Tabular deep learning: a comparative study applied to multi-task genome-wide prediction.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Accurate phenotype prediction is crucial for genomic selection in breeding and disease risk assessment.
• Traditional linear models struggle with complex genetic interactions.
• Deep learning offers advanced non-linear modeling capabilities, but applications in tabular genomic data are emerging.

Purpose of the Study:

• To overview recent deep learning architectures for tabular data.
• To apply these architectures to genome-wide prediction (GWP) tasks, including multi-trait regression and multi-class classification.
• To benchmark deep learning methods against traditional approaches for GWP.

Main Methods:

• An extensive review of deep learning architectures for tabular data: NODE, TabNet, TabR, TabTransformer, FT-Transformer, AutoInt, GANDALF, SAINT, and LassoNet.
• Application of these models to multi-trait GWP using real genomic datasets.
• Comprehensive benchmarking against traditional and tree-based methods like LightGBM.

Main Results:

• LassoNet demonstrated superior performance in both prediction accuracy and computational efficiency across multiple genomic datasets.
• It outperformed other tabular deep learning models and the LightGBM benchmark.
• Experimental results were obtained for three multi-trait regression and two multi-class classification GWP tasks.

Conclusions:

• LassoNet is identified as a leading deep learning architecture for GWP, offering significant advantages over existing methods.
• Its effectiveness in predictive accuracy and computational efficiency was validated on real-world genomic data.
• LassoNet's built-in variable selection capability aids in identifying important genetic markers influencing phenotype.