flexcraft

flexcraft is a library for combining protein generative models, sequence design and structure predictors to write end-to-end, single-script protein design pipelines à la BindCraft. Currently, flexcraft mostly relies on salad for backbone generation, ProteinMPNN-variants for sequence design and AF2 for structure prediction and filtering. It also provides access to PyRosetta for additional physics-based filtering.

flexcraft is currently in the early stages of development. Documentation is sparse and APIs might change, however, flexcraft already provides a number of end-to-end protein design pipelines that are ready to use.

Getting started

Installation

To install flexcraft and salad, simply clone this repository and follow the steps:

# clone repository
git clone https://github.com/mjendrusch/flexcraft.git
# set up environment
conda create -n flexcraft python=3.10
conda activate flexcraft
# install flexcraft, salad & dependencies
cd flexcraft
pip install -e .
# download model parameters
bash download_params.sh
# optionally install PyRosetta
pip install pyrosetta-installer
python -c 'import pyrosetta_installer; pyrosetta_installer.install_pyrosetta()'

Getting your 1st protein

Check if flexcraft is working properly on your machine / GPU by having it design a protein de novo:

python -m flexcraft.pipelines.denovo --num_aa "100" --num_designs 10 --out_path test_runs/denovo-100-1/

If your machine has more than 1 GPU, consider using CUDA_VISIBLE_DEVICES=0 to restrict the script to the first GPU on the machine.

This will create a directory test_output with the following contents:

• attempts/*.pdb: output designs from salad prior to sequence design
• fail/*.pdb: designs that failed AF2 filters after ProteinMPNN sequence design
• success/*.pdb: designs that successfully passed AF2 filters after ProteinMPNN sequence design
• all.csv: AF2 scores and sequences for all ProteinMPNN-designed sequences tested
• success.csv: AF2 scores and sequences for all designs passing AF2 filters

This directory structure is common to all pipelines currently implemented in flexcraft.

Pipelines

flexcraft implements customizable pipelines for a number of protein design tasks:

• motif scaffolding:
  • flexcraft.pipelines.graft.simple: pipeline for simple motif scaffolding tasks with small motifs, no symmetry requirements and no large inter-motif distances
  • flexcraft.pipelines.graft.general: pipeline for more involved motif scaffolding tasks, with shape or symmetry requirements, large motifs (>50 residues)
• binder design:
  • flexcraft.pipelines.binders
• unconditional de novo design:
  • flexcraft.pipelines.denovo

How to write flexcraft pipelines

For now, flexcraft is rather sparsely documented. Have a look at scripts/ and flexcraft/pipelines for examples on how to write pipelines. In general, flexcraft is built on top of the DesignData class in flexcraft.data, which can be used as input for all the models wrapped in flexcraft.

E.g. for running AlphaFold predictions (see the script at the bottom of flexcraft.structure.af._model):

# given a DesignData object
af_data: DesignData = ...
# make an AlphaFold input with initial guess and compute plddt and pae:
features = AFInput.from_data(af_data).add_guess(af_data)
result: AFResult = predictor(params, key(), features)
plddt = result.plddt.mean()
pae = result.pae.mean()

In the scripts folder, you will find the following example scripts.

• simple_salad.py: generate structures using salad with random secondary structure conditioning, design sequences with ProteinMPNN and predict structures with AlphaFold 2
• test_hal.py: generate structures using AF2 hallucination.