#31 · ML / AI methods

Uncertainty-aware statistical inference for heterogeneous biological populations

The dominant thread is explicit probabilistic modeling of heterogeneity, where uncertainty is propagated into the target quantity rather than reported as a point estimate. The preprints extend this to distributional and mixture settings: a varying-coefficient mixture-of-experts model with label-consistent EM, identifiability results and confidence bands; Bayesian density-on-density regression built on sliced Wasserstein barycenters and generalized Bayes inference; and weighted fuzzy c-means with a normalized density model, MM estimation, likelihood-ratio tests and bootstrap intervals. The published papers apply analogous uncertainty handling to causal and compositional questions. UBD models uncertainty in cell type proportions from bulk samples, while MACFIV, TLMR and GAN-IV address Mendelian randomization with nonlinear effects, population heterogeneity, or distribution-free instruments. What is new is the joint treatment of estimation error, model specification and inference validity in omics and population data.