CoChat

Multi-modal deep learning improves grain yield prediction in wheat breeding by fusing genomics and phenomics

Open access

The DOI matches its Crossref record; 38 of the 38 references it cites were found in a registry or an index.

Abstract

Motivation: Developing new crop varieties with superior performance is highly important to ensure robust and sustainable global food security. The speed of variety development is limited by long field cycles and advanced generation selections in plant breeding programs. While methods to predict yield from genotype or phenotype data have been proposed, improved performance and integrated models are needed. Results: We propose a machine learning model that leverages both genotype and phenotype measurements by fusing genetic variants with multiple data sources collected by unmanned aerial systems. We use a deep multiple instance learning framework with an attention mechanism that sheds light on the importance given to each input during prediction, enhancing interpretability. Our model reaches 0.754 ± 0.024 Pearson correlation coefficient when predicting yield in similar environmental conditions; a 34.8% improvement over the genotype-only linear baseline (0.559 ± 0.050). We further predict yield on new lines in an unseen environment using only genotypes, obtaining a prediction accuracy of 0.386 ± 0.010, a 13.5% improvement over the linear baseline. Our multi-modal deep learning architecture efficiently accounts for plant health and environment, distilling the genetic contribution and providing excellent predictions. Yield prediction algorithms leveraging phenotypic observations during training therefore promise to improve breeding programs, ultimately speeding up delivery of improved varieties. Availability and implementation: Available at https://github.com/BorgwardtLab/PheGeMIL (code) and https://doi.org/doi:10.5061/dryad.kprr4xh5p (data).

