Integrative co-expression and deep learning reveal stage-specific regulators of Apis cerana drone embryogenesis

Stage-resolved regulators in Apis cerana drone embryos remain unclear. Here, RNA-seq sampled at 12/24/36/48/60 h was analyzed by WGCNA to build co-expression modules, refined with k-nearest neighbors on five-point trajectories, and benchmarked across 19 ML/DL classifiers (including ResNet,SVM,ExtraTrees,KNN and RandomForest). To balance accuracy and interpretability, we combined model benchmarking, consensus voting, and feature attribution (e.g., gradient/SHAP) and report parameter and sampling sensitivities. Expression co-regulated into 6 clusters aligned with developmental stages—germline/epigenetic control (12 h), ncRNA/rRNA-centric regulation (24 h), neurogenesis/morphogenesis (36–48 h), and organ specialization (60 h). ResNet achieved ~ 97% accuracy for gene-pattern classification; consensus reduced > 5,800 candidates to ~ 2,000 high-confidence stage-specific genes and hub PPI subnetworks. The WGCNA→kNN(unsupervised cluster analysis) →ML/DL pipeline thus resolves stage-specific regulators while addressing the “black-box” trade-off, and its performance generalizes with attention to module granularity (β, minModuleSize), k, model capacity/regularization, and temporal sampling density.