Recent studies have brought to light the potential of combining artificial intelligence technologies to integrate digital pathology and transcriptomic data for enhancing the diagnosis and prognosis of cancer. However, directly merging these two types of information is not feasible in real-world medical settings, where traditional histopathology methods are still widely employed and transcriptomic analyses are not commonly utilised in public healthcare systems.

Our research involved the analysis of two publicly accessible datasets, namely The Cancer Genomic Atlas and the Clinical Proteomic Tumour Analysis Consortium, covering various cancer types, including glioma-glioblastoma, renal, uterine, and breast cancers across four distinct patient groups. We found substantial improvements in cancer grading and risk prediction accuracy (with statistical significance indicated by a p-value of 0.05), consistently demonstrated across all patient cohorts.

Our study showcases the efficacy of our innovative crossmodal generative artificial intelligence model, PathGen, in utilising gene expressions derived from digital histopathology to effectively forecast cancer grading and patient survival risk with exceptional precision (exhibiting state-of-the-art performance), reliability (backed by a conformal coverage guarantee), and interpretability (facilitated by distributed co-attention maps). The PathGen software code is openly accessible on GitHub at https://github.com/Samiran-Dey/PathGen for broader utilisation.