PubMed2026
MOTIVATION: Modelling long genomic sequences remains challenging due to extreme sequence length, high redundancy, and the need for biological interpretability. Although Transformer-based architectures have achieved strong performance across genomic tasks, their high computational cost and reliance on fixed tokenization strategies limit their scalability and ability to focus on biologically informative regions. RESULTS: We propose ATSFormer, a token-pruning Transformer framework for efficient and biologically informed genomic sequence modelling. ATSFormer incorporates an attention-guided and parameter-free Adaptive Token Sampling (ATS) module into Transformer layers. Guided by attention-derived importance scores, ATS dynamically retains informative tokens while probabilistically discarding redundant ones, thereby reducing sequence length, FLOPs, and memory usage without introducing additional learnable parameters or extra training procedures. Importantly, the retained tokens correspond to key contributors to model predictions, enabling ATSFormer to highlight biologically meaningful sites and sequence motifs. We evaluated ATSFormer on four benchmark RNA modification datasets derived from RMVar 2.0, covering A-to-I, m1A, m5C, and m7G. Experimental results show that ATSFormer consistently outperforms existing state-of-the-art methods while achieving substantial computational savings. Furthermore, structural analysis using AlphaFold3 supports the biological relevance of the motifs identified by ATSFormer. AVAILABILITY AND IMPLEMENTATION: The source data and code are freely available at GitHub (https://github.com/1gao2/ATSFormer) and Zenodo (https://doi.org/10.5281/zenodo.21813541).