molforge

The connective tissue for computational biology — fold, dock, simulate, and design across sixteen engines with one data model and no format-conversion tax.

Get started Quickstart GitHub

pip install molforge

It is a library, not a framework: no orchestrator, no DAG runtime, no decorators to import. Bring your structures and sequences in, plug in your engines of choice, and walk out with a coherent pipeline instead of five incompatible Python environments and a graveyard of conversion scripts.

What it does

•   Folding

  ---

  ESMFold, AlphaFold/ColabFold, Boltz, Chai-1, RoseTTAFold — with multi-component cofolding on the AF3-style engines.

•   Docking

  ---

  AutoDock Vina, Gnina (CNN rescoring), DiffDock — with automatic meeko/RDKit ligand prep.

•   Molecular dynamics

  ---

  OpenMM, GROMACS, AMBER behind one prepare → minimize → run interface, plus trajectory I/O and analysis.

•   Generative design

  ---

  RFdiffusion for backbones, ProteinMPNN and ESM-IF1 for sequence design — the full de novo loop in one library.

•   Binding free energy

  ---

  MM-PB(GB)SA via AmberTools and gmx_MMPBSA, FEP/TI ingestion through alchemlyb and cinnabar, plus Boltz-2 binding-affinity prediction.

•   Pocket detection

  ---

  fpocket (geometric) and P2Rank (ML) surface-pocket detection, feeding straight into the docking workflow.

The glue — an engine-agnostic layer

•   Cross-engine ensembles

  ---

  cross_engine_fold folds one sequence with several engines and returns the pairwise TM/RMSD spread, a consensus, and a per-residue map of where they disagree — trust the regions your methods agree on.

•   Design loop

  ---

  DesignLoop runs generate → fold → dock → score → iterate with a ranked design table — the protein-engineering loop in one object.

•   Unified scoring

  ---

  molforge.scoring gives every score an explicit direction, so docking affinity, folding confidence, and learned scores rank and compare uniformly — and plug into the design loop.

•   Reproducibility

  ---

  Every output carries a Provenance chain; emit it as a citable pipeline.yaml, then replay() re-runs the whole workflow.

Why molforge

• Workflows over silos

  ---

  Every design decision is judged by one question: does this make it easier to chain N tools together?

• Wrappers, not reimplementations

  ---

  We don't rebuild OpenMM or AutoDock. We give them a shared vocabulary — one Protein, one Provenance, one cache.

• One data model, two views

  ---

  Hierarchical (protein.chains["A"].residues[42]) for biology, linear (protein.atom_array.coords) for ML — same data, no conversion.

• Typed, tested, documented

  ---

  Strict mypy, ruff-clean, a large test suite on a 3-OS × 3-Python matrix, every public symbol with a Google-style docstring.

Where to go next

• New here? Installation and the Quickstart.
• Trying to do something specific? The Cookbook has task-oriented recipes and decision tables for choosing engines.
• Want the design rationale? The Architecture overview and Roadmap.
• Looking for a symbol? Browse the API reference or hit the search box.

---

molforge is MIT-licensed. Issues and pull requests are welcome at github.com/DoctorDean/molforge.