Hist2Vec: A histogram and kernel-based embedding method for molecular sequence analysis

Abstract

Due to the huge surge in genomic data, there is an increasing need for better and more efficient molecular sequence classification techniques. There has been plenty of work proposed by researchers using machine learning models for promising classification results. However, they face few limitations in capturing hierarchical structures and relationships in the molecular sequences. To overcome such limitations, we propose Hist2Vec, a novel kernel-based technique for embedding generation that captures the sequence similarities by constructing histogram-based kernel matrices and Gaussian kernel functions. By building histogram-based representations from the distinct $k$-mers and minimizers found in each sequence, Hist2Vec is able to identify similarities between sequences. The sequence information is preserved by converting these representations to higher dimensional feature spaces using Gaussian Kernel functions. Then we apply kernel Principal Component Analysis to obtain the final embedding for the molecular sequences. These embeddings are then used as input to classical machine learning models for supervised analysis. We also establish the theoretical properties of Hist2Vec, ensuring the validity and effectiveness of the method. The experimental evaluation of our method shows that Hist2Vec outperforms all other state-of-the-art methods demonstrating high accuracy of $>76\text{\%}$ for the Human DNA dataset, $>83\text{\%}$ for the Coronavirus Host dataset, and high precision in the case of t-Cell dataset.