基于生物学知识库的临床预测深度学习算法研究进展  

Research progress in deep learning algorithms for clinical prediction based on biological knowledge base

作　　者：曾园媛 游至宇 周颖 李奇渊 ZENG Yuanyuan;YOU Zhiyu;ZHOU Ying;LI Qiyuan(National Institute for Data Science in Health and Medicine,Xiamen University School of Medicine,Xiamen 361102,Fujian Province,China;The First Affiliated Hospital of Xiamen University)

机构地区：[1]厦门大学医学院健康大数据国家研究院,福建厦门361102 [2]厦门大学附属第一医院

出　　处：《中国数字医学》2023年第8期42-50,66,共10页China Digital Medicine

摘　　要：深度学习在生物医学领域具有较高准确率和自动提取特征的优势,已广泛应用于疾病预测。由于网络复杂且参数众多,为确保预测结果的稳定性和可靠性,进一步了解疾病的发生机制,深度学习模型的可解释性成为亟待解决的关键问题。通过将现有的知识框架,如信号通路调控网络,与深度神经网络相结合,构建具有生物学可解释性的深度学习模型,即可视化神经网络。本文总结了近5年来关于可解释生物深度模型的研究成果,并阐述了可解释模型(可视化)的原理,这些模型主要应用于肿瘤、遗传疾病和药物合成等领域。可视化神经网络降低了模型的复杂度和计算成本,逐步建立起一种推动疾病诊断、治疗和药物发现的生物信息学新范式。Deep learning,with its advantages of high accuracy and automatic feature extraction,has been widely applied in the field of biomedicine for disease prediction.However,because of the complexity and numerous parameters of the network,the interpretability of the deep learning model has become a critical issue to be addressed in order to ensure the stability and reliability of the prediction results and further understand the mechanism of the disease.By integrating the existing biological knowledge frameworks,such as signaling pathway regulatory networks,with deep neural networks,a deep learning model with biologically interpretability,known as visible neural network,was constructed.This paper summarizes the research progress on interpretable biological deep learning models in the past five years and elaborates on the principles of interpretable(visualized)models,which have been extensively applied in various fields including tumors,genetic diseases and drug generation.Visualized neural networks can reduce the complexity and computational cost of models,gradually establishing a new bioinformatics paradigm that advances disease diagnosis,treatment and drug development.

关 键 词：深度学习可解释性 可视化神经网络 生物信息学新范式 临床预测 

分 类 号：R319[医药卫生—基础医学] R857.3