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1093/bioinformatics/btad336

Cites

38 of 38 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 PheGeMIL leverages multi-channel inputs, combining dynamic phenotypic observations with genotypes. (a) Advanced spring wheat (Triticum aestivum L.) breeding lines from the International Maize and Wheat Improvement Center (CIMMYT) breeding program were sown during season 2017–2018. Multispectral and thermal images were collected by UAV flights over the experiment field and digital elevation model (DEM) developed from the images. Lines were profiled using genotyping-by-sequencing with calling of single nucleotide variants (SNPs). Final crop yield was measured by combine harvest in tons per hectare (t/ha) on a per plot basis. (b) The model flexibly enables the fusion of information across channels via a single deep learning architecture that is fully differentiable. Encoder networks for each data type transform images or genotypes into vectors of fixed dimensionality (256 dimensions). The attention mechanism then combines these vectorial representations in a single 256-dimensional vector, independently of the number of input images provided. The unique representation is then fed in a final fully connected network to predict plot-level yield. nMS, number of multispectral images for a given sample; nTH, number of thermal images for a given sample; nDEM, number of digital elevation model images for a given sample; ResNet-18, a convolutional residual neural network architecture for image feature extraction; FCN, fully connected neural network architecture; SNPs, single nucleotide polymorphisms.
Figure 1 PheGeMIL leverages multi-channel inputs, combining dynamic phenotypic observations with genotypes. (a) Advanced spring wheat (Triticum aestivum L.) breeding lines from the International Maize and Wheat Improvement Center (CIMMYT) breeding program were sown during season 2017–2018. Multispectral and thermal images were collected by UAV flights over the experiment field and digital elevation model (DEM) developed from the images. Lines were profiled using genotyping-by-sequencing with calling of single nucleotide variants (SNPs). Final crop yield was measured by combine harvest in tons per hectare (t/ha) on a per plot basis. (b) The model flexibly enables the fusion of information across channels via a single deep learning architecture that is fully differentiable. Encoder networks for each data type transform images or genotypes into vectors of fixed dimensionality (256 dimensions). The attention mechanism then combines these vectorial representations in a single 256-dimensional vector, independently of the number of input images provided. The unique representation is then fed in a final fully connected network to predict plot-level yield. nMS, number of multispectral images for a given sample; nTH, number of thermal images for a given sample; nDEM, number of digital elevation model images for a given sample; ResNet-18, a convolutional residual neural network architecture for image feature extraction; FCN, fully connected neural network architecture; SNPs, single nucleotide polymorphisms.CC BY · PubMed Central
Figure 2 Performance of deep multiple instance learning (MIL) and genomic selection methods models for yield prediction measured using Pearson correlation coefficient over a 5-fold cross validation scheme. (a) Performance when relying on multispectral images alone. The baseline models use summary statistics of the multispectral channels and of notable vegetation indices computed on the images. Moreover, they are trained on images from different date groups, as images of late-growth stage carry more predictive power. (b) Performance when relying on a combination of multispectral and genotypic input for yield prediction for a nonlinear baseline (Random forest), a linear baseline (Lasso), and PheGeMIL. For the two baseline models, the most informative images were preselected and used for prediction (dates 01 February 2018–02 March 2018, as opposed to all for PheGeMIL). (c) Performance evaluation when adding new data channels in PheGeMIL prediction. Multispectral images are always kept in the comparison, given that they are the basis of the approach. Only a subset of the possible channel combinations is explored, as the focus lies in combining images and genotypic information. MIL, multiple instance learning; DEM, digital elevation models; Multispec., multispectral; Therm., thermal.
Figure 2 Performance of deep multiple instance learning (MIL) and genomic selection methods models for yield prediction measured using Pearson correlation coefficient over a 5-fold cross validation scheme. (a) Performance when relying on multispectral images alone. The baseline models use summary statistics of the multispectral channels and of notable vegetation indices computed on the images. Moreover, they are trained on images from different date groups, as images of late-growth stage carry more predictive power. (b) Performance when relying on a combination of multispectral and genotypic input for yield prediction for a nonlinear baseline (Random forest), a linear baseline (Lasso), and PheGeMIL. For the two baseline models, the most informative images were preselected and used for prediction (dates 01 February 2018–02 March 2018, as opposed to all for PheGeMIL). (c) Performance evaluation when adding new data channels in PheGeMIL prediction. Multispectral images are always kept in the comparison, given that they are the basis of the approach. Only a subset of the possible channel combinations is explored, as the focus lies in combining images and genotypic information. MIL, multiple instance learning; DEM, digital elevation models; Multispec., multispectral; Therm., thermal.CC BY · PubMed Central
Figure 3 Generalization performance on a new environment and utilization for crop selection. (a) Comparison of yield prediction performance of a linear baseline (Lasso), two nonlinear baselines (Random forest and FCN) and our model (PheGeMIL) for prediction on a new, unseen environment using genotypic or phenotypic data. Multiple scenarios are evaluated. In all cases, training is done on data from environment A (2018 YT, see Table 1) and testing is done on data from environment B (2018 EYT). A set of experiments is conducted by training and evaluating on both multispectral images and genotypic data (first three rows). A second set of experiments is conducted by evaluating on genotypes alone (last three rows), to mimic prediction before sowing in breeding program scenarios. For baselines, training and testing must be done on the same data types and training can only be done on genotype alone. PheGeMIL, on the contrary, is trained with phenotypic data too, while still being evaluated on genotypes alone, thanks to the MIL framework. Distributions represent the performance in terms of Person correlation coefficient obtained on models trained on the 5 different splits of the training set. Ensembled performance for genotype-only predictions represents the prediction performance obtained when averaging the predicted values for a given sample across the 5 trained models. (b) Average yield obtained from a prediction-driven line selection of varying sizes (binned) using rankings derived from the values predicted by different ensembled methods. Lines are selected based on the predicted yield, their effective yield is then averaged across the set of selected lines and reported for an increasing selection size, ranging from 5% to 40% of the lines in the test set. MIL, multiple instance learning; FCN, fully connected network.
Figure 3 Generalization performance on a new environment and utilization for crop selection. (a) Comparison of yield prediction performance of a linear baseline (Lasso), two nonlinear baselines (Random forest and FCN) and our model (PheGeMIL) for prediction on a new, unseen environment using genotypic or phenotypic data. Multiple scenarios are evaluated. In all cases, training is done on data from environment A (2018 YT, see Table 1) and testing is done on data from environment B (2018 EYT). A set of experiments is conducted by training and evaluating on both multispectral images and genotypic data (first three rows). A second set of experiments is conducted by evaluating on genotypes alone (last three rows), to mimic prediction before sowing in breeding program scenarios. For baselines, training and testing must be done on the same data types and training can only be done on genotype alone. PheGeMIL, on the contrary, is trained with phenotypic data too, while still being evaluated on genotypes alone, thanks to the MIL framework. Distributions represent the performance in terms of Person correlation coefficient obtained on models trained on the 5 different splits of the training set. Ensembled performance for genotype-only predictions represents the prediction performance obtained when averaging the predicted values for a given sample across the 5 trained models. (b) Average yield obtained from a prediction-driven line selection of varying sizes (binned) using rankings derived from the values predicted by different ensembled methods. Lines are selected based on the predicted yield, their effective yield is then averaged across the set of selected lines and reported for an increasing selection size, ranging from 5% to 40% of the lines in the test set. MIL, multiple instance learning; FCN, fully connected network.CC BY · PubMed Central
Figure 4 The predictions of the presented model can also be interpreted thanks to the attention mechanism. (a) Distribution of the attention given by the model to each of the input channels for a given sample: the distribution can be interpreted as the relative importance of each data source for the yield prediction of that sample. The selected architecture has 8 different attention heads, each represented by a column in the plot, and they all consider different aspects of the input, hence giving varying importance to each channel. On average, multispectral images is the channel most considered, indicating that the model prioritizes its input well, as multispectral images are known to better correlate with yield than the other modalities. (b and c) Attention of a heads 1 and 4 respectively for 50 samples, the attention is consistent across samples, indicating reproducible behavior. (d) The attention distribution can also be evaluated across time for images collected at different stages of the plant growth. The model focuses on multispectral images at the beginning of growth, multispectral, and Digital Elevation Models at mid-growth and thermal images at the end of the growth. Interestingly, the model gives more importance to images obtained at mid-growth, confirming results found in Fig. 1. MIL, multiple instance learning; DEM, digital elevation models; Multispec., multispectral; Therm., thermal.
Figure 4 The predictions of the presented model can also be interpreted thanks to the attention mechanism. (a) Distribution of the attention given by the model to each of the input channels for a given sample: the distribution can be interpreted as the relative importance of each data source for the yield prediction of that sample. The selected architecture has 8 different attention heads, each represented by a column in the plot, and they all consider different aspects of the input, hence giving varying importance to each channel. On average, multispectral images is the channel most considered, indicating that the model prioritizes its input well, as multispectral images are known to better correlate with yield than the other modalities. (b and c) Attention of a heads 1 and 4 respectively for 50 samples, the attention is consistent across samples, indicating reproducible behavior. (d) The attention distribution can also be evaluated across time for images collected at different stages of the plant growth. The model focuses on multispectral images at the beginning of growth, multispectral, and Digital Elevation Models at mid-growth and thermal images at the end of the growth. Interestingly, the model gives more importance to images obtained at mid-growth, confirming results found in Fig. 1. MIL, multiple instance learning; DEM, digital elevation models; Multispec., multispectral; Therm., thermal.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 38 references

