基于深度学习的功能性非编码变异预测研究进展OA

Functional non-coding variants play a pivotal role in genetic susceptibility to human complex diseases.However,due to the complexity and concealment of its regulatory mechanism,accurately predicting its functional effects remains a core challenge in genetics research.In recent years,artificial intelligence technologies represented by deep learning have provided a brand-new computational approach.This paper synthesizes recent advances in deep-learning methods for pre-dicting the function of non-coding variants.It traces model evolution from convolutional neural networks that detect local sequence motifs,through hybrid architectures that combine convolutional and recurrent layers,to attention-based models that employ self-attention to capture long-range dependencies.It also examines the progression of task formulations,including binary classification of regulatory activity,quantitative regression of effect sizes,and multitask learning across cell types and assays.This paper focuses on elaborating the"pre-training-fine-tuning"new framework represented by large genomic language models and its potential for application.It further summarizes the application of the computational methods in downstream tasks such as assisting in causal variation mapping and elucidating the molecular mechanisms of diseases.Finally,the paper discusses current challenges and proposes directions for future research.