Decoding Significant Genomic Loci for Traits from Image Encoded DNA

Abstract

Understanding the complex relationship between genotype and phenotype is a critical goal in biological research, with significant implications for fields such as medicine, agriculture, and biosciences. The ability to predict phenotypes from genetic information can advance crop improvement, precision medicine, and disease prevention. However, genotype-to-phenotype prediction presents substantial challenges due to the high dimensionality of genomic data, complex gene-environment interactions, and the difficulty of pinpointing key genomic loci that drive phenotypic variation. Traditional approaches often struggle to identify the precise genomic regions that are crucial for specific traits. To address these challenges, this study explores the use of image-based encoding of genomic data, which transforms linear DNA sequences into spatial representations amenable to deep learning models. This approach facilitates pattern recognition and spatial feature extraction, capabilities that are unparalleled by traditional methods. Additionally, we employ Gradient-weighted Class Activation Mapping (Grad-CAM) to interpret model predictions, providing a mechanism to trace significant features back to specific DNA sequences. By leveraging Grad-CAM, we aim to not only improve predictive accuracy but also enhance biological interpretability by mapping key predictive signals to their genomic origins. This study demonstrates an innovative pathway toward more accurate and interpretable genotype-to-phenotype prediction models, advancing our understanding of the genetic basis for complex traits.