Improved Cancer miRNA Biomarker Classification through Adaptive Multi-Resolution Temporal Attention Network (AMRTA Net) with Efficient Transfer Learning

Cancer diagnosis and treatment rely on biomarker discovery, and miRNA profiles are significant tools. The Adaptive Multi-Resolution Temporal Attention Network (AMRTA Net) is a novel miRNA biomarker classification method. AMRTA Net uses better multi-resolution feature extraction and fine-tuned temporal attention to represent miRNA data with fine-grained and global characteristics. AMRTA Net's Hierarchical Resolution Analysis Module analyzes biomarker data at several resolutions while concentrating on critical aspects for cancer classification. A Customizable Transfer Learning (CTL) architecture improves model performance. To effectively learn biomarker-related information, it adaptively modifies just the final few layers of a pre-trained deep neural network. The method uses a Temporal Stability Optimization (TSO) loss function to reduce feature extraction shift at various time points to counteract temporal volatility in miRNA data sets. Biomedical researchers typically lack labeled samples, hence a successful Hybrid Semi-Supervised Learning Strategy adds pseudo-labelled data to the training set. In experimental investigation, AMRTA Net outperformed other approaches in cancer miRNA dataset categorization with 97.2% accuracy. The model is ideal for biomedical diagnostics because to its noise tolerance and strong scalability for big data sets. AMRTA Net, a precision medicine breakthrough, classifies miRNA biomarkers intelligently and accurately using adaptive learning, multi-resolution analysis, and temporal attention.