This project aims to develop and train deep learning models to predict the interaction of proteins by using protein sequences as inputs. The goal is to provide accurate predictions that can be used to study the complex interactions between proteins. The models are based on large Language Models as encoders. If you find this reseach interesting and have questions please reach out to me (e.g. through github issues).
This project is based on Python 3.11 and built using Poetry.
Please install Poetry and clone the github repository. Poetry will handle the creation of a vitual enviroment and the installation of all dependencies.
To install the dependencies using poetry
poetry install
in the directory of the repository. If this fails on our LINUX machines try export PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring.
You can activate the virtual env using poetry shell. If you want to use GPU acceleration please install pytorch in this enviroment.
After installation use poetry run python main.py to run the main.py file.
The model uses a pretrained encoder model (ESM) to transform the two protein sequences into embeddings. The resulting embeddings are concatinated and fed into a head. The head is a 2 layer deep fully connected network. Softmax is applied to the logits of the head to generate the probabilites of interaction between the two input proteins.