Expanding drug targets for 112 chronic diseases using a machine learning-assisted genetic priority score

Gene-phecode pairs in the top 10% of ML-GPS scores can be accessed interactively at https://rstudio-connect.hpc.mssm.edu/mlgps/.

This respository contains the following data:

All ML-GPS and ML-GPS DOE predictions (Predictions for all gene-phecode pairs.zip)

Summary statistics for all 112 phecodes (Summary statistics.zip)

Cleaned genetics and drug datasets (Cleaned files to generate Open Targets and SIDER datasets.zip)

Input datasets for ML-GPS and ML-GPS DOE training and prediction (Inputs for ML-GPS training and prediction.zip)

Code to clean data and train models (Jupyter Notebooks.zip)

Dependencies can be found within each Jupyter Notebook; these analyses were performed using Python 3.12.

----------

New for version 3 (December 16, 2024): We have corrected an error in the direction-of-effect (DOE) model where effect directions for rare and ultrarare variants were not correctly incorporated into the model. We use the following framework to assign predicted mechanisms to variants/genes:

Gain of function and beta > 0: inhibitor drug

Gain of function and beta < 0: activator drug

Loss of function and beta > 0: activator drug

Loss of function and beta < 0: inhibitor drug

We updated the code and outputs in this version, resulting in an approximately 30% increase in AUPRC for ML-GPS DOE compared to the original results reported in our paper.