MuSE‐Promoter: a multi-scale feature fusion and weighted ensemble learning method for identifying promoters across multiple cell lines

Promoters are central to regulating gene transcription by orchestrating cell-type- and developmental-stage-specific expression programs. However, the intrinsic heterogeneity of transcription factor binding sites, characterized by variable lengths, complex combinatorial patterns, and substantial sequence diversity across cell types, poses significant challenges to the robustness and generalizability of models relying on single-feature representations. To address these limitations, we propose MuSE-Promoter, a deep ensemble framework that integrates multi-scale feature fusion with weighted ensemble learning for accurate promoter identification across diverse cell lines. MuSE-Promoter constructs parallel feature extraction channels that combine contextual sequence embeddings from the DNABERT model and Word2Vec embeddings with handcrafted descriptors, including tri-nucleotide physicochemical properties (TPCP) and reverse-complement k-mer frequencies (RCKmer). A multi-scale convolutional neural network augmented with squeeze-and-excitation (SE) attention captures hierarchical motif patterns while effectively suppressing noise, followed by a Transformer module to model long-range dependencies. Furthermore, the deep learning branch is integrated with a Random Forest classifier through a learnable weighted ensemble strategy, thereby enhancing cross-domain robustness and prediction stability. Comprehensive evaluations across human cell lines from multiple tissues and Arabidopsis thaliana datasets demonstrate that MuSE-Promoter consistently outperforms state-of-the-art methods. Notably, it achieves superior generalization performance in challenging scenarios, including cross-cell-line transfer and enhancer–promoter discrimination. Collectively, MuSE-Promoter provides a powerful computational framework for large-scale promoter annotation, offering new insights into the regulatory mechanisms underlying complex transcriptional regulation. • We propose a multi‑scale fusion of embeddings and handcrafted features. • We introduce a multi‑scale CNN and SE attention to capture variable motifs. • We employ a Transformer to model long‑range dependencies in promoter regions. • We achieve robust cross‑cell‑line generalization and promoter‑enhancer discrimination.