基因与细胞图共驱动的协同自编码插补算法  

Imputation algorithm based on collaborative autoencoder network propelled by gene-graph and cell-graph

作　　者：潘祥 丁龙珍 汪佩 厉力华[2] PAN Xiang;DING Longzhen;WANG Pei;LI Lihua(School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi 214122,China;Intelligent Health Research Institute,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]江南大学人工智能与计算机学院,无锡214122 [2]杭州电子科技大学智慧健康研究院,杭州310018

出　　处：《东南大学学报(自然科学版)》2025年第2期575-584,共10页Journal of Southeast University:Natural Science Edition

基　　金：国家自然科学基金资助项目(W2411054,62271178).

摘　　要：为解决单细胞RNA测序数据中存在大量丢失值的问题,提出了一种基因与细胞图共驱动的协同自编码网络插补(CoImpute)算法。该算法利用基因-细胞图神经网络模块,提取基因维度和细胞维度上的生物相似和差异信息,协同辅助基于零膨胀负二项分布的自编码网络,系统全面地恢复单细胞RNA测序数据中的丢失值,且辅助过程创新性地采取图空间上下文感知策略。结果表明,CoImpute算法在14个数据集上的平均绝对误差为0.01~0.21,皮尔森相关系数为0.91~0.97,准确预测了基因丢失值。在不同的稀疏度条件下,CoImpute算法均能保持最小的平均绝对误差和最大的皮尔森相关系数,展示了其处理高噪声数据的优越性能。该算法不仅可以有效恢复丢失值,而且能够提供更准确的数据分析结果和生物学解释,促进了单细胞转录组动力学。To address the issue of a large number of dropout values in single‑cell ribonucleic acid(RNA)sequencing data,a collaborative imputation autoencoder network(CoImpute)algorithm is proposed.This algorithm utilizes a gene‑cell graph neural network module to extract biological similarities and differences in both gene and cell dimensions.An autoencoder network based on the zero‑inflated negative binomial distribution is synergistically aided.Dropout values in single‑cell RNA sequencing data are systematically and comprehensively restored.The auxiliary process innovatively adopts a graph‑space context‑aware strategy.The results show that the CoImpute algorithm achieves a mean absolute error(MAE)of 0.01 to 0.21 and a Pearson correlation coefficient(PCC)of 0.91 to 0.97 across 14 datasets,accurately predicting the missing gene values.Under various sparsity conditions,the CoImpute method can maintain the lowest MAE and the highest PCC,demonstrating its superior performance in handling high‑noise data.This algorithm can not only effectively restore dropout values,but also provide more accurate data analysis results and biological interpretations,promoting the dynamics of single‑cell transcriptomics.

关 键 词：单细胞RNA测序数据 丢失值 插补 基因-细胞图神经网络 协同自编码网络 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]