A Linear-Transformer Hybrid for SNP-Based Genotype-to-Phenotype Prediction in Grapevine

Researchers developed LiT-G2P, a hybrid machine learning model combining linear genetic effects with Transformer-based neural networks to predict plant traits from DNA sequences in grapevines. The approach achieved superior prediction accuracy for leaf and trichome density across multiple years, demonstrating practical applications for genomic selection in agricultural breeding.

This research addresses a fundamental challenge in computational biology: predicting observable traits from genetic data with sufficient accuracy to guide breeding decisions. The LiT-G2P framework represents a methodological advance by combining two complementary modeling approaches—traditional additive genetic variance (linear) effects with modern deep learning (Transformer) to capture complex gene interactions. This hybrid strategy acknowledges that trait inheritance involves both straightforward genetic contributions and subtle nonlinear interactions that neither method captures alone.

The grapevine case study demonstrates practical utility through rigorous evaluation across temporal conditions. The model maintains prediction robustness even when tested on data from different growing years, a critical real-world requirement that many ML models fail to achieve. RMSE values around 0.45-0.47 and tolerance accuracies exceeding 74% represent meaningful improvements over baseline approaches, suggesting the method could accelerate breeding cycles by reducing uncertainty in trait prediction.

For the agricultural biotechnology sector, this work validates that attention-based mechanisms can identify genomic regions most influential for trait expression, enabling researchers to prioritize candidate genes and SNP markers for downstream biological validation. This interpretability layer distinguishes the approach from black-box deep learning models, making findings actionable for plant scientists.

The implications extend beyond grapes to any crop requiring accelerated genetic improvement under environmental variability. As genomic data becomes cheaper and more abundant, hybrid ML architectures that combine domain knowledge with learned patterns will likely become standard in precision agriculture. Future developments should test scalability across larger genomic datasets and additional crop species.

• →LiT-G2P integrates linear additive genetic effects with Transformer networks to improve genotype-to-phenotype prediction accuracy in grapevines.
• →The model achieved RMSE of 0.454 for cross-year predictions with 74.6% tolerance accuracy, outperforming baseline methods.
• →Attention weights enable identification of candidate SNP markers for biological validation, adding interpretability to predictions.
• →Cross-year robustness addresses a critical gap where many ML models fail on temporal data variability.
• →Hybrid approaches combining domain knowledge and deep learning offer practical advantages over purely data-driven methods in genomic selection.