基于复杂网络的合成致死预测方法研究综述

Review of Network-Based Methods for Synthetic Lethality Prediction

合成致死(Synthetic Lethality,SL)是一种负遗传相互作用,描述的是两个非必要基因之间的相互关系:其中任何一个基因的突变对细胞存活的影响很小,但两个基因的共同突变会导致细胞死亡或其他有碍细胞存活的表型.SL对于解释复杂生物过程、推动癌症的临床诊治有着重要的意义.因此,利用海量的高通量数据,通过构建数据分析模型和计算方法,从计算的角度进行SL对的挖掘和预测,是计算生物学研究的一个重要方向.本文首先对于SL预测所使用的相关数据进行了详细的综述,然后从生物网络这一全新视角出发,重点讨论了基于网络分析的SL预测方法.从网络上的统计学方法、基于网络结构变化的方法、基于网络特征学习的方法、基于图表示学习的方法四个方面综述了相关预测模型和研究的最新进展,详细地比较了各类方法的算法思路、应用场景和优缺点,最后针对SL预测的结果评估和验证方法的研究进展进行了论述.在此基础上,论文进一步总结出SL预测研究中所面临的几项挑战,并针对性的对未来发展方向进行展望,希望为今后的相关研究提供一些有用的参考和思路.

Tumors usually arise from loss-of-function mutations or gain-of-function mutations in oncogenes,which are usually found only in cancer cells,and it has been of great interest to exploit this weakness of cancer cells to develop more effective means of fighting cancer.Synthetic lethality(SL)has been proposed in this context for cancer treatment.Synthetic lethal is a negative genetic interaction that describes the interrelationship between two non-essential genes:the loss of either gene has little effect on cell survival,but the joint loss of both genes leads to cell death or other phenotypes that are poor for cell survival.Synthetic lethal interactions have an important role in explaining complex biological processes and the treatment of human cancers.In order to identify more synthetic lethal pairs and make the concept of synthetic lethality benefit the cancer population,researchers have gone through a process from SL screening in model organisms to SL screening in human cells.However,the high cost of experimental screening and the frequent off-target problems make in vitro screening of synthetic lethal difficult.Therefore,how to mine and predict synthetic lethal pairs through massive high-throughput data analysis is an important direction of synthetic lethal-related research in recent years.In this paper,we focus on synthetic lethality prediction based on network analysis and review recent advances in relevant prediction methods and models in four areas:statistical methods on networks,methods based on the variation of network structure,methods based on network feature learning,and methods based on graph representation learning.We compare in detail the computational ideas,application scenarios,advantages and disadvantages of various methods,and analyze and summarize the main challenges and possible directions of development for synthetic lethality prediction.The statistical methods on networks are transformed from experimental screening,and the biological characteristics of SL distinguished from non-SL pairs are presented in the form of statistical judgments as the screening conditions to complete the prediction task;the prediction methods based on network structure changes take the main idea of simulating the occurrence of synthetic lethal in the network and quantifying its impact on the whole network;the methods based on network feature learning use traditional machine learning methods in the prediction task,and the methods based on graph representation learning,which would be the main direction of future research,represent the network nodes as an low-dimensional vector for the synthetic lethal prediction task.In general,biological network-based synthetic lethal prediction methods are still in the developmental stage,especially for human cells.And the synthetic lethal research faces many challenges:first,SL data consists of a small number of positive samples and a large number of unlabeled samples,and the extremely uneven nature of the data is not a small challenge for all types of computational methods;second,starting from the biological background of SL itself,the same pair of SL presents different pheno-types in different cancers,and this specificity makes its clinical use limited,so it needs special attention from researchers.This specificity makes its clinical use limited and therefore requires special attention from researchers.To address the above challenges,we propose several possible research directions,such as finding reliable negative samples from unlabeled samples for prediction tasks by PU learning heuristics;modeling SL data by knowledge graphs and graph neural networks to use small amount of reliable SL data and other auxiliary biological information as efficiently as possible;minimizing synthetic lethal specificity in research by using single-cell data,etc.This paper may offer some ideas and guidelines for the research of synthetic lethal,and we hope that it will draw increasing interdisciplinary attention from computer scientists,biologists,physicists and so on.