DecBin: Bayesian-optimized deep clustering for long-read metagenomic binning

Accurate metagenomic binning is essential for reconstructing microbial genomes, thereby facilitating the discovery of novel taxa, the identification of pathogens, and the characterization of antimicrobial resistance genes. Most existing methods are designed for assembled contigs and cannot be directly applied to raw long reads. Although a few dedicated long-read binning methods have emerged, challenges remain in bin-number estimation and representation learning. Here, we propose Deep Embedded Clustering for Long-read Binning (DecBin), an adaptive method for directly binning raw long reads through integrating a variational autoencoder (VAE), Bayesian optimization, and deep embedded clustering within a unified framework. Computational experiments on large-scale simulated and mock-community datasets demonstrate that DecBin outperforms state-of-the-art methods in both long-read binning and metagenome-assembled genome recovery, while also achieving strain-level resolution in a low-complexity natural microbiome. Further analysis shows that the advantage of DecBin is partly attributable to its Davies-Bouldin index-guided Bayesian optimization, which explicitly evaluates candidate cluster numbers using a global cluster-validity criterion, thereby improving bin-number estimation across communities of varying complexity. Meanwhile, joint fine-tuning of the VAE encoder and clustering layer enables clustering feedback to refine the latent space, yielding representations better aligned with genome partitions. These findings highlight the potential of adaptive deep clustering for metagenomic binning directly from raw long reads without prior assembly. The source code of DecBin is freely available at https://github.com/LorMeBioAI/DecBin.