Theory and infrastructure to expand the universe of statistical genetic models - Project Summary/Abstract Complex traits reflect both biological mechanisms and social structure, yet the additive models that dominate biobank and whole-genome sequence (WGS) studies often conflate mechanism with population dynamics, environmental transmission, and measurement artifacts. This program expands the statistical-genetics toolkit so biological signal can be cleanly separated from assortative mating (AM), vertical transmission (VT), and scale effects, turning long-recognized “nuisance” processes into measurable structure that improves inference, interpretability, portability, and computational scalability across cohorts. Over the award period, our lab will develop multivariate, inter-generational theory that characterizes how genotype-phenotype relationships reorganize across generations and derive corrected estimators for SNP- heritability, genetic correlation, and annotation-partitioned parameters that remain valid under realistic mating and environmental transmission patterns. We will provide diagnostics that quantify AM- and VT-induced covariance, clarify when cross-trait associations reflect shared biology versus social structure, and develop inference procedures robust to monotone outcome rescaling. To accelerate the pace of scientific development, we will build field-level infrastructure: A public benchmarking suite spanning realistic genetic architectures and inter-generational dynamics, with versioned datasets, leaderboards, and ground-truth parameters; Genetics-focused probabilistic-programming templates that accelerate the jump from model specification to biobank-scale distributed inference on GPUs/AI accelerators, with reproducible workflows for validation, sensitivity analysis, and uncertainty calibration; and Graph-based linear-operator toolkits that provide faster, memory-efficient implementations of core genetics analyses via compact genotype representations—enabling analysis of WGS data at biobank scale on a workstation. Methodologically, our work emphasizes estimators with finite-sample guarantees, principled uncertainty quantification, and stress-testing against known sources of bias (phenotype rescaling, population structure). All software, benchmarks, and exemplar analyses will be open and well-documented, with training materials that lower barriers for non-specialists. By expanding the universe of workable models and standardizing how they are evaluated, this program aims the bridge the gap between scalable statistical frameworks and realistic models of complex traits to maximize biological insights.
