Abstract
BACKGROUND/OBJECTIVES: Early larval development is critical for caste and sex differentiation in honeybees. This study investigates molecular divergence in 4-day-old Apis mellifera larvae and introduces a customized deep learning model for hub-gene discovery. METHODS: Genome-guided RNA-seq, DEGs, WGCNA, and splicing analyses were integrated. A hybrid convolution-attention model, ACmix-Swin, combined with WGAN-GP augmentation, was developed to classify larvae and prioritize caste-biased genes. Selected genes were validated by qPCR. RESULTS: Significant caste- and sex-specific divergence was detected in cuticle formation, hormone metabolism, and reproductive signaling. ACmix-Swin achieved the highest accuracy among baseline models and consistently identified key regulators, including Vg, LOC725841, LOC412768, and LOC100576841. qPCR confirmed RNA-seq trends. CONCLUSIONS: Caste- and sex-specific transcriptional programs are established early in larval development. The ACmix-Swin framework provides an effective strategy for high-dimensional transcriptome interpretation and robust hub-gene identification.