iEnhancer-XLNet3D: an enhancer identification method based on 3-mer tokenization optimization and encoder deep perception

Enhancers are non-coding regulatory elements whose sequence patterns are diverse and context-dependent, making accurate identification from DNA sequence alone challenging. This study presents iEnhancer-XLNet3D, an enhancer prediction framework that combines global contextual encoding with local pattern refinement under a unified fine-tuning pipeline. Given a fixed-length DNA sequence, we apply overlapping 3-mer tokenization and reconstruct a compact 3-mer embedding to adapt XLNet-Base for genomic input (DNA3_XLNet). To better exploit complementary information across network depths, we introduce FUSE_ENCODER to aggregate full-layer representations, and refine the fused features using a lightweight dual depthwise-separable convolution module (DDS_CNN) before classification. The model is evaluated on the canonical enhancer benchmark for Stage-1 (enhancer vs. non-enhancer) and Stage-2 (strong vs. weak enhancer). Under a unified protocol, we compare against representative prior methods as well as modern pretrained baselines (including DNABERT-2 and a small Nucleotide Transformer) fine-tuned under the same conditions. On the independent test set, iEnhancer-XLNet3D attains 86.7% AUC on Stage-1 and 96.6% AUC on Stage-2. Ablation analyses suggest that DNA3_XLNet, layer fusion, and convolutional refinement provide complementary contributions. Model weights are publicly available at: https://github.com/tilerons/iEnhancer-XLNet3D.