Large-scale systematic feasibility study on the pan-cancer predictability of multi-omic biomarkers from whole slide images with deep learning

Abstract

We assessed the pan-cancer predictability of multi-omic biomarkers from haematoxylin and eosin (H&E)-stained whole slide images (WSI) using deep learning (DL) throughout a systematic study. A total of 13,443 DL models predicting 4,481 multi-omic biomarkers across 32 cancer types were trained and validated. The investigated biomarkers included a broad range of genetic, transcriptomic, proteomic, and metabolic alterations, as well as established markers relevant for prognosis, molecular subtypes and clinical outcomes. Overall, we found that DL can predict multi-omic biomarkers directly from routine histology images across solid cancer types, with 50% of the models performing at an area under the curve (AUC) of more than 0.633 (with 25% of the models having an AUC larger than 0.711). A wide range of biomarkers were detectable from routine histology images across all investigated cancer types, with a mean AUC of at least 0.62 in almost all malignancies. Strikingly, we observed that biomarker predictability was mostly consistent and not dependent on sample size and class ratio, suggesting a degree of true predictability inherent in histomorphology. Together, the results of our study show the potential of DL to predict a multitude of biomarkers across the omics spectrum using only routine slides. This paves the way for accelerating diagnosis and developing more precise treatments for cancer patients.