Abstract
Accurate genomic prediction is vital for agricultural breeding and personalized medicine. While linear models are standard, they often overlook non-linear genetic interactions. We propose GpNet, a deep learning framework utilizing locally connected layers (LCL) to capture linkage disequilibrium (LD) block effects. By mimicking the directional nature of biological translation, GpNet achieved a predictive ability of 0.721 for carcass weight in Korean native cattle, outperforming traditional linear methods like GBLUP and BayesA. Furthermore, GpNet identified key quantitative trait loci (QTL) on chromosomes 6 and 14, demonstrating its utility as a discovery tool.
Introduction
Genomic prediction (GP) utilizes DNA marker information to estimate the genetic merit of individuals. In cattle breeding, predicting traits such as carcass weight (CWT), backfat thickness (BF), and eye-muscle area (EMA) allows for optimized selection. Historically, linear models like Genomic Best Linear Unbiased Prediction (GBLUP) have dominated the field. These models assume additive effects, where each marker contributes independently to a trait. However, biological systems are characterized by epistasis—complex interactions between neighboring genetic markers.
Deep learning models, particularly Convolutional Neural Networks (CNNs), are designed to capture local patterns. However, standard CNNs employ weight sharing (translation invariance), which is unsuitable for genomic data where the biological function of a sequence is dependent on its specific chromosomal location. GpNet addresses this by employing Locally Connected Layers (LCL), which use unique weights for each genomic position.
Methodology
Model Design
GpNet is inspired by the biological process of translation, where RNA polymerase reads DNA sequences directionally. The network architecture consists of:
- Locally Connected Layers (LCL): Unlike CNNs, LCLs do not share weights across the input, allowing the model to learn the specific impact of LD-blocks at their exact locations.
- Residual Connections: Skip connections are utilized to enable the training of deeper stacks of LCLs without signal degradation.
- Heritability-Adjusted Loss: The loss function incorporates a heritability ($h^2$) term, directing the model to prioritize the genetic component of phenotypic variance over environmental noise.
Data and Training
The study utilized a dataset of 10,000 commercial Korean native cattle (Hanwoo). Genotypes consisted of 44,314 high-quality SNPs. The model was evaluated on three key economic traits: CWT, BF, and EMA, using an 80/10/10 split for training, validation, and testing.
Results and Discussion
Performance Comparison
GpNet demonstrated superior predictive ability for Carcass Weight (CWT) compared to state-of-the-art linear models.
As shown in the table above, GpNet achieved a predictive ability of 0.721 for CWT, outperforming GBLUP (0.714) and BayesA (0.719). This suggests that LCLs effectively capture the local epistatic interactions relevant to carcass weight. However, GpNet showed lower performance for BF and EMA, indicating that these traits might be influenced by distant epistasis or dominance effects not currently captured by the model’s architecture.
QTL Discovery
Beyond prediction, GpNet successfully identified significant genomic regions associated with the target traits.
The model identified a shared QTL on chromosome 6 (38.4–39.8 Mb) across all traits, highlighting the SLIT2 gene. For CWT and EMA, a region on chromosome 14 (25.3–29.9 Mb) containing the RAB2A gene was also identified. These findings align with known biological drivers of muscle and carcass development in cattle.
Conclusion
GpNet introduces a biologically inspired deep learning approach to genomic prediction. By utilizing locally connected layers, it successfully captures local genetic interactions that linear models miss, particularly for carcass weight. While limitations remain regarding distant epistasis and dominance, GpNet provides a robust foundation for the next generation of genomic selection tools. Future work will explore dilated convolutions and alternative activation functions to further enhance predictive accuracy across a broader range of complex traits.
Summary updated