LSTM-Attention-Guided Graph Neural Networks for Integrated Genotype–Environment Modeling in Maize Yield Prediction

Amir Morshedian, Michael DomaratzkibioRxiv 2025 · DOI

TL;DR

This paper introduces a novel deep learning framework that integrates Long Short-Term Memory (LSTM) networks and Graph Neural Networks (GNNs) with attention mechanisms to predict maize yield by modeling complex genotype-environment (GxE) interactions. Their best model significantly outperforms previous methods, achieving a lower RMSE and higher Pearson correlation on unseen genotypes and environments, demonstrating the power of this unified approach.

What Problem Does This Paper Address?

Predicting crop yield is crucial for global food security, but it's a complex challenge due to the intricate interplay between a plant's genetic makeup (genotype) and the environmental conditions it grows in (environment) – known as genotype-environment (GxE) interaction. Traditional methods often struggle to capture these non-linear interactions, especially when dealing with vast amounts of genomic data (millions of Single Nucleotide Polymorphisms or SNPs) and dynamic environmental factors like weather. Existing deep learning approaches have also faced limitations in effectively encoding both temporal environmental data and complex genomic relationships simultaneously. This paper aims to bridge this gap by proposing a unified framework that can model these interactions more comprehensively, leading to more accurate yield predictions, even for novel genotypes and environments.

Methods

The core of this work is a deep learning architecture that combines three key components:

  • LSTM for Environmental Encoding: Daily weather data (temperature, precipitation, solar radiation, etc.) over a growing season is fed into a unidirectional LSTM. This allows the model to capture temporal dependencies and summarize the complex environmental conditions into a compact, 21-dimensional "weather fingerprint" or embedding. This is a departure from methods that rely on static summary statistics, as it accounts for the timing and duration of weather events.
  • PCA for Genotype Representation: High-dimensional genomic data (over 437,000 SNPs) is first dimensionality-reduced using Principal Component Analysis (PCA) to 548 components. These components are then treated as individual nodes in a graph, representing the genetic makeup of different maize varieties.
  • Graph Neural Network (GNN) with Attention for GxE Interaction: The genotype and environment embeddings are integrated into a GNN. The GNN uses message-passing to allow nodes (representing genotypes and environments) to exchange information. Crucially, they employ multi-head attention mechanisms to dynamically weight the importance of these messages, allowing the model to focus on the most relevant interactions.

Three GNN architectures were explored:

  • Architecture A: A basic fully connected bipartite graph connecting all genotype nodes to all environment nodes.
  • Architecture B: Extends Architecture A by adding "intra-set" connections, allowing genotypes to interact with other genotypes and environments with other environments based on similarity. This captures local relationships within each domain.
  • Architecture C: Builds upon Architecture B by introducing a "supernode" that acts as a global aggregator. This supernode uses attention to attend to all genotype and environment nodes, creating a single, comprehensive representation of the GxE interaction for prediction.

Datasets

The study utilizes the Genomes to Fields (G2F) Genotype × Environment Prediction Competition dataset. This is a large, publicly available dataset containing maize yield data from multiple years and locations. The dataset includes:

  • Phenotypic data: Grain yield, days to anthesis, days to silking, and anthesis-silking interval.
  • Genotypic data: Over 437,000 high-quality SNPs from approximately 4,928 distinct hybrids.
  • Weather data: Daily meteorological records for each site-year combination, including temperature, precipitation, solar radiation, and humidity.

The dataset was split using a forward-time validation strategy, training on data from 2014-2021 and testing on the unseen 2022 data, which included both novel genotypes and environments.

Key Insights and Ideas

  • Unified GxE Modeling is Key: The paper strongly suggests that effectively modeling GxE interactions requires a unified approach that simultaneously considers both genomic and environmental information, rather than treating them separately.
  • Temporal Environmental Data Matters: Using sequential daily weather data encoded by an LSTM, rather than static averages, allows the model to capture the impact of specific weather events (like heatwaves) on yield, which is crucial for accurate prediction.
  • Graph Structure Enhances Interaction Modeling: Representing genotypes and environments as nodes in a graph, and using GNNs with attention, allows for the explicit modeling of complex, non-linear relationships between them.
  • Progressive Graph Densification Improves Performance: The progression from a simple bipartite graph (A) to one with intra-set connections (B) and finally to a global supernode readout (C) consistently improved predictive accuracy. This highlights the benefit of progressively incorporating more complex relational information.
  • Global Context is Crucial for Prediction: Architecture C, with its global supernode attention, achieved the best performance, indicating that a holistic view of all GxE interactions is more powerful than just local or bipartite connections for yield prediction.

Key Results

The proposed models were evaluated on unseen genotypes and environments using RMSE (Root Mean Squared Error), PCC (Pearson Correlation Coefficient), and R² (Coefficient of Determination).

  • Monotonic Improvement: Performance improved consistently across the three architectures:
  • Architecture A: RMSE = 2.7749, PCC = 0.4115
    • Architecture B: RMSE = 2.3683, PCC = 0.6622 (20.3% RMSE reduction, 68.8% PCC increase vs. A)
    • Architecture C: RMSE = 2.2120, PCC = 0.6945 (20.3% RMSE reduction, 68.8% PCC increase vs. A; 6.6% RMSE reduction, 4.9% PCC increase vs. B)
  • State-of-the-Art Performance: Architecture C achieved the best results, outperforming the winner of the G2F GxE Competition (CLAC) with an RMSE of 2.2120 and PCC of 0.6945, compared to CLAC's RMSE of 2.458 and PCC of 0.631.
  • Robustness: The model's performance remained consistent regardless of the number of genotypes per environment and showed strength with unbalanced genotype sampling.

Limitations and Future Work

  • Sensitivity to Extreme Climates: The study notes that while the model captures average environmental patterns well, its ability to explicitly account for extreme climatic conditions appears limited, as most test environments clustered in a PCA space without clear performance differentiation in extreme positions. Future work could explore strategies to better capture these extreme conditions.
  • Genotype-Level Analysis: The current work focuses on GxE interaction modeling. The authors suggest that future research could delve into more detailed genotype-level analysis, such as identifying specific influential genetic markers, to provide greater biological interpretability.
  • Alternative Genomic Representations: The current framework uses PCA to compress genomic information. Future work could explore using original marker sequences or constructing genotype-level nodes that preserve more detailed genetic information.
  • Sample-Level Graphs: The current graph represents features shared across all samples. Defining each node as a specific genotype or environment instance and modeling GxE interactions through edges could offer a more direct representation of biological entities.

Summary updated

Abstract

Abstract We present a deep-learning framework that combines an LSTM, a graph neural network (GNN), and transformer-style attention to model genotype–environment (G × E) effects for maize yield prediction. Weather data for a growing season is summarized using LSTM and encoded into a 21-dimensional embedding that is used as the environment node feature; 437,214 SNPs are summarized into 548 principal components that instantiate genotype nodes. Multi-head attention dynamically weights the edges during message passing. We compare three architectures: A (fully bipartite graph), B (A with intra-set top- k similarity within genotype and within environment), and C (B with a single learnable supernode readout that attends over all nodes after message passing). The joint representations feed a compact MLP for yield prediction. Using a forward-time split (2014–2021 train; 2022 test with unseen genotypes and unseen environments), performance improves monotonically from A to C: A (RMSE 2.7749, PCC 0.4115, R 2 0.1693), B (2.3683, 0.6622, 0.4385), C (2.2120, 0.6945, 0.4823). Compared to A, C has a reduction in RMSE by 0.5629 (∼20.3%) and an increase in PCC by 0.283 (∼68.8%), indicating that global, content-adaptive aggregation promotes local G × E propagation. Performance of our approach remains consistent regardless of the number of genotypes per environment and has strong performance under variable or unbalanced genotype sampling expression across environments.
BibTeX
@misc{morshedian2025lstm,
  author = {Amir Morshedian and Michael Domaratzki},
  title = {{LSTM-Attention-Guided Graph Neural Networks for Integrated Genotype–Environment Modeling in Maize Yield Prediction}},
  year = {2025},
  doi = {10.1101/2025.11.09.687529},
  howpublished = {bioRxiv},
}
Figure