cmxflow 🧪

Build cheminformatics and computational chemistry pipelines with composable blocks. Tune end-to-end with Bayesian Optimization. Or ask an LLM agent to do it.

Quick examples

Prepare ligands for docking

server.ts

from cmxflow import Workflow
from cmxflow.sources import MoleculeSourceBlock
from cmxflow.operators import (
    MoleculeStandardizeBlock,
    IonizeMoleculeBlock,
    EnumerateStereoBlock,
    ConformerGenerationBlock,
)
from cmxflow.sinks import MoleculeSinkBlock

# Standardize → ionize (pH 6.4–8.4) → enumerate stereo → generate 3D conformers
workflow = Workflow()
workflow.add(
    MoleculeSourceBlock(),
    MoleculeStandardizeBlock(),
    IonizeMoleculeBlock(),
    EnumerateStereoBlock(),
    ConformerGenerationBlock(),
    MoleculeSinkBlock(),
)
workflow("library.smi", "prepared.sdf")

Dock a congeneric series

Pure-Python docking. Free docking is the default (index_poses=False); scaffold-indexed mode caches poses by Bemis–Murcko scaffold for ~3× faster throughput on congeneric series with consistent pose alignment.

server.ts

from cmxflow import Workflow
from cmxflow.sources import MoleculeSourceBlock
from cmxflow.operators import ConformerGenerationBlock, MoleculeDockBlock
from cmxflow.sinks import MoleculeSinkBlock
from cmxflow.utils.parallel import make_parallel

workflow = Workflow()
workflow.add(
    MoleculeSourceBlock(),
    ConformerGenerationBlock(),
    make_parallel(
        MoleculeDockBlock(
            receptor="receptor.pdb",
            site_reference="crystal_ligand.sdf",
            index_poses=True,  # omit for free docking
        )
    ),
    MoleculeSinkBlock(),
)
workflow("library.smi", "docked.sdf")

Tune a ligand-based virtual screen

server.ts

from cmxflow import Workflow
from cmxflow.sources import MoleculeSourceBlock
from cmxflow.operators import MoleculeSimilarityBlock
from cmxflow.scores import EnrichmentScoreBlock
from cmxflow.opt import Optimizer

# Rank a library by 2D similarity to a known active, then tune the
# fingerprint end-to-end to maximize enrichment AUC.
workflow = Workflow()
workflow.add(
    MoleculeSourceBlock(),
    MoleculeSimilarityBlock(queries="crystal_ligand.sdf"),
    EnrichmentScoreBlock(target="active"),
)

opt = Optimizer(workflow, "benchmark.csv")
opt.optimize(n_trials=30, direction="maximize")

print(f"Best enrichment AUC: {opt.best_score:.3f}")
print(opt.best_params)
# Best enrichment AUC: 0.836
# {'fingerprint_type': 'morgan', 'similarity_metric': 'sokal', 'radius': 2, 'nbits': 2545}

The four fingerprint parameters above are searched automatically — every block exposes its mutable parameters to the optimizer.

Or build it conversationally via an LLM agent

Terminal

claude mcp add cmxflow -- cmxflow-mcp

"How many of the molecules in library.csv pass Lipinski's rules?"

"I need to build a ligand-based virtual screening workflow. I'm not sure if 2D or 3D is better. Can you optimize two workflows?"

"Dock the molecules in hits.csv against receptor.pdb with crystal_ligand.sdf as a reference."

The agent can build, run, and optimize workflows. See Using with Claude for full transcripts.

What's in the box

15+ blocks for sourcing, transforming, filtering, clustering, scoring, and docking molecules
Bayesian optimization of pipeline parameters via Optuna
Parallel execution for compute-heavy blocks (conformer generation, docking)
Workflow serialization for save / load / reuse
An MCP server with five tools: build_workflow, run_workflow, optimize_workflow, manage_workflows, view_structures

Install

Terminal

pip install cmxflow

MCP server

Terminal

claude mcp add cmxflow -- cmxflow-mcp

Optional: PyMOL

Required only for the view_structures MCP tool (3D visualization):

bash

conda install -c conda-forge pymol-open-source

Documentation

Docs site
Block catalog
Using with Claude — agent transcripts
examples/basic_usage.ipynb — full tutorial
examples/docking/docking.ipynb — docking walkthrough (ILS, scaffold-indexed, and template modes)

Project

MIT licensed. See CONTRIBUTING.md and RELEASING.md.