GenoDrawing: An Autoencoder Framework for Image Prediction from SNP Markers

Area of Science:

• Plant genomics and bioinformatics
• Machine learning applications in agriculture
• High-throughput phenotyping

Background:

• Genome sequencing provides vast genotypic data for plant analysis.
• Genomic selection and neural networks (NNs) predict complex traits.
• Autoencoders are effective NN models for unsupervised feature extraction in plant phenotyping.

Purpose of the Study:

• Introduce GenoDrawing, an autoencoder framework for predicting and retrieving apple images from single-nucleotide polymorphism (SNP) data.
• Assess the utility of GenoDrawing for predicting traits that are difficult to define.
• Explore the impact of SNP selection on image prediction accuracy.

Main Methods:

• Developed a novel autoencoder framework named GenoDrawing.
• Utilized a low-depth SNP array for apple image prediction.
• Trained and evaluated the model on a dataset of shape-related SNPs.

Main Results:

• GenoDrawing successfully predicted apple images from SNP data, demonstrating proficiency with a small dataset.
• SNPs associated with visual traits significantly impacted generated image quality, aligning with biological expectations.
• Performance degraded when unrelated SNPs were included in simpler NN architectures.

Conclusions:

• GenoDrawing offers a practical framework for genomic prediction in fruit tree phenotyping, especially for smaller breeding companies.
• The study highlights the importance of selecting relevant SNPs for accurate image prediction.
• Further research should focus on advanced models and balanced datasets for improved outcomes in genomic-based image prediction.