deepGBLUP: joint deep learning networks and GBLUP framework for accurate genomic prediction of complex traits in Korean native cattle

Hyo-Jun Lee, Jun Heon Lee, Cedric Gondro, Yeong Jun Koh, Seung Hwan LeeGenetics 2023 · DOI

Abstract

We introduced a novel genomic prediction algorithm, deepGBLUP, which successfully integrates deep learning networks and GBLUP framework. Through comprehensive evaluations on the Korean native cattle data and simulated data, deepGBLUP consistently achieved superior performance across various traits, marker densities, training sizes, heritabilities, and QTL effects. Therefore, deepGBLUP is an efficient method to estimate an accurate genomic value. The source code and manual for deepGBLUP are available at https://github.com/gywns6287/deepGBLUP .

Introduction

In animal and plant breeding, genomic prediction uses DNA markers (SNPs) to estimate the genetic merit of individuals. The industry standard, Genomic Best Linear Unbiased Prediction (GBLUP), is robust but primarily relies on linear assumptions. It often struggles to capture complex, non-linear interactions such as dominance and epistasis (gene-gene interactions).

While Deep Learning (DL) offers a potential solution for modeling these non-linearities, standard architectures like Convolutional Neural Networks (CNNs) often underperform in genomic tasks. CNNs utilize "weight sharing," assuming a pattern found in one part of an image has the same meaning elsewhere. In genetics, this assumption is flawed: a specific DNA sequence on one chromosome has a functional effect entirely different from the same sequence on another. The deepGBLUP model addresses this by incorporating locus-specific context while leveraging the pattern-recognition power of DL.

The deepGBLUP Architecture

The deepGBLUP model is a two-pronged ensemble designed to capture both broad genetic relationships and fine-grained marker effects.

The deepGBLUP framework integrates a deep learning network with locally-connected layers and a GBLUP framework to predict phenotypes from SNP markers.
  1. The Deep Learning Branch: This branch utilizes Locally-Connected Layers (LCL). Unlike CNNs, LCLs assign unique, unshared weights to different regions of the genome, allowing the model to learn the specific effects of adjacent markers based on their actual genomic position. These features are processed through fully-connected layers to estimate an initial genomic value.
  2. The GBLUP Branch: This branch follows the classical mixed-model approach, calculating three distinct Genomic Relationship Matrices (GRMs) to account for additive, dominance, and epistatic effects.

The final prediction is the sum of outputs from both branches, creating a synergy that outperforms either component alone.

Locus-Specific Learning with LCL

The shift from CNNs to LCLs represents a critical biological alignment. By breaking the weight-sharing constraint of standard CNNs, the model respects the biological reality that the genome is not translationally invariant; location is everything.

Comparison between (a) standard CNN architecture with weight sharing and (b) the LCL architecture used in deepGBLUP, which employs locus-specific weights.

Experimental Validation and Results

The authors validated deepGBLUP using two primary datasets:

  • Korean Native Cattle (Hanwoo): A dataset of 10,000 individuals with phenotypes for carcass weight (CWT), eye-muscle area (EMA), backfat thickness (BF), and marbling score (MS).
  • Simulated Data: 10,000 genotypes used to test varying levels of heritability and genetic effects.

Key Findings

  • Superior Accuracy: In the Hanwoo dataset (50K SNPs), deepGBLUP consistently outperformed GBLUP, its variants (DGBLUP, EGBLUP), and Bayesian methods. For Carcass Weight, deepGBLUP achieved a predictive ability of 0.752 compared to GBLUP’s 0.729.
  • Robustness in Low-Heritability Scenarios: The model showed significant improvements in "noisy" traits where the genetic signal was weakest (h² = 0.1).
  • Efficiency with Small Data: Unlike many DL models that require massive datasets, deepGBLUP maintained superior performance even with training sizes as small as 1,000 individuals.
  • Computational Performance: The model is practical for real-world use, training in approximately 3.24 seconds per epoch on a modern GPU for the 50K SNP dataset.

Conclusion and Future Outlook

deepGBLUP demonstrates that deep learning should complement rather than replace established statistical frameworks like GBLUP. While the current results are promising, future work will focus on multi-generational testing to ensure the model's predictive merit holds across several generations of offspring. Additionally, the authors aim to develop a "pedigree module" to incorporate non-genotyped animals, further increasing the model's utility in commercial breeding programs.

Summary updated

Abstract

We introduced a novel genomic prediction algorithm, deepGBLUP, which successfully integrates deep learning networks and GBLUP framework. Through comprehensive evaluations on the Korean native cattle data and simulated data, deepGBLUP consistently achieved superior performance across various traits, marker densities, training sizes, heritabilities, and QTL effects. Therefore, deepGBLUP is an efficient method to estimate an accurate genomic value. The source code and manual for deepGBLUP are available at https://github.com/gywns6287/deepGBLUP .
BibTeX
@inproceedings{lee2023deepgblup,
  author = {Hyo-Jun Lee and Jun Heon Lee and Cedric Gondro and Yeong Jun Koh and Seung Hwan Lee},
  title = {{deepGBLUP: joint deep learning networks and GBLUP framework for accurate genomic prediction of complex traits in Korean native cattle}},
  booktitle = {Genetics},
  year = {2023},
  doi = {10.1186/s12711-023-00825-y},
}
Figure