The 18th Eucarpia Biometrics in Plant Breeding Conference is taking place in Paris-Saclay University Campus, France on 21-23 September 2022.
The main scientific topics of the meeting are:
Computomics is a highly innovative bioinformatics company focussing on agricultural challenges in plant breeding. We work together with breeders and growers across all crops and breeding programs to develop new varieties by applying ⨉SeedScore®, our machine learning-based predictive breeding technology. Our innovative technology transforms traditional plant breeding processes by integrating additional data such as information on the genotype, the cultivation environment, climate, soil microbiome, and the used field management to identify and predict the best-performing plants for any specific location. This allows breeders to significantly accelerate their breeding program, by identifying up to 10x more candidates to escalate into the commercial pipeline, which can reduce time to market by 3-6 years.
On Wednesday, 21 September 2022, Rupashree Dass will present a poster on how machine learning can be used for more accurate phenotype predictions.
Title: Application of Machine Learning models for more accurate phenotype predictions in two real world commercial breeding programs
When: Wednesday 21 September 2022, 18:00 - 20:00
Rupashree Dass, Machine Learning Scientist at Computomics
Genomic selection (GS) is an integral tool for plant breeders to accurately select crosses directly from genotype data, leading to faster and more resource-efficient breeding programs. GS uses genome-wide markers to predict phenotypes and breeding values. Several models for phenotype prediction have been established to date and have been used extensively in animal and plant breeding. These range from classical linear mixed models to complex non-linear machine learning approaches. In this work, we have systematically implemented eight different phenotype prediction models, including basic genomic selection methods such as RRBLUP to more advanced deep learning-based techniques. These range from classical models, such as regularized linear regression over ensemble learners, e.g. XGBoost, to deep learning-based architectures, such as Convolutional Neural Networks (CNN). We have compared the results from these models with our machine learning technology xSeedScore. The performance of these models has been compared across five phenotypes for commercial soy line breeding and corn hybrid breeding programs. We observe that Least Absolute Shrinkage and Selection Operator (LASSO), Support Vector regression (SVR) and Random Forest (RF) have similar performance for each of the phenotypes. Although xSeedScore had different data splits during the five fold cross-validation, it performed similarly to the other models and even 15% better in two out of five phenotypes. We also see that there was no superiority of complex neural network-based architectures for phenotype prediction compared to well established methods. Currently, we have extended xSeedScore to account for gene-environment interactions. Since phenotypes in commercial breeding programs are typically influenced by environmental conditions, the integration of environmental features to phenotype prediction models is highly relevant. We show that by integrating data from various environments we are able to further increase the overall prediction performance of all the phenotypes by more than 4% and provide predictions per environment. Even though additional research and method development is still required to integrate environmental conditions with the other ML models efficiently, through this work we deliver first insights on phenotype prediction using ML methods in multi-environment trials and in predicting performance in new environments.
Don't miss this opportunity to connect with us in person! Rupa will be onsite for the full symposium. In case of questions regarding the poster, machine learning models for phenotype predictions or to discuss your plant breeding challenges please feel free to contact Rupa directly.