Adding Muscle RNA Transcripts Improves Swine Production Trait Prediction Accuracies

Incorporating gene expression measurements from muscle tissue into linear mixed models markedly increases the accuracy of phenotype prediction for swine production traits over conventional DNA-marker-based approaches, according to a multi-breed investigation of commercial boars. Evaluating predictive performance across four economically significant swine traits revealed that muscle transcriptomic profiles served as the primary driver of performance gains, frequently outperforming genomic and DNA methylation layers when used alone or in multi-omics combinations. While genetic improvement in livestock has historically relied on pedigree records and single-nucleotide polymorphism arrays to forecast breeding performance, modern molecular profiling technologies allow researchers to capture downstream regulatory layers directly from target tissues. The study demonstrated that incorporating RNA transcript abundance and cytosine methylation levels into best linear unbiased prediction frameworks yielded trait-dependent enhancements, with the strongest predictive gains emerging for residual feed intake and average daily gain.

The study evaluated a population of 443 male, non-castrated pigs born in 2021 and 2022 from the DanBred commercial breeding programme in Denmark, comprising three pure breeds: 143 Duroc, 148 Landrace, and 152 Yorkshire boars. All animals were comprehensively characterized with performance phenotypes, deep pedigree files, dense single-nucleotide polymorphism genotypes, whole-transcriptome sequencing, and genome-wide DNA methylation data. Biological samples were acquired through biopsies of the longissimus dorsi loin muscle collected when individual boars attained a live body weight of approximately 100 kilograms. By gathering both intermediate molecular layers from identical muscle samples in the same individuals, the researchers established an empirical multi-omics dataset designed to assess how molecular variation downstream of DNA sequence influences phenotypic variation and predictive accuracy across multiple swine breeds.

Four distinct production and conformation traits recorded across the test weight interval from 30 to 100 kilograms of body weight served as the phenotypic targets: average daily gain, backfat thickness, residual feed intake, and an overall strength conformation score. Average daily gain was recorded in grams per day using automated electronic feeding stations and was statistically preadjusted within each breed for test section, sex, dam parity, and starting body weight. Residual feed intake, also measured in grams per day, was derived as the residual term from a linear regression of daily feed intake on test section, dam parity, average daily gain, and metabolic body weight calculated from starting and final body weights. Backfat thickness was recorded as the arithmetic mean of four ultrasonic measurements taken in millimetres at the conclusion of testing, adjusted for section, sex, dam parity, and final body weight. Strength score was evaluated as an integer conformation metric ranging from 1 to 14, combining visual assessments of back strength, leg soundness, and skeletal structure, adjusted for identical environmental covariates.

To model molecular relationships, the investigators generated standardized relationship matrices corresponding to each biological layer and integrated them as independent random effects within a classical linear mixed model architecture. Additive pedigree relationships were established using breed-specific pedigree databases spanning 44 generations for Duroc, 53 generations for Landrace, and 49 generations for Yorkshire, containing 8,942, 7,587, and 6,775 registered animals, respectively. High-throughput genotyping was executed on a commercial 50k porcine beadchip, filtered to a common panel of 11,822 single-nucleotide polymorphisms shared across breeds after filtering for minor allele frequencies above 1.5 percent. For multi-breed analyses, block-structured genomic relationship matrices were computed using centred genotype counts scaled by breed-specific allele frequencies, establishing standardized covariance structures for additive genomic effects across both within-breed and multi-breed evaluation frameworks.

Transcriptomic profiles were acquired by paired-end polyadenylated RNA sequencing on an Illumina NovaSeq 6000 instrument, aligning high-quality reads against the domestic pig reference assembly Sscrofa11.1. After retaining genes exhibiting counts per million exceeding 0.5 in at least 20 percent of samples and performing trimmed mean of M-values normalization and logarithmic transformation, the researchers filtered the expression matrix to 14,293 active genes. Genome-wide DNA methylation was profiled to approximately 40-fold sequencing coverage using reduced representation bisulphite sequencing, targeting cytosine-phosphate-guanine dinucleotides across autosomal genomic regions. Following quality filtering, removal of coverage artefacts above 783-fold depth, and beta-binomial imputation of low-coverage positions, the methylation dataset was condensed to 16,074 highly variable autosomal sites shared across breeds, with a full set of 1,095,304 sites maintained for empirical sensitivity comparisons.

The team evaluated both single-omics and combined multi-omics linear mixed models, estimating variance components via restricted maximum likelihood using the ASReml-R package before solving mixed model equations. Evaluated formulations encompassed baseline pedigree prediction, genomic prediction, transcriptomic prediction, and methylomic prediction, alongside paired combinations and a comprehensive three-layer model incorporating genomic, transcriptomic, and methylomic effects simultaneously. In multi-breed leave-one-out cross-validation across all 443 animals, prediction accuracy was quantified as the Pearson correlation coefficient between model-predicted values and preadjusted phenotypes. For average daily gain, predictive accuracy rose from 0.12 with genomic prediction to 0.37 with combined transcriptomic-methylomic or three-omics prediction, whereas transcriptomic prediction alone achieved 0.35. For backfat thickness, prediction accuracy ranged from 0.18 under methylomic prediction to 0.39 with genomic-transcriptomic and three-omics models, improving noticeably over the 0.31 accuracy produced by standard genomic prediction.

