agentleFS
K-Dense-AIK-Dense-AI/drug-discovery-agent-skills/skills/chemical-space/SKILL.md

chemical-space

Skill56 starsChanged 3 months agoMITView on GitHub

Navigate make-on-demand catalogues — ZINC-22 through CartBlanche and Enamine REAL Space — to find compounds that can actually be ordered. Use this skill to look substances up by ZINC identifier or structure, understand tranche partitioning by heavy-atom count and logP, and choose between screening an enumerated subset and searching a combinatorial synthon space with a fragment-growing method such as V-SYNTHES. Also trigger on ZINC22, CartBlanche, Enamine REAL, make-on-demand, tangible library, synthon, tranche, giga-scale enumeration, or ultra-large virtual screening.

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Read references/zinc22-and-cartblanche.md before calling the API or downloading a tranche, references/combinatorial-spaces.md before …

    line 29

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Purchasable Chemical Space

Make-on-demand catalogues turned virtual screening from "the million compounds we own" into "the tens of billions a vendor will synthesise". ZINC-22 holds about 54.9 billion 2D structures drawn from Enamine, WuXi, and Mcule; Enamine REAL Space defines around 94 billion combinatorially. This skill is about finding compounds in that space that can actually be ordered — and knowing when the space is too large to enumerate at all.

Services: https://cartblanche.docking.org (per-substance JSON) · https://files.docking.org/zinc22 (bulk tranche tree). Both unauthenticated. Checked against: the live services, August 2026.

Read references/zinc22-and-cartblanche.md before calling the API or downloading a tranche, references/combinatorial-spaces.md before assuming a space can be enumerated, and references/screening-strategy.md before designing a cascade — that one is judgement, not syntax.

The two scripts

ScriptAnswers
cartblanche_lookup.pyWhat is this ZINC id, and can I buy it?
space_plan.pyWhich tranches do I need, and what will the screen cost?

The API lies about its own responses

This is the thing to get right. CartBlanche answers unknown routes with HTTP 200, a Content-Type: application/json header, and its React app's HTML:

RouteStatusContent-TypeBody
/substance/ZINC000000000053.json200application/jsonJSON
/substance.json?zinc_id=…200application/jsonHTML shell
/tranches.json200application/jsonHTML shell

Neither the status code nor the header can be trusted; only the body can. get_json() raises a named error when a response begins with <. The one reliable JSON route is /substance/<ZINC id>.json — treat CartBlanche as per-identifier lookup and use the file tree for anything bulk.

Identifiers are zero-padded to twelve digits: ZINC53 does not resolve, ZINC000000000053 does. The scripts accept either.

Can I buy it?

python skills/chemical-space/scripts/cartblanche_lookup.py substance ZINC19632618
zinc_id           smiles                              mwt      logp  heavy_atoms  purchasable  catalogs  min_price
ZINC000019632618  Cc1ccc(NC(=O)c2ccc(CN3CCN(C)CC3)...  493.615  4.59  37           true         281       240

That is imatinib, offered by 281 catalogues from 240 up. catalogs shows supplier, quantity, and lead time per entry.

Existing in ZINC is not the same as being purchasable. An empty catalogs block is a compound that was enumerated but that nobody offers. And a catalogue entry is an order, not a vial: make-on-demand synthesis fails roughly one time in five, with lead times of several weeks.

Which tranches, and what will it cost?

The bulk tree is addressed by heavy-atom count and logP, not molecular weight — H25P200 is 25 heavy atoms and logP 2.0, H25M100 is logP −1.0. A target profile expressed in MW cannot address the directories:

python skills/chemical-space/scripts/space_plan.py tranches --hac-min 18 --hac-max 26 \
    --logp-min 1 --logp-max 4

Then make the funnel explicit before committing compute:

python skills/chemical-space/scripts/space_plan.py cascade --library-size 1e9
stage              input          keep  survivors    core_hours  wall_days_at_cores
property filter    1000000000     0.3   300000000    138.9       0.01
fast dock          300000000      0.01  3000000      250000      10.42
standard dock      3000000        0.1   300000       25000       1.04
rescore / MM-GBSA  300000         0.1   30000        25000       1.04
visual triage      30000          0.2   6000         500         0.02

# 1,000,000,000 in, 6,000 out
# 300,639 core-hours total; 12.5 days on 1000 cores

The property filter is free and removes 70% — always first. Fast docking is 83% of the compute, so that is the only stage where a speedup matters. And 6000 survivors is still far more than anyone buys, so a real campaign ends with diversity selection and a budget cut.

Above 10⁸ compounds, stop enumerating

strategy reports the boundary, and the arithmetic behind it is unforgiving: 94 billion SMILES at 100 bytes is about 9 TB before conformers, and docking each for three seconds is roughly 78 000 core-years.

A combinatorial space is reagents plus reaction rules, not a list — so search it without enumerating it. Dock a minimal fragment library covering every scaffold and synthon, keep the best, and enumerate only those. V-SYNTHES2 reports this over 36 billion REAL Space compounds.

Four ways this misleads

  1. Bigger is sublinearly better. A thousandfold larger library buys closer to one log of affinity, and it does not improve enrichment — false positives scale with N too.
  2. Giga-scale finds your scoring function's blind spots reliably. High molecular weight, over-buried hydrophobics, and strained conformers presented as favourable are systematic artefacts, and a thorough search finds them precisely because it is thorough.
  3. Receptor quality does not scale away. A wrong protonation state costs the same fraction of the answer at any N. Preparing the structure beats enlarging the library.
  4. REAL Space is narrower than its size suggests. It is a product over amide couplings, Suzukis, and reductive aminations; unusual scaffolds are systematically absent.

When to stop using these tools

For similarity and substructure search across REAL Space without enumeration, use SmallWorld (https://sw.docking.org) and Arthor (https://arthor.docking.org). For the full synthon definition of REAL Space you need an agreement with Enamine — identifiers are public, the reagent and reaction files are not.

Composing with the rest of the bundle

  • binding-site-analysis → before: is the pocket worth a billion compounds at all?
  • medchem / rdkit → before: property filters and PAINS removal, at the top of the funnel.
  • autodock-vina → alongside: the docking engine the cascade is costed for.
  • retrosynthesis → alongside: for anything not in a catalogue, can it be made?
  • chembl → after: is a hit series already known against this target?

Reporting results honestly

Give the library actually screened, the keep-fraction at every stage, and whether a redocking control reproduced a known pose. A docking score is not an affinity and a rank is not a potency prediction. Report compounds ordered and compounds received — a hit rate computed against orders rather than deliveries is wrong by about a fifth.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.