Improving Trait Prediction in Plants with a Compositional Autoencoder
Author Information
Author(s): Powadi Anirudha, Jubery Talukder Zaki, Tross Michael C., Schnable James C., Ganapathysubramanian Baskar
Primary Institution: Iowa State University
Hypothesis
Disentangling genotype-specific and environment-specific features can enhance predictive models for plant traits.
Conclusion
The compositional autoencoder significantly improves trait prediction models for key agricultural traits like 'Days to Pollen' and 'Yield'.
Supporting Evidence
- The CAE outperformed traditional methods like PCA and vanilla autoencoders.
- R-squared values for 'Days to Pollen' reached 0.74 with CAE.
- Yield prediction accuracy improved significantly compared to baseline models.
- Disentangled representations showed consistent performance across different initializations.
- KL-divergence indicated effective separation of environmental influences.
Takeaway
This study shows that separating genetic and environmental factors helps scientists better predict how plants will grow.
Methodology
A compositional autoencoder was developed to separate genotype-specific and environment-specific features from hyperspectral data of maize.
Potential Biases
Potential biases in data collection and model training could affect results.
Limitations
The study focused on two specific traits and a single maize diversity panel, limiting generalizability.
Participant Demographics
The study involved 578 inbred maize genotypes.
Statistical Information
P-Value
0.01
Statistical Significance
p<0.05
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website