#9 · ML / AI methods

Deep learning links routine histology and multi-omics to cancer biomarkers

The common thread is deep learning applied to routine histopathology or high-dimensional cancer data to produce biomarkers with clinical framing. Several items operate on H&E slides: PanSubNet predicts PDAC basal-like and classical subtypes with AUC 84 to 88.5% across PANCAN and TCGA, PROTAS quantifies reactive stroma with prognostic framing across cohorts and biopsies, and HEX predicts spatial protein expression from H&E for lung cancer prognosis and immunotherapy response. Other items integrate modalities: TSTScope couples scRNA-seq with TCR sequencing to define functional tumor-specific CD8 T cell states, and BiChemoCLAM fuses whole-slide images with gene expression for chemotherapy response. Segmentation work, including CSRefiner for few-annotation fine-tuning, SegJointGene, and a benchmark of cell segmentation methods, supports the upstream tooling. The newer elements are interpretable architectures and weakly or semi-supervised training strategies applied to clinically oriented endpoints.