PSSM2Vec: A Compact Alignment-Free Embedding Approach for Coronavirus Spike Sequence Classification

Abstract

The coronavirus SARS-CoV-2 is the cause of the COVID-19 disease in humans. Like many coronaviruses, it can adapt to various hosts and evolve into different variants. It is well-known that the major SARS-CoV-2 variants are characterized by mutations that happen predominantly in the spike protein. Understanding the spike protein structure and determining its perturbations are vital for predicting coronavirus host specificity and determining if a variant is of concern. These are crucial to identifying and controlling current outbreaks and preventing future pandemics. Machine learning (ML) methods are a viable solution to this effort, given the volume of available sequencing data, much of which is unaligned or even unassembled. However, such ML methods require fixed-length numerical feature vectors in Euclidean space to be applicable. For this purpose, we design two feature embedding techniques to convert spike sequences into a compact representation, which serves as input to various ML classifiers. Such embeddings are alignment-free, unlike some previous approaches, avoiding computationally expensive alignment and assembly pipelines. Our proposed embeddings, PSSM2Vec and PSSMFreq2Vec, combine the power of the position weight matrix for compactness, and k-mers to be alignment-free. Experiments on both SARS-CoV-2 and more general coronavirus sequence data show that the proposed embeddings yield better predictive performance, in most cases, than the baseline and state-of-the-art methods, and are also scalable to millions of spike sequences. We also show that in terms of runtime, PSSM2Vec is extremely efficient, which makes it applicable on datasets composed of millions of sequences. Using statistical analysis, we also show the compactness of the proposed feature embeddings compared to the existing methods.