Ding, Ke2026-09-292026-09-29https://hdl.handle.net/1885/733815881RNA-mediated gene regulation spans post-transcriptional mechanisms in which non-coding RNAs (e.g., microRNAs) silence targets, RNA-binding proteins control splicing, stability, localization, and translation, RNA secondary structures shape processing and interactions, and chemical modifications such as methylation and editing modulate RNA function and target recognition. Together, these factors determine translation efficiency and influence development, cancer, neurodegeneration, and emerging RNA-based therapeutics. The vast sequence space and combinatorial complexity outstrip experimental throughput, motivating deep learning approaches to model sequence-function relationships. This thesis advances RNA biology and deep learning through four major contributions that address long-range dependency modeling, multimodal integration, interpretability, and computational efficiency. First, DeepMirtron, a multimodal bidirectional LSTM model, integrates nucleotide sequence and secondary structure to identify non-canonical microRNAs, mirtrons, with 95.4% accuracy on canonical versus mirtron classification and 99% accuracy on an independent positive-only set of 286 mirtrons, outperforming prior CNN-based methods like CNN-concat and CNN-LSTM, and discovering 3,642 novel high-confidence candidates for experimental validation through genome-wide screening. Second, CRMnet, a transformer-encoded U-Net with squeeze-and-excitation and skip connections, predicts gene expression from yeast promoters with Pearson r = 0.971 on native and 0.987 on random sequences, surpassing existing models; mixed-precision distributed training shortens runtime by over 30-fold. Third, NucEL applies ELECTRA-style pretraining to genomic sequences, using single-nucleotide tokenization and hybrid local-global attention in a 93M-parameter model. It achieves state-of-the-art performance on regulatory benchmarks with 27 times fewer parameters than 2.5 B-parameter MLM models such as Nucleotide Transformer. We advance foundation models through efficient replaced token detection (RTD) rather than standard masked language modeling, replacing sparse 15% supervision with dense generator-discriminator feedback that provides roughly 7-fold greater supervisory signal, enabling compact yet powerful representations. Fourth, RegLLM continues pretraining from NucEL on transcriptomes, demonstrating efficient DNA-to-RNA transfer learning. It achieves state-of-the-art or competitive results across eight RNA-centric task domains, including non-coding RNA classification, RNA-RNA interaction prediction, m6A/m5C modification detection, mRNA stability, and translation efficiency using only raw sequences while reducing total training time from months to days. Interpretability analyses, saliency maps, attention visualization, and motif recovery confirm that learned representations align with biological priors such as regulatory motifs, thermodynamic stability, and position-specific signatures. Our findings establish key principles: hybrid architectures combining complementary components outperform single-paradigm approaches; multimodal integration yields more robust predictions; transfer learning offers substantial efficiency advantages; and interpretability mechanisms are essential for trust and actionable insights. By integrating biological knowledge with deep learning, this work advances our ability to decode genomic regulatory logic, establishing methodological foundations for continued progress in computational gene regulation understanding.en-AU202610.25911/QH0M-YF70