Image-based Genotype Encoding for Phenotype Prediction and Exploration for Result Explainability

Genome-to-phenome prediction (G2P) aims to capture the informative features from the extremely long genotype sequence and predict the observable traits. High-dimensional genotype sequences are usually converted to low-dimensional latent representations using dimensionality reduction techniques such as principal component analysis. However, such dimensionality reduction methods may compromise the informative features due to the conversion. This study introduces two image-based encoding techniques of the high-dimensional genotype sequence, 1) K-Skip encoding, and 2) Multi-Channel Full-Sequence (MCFS) encoding. K-Skip encoding utilizes the skipping method to reduce the dimension of the long sequence and then convert it to image representation. On the other hand, MCFS converts the full genotype sequence into multi-channel image data that ensures data integrity. The encoded images are then employed to train a state-of-the-art deep neural network to build a phenotype trait classifier. Results demonstrate that the MCFS encoding achieved higher accuracy compared to the K-Skip encoding method. Moreover, to pinpoint the most significant single nucleotide polymorphisms (SNPs), we propose an iterative training procedure using a Long Short Term Memory (LSTM) network with sequential genomic data. Our results demonstrate the remarkable potential of the image-based genotype encoding approach for G2P and result explainability.