Comparative Analysis of Classical Machine Learning and Deep Learning Regression Models for Codon-Level k-mer Composition Analysis of the Nipah Virus Genome

The Nipah virus (NiV) is highly virulent zoonotic paramyxovirus, where a complete genome sequence resource is relatively limited compared to other emerging pathogens. In this study, the codon level (3-mer) k-mer composition analysis of 93 complete Nipah virus genome sequences retrieved from the NCBI Nucleotide database was presented. The genomes were encoded into a vector of 64 elements representing the number of appearances of each of the 64 three-nucleotide combinations. The feature set was subjected to descriptive statistics, covariance and Pearson correlation between each feature, which showed very high degrees of pairwise collinearity in the set of codon-count features. Using a correlation-based feature-selection procedure (|r| > 0.99), the feature space was reduced from 64 to 5 almost independent codon counts (AAA, TAG, TCG, CGC and CCG). We considered seven regression algorithms across three learning paradigms (classical models: Linear Regression, K-Nearest Neighbors, Support Vector Machine/SVR, an ensemble model: Random Forest, deep sequential models: Long Short-Term Memory network and Simple Recurrent Neural Network, and a Decision Tree baseline) and compared their performance using the number of codons CCG as a regression target and the remaining four codon counts as predictors. The 70/20/10 train/validation/test split was used to train all models, and they were compared using the coefficient of determination (R²), root mean squared error (RMSE), and mean absolute error (MAE). Decision Tree Regression achieved the best performance (R² = 0.9945), followed by Linear Regression (R² = 0.9908), Random Forest (R² = 0.9892), LSTM (R² = 0.9731), and SimpleRNN (R² = 0.9718), while KNN (R² = 0.8038) and SVM/SVR (R² = 0.1038 with feature scaling) underperformed. The results are evaluated with the knowledge of the large compositional constraints on codon usage over a set of Nipah virus genomes, the known sensitivity of instance- and margin-based regressors to unscaled, collinear small-sample data, and the limited benefit of deep sequential architectures on a pseudo-sequence of a very small tabular sample (no temporal dimension). This paper presents in a transparent manner the pipeline results, statistical results and comparative modeling results, along with the limitations of the dataset, as a baseline for future codon-composition based computational studies of Nipah virus.