#25 · ML / AI methods

Learned interaction and noise structure in high-dimensional omics regression and reduction

The four preprints share a focus on inferring structure that is not specified in advance. BaGGLS fits a Bayesian group global-local shrinkage binary regression with variational inference to detect sparse interactions in high-dimensional biological data. CBKMR uses a copula-based Bayesian kernel machine regression for discrete omics outcomes, with NNCBKMR for near-linear scaling, and yields marker panels from scRNA-seq. AGL estimates spatial expression for low-expression genes as auxiliary tasks, selecting them using DkGSB. The Rayleigh quotient preprint shows that contrastive PCA approximates a Rayleigh quotient and proposes rho PCA, a faster and more accurate generalized-eigenvector dimension reduction. The methods are heterogeneous, so the common thread is dominant rather than universal. What is new is treating noise and interaction structure as learned quantities rather than as fixed assumptions.