Multi-omics analysis

A biologically motivated nonlinear latent variable model for genetic evaluation using genomic sequence

---

Abstract

Current genetic evaluation models are strictly linear and SNP-based, typically assuming independence among variants and ignoring the biological structure of genes and the nonlinear processes that link genotype to phenotype. These models have been effective for livestock improvement over the past 15 years, particularly when using genotypes from SNP arrays designed for intermediate allele frequencies. However, they are limited in their ability to leverage the full information content of low-coverage sequencing and cannot accommodate the complexity of gene action. We propose a biologically motivated, hierarchical latent variable model that uses genomic sequence to more accurately capture the causal architecture of complex traits. The model distinguishes between two classes of haplotypes for each gene: those affecting transcription (expression haplotypes) and those affecting translation and protein function (functionality haplotypes). The effects of these haplotypes are combined multiplicatively to define a latent variable termed gene activity, reflecting the biological requirement that both transcription and translation must occur for a gene to exert physiological effect. Gene activity from maternal and paternal chromosomes is then aggregated additively to form individual gene activity. Gene effects are modeled as nonlinear transformations of gene activity, accommodating a range of gene action modes from additivity to dominance and over-dominance. Although not required, the model supports integration of RNA-seq data to estimate expression haplotype effects through a cis-eQTL component. Because gene expression is primarily influenced by local (cis-acting) variants near the gene being expressed, RNA-seq enables powerful single-gene association analyses for estimation of expression haplotype effects, reducing the number of parameters needing estimation from phenotypic data alone. This external estimation simplifies computation and mitigates overparameterization. A default assumption is that gene-level properties, such as haplotype effects and mode of gene action, are shared across traits unless sufficient evidence supports trait-specific effects. This implies that differences in relative breeding values between traits reflect different weightings of gene effects rather than genomic variants. It also suggests that phenotypes from multiple traits can contribute to estimating fundamental gene properties, further reducing dimensionality compared to models treating millions of variants as independent potentially causal effects. The model can incorporate an intermediate layer between gene effects and economically important traits to integrate high-dimensional -omics traits. We developed a software package for simulation and estimation of haplotype effects, mode of gene action, and regression of traits on gene effects under multi-gene, multi-trait models. In current validation on simulated data, the EM-based algorithm recovers simulated values. Ongoing development is increasing performance as model complexity increases. By aligning statistical modeling with molecular biology, this approach offers a path to extract more information from genomic sequence than traditional linear models by modeling gene-level latent variables and their interactions. The USDA is an equal opportunity provider and employer.

Keywords: 2026

How to Cite:

Thallman, R., Engle, B., Borgert, J., Snelling, W., Keele, J. & Kuehn, L., (2026) “A biologically motivated nonlinear latent variable model for genetic evaluation using genomic sequence”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2287066. doi: https://doi.org/10.31274/wcgalp.24203

Rights: 1