#18 · ML / AI methods

Deep models encode priors and align modalities for transcriptomic prediction

The dominant thread is encoding structure beyond the raw expression matrix into learned representations. scLLM-DSC fuses NCBI gene-prior LLM embeddings with a graph-guided topological view through cross-modal contrastive alignment, while scGTN pairs a Siamese graph transformer with dual augmented graph views and optimal transport for scRNA-seq clustering. TopoMIL adds sample-level cell-distribution topology to multiple instance learning on microscopic images. Other items align modalities directly: MicNet contrastively links spatial transcriptomics with pathology images, and T2Pdecoder predicts broad protein abundance from RNA via a shared protein-RNA embedding. HoloCell and OCOO-T extend this toward generative and perturbation-response modeling, and TifBERT applies masked gene modeling to bulk RNA-seq. What is new is the explicit choice of which prior is encoded and how modalities are aligned. Reported gains are method-level and confined to the datasets named in each summary.