$Astrazeneca(AZN.US)

AstraZeneca develops AI for predictive biomarker discovery — PBMF

Yesterday, Etai Jacob and Gustavo Arango-Argoty from the Oncology Data Science Department of AstraZeneca co-authored a paper titled "AI-driven predictive biomarker discovery with contrastive learning to improve clinical trial outcomes" in Cancer Cell. The paper introduces a Predictive Biomarker Modeling Framework (PBMF) based on contrastive learning, a neural network-based approach that systematically and unbiasedly explores potential predictive biomarkers.

Background: In modern medicine, clinical trials generate vast amounts of clinical and genomic data for each individual. However, the discovery of predictive biomarkers, which are crucial for personalized treatment, remains a significant challenge. Unlike prognostic biomarkers, which relate to overall disease outcomes regardless of treatment, predictive biomarkers identify patients likely to benefit from specific therapies.

PBMF has undergone comprehensive evaluation across various datasets, yielding remarkable results. In synthetic datasets simulating real-world scenarios, PBMF outperformed existing methods such as SIDES and Virtual Twins (VT). When compared against these methods in nine diverse clinical studies—spanning real-world data, different cancer and non-cancer indications, and various stages of immuno-oncology (IO) clinical trials—PBMF consistently identified predictive biomarkers.

In survival analyses of breast cancer and diabetic retinopathy studies, PBMF demonstrated superiority. In the breast cancer study, while identifying biomarkers predictive of longer survival with hormone therapy combined with chemotherapy, PBMF successfully generalized as a predictive biomarker in the test dataset, whereas VT and SIDES failed to do so. In the diabetic retinopathy study, PBMF also identified predictive biomarkers, while VT and SIDES primarily detected prognostic biomarkers.