Case Study: AstraZeneca predicts drug toxicity with Julia Computing and Bayesian neural networks

AstraZeneca predicts drug toxicity with Julia Computing using 10,000+ companies’ trusted Julia ecosystem

AstraZeneca faced the challenge of predicting drug-induced liver injury preclinically to reduce development costs and increase patient safety. While deep neural networks were a popular tool for such toxicity predictions, they were prone to overfitting and could not quantify the uncertainty in their forecasts, potentially leading to overconfident but erroneous results. To address this, AstraZeneca partnered with Julia Computing.

The solution implemented by Julia Computing was a Bayesian neural network (BNN) built using the Julia programming language, specifically leveraging the Flux.jl and Turing.jl packages. This BNN model describes each weight with a distribution rather than a point estimate, which prevents overfitting and, crucially, provides a measure of uncertainty for its predictions. This allows AstraZeneca to make more reliable and informed decisions on drug toxicity, reducing risks and costs in the drug development process.

AstraZeneca