FoldCraft

Note: this project is WORK IN PROGRESS.

Implementation of AlphaFold in independent modules. AlphaFold became a sandbox in FoldCraft, and you can play with it as you wish.

Installation

CPU installation

Using conda

conda create -n foldcraft python=3.10
conda install ipykernel   # recommendation
pip install ml-collections==0.1.1
conda install numpy jax absl-py
conda install dm-haiku dm-tree -c conda-forge
conda install tensorflow-cpu -c conda-forge
conda install biopython -c conda-forge

Or you can just (Not ready for now):

conda env create -f environment/environment_cpu.yml

GPU installation

Using NVIDIA A100 as example

conda create -n foldcraft python=3.10
conda install jax absl-py ipykernel dm-haiku dm-tree biopython -c conda-forge
conda install tensorflow-cpu -c conda-forge
pip install ml-collections==0.1.1
pip install --upgrade "jax[cuda11_pip]" -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.html

Or you can just:

conda env create -f environment/environment_gpu.yml

Apple silicon installation (GPU? No, it's still CPU)

Apple GPU might be supported in the future. (wait for jax-metal updates)

conda install ipykernel matplotlib
conda install jax absl-py -c conda-forge
conda install dm-haiku dm-tree -c conda-forge
conda install biopython -c conda-forge
conda install ml-collections -c conda-forge
conda install tensorflow-cpu -c conda-forge
# pip install jax-metal

In this environment, dm-haiku and tensorflow-cpu can bring in a lot of conflicts

conda env create -f environment/environment_m1.yml

Installed package list

python=3.10
ml-collections==0.1.1 # only pypi
absl-py
jax
dm-haiku
dm-tree
numpy
biopython=1.82
tensorflow=2.15.0 # required in alphafold.model.features
[?]pandas
[?]immutabledict

Run

Step 1: Create empty MSA feature (input feature of alphafold)

python foldcraft/parafold/create_empty_feature.py --fasta_paths input/test.fasta --output_dir output

Step 2: Run FoldCraft

python run_foldcraft.py

Reference

Jax installation: link

Zuricho/FoldCraft