Dynamic lncRNA-miRNA-mRNA-Drug Regulatory Networks from Time-Series Transcriptomic Data

The cellular response to pharmacologic agents is orchestrated by a complex, time-dependent interplay between long non-coding RNAs (lncRNAs), microRNAs (miRNAs), and messenger RNAs (mRNAs). Prevailing network models offer an incomplete view, typically modeling these interactions as static, overlooking the critical non-coding RNA layers, and lacking temporal resolution, drugspecific context, and causal inference. To overcome these limitations, we propose a novel computational framework for reconstructing dynamic, drug-perturbed lncRNA-miRNA-mRNA regulatory networks from time-series transcriptomic data. Our method employs a Gaussian Process encoder for robust trajectory smoothing and a Neural Ordinary Differential Equation (Neural ODE) network to model the latent temporal dynamics. A temporal Graph Neural Network (GNN) subsequently generates a time-evolving graph that captures the druginduced rewiring of molecular interactions. We further integrate causal discovery techniques to infer interaction directionality and attention mechanisms to quantify the importance of specific nodes and edges. The framework demonstrates superior performance in denoising high-dimensional data and preserving temporal dependencies. It consistently outperforms state-of-the-art benchmarks, achieving a mean area under the curve (AUC) of 0.93 for network inference accuracy and strong predictive power (AUC > 0.90) for drug response classification. By generating interpretable, dynamic regulatory maps, our approach provides a powerful systems-level tool for uncovering key mechanistic drivers of drug response, with direct applications in pharmacogenomics and therapeutic discovery.