#4 · ML / AI methods

Shared embeddings fuse tissue images, spatial transcriptomes and molecular profiles

The common thread is learned representations that bind a second information source or structural prior into one embedding. MicNet projects spatially resolved transcriptomes and pathology images into a shared space via unsupervised contrastive learning, and TopoMIL adds sample-level topology of cell distributions to multiple instance learning, reporting AUCROC gains across four pooling schemes. T2Pdecoder learns a shared protein-RNA embedding to predict protein abundance from transcriptomic data, while PathTIGR combines pathway topology with cancer genomic variants through attention and transformer modules. Outside spatial settings, TifBERT applies masked gene modeling to bulk RNA-seq for normalization robustness, MOLAR learns molecular representations from graph and text views under noisy labels, and the arXiv benchmark tests image and transcriptomic foundation-model embeddings with fusion and conformal prediction. What is new is joint treatment of tissue morphology, spatial organization and molecular profiles within one representation, most evident in MicNet and TopoMIL.