FragNet

FragNet is a molecular property prediction model with intrinsic Interpretability. FragNet has been pretrained using about 2 million drug-like molecules. Finetuned checkoints are also available for solubility and cancer drug response prediction.

Domain of Applicability Exploration (DoA)

DoA is a tool to find domains of varying prediction accuracies of machine learning models. The domains are defined using molecular descriptors. This approach can be used with any machine learning model gihen that there exits a feature space.

Cancer Drug Response Prediction

This repository contains model that can use four different drug encoding methods: molecular descriptors, Morgan fingerprints, Graph Neural Networks and Transformers. We used this model to compare the effect of different drug representations to cancer drug response prediction.

Molecular Property Prediction

This repository can be used to deelop molecular property prediction models using different molecular representations. Currently contains instructions to create Graph Neural Networks using Pytorch Geometric, Feed Forward Neural Networks, SchNet, and RNNs.