Research works printed but not linked: 1

  1. Falcon, 2019, PyTorch Lightning

Cited by

See all 49

Metadata

Authors
Togninalli, M., Wang, X., Kucera, T., Shrestha, S., Juliana, P., Mondal, S., Pinto, F., Govindan, V., Crespo-Herrera, L., Huerta-Espino, J., Singh, R. P., Borgwardt, K., & Poland, J.
Affiliations
  • SIB Swiss Institute of Bioinformatics (CH)
  • Kansas State University (US)
  • Centro Internacional de Mejoramiento de Maíz Y Trigo (MX)
  • Global Wheat Program, International Maize and Wheat Improvement Center , Texcoco, Estado de Mexico, Mexico
Corresponding author
  • Togninalli, Matteo
  • Borgwardt, Karsten
  • Poland, Jesse
Venue
Bioinformatics
Published
2023-05-23
Volume
39
Issue
6
ISSN
  • 1367-4803 (ISSN-L)
  • 1367-4811
OpenAlex
W4377691294
Cited by
49
References
38
Citations by year
2026: 22 · 2025: 17 · 2024: 10
FWCI
7.72
Type
journal article
Language
en
Open access
Locations
  • Bioinformatics (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • King Abdullah University of Science and Technology Repository (King Abdullah University of Science and Technology) (submittedVersion, cc-by)
  • Repository for Publications and Research Data (ETH Zurich) (submittedVersion, cc-by)
  • ETH Zürich Research Collection
Funders
National Science Foundation (1238187)
Topics
  • Life Sciences › Agricultural and Biological Sciences › Plant Science › Smart Agriculture and AI
  • Life Sciences › Agricultural and Biological Sciences › Plant Science › Genetics and Plant Breeding
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genetic and phenotypic traits in livestock
Keywords
wheat breeding, grain yield prediction, multimodal deep learning, unmanned aerial systems, plant phenotyping, attention mechanism, interpretability, plant health, cultivar development
MeSH terms
Plant Breeding, Deep Learning, Phenomics, Edible Grain, Genotype, Phenotype, Selection, Genetic, Triticum, Genomics

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Multi-modal deep learning improves grain yield prediction in wheat breeding by fusing genomics and phenomics. https://app.cochat.ai/papers/10.1093/bioinformatics/btad336
The paper
Togninalli, M., Wang, X., Kucera, T., Shrestha, S., Juliana, P., Mondal, S., Pinto, F., Govindan, V., Crespo-Herrera, L., Huerta-Espino, J., Singh, R. P., Borgwardt, K., & Poland, J. (2023). Multi-modal deep learning improves grain yield prediction in wheat breeding by fusing genomics and phenomics. Bioinformatics. https://doi.org/10.1093/bioinformatics/btad336

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.