Deep Learning for Predicting Gene Splice Sites: Technical Breakthroughs and Biomedical Significance of splice-site-predictor

The splice-site-predictor project uses a dilated pre-activated residual convolutional neural network to predict classic GT-AG splice donor and acceptor sites in human DNA sequences. Trained on the HS3D dataset, it demonstrates the application potential of deep learning in genomics and holds important biomedical significance for rare disease diagnosis, cancer research, synthetic biology, and other fields.

In the process of gene expression, splicing is a key step of removing introns and joining exons, carried out by the spliceosome. Accurate identification of splice sites determines the correctness of gene products; aberrant splicing leads to abnormal protein function and is closely associated with cancer, neurodegenerative diseases, genetic disorders, and more.

To address challenges in splice site prediction such as weak signals, context dependence, and long-range interactions, the project uses dilated convolution (expands receptive field to capture long-range dependencies) and pre-activated residual blocks (more direct gradient flow, better regularization, higher training efficiency). The network architecture roughly includes an input layer (one-hot encoded DNA sequences), an initial convolutional layer, stacked dilated residual blocks, a global pooling layer, a fully connected layer, and an output layer.

HS3D is a benchmark dataset for splice site prediction, containing real splice sites (positive samples with surrounding sequence context) and pseudo-sites that match the GT-AG pattern (negative samples with features similar to positive samples), ensuring the model learns to distinguish key features of real splice sites.

This tool can aid in rare disease diagnosis (pathogenic variant annotation, aberrant splicing detection, drug target discovery), cancer research (diagnostic markers, prognostic indicators, therapeutic targets), synthetic biology and gene therapy (optimizing gene expression cassettes, designing regulatable splicing systems, improving gene therapy vectors).

Current limitations: Only predicts classic GT-AG sites, sequence length constraints, ignores tissue specificity, and does not focus on elements like branch points. Future directions: Multi-task learning (predicting multiple splicing elements simultaneously), introducing attention mechanisms, tissue-specific models, transfer learning, and improving interpretability.