#4 · ML / AI methods

Transfer of learned structure across resolutions, domains, and biological priors

The common thread is learned models that carry structure from one data layer, resolution, or domain into another. The autoencoder study adapts cell-line drug-response models to patient tumors with few-shot supervision under strong biological domain shift. DeSCENT deconvolves bulk RNA-seq into subject-level single-cell profiles and combines them with bulk data for multimodal survival analysis. A second strand encodes biological priors in architecture: BKDRP embeds protein complex and pathway priors for drug response from multi-omics GDSC data, and RNAGAN includes an embedded pathway layer. scFoundation embeddings derived from TCGA bulk RNA-seq are applied to survival prediction across 25 cancers. The remaining items cover single-cell representation, pathology foundation models, and benchmarking; the semi-supervised integration benchmark reports that gains often degrade under realistic label imperfections. What is new is transfer across resolutions and domains, with priors moved into model design rather than only into training data.