Software

NNMM: Software for Mixed-Effects Neural Networks Incorporating Intermediate Omics Data in Genome-to-Phenome Analysis

---

Abstract

Advances in high-throughput sequencing technology have generated increasingly large and diverse multi-omics data (e.g., gene expression, protein abundance). These intermediate omics layers can mediate the effects of genotypes on phenotypes. Therefore, incorporating intermediate omics data into genome-to-phenome analyses, such as genomic prediction and genome-wide association studies, can improve prediction accuracy and provide deeper biological insights into multi-layer regulatory processes This multi-layer regulatory system connects genetic variants to traits through intermediate omics features, and the relationships across layers are often complex, involving interactions and nonlinear effects. The NNMM (mixed effect neural networks) model has been proposed to hierarchically model the genotype-omics-phenotype relationship as a unified, multilayer regulatory network. By allowing nonlinear activation functions within a mixed-model framework, NNMM can mimic the complex dependencies between data layers. Previous studies have demonstrated this model's ability to significantly improve genomic prediction accuracy for traits such as methane emissions in sheep and feed efficiency in both sheep and beef cattle. In this study, we developed an open-source software package that implements the NNMM framework with a user-friendly interface. Compared with general-purpose machine-learning tools, our software is optimized specifically for genome-to-phenome analyses, including genomic prediction and genome-wide association studies. First, it enables users to choose from a collection of widely used Bayesian mixed effects models optimized for high-dimensional genomic data, such as GBLUP, BayesA, BayesB, and BayesC, to hierarchically model genotype-intermediate omics-phenotype relationships. The software also allows easy incorporation of fixed effects (e.g., herd, year, age, sex), non-genetic random effects (e.g., litter, pen), and pedigree-based random effects. Second, build-in functions are provided to calculate significance measures for association studies, including posterior inclusion probability and window posterior probability of association. In addition, the software supports flexible network architectures based on biological prior knowledge (e.g., functional annotations of genetic variants). Users can pre-define a partially connected neural network, for example, where specific genetic variants influence only subsets of intermediate omics features. Common activation functions in neural networks (e.g., linear, sigmoid, ReLU, leaky ReLU) are supported, and with prior knowledge of the relationships among different layers, users can also pre-define custom activation functions. In summary, our open-source software has the potential to greatly expand the adoption of the NNMM framework in genome-wide prediction and association studies due to its flexibility, interpretability, and computational efficiency.

Keywords: 2026

How to Cite:

Cheng, H. & Zhao, T., (2026) “NNMM: Software for Mixed-Effects Neural Networks Incorporating Intermediate Omics Data in Genome-to-Phenome Analysis”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286778. doi: https://doi.org/10.31274/wcgalp.24118

Rights: 1