Alignment-free sequence composition based features can distinguish paralogous proteins

Abstract

Paralogous proteins have a common ancestor but have diverged in functionality. Homology- and similarity-based methods often fail to distinguish paralogous proteins, due to a nuanced sequence-function relationship that remains largely intractable. Going beyond these methods, we extract various sequence composition based feature representations from proteins and use the same to train classifiers for distinguishing paralog pairs. We consider 20 paralog pair datasets and train a linear support vector machine (SVM) classifier for each dataset to distinguish among them. Amino acid composition (AAC) and 2-ordered amino acid composition (2-OAAC) features achieve high predictive performance across these datasets, despite the low sample size and potential label noise. Furthermore, we utilise the marginal contribution feature importance (MCI) to identify specific amino acids driving classification. We show that retaining only the top-$k$ MCI-ranked amino acids preserves nearly all predictive power while acting as a regularizer. Structural mapping of these top-ranked features in families such as globins, tubulins, and histones reveals that they frequently localise to critical functional and oligomerisation interfaces. Ultimately, our interpretable machine learning framework successfully bridges sequence composition and protein activity, providing a scalable computational method for identifying potential functional determinants that can guide future wet-lab site-directed mutagenesis experiments to study protein function.