Alignment-Free DNA Sequence Clustering Using Deep Siamese BiLSTM and Attention Mechanisms

Abstract

Clustering of genomic sequences is critical for elucidating biological relationships, yet unsupervised methods often compromise between accuracy, generalizability, and biological interpretability. We propose a Siamese Bidirectional Long ShortTerm Memory (BiLSTM) network with an attention mechanism, trained via contrastive loss to learn biologically meaningful representations of DNA sequences, entirely without supervision, augmentation, or hand-crafted <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$k$</tex>-mer features. We evaluate our model on a challenging benchmark of Betacoronavirus genomes. The resulting embeddings capture both sequence-level similarity and evolutionary patterns. When paired with standard clustering algorithms, these embeddings yield highquality partitions: silhouette score of 0.814, Adjusted Rand Index (ARI) of 0.713, and purity of 0.864. Our model also reduces intra-cluster embedding variance by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2} \boldsymbol{-} \mathbf{4} \boldsymbol{\times}$</tex> compared to baselines, aligning closely with known taxonomic subgenera. Overall, this framework offers a scalable, interpretable, and unsupervised solution for genomic sequence clustering, with strong potential in comparative genomics and evolutionary biology.