Awesome

ShapeLinker

ShapeLinker is a method for the shape-conditioned de novo linker design for PROTACs. It is based on Link-INVENT and uses reinforcement learning to steer the linker generation towards a query shape with desired physicochemical properties. Shape alignment is performed with a novel, fast attention-based point cloud alignment method.

Preprint: Reinforcement Learning-Driven Linker Design via Fast Attention-based Point Cloud Alignment

Requirements

• Multi-parameter optimization using shape alignment requires two different conda environments (see below).
• Only works on Cuda-enabled GPU.
• The code was tested on Debian 10 only.

Installation

1. Create ShapeLinker conda environment

conda env create -f env.yml

2. Create shape_align environment:

conda install -c conda-forge mamba
mamba create -n shape_align python=3.9 pytorch=1.13.0 torchvision pytorch-cuda=11.6 fvcore iopath nvidiacub pytorch3d -c bottler -c fvcore -c iopath -c pytorch -c nvidia -c pytorch3d
conda activate shape_align
pip install pykeops biotite open3d plyfile ProDy pykeops rdkit==2022.9.5 tqdm==4.49.0 unidip pytorch-lightning
pip install git+https://github.com/hesther/espsim.git

Data

Download data and models from https://storage.googleapis.com/vantai-public-archive/shapelinker. This data dump includes:

• folder data
  • folder xtal_poses: Processed and fragmented crystal structures
  • protacdb_extended_linkers.csv: Processed PROTAC-DB data
  • pdb_systems_data.csv: Processed data for the investigated crystal structures
• folder models
  • protacdb_extlinker_model_align.pth: Trained model for shape alignment
  • folder agents: Trained RL agents for the different crystal structures

The Link-INVENT prior, which is needed for any RL run, can be accessed here.

Steps to get directory structure used in notebooks:

1. Store folder data in ShapeLinker/utils

cd ShapeLinker/utils
gsutil cp -r gs://vantai-public-archive/shapelinker/data .

2. Store folder models in ShapeLinker

cd ShapeLinker
gsutil cp -r gs://vantai-public-archive/shapelinker/models . # includes trained RL agents

3. Dump linkinvent.prior in ShapeLinker/models

cd ShapeLinker/models
wget https://github.com/MolecularAI/ReinventCommunity/raw/master/notebooks/models/linkinvent.prior

Usage

The notebooks (folder ShapeLinker/notebooks) used here were adapted from ReinventCommunity and help with preparing runs for RL or sampling. There is also a notebook to help with training a shape alignment model (notebooks/train_shape_alignment_model.ipynb). We recommend training a new model for poses different from the crystal structures investigated here (of which the extended linkers were used).

The folder utils/postprocessing contains more useful jupyter notebooks allowing the postprocessing and evaluation of the generated data.