#23 · ML / AI methods

Histology-based models predict gene expression using spatial and microenvironment context

The dominant thread is conditioning histology-to-expression prediction on structure beyond the target gene profile. MoSPR combines morpho-spatial macrostates with a low-rank molecular basis to predict gene expression from whole-slide images. TMEvolve models whole-slide images as reaction-diffusion tumor microenvironment fields on patch graphs, targeting survival, gene expression, and subtype prediction. GATE-ST fuses LLM-generated gene text summaries with histology embeddings through cross-attention. CCMIL aligns H&E features with RNA signatures via cross-modal contrastive multiple instance learning. What is new is the modelling target: microenvironment fields and gene-level text priors rather than image features alone. NexuST and CellMSA extend the same logic to transcriptomic foundation models, adding hierarchical gene-cell interleaving and cross-cell retrieval context.