Regression Models for Estimating Similarity Score of the Optimal Expected Multiple Sequence Alignments

Abstract

Sequence alignment is a crucial tool in biological research. Alignment quality is typically assessed using scores, but these do not guarantee optimality. Estimating the optimal alignment score can help improve and optimize alignment algorithms. This study addresses this gap by presenting three regression models designed to estimate the score of the optimal expected alignments for a set of unaligned sequences. Due to the challenge of obtaining data on the best alignments - since such optimal alignments are infeasible to determine in practical scenarios - this study relies on simulated data where true alignments are known and serve as a standard for learning. We trained and evaluated three regression models to predict entropy-based alignment quality scores: (i) CNN-OHE-Model, a fully alignment-free convolutional neural network model using one-hot encoded sequences; (ii) ANN-Stat-Model, an artificial neural network trained on statistical descriptors derived from pairwise alignments; and (iii) POLY-Stat-Model, a polynomial regression model using the same statistical features. The models achieved high predictive performance with <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</i><sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> values of 0.9444, 0.9892, and 0.9720, respectively. Notably, the CNN-OHE-Model demonstrated strong generalizability when evaluated on real-world datasets from the BALIBASE and HOMSTRAD benchmarks, despite being trained on only 2000 samples compared to 50000 for the ANN-Stat-Model. These findings highlight the robustness of the proposed models and their potential to assist in improving the evaluation and optimization of sequence alignment.