MGS-HCT enables resistance gene discovery and functional sequence design in forest trees

Resistance gene discovery in forest trees remains challenging because of their large, repeat-rich genomes and the sparse distribution of functional elements, which restrict the direct application of deep learning approaches developed for crop species. While genomic language models have shown promise in crop species, their application to forest trees remains limited. Here, we present MGS-HCT (Multi-scale Gated Sparse Encoder-Hybrid Convolutional Transformer), a deep learning framework tailored for forest tree genomic sequences. The model integrates multi-scale gated units with structure-aware sparse attention for efficient sequence compression, and combines convolutional modules with transformer layers to capture both local motifs and long-range regulatory dependencies. We show that MGS-HCT supports two key applications in forest tree biotechnology: (1) the conditional generation and prioritization of putative NLR candidates for Alnus glutinosa (common alder) and (2) the de novo generation of synthetic DNA sequences containing promoter-like and splice-site-like elements. These results provide a computational foundation for accelerating disease-resistance gene discovery and genomic-assisted breeding in forest tree species, expanding the scope of deep learning from crop plants to forest genomics. By prioritizing candidates supported by multiple computational criteria for subsequent experimental validation, the framework is intended to narrow the experimental search space and help shorten the long breeding cycles of forest trees.