FuseMedML based molecular biochemistry library for drug discovery/repurposing
fuse-drug contains generic tools to facilitate working with datasets and representations for proteins, small molecules, and interactions. These include data loaders, processing and augmentation fuse style ops, utilities and more.
It also contains end to end examples of data and model training pipelines, currently focused on the protein-ligand affinity prediction task.
fuse-drug is a work in progress. It will gradually expand to cover more representations and tasks such as molecule property prediction, protein-protein interaction, generation and more. It will also extend the fuse.eval package to cover evaluation metrics specific to the biochemistry domain.
Coming soon: DrugDiscoveryFoundationBenchmarks - A repository for biochemical ML benchmarks, which includes tools for data curation and creation, creating different types of splits for model training, and application of evaluation metrics.
-
Install FuseMedML and its dependencies as described here.
-
Install Fuse-Drug only (without examples) by running:
pip install -e .
# to also install development deps use:
pip install -e .[dev]
or:
Install Fuse-Drug with examples by running:
pip install -e .[examples]
In case of a CUDA related error, we recommend working in a conda environment with Python>=3.9 (Create one by running conda create -n ENV_NAME python=3.9) and updating PyTorch following the official PyTorch installation instructions after completing the above steps.
- [Optional] Install
abnumberpython package. This package is used for antibody numbering and alignment (see antibody.py)
conda install -c bioconda abnumber