an active-site lysine, and the score, the pose, and every downstream interpretation are wrong. The effect is not marginal — protonation and tautomeric state are well documented to swing virtual-screening enrichment and pose ranking (Kim et al., J. Comput. Aided Mol. Des. 2013).

The recurring, unglamorous prep step is therefore: for each ligand, decide the dominant microspecies at pH ≈ 7.4, assign the right formal charges, and note the pKa that justifies it (so a reviewer can see why a group was protonated). Solved looks like: a list of SMILES in, a table out with each compound’s physiological-pH SMILES, net charge, and governing macro-pKa values — reproducibly, on a laptop, before a single docking job is launched.

  1. Install the Rowan skill and set an API key. Rowan runs the pKa/macro-pKa models as cloud workflows; the free tier at labs.rowansci.com covers this recipe.

    npx skills add K-Dense-AI/scientific-agent-skills
export ROWAN_API_KEY="your_api_key_here"

    Enable the rowan skill when prompted. Confirm the key is visible to the session before running any workflow.

  2. (Optional, local) canonicalize the input first with the Datamol skill. Strip salts, keep the largest fragment, and canonicalize tautomers so the pKa call operates on a clean parent. This step is laptop-local and free; do it when the input SMILES come from a messy vendor export. Skip it for hand-drawn, already-clean leads.

  3. Write the protonation-prep step to a versioned script. Have Claude Code author prepare_protonation.py — do not leave the workflow as an interactive chat. A minimal prompt:

    Input: ligands.csv with columns id, smiles.

Write prepare_protonation.py that, for each row:
  1. (if datamol available) standardize the SMILES: largest
     fragment, canonical tautomer, neutralize, canonical SMILES.
  2. Use the Rowan skill: submit rowan.submit_macropka_workflow
     over pH 0-14; wait; retrieve. From the result take the
     dominant microspecies at pH 7.4 and its SMILES, and the
     macro-pKa values (acidic and basic) that bracket it.
  3. Record: id, input_smiles, prepared_smiles_pH7.4, net_charge,
     macro_pka_acidic, macro_pka_basic, dominant_fraction.
Emit prepared_ligands.csv sorted by id. Print a one-line summary
per ligand: id -> net_charge (governing pKa).

    Have Claude submit workflows in a small batch with polite waits; macro-pKa runs take 5–15 minutes each, so cache retrieved results by input InChIKey so a re-run does not re-submit.

  4. Record provenance. Emit provenance.json capturing the Rowan skill version/tag, the macro-pKa workflow/model identifier and its Rowan-side workflow IDs, the rowan-mcp/skill package version, the pH used (7.4), the run date, the sha256 of ligands.csv and prepared_ligands.csv, and the model id driving the session. Cloud workflow results are not byte-reproducible — the recorded workflow IDs and date are what make a later divergence visible.

  5. Hand off the prepared table. prepared_ligands.csv is the durable artifact. Feed prepared_smiles_pH7.4 straight into the Dock a ligand library with DiffDock or Rank a compound library by predicted binding affinity recipes, or into Set up a protein MD simulation in GROMACS. Commit prepare_protonation.py, a pinned requirements.txt, prepared_ligands.csv, and provenance.json to version control. Follow the reproducibility guide for the artifact layout.

Why this assembly

Rung 2 of the simplicity ladder. Physiological-pH protonation-state assignment needs a real pKa model — the RDKit/Datamol standardization step (rung 1) canonicalizes charges and tautomers but does not predict pH-dependent ionization, so it cannot tell you a carboxylate is anionic at 7.4. The Rowan skill exposes exactly one primitive that closes the gap: submit_macropka_workflow, which returns the pH-resolved microspecies distribution and macro-pKa values. One skill, one workflow — no orchestration across heterogeneous APIs (rung 3) and no autonomous chemist (rung 4) is warranted. The Datamol pre-clean is an optional laptop-local convenience, not a second essential tool.

Availability

Fully open. The Rowan skill ships in the K-Dense-AI/scientific-agent-skills collection (community OSS); the optional Datamol skill is MIT. Rowan’s pKa/macro-pKa models run as cloud workflows and require a free Rowan API key (ROWAN_API_KEY; free tier at labs.rowansci.com) — no subscription or institutional licence, but the compute runs on Rowan’s servers, so SMILES leave your machine. If ligand structures are confidential, treat this as a data-residency caveat and confirm your project’s policy before submitting. The rowan-mcp repo does not publish a LICENSE file; confirm licensing before redistributing the wrapper.

Compute requirements

Laptop. All heavy computation runs cloud-side on Rowan; your machine only submits and polls. Budget wall-clock per ligand: microscopic pKa 2–5 min, macro-pKa 5–15 min, at roughly 5–15 Rowan credits each. For a 100-ligand batch, submit concurrently and cache by InChIKey; the optional Datamol standardization pass is sub-second per molecule on a single CPU core. No GPU, negligible local RAM.

Evidence

Proposed. No published benchmark of an LLM-driven Rowan protonation-prep workflow is known. The closest documented evidence is component-level and problem-level: Rowan’s own platform documents microscopic and macro-pKa prediction workflows (Rowan Labs), and the impact this step addresses is well established — protonation, tautomeric, and receptor ionization states are repeatedly shown to change virtual-screening enrichment and pose ranking (Kim et al., J. Comput. Aided Mol. Des. 27:235, 2013). Each component (a pKa model; a docking or MD downstream) has independent validation; the agent-orchestrated prep assembly does not. Treat prepared_ligands.csv as a defensible starting state to be sanity-checked (does the net charge match chemical intuition for each ionizable group?), not an oracle.

Alternatives considered

  • Datamol / RDKit standardization alone (rung 1). Neutralizes charges and canonicalizes tautomers but is pH-blind — it will happily hand a neutral carboxylic acid to docking. Use it as the pre-clean here, not as the protonation decision. See Enumerate analogs around a lead for that standardization surface.
  • A local pH-aware protonator (e.g. Dimorphite-DL, OpenBabel -p). These enumerate plausible protonation states offline and keep SMILES on your machine — attractive when structures are confidential. None is currently a Claude-installable skill or MCP in catalog/tools/, so the recipe cannot recommend them yet; a note is filed for the catalog curator. Reach for one manually if data residency forbids cloud submission.
  • Skipping prep and docking the drawn state. Defensible only for neutral, non-ionizable scaffolds. For anything with a carboxylate, a basic amine, or an ambiguous tautomer, skipping this step silently biases every downstream score.

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.