The most pronounced predictive performance emerged for residual feed intake, where accuracy escalated from 0.19 under pedigree, genomic, or methylomic prediction models to 0.56 with transcriptomic prediction and 0.58 when pairing genomic and transcriptomic information. Conversely, the subjective skeletal conformation trait, strength score, exhibited substantially lower predictive ability across all model formulations, yielding correlation values spanning from negative 0.27 under methylomic prediction to 0.11 across transcriptomic and multi-omics architectures. Across all four investigated traits, incorporating muscle RNA transcript levels delivered the vast majority of predictive improvements over standard genomic prediction, whereas adding DNA methylation to transcriptomic models yielded negligible changes in cross-validation performance. Genomic-methylomic combinations exhibited modest intermediate predictive performance, demonstrating that DNA methylation alone captured limited phenotypic variation beyond what was already represented by genomic markers or gene expression profiles.

Variance component partitioning demonstrated that transcriptomic random effects explained substantially larger proportions of phenotypic variance than genomic or epigenomic components whenever gene expression was included. In single-omics models, transcriptomic effects accounted for proportions of 0.68 for average daily gain, 0.78 for backfat thickness, 1.00 for residual feed intake, and 0.09 for strength score, whereas DNA methylation accounted for 0.41, 0.41, 0.57, and 0.01, respectively. In models pairing genomic and transcriptomic effects, fitting transcript abundance reduced the proportion of variance attributed to genomic markers from 0.15 to 0.08 for average daily gain, from 0.31 to 0.23 for backfat thickness, from 0.23 to 0.13 for residual feed intake, and from 0.04 to 0.03 for strength score. Residual feed intake yielded restricted maximum likelihood solutions at the boundary of the parameter space, resulting in zero residual variance estimates in transcriptomic models and requiring cautious biological interpretation of those specific variance proportions.

Comparing feature panel sizes showed that selecting the 16,074 most variable cytosine-phosphate-guanine sites provided predictive accuracies equivalent to or slightly better than utilizing the comprehensive set of 1,095,304 sites, with an off-diagonal matrix correlation of 0.91 indicating high biological conservation. In across-breed evaluations utilizing leave-one-breed-out validation, predictive performance remained portable across genetic backgrounds, with transcriptomic and multi-omics models attaining robust accuracies despite greater genetic divergence between training and validation cohorts. Duroc animals exhibited slightly lower across-breed transferability than Landrace and Yorkshire boars, consistent with the closer phylogenetic relationship between the latter two breeds. Five-fold stratified cross-validation corroborated the leave-one-out findings, with the three-omics model attaining median correlation values of 0.39 for average daily gain, 0.39 for backfat thickness, 0.59 for residual feed intake, and 0.17 for strength score.

The authors emphasized critical methodological distinctions between predicting realized phenotypes and estimating transmissible additive genetic breeding values for selection decisions. While genomic profiles remain constant across lifetime environments and reflect inherited genetic variation transmitted stably to offspring, transcriptomic and epigenomic measurements reflect dynamic cellular states influenced by both genetics and transient microenvironments. Consequently, high phenotypic prediction accuracy achieved with muscle gene expression partly captures non-heritable environmental variation, meaning that phenotypic predictive gains do not directly guarantee accelerated genetic progress across generations without validation against progeny performance. Furthermore, estimating variance components from the complete dataset prior to cross-validation constituted a mild form of data leakage, and assuming independence among molecular layers oversimplifies documented biological relationships where genetic variants regulate methylation and gene transcription. Nevertheless, the researchers concluded that multi-omics linear mixed models provide powerful statistical tools for precision livestock management, where accurate forecasting of individual phenotypic outcomes remains commercially valuable.

Subject of Research: Multi-omics linear mixed models for predicting swine production traits

Article Title: Multi-omics prediction of swine production traits with linear mixed models

Article References: Ayres, L., Bovenhuis, H., Vourlaki, I.-T., Piles, M., Luigi-Sierra, M. G., Karlskov-Mortensen, P., & Calus, M. P. L. (2026). Multi-omics prediction of swine production traits with linear mixed models. Genetics Selection Evolution, 58(1), Article 84. https://doi.org/10.1186/s12711-026-01100-6

Image Credits: AI Generated

DOI: 10.1186/s12711-026-01100-6

Keywords: swine genomics, multi-omics prediction, linear mixed models, transcriptomics, DNA methylation, GBLUP, TBLUP, MBLUP, average daily gain, backfat thickness, residual feed intake, strength score, longissimus dorsi, RNA sequencing, reduced representation bisulphite sequencing, cross-validation, variance components, quantitative genetics