scMAF: Single-Cell Multi-Omics Clustering via Adaptive Modality Fusion

Single-cell multi-omics sequencing has greatly advanced the characterization of cellular heterogeneity by jointly profiling multiple molecular modalities. However, since different modalities exhibit varying discriminative signals across cell types, existing integration strategies may dilute or even obscure the signals carried by more informative modalities when fused with less informative ones, thereby hindering accurate multi-omics clustering. To address this limitation, this work presents a novel adaptive fusion framework for unsupervised clustering of single-cell multi-omics data. The core of our approach is an attention-based fusion module that dynamically assigns modality-specific weights within a shared latent space, allowing the model to naturally focus on the most informative signals from each modality. Extensive experiments on eight real-world multi-omics datasets encompassing gene expression, chromatin accessibility, and protein modalities demonstrate that our method achieves the state-of-the-art single-cell clustering performance over fourteen competitive baselines. The code will be released publicly upon acceptance.