Lee W H, Narang V, Xu H, Lin F, Chin K C, Sung W K
Department of Computer Science, National University of Singapore, Singapore.
Ann N Y Acad Sci. 2009 Mar;1158:196-204. doi: 10.1111/j.1749-6632.2008.03755.x.
In the Dialogue for Reverse Engineering Assessments and Methods Conference (DREAM2) BCL6 target identification challenge, we were given a list of 200 genes and tasked to identify which ones are the true targets of BCL6 using an independent panel of gene-expression data. Initial efforts using conventional motif-scanning approaches to find BCL6 binding sites in the promoters of the 200 genes as a means of identifying BCL6 true targets proved unsuccessful. Instead, we performed a large-scale comparative study of multiple expression data under different conditions. Specifically, we employed a supervised learning approach that learns and models the expression patterns under different conditions and controls from a training collection of known BCL6 targets and randomly chosen decoys. Genes in the given list whose expression matches well with that of the training set of known BCL6 targets are more likely to be BCL6 targets. Using this approach, we are able to identify BCL6 targets with high accuracy, making us joint best performers of the challenge.
在逆向工程评估与方法对话会议（DREAM2）的BCL6靶点识别挑战赛中，我们得到了一份包含200个基因的列表，并被要求使用一组独立的基因表达数据来确定哪些基因是BCL6的真正靶点。最初尝试使用传统的基序扫描方法在这200个基因的启动子中寻找BCL6结合位点，以此来识别BCL6真正靶点，但结果并不成功。相反，我们对不同条件下的多个表达数据进行了大规模的比较研究。具体而言，我们采用了一种监督学习方法，该方法从已知BCL6靶点的训练集和随机选择的诱饵中学习并模拟不同条件和对照下的表达模式。给定列表中那些表达与已知BCL6靶点训练集