Gradient responsive regularization: a deep learning framework for codon frequency based classification of evolutionarily conserved genes.

backgroundIdentifying conserved genes among major crops like Triticum aestivum (wheat), Oryza sativa (rice), Hordeum vulgare (barley), and Brachypodium distachyon (BD) is essential for understanding shared evolutionary traits and improving agricultural productivity. Traditional bioinformatics tools, such as BLAST, help detect sequence similarity but often fall short in handling large-scale genomic data effectively. Recent advances in deep learning, particularly Multilayer Perceptrons (MLPs), offer powerful alternatives for uncovering complex genomic patterns. However, optimizing these models requires advanced regularization methods to ensure reliability. Integrating bioinformatics with adaptive deep learning techniques provides a robust approach to reveal conserved genes and enhance our understanding of plant genome evolution and function. This study addresses the genomic conservations across four agriculturally vital species wheat, rice, barley and BD by integrating bioinformatics and deep learning to identify genes conserved evolutionarily.

resultsThe whole genome data for four species were downloaded from Ensembl (253,076 genes). Reciprocal best hits (RBH) via BLASTn reduced the data set to 25,152 highly similar sequences, highlighting shared ancestry, across four species. A novel Multilayer Perceptron (MLP) framework, enhanced with Gradient Responsive Regularization (GRR), was bench-marked against MLP penalized variants (L1, L2, Elastic Net, Adaptive) using learning rates (0.01, 0.001, 0.0001) and batch sizes (16, 32, 64, 128). All models achieved > 99% accuracy, precision, recall, F1-score and Matthews Correlation Coefficient (MCC), with the novel GRR performing comparably, e.g. 0.9992 accuracy at Learning Rate (LR) = 0.0001 based all genes (full data).

conclusionThe Novel Gradient Responsive Regularization (GRR) framework achieved state-of-the-art performance across all evaluated metrics (accuracy, precision, recall, F1-score, and MCC). Specifically, at a learning rate of 0.0001, Novel GRR attained peak performance with larger batch sizes (128 and 256) on the RBH filtered dataset (25152 genes), while demonstrating exceptional scalability on the full genomic dataset (253076 genes) at a batch size of 32. Statistical validation via Kruskal Wallis tests (p < 0.05) confirmed the method’s significant advantage over conventional regularization approaches. Notably, Novel GRR maintained consistent superiority across all evaluation phases training, validation, and testing highlighting its robustness for both feature refined (RBH) and genome wide analyses. These results highlights the method’s capacity to adapt to varying data complexities, offering a powerful tool for genomic prediction and evolutionary studies.