#3 · ML / AI methods

Models infer unmeasured regulatory states from single-cell data, tested on held-out sets

The dominant thread is inference of unmeasured regulatory or cellular states from transcriptomic and multi-omic data, with predictions evaluated against withheld or external data. CardioChrom predicts chromatin accessibility and H3K27ac/H3K27me3 from snRNA-seq and is benchmarked against controls. A benchmark of ten general and three aging-specific single-cell foundation models across more than 2.5 million transcriptomes spans five aging questions. PASCode scores single-cell phenotype associations in roughly 6 million AD/PsychAD nuclei, and Dreamlet applies precision-weighted linear mixed models to pseudobulk data for differential expression. Explicit causal layers appear in the LSD pipeline, which pairs LASSO and structural equation modeling with DoWhy to map miRNA-target-TF chains, and in DoAtlas-2, which generates and tests causal hypotheses across 771 cohort resources. What is new is the pairing of predictive single-cell models with causal frameworks and held-out validation.