mdprep

mdprep 0.4.0 prepares reproducible Amber molecular-dynamics systems from a PDB structure and a validated YAML manifest. It supports standard proteins, independent ligands and cofactors, explicit metal models, optional active-site QM/MM refinement, final tleap assembly, validation, and an optional OpenMM equilibration/production stage.

The manifest is the source of truth. Chemistry-sensitive decisions—ligand charge and multiplicity, catalytic protonation, metal oxidation/spin state, coordinating bonds, and QM components—must be supplied explicitly and are recorded in the output reports.

Highlights

• ff14SB or ff19SB proteins with TIP3P or OPC water.
• PropKa protonation, manual catalytic overrides, disulfides, and optional GFN2-xTB/g-xTB HID–HIE selection.
• Multiple GAFF or GAFF2 ligands/cofactors with AM1-BCC, user mol2/frcmod, PySCF gas-phase RESP, or electrostatically embedded PySCF QMMESP charges.
• Amber nonbonded metal models or bonded MCPB.py models with explicit coordinating atoms, charges, and multiplicities.
• Dimers, trimers, tetramers: every protein chain in one system. Copies of one protein can share parameters -- asked about once, matched residue for residue by sequence and position rather than number, and a bonded metal site fitted once and reused in every copy (docs/oligomers.md).
• MCPB Seminario Hessians from PySCF, GFN2-xTB, g-xTB, or MACE-POLAR-1; PySCF supplies the embedded large-model ESP/RESP calculation.
• Optional ASH/OpenMM active-site QM/MM refinement before final parameterization.
• Final prmtop, inpcrd, and PDB generation through tleap, followed by topology, coordinate, charge, and optional OpenMM energy validation.
• Optional Roe–Brooks OpenMM minimization, heating, density equilibration, and manifest-sized production MD with GPU-first/CPU-fallback execution.
• Predicted models: the closest X-ray structure from the RCSB, and its crystal waters transplanted onto the model with PyMOL -- a water on the metal moves with it.
• An official Colab notebook that prepares, equilibrates, and shows every step: charges, metal-site parameters, the equilibration, the trajectory.

Installation

Conda or Mamba is recommended because the real workflows use AmberTools, PropKa, xTB, OpenMM, ParmEd, and PySCF:

git clone https://github.com/miqueleg/mdprep.git
cd mdprep
mamba env create -f environment.yml
conda activate mdprep
mdprep selftest --quick

For a lighter Python-only installation, which does not install external chemistry executables:

python -m pip install .

Optional Python extras are available as .[md] for OpenMM and .[qm] for PySCF/ParmEd. External commands are discovered through PATH; manifests may also name a reviewed relative executable. No developer-machine path is embedded in the distributed examples.

Basic example

Start from a standard protein PDB:

mdprep inspect input.pdb
mdprep plan input.pdb                           # what must you decide?
mdprep init input.pdb --interactive -o system.yaml

mdprep plan lists only the manifest decisions this structure actually requires, and --interactive asks them and writes a validated manifest. See Guided manifest generation. The non-interactive mdprep init input.pdb -o system.yaml still writes the commented starter manifest.

A minimal protein-only system.yaml is:

project:
  name: my_system
  input_structure: input.pdb
  output_dir: prepared/my_system

structure:
  keep_crystal_waters: true
  altloc_policy: highest_occupancy
  remove_unknown_heterogens: false
  preserve_chain_ids: true
  remove_input_hydrogens: true

protein:
  forcefield: ff14SB
  water_model: TIP3P

protonation:
  ph: 7.0
  method: manual_only
  overrides: []

disulfides:
  auto_detect: true
  detection_cutoff_angstrom: 2.2
  force: []
  forbid: []

ligands: []
metals: []

solvation:
  enabled: true
  box: truncated_octahedron
  buffer_angstrom: 10.0
  neutralize: true
  salt_concentration_molar: 0.15
  positive_ion: Na+
  negative_ion: Cl-

validation:
  run_openmm_energy_check: true
  fail_on_warnings: false
  fail_on_missing_parameters: true
  fail_on_noninteger_ligand_charge: true

molecular_dynamics:
  enabled: false

The same validated manifest is available as examples/00_basic_protein.yaml.

Validate and prepare it:

mdprep config-check system.yaml
mdprep prepare system.yaml
mdprep validate \
  prepared/my_system/final/system.prmtop \
  prepared/my_system/final/system.inpcrd

Relative paths are resolved from the manifest directory. Unknown heterogens are not deleted or guessed: configure them under ligands or metals, or explicitly request their removal.

For a guided workflow, use the web input builder: it shows the structure in 3D, asks only the questions your structure needs, and writes a validated manifest and a bundle. Then open the official Colab notebook (examples/mdprep_colab.ipynb), upload the bundle, and it installs what the manifest needs, prepares the system, equilibrates it, and lets you investigate every step: the structure at each stage, the protonation states, the charge on every atom, the metal site's force constants, how the box equilibrated, and the MD trajectory.

No crystal structure of your protein? Predict the complex with its cofactors (Boltz or AlphaFold3 in ColabFold), then let mdprep find the closest X-ray structure and bring its crystal waters across -- including a water bound to the metal, which a prediction never has:

mdprep closest-xray model.cif
mdprep transplant-waters model.cif --reference 2FCT

See Predicted structures.

Common workflows

Generate a starter manifest using PropKa and ff19SB/OPC:

mdprep init input.pdb -o system.yaml \
  --forcefield ff19SB \
  --water-model OPC \
  --protonation-method propka

Stop after a stage while reviewing a complex system:

mdprep prepare system.yaml --stop-after structure
mdprep prepare system.yaml --stop-after protonation --overwrite
mdprep prepare system.yaml --stop-after refinement --overwrite
mdprep prepare system.yaml --stop-after ligands --overwrite
mdprep prepare system.yaml --stop-after metals --overwrite
mdprep prepare system.yaml --stop-after tleap --overwrite

Run the manifest-configured Roe–Brooks protocol from an existing Amber build:

mdprep run-md system.yaml \
  --prmtop prepared/my_system/final/system.prmtop \
  --inpcrd prepared/my_system/final/system.inpcrd

When molecular_dynamics.enabled: true, an ordinary mdprep prepare continues through MD. Production length and trajectory/state/checkpoint intervals are specified in steps in the manifest.

Where to go next

All public example manifests are schema-validated in the test suite.

Outputs and reproducibility

A complete run preserves:

• the input and locked manifests;
• normalized/protonated/refined intermediate PDBs;
• ligand, QM, MCPB.py, and tleap inputs and logs;
• final Amber topology, coordinates, and PDB;
• structure, protonation, ligand, metal, tleap, and validation reports;
• executable and Python-package version information;
• optional MD PDBs, density history, trajectory, state log, and restarts.

External failures are reported with command, working directory, return code, stdout, stderr, and runtime. mdprep does not silently fall back to another charge model, protonation, force field, QM method, or metal model.

Current limitations

mdprep does not support noncanonical amino acids inside peptide chains, covalent ligands, automatic loop modeling (a missing loop is left as two termini), a metal-bound hydroxide, ORCA, Multiwfn, more than one independently fitted bonded MCPB site (copies of one site in an oligomer share a single fit), or a bonded MCPB site bound by two chains. Unsupported chemistry fails explicitly.

Development and release checks

mamba env create -f environment-dev.yml
conda activate mdprep-dev
pytest -q
python -m mdprep.cli --help
python -m mdprep.cli config-check examples/*.yaml
python -m mdprep.cli selftest --quick
python -m hatchling build

See CONTRIBUTING.md, CHANGELOG.md, and CITATION.cff. mdprep is distributed under the MIT license.