LSTM-attention-guided graph neural networks for integrated genotype-Environment modeling in maize yield prediction

Area of Science:

• Agricultural Science
• Computational Biology
• Genetics

Background:

• Genotype-environment (G×E) interactions significantly impact crop yield, making accurate prediction crucial for food security.
• Traditional methods often struggle to capture complex G×E effects, limiting predictive performance.
• Advanced machine learning offers potential for improved modeling of these interactions.

Purpose of the Study:

• To develop and evaluate a deep-learning framework for maize yield prediction that explicitly models genotype-environment (G×E) effects.
• To compare the performance of three distinct deep-learning architectures incorporating graph neural networks and attention mechanisms.
• To assess the framework's robustness across different genotype sampling scenarios and compare it against existing state-of-the-art methods.

Main Methods:

• A deep-learning framework combining Long Short-Term Memory (LSTM), Graph Neural Network (GNN), and transformer-style attention was developed.
• LSTM summarized seasonal weather data into environment node features, while Single Nucleotide Polymorphisms (SNPs) were reduced to genotype nodes.
• Multi-head attention dynamically weighted graph edges, with three architectures (A, B, C) evaluated for yield prediction using a compact MLP.

Main Results:

• Performance improved monotonically across architectures A, B, and C, with architecture C achieving the best results (RMSE: 2.2120, PCC: 0.6945, R2: 0.4823).
• Architecture C demonstrated a ~20.3% reduction in RMSE and a ~68.8% increase in PCC compared to architecture A.
• The proposed framework showed consistent performance regardless of genotype sampling and outperformed the winning model from the Global G×E Prediction Competition.

Conclusions:

• The developed deep-learning framework effectively models complex G×E interactions for enhanced maize yield prediction.
• Global, content-adaptive aggregation via attention mechanisms is key to improving G×E propagation and predictive accuracy.
• The superior performance of architecture C highlights the potential of advanced deep learning for precision agriculture and crop breeding.