Multi-View Genomic Prediction

Introduction

Genomic prediction aims to predict phenotypic traits from high-dimensional genomic data and has become an important tool for supporting efficient breeding decisions. With the growing availability of genomic data, deep learning models, particularly Transformer-based architectures, provide new opportunities to capture complex relationships among genetic markers.

However, directly processing complete genomic sequences can be computationally expensive due to their high dimensionality. This challenge becomes more significant when the number of available samples is limited, potentially increasing model complexity and the risk of overfitting.

To address these challenges, we propose SAMoE, an efficient genomic prediction framework that leverages diverse genomic information while avoiding the computational burden of processing complete genomic sequences. SAMoE aims to achieve accurate and efficient phenotype prediction with reduced model size, training cost, and inference latency.

Method

SAMoE adopts a two-stage framework to efficiently learn from high-dimensional genomic data. In Stage 1, multiple models are trained on different genomic views generated through random position selection. Each model processes only a subset of genomic positions, reducing the input dimensionality while capturing diverse genomic patterns.

In Stage 2, the trained Stage 1 models are treated as experts and integrated using an Adaptive Mixture-of-Experts (AMoE) module. A routing mechanism dynamically evaluates and selects relevant experts for each input, then combines their predictions to produce the final phenotype prediction. This design enables SAMoE to leverage complementary information across genomic views while maintaining computational efficiency.

Experiment Result

• SAMoE achieves 0.516 avgerage ACC, outperforming baselines ESAET and transformer by 0.0221 (+4.21%) and 0.013 (+2.6%).

• SAMoE achieves 0.683 avgerage PCC, outperforming baselines ESAET and transformer by 4.76% and 2.32%, respectively. 

• SAMoE obtains 0.692 avgerage RMSE, reducing error over ESAET by 3.92% and transformer by 2.98%.

Publication

• Te-Yang Lee, Van Thanh Duy Tran, Huy Hoang Pham, Yi-Wei Chen, and Chih-Yi Chiu, "A multi-view stochastic subsampling framework for efficient genomic selection," Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), Bangkok, Thailand, Nov. 8-12, 2026.