Driven by aging populations and increasing insight into disease burden, we need treatments that can directly improve human health. Several thousand diseases affect humans of which only about 500 have any U.S. Food and Drug Administration-approved treatment (source: NIH’s National Center for Advancing Translational Sciences). Even for diseases with treatments, finding new drugs can give alternative treatment options with less severe side effects in patients and replacements for drugs that are ineffective in subsets of patients. Predicting therapeutic use of candidate drugs remains, however, a difficult task that current algorithms often fail to perform accurately, leaving many diseases with no existing treatments inaccessible to machine learning.

Here we leverage recent advances in geometric deep learning and human-centered AI to introduce TxGNN, a flexible model to identify therapeutic opportunities for diseases with few or no treatments available and limited molecular understanding. TxGNN is a graph neural network pre-trained on a comprehensive knowledge graph of 17,080 clinically-recognized diseases and 7,957 therapeutic candidates that can process various therapeutic tasks in a unified formulation. Once trained, we show that TxGNN can perform zero-shot inference on novel diseases without introducing additional parameters or requiring fine-tuning on ground truth labels.

TxGNN improves by up to 49.2% on indication tasks and 35.1% on contraindication tasks over existing methods and predicts therapeutic use for new therapies developed since June 2021. We develop an interpretable human-AI explorer for TxGNN, evaluate its usability in a study with human experts, and use it to examine TxGNN's novel predictions in a large electronic health record system.. TxGNN paves the way to improvements on many therapeutic tasks for neglected diseases.