Impact of Encoding Techniques on Convolutional Neural Network Performance in DNA Sequence Classification: A Comparative Study of One-Hot and TF-IDF

Abstract

The Convolution Neural Networks (CNNs) model has been a popular choice for genome data identification because it automatically extracts relevant features; however, model selection and the encoding technique employed are very crucial to the performance of the model. This is because different encoding methods capture different features of sequence data, including motif frequency and positional information, which affect how well a model learns biologically relevant patterns. This study examines the impact of two popular encoding techniques on CNNs' performance in DNA sequence classification tasks, viz, one-hot encoding and term frequency-inverse document frequency (TF-IDF). The datasets used for the study are publicly available histone datasets, which comprise DNA sequences that wrap around histone proteins. The performance of the models was measured in terms of accuracy and computational cost. One-Hot Encoding attains higher classification accuracy and better generalization across histone types, whereas TF-IDF shows faster training and lower resource consumption. These findings provide important insights for creating efficient and scalable bioinformatics pipelines by highlighting the trade-off between computational efficiency and predictive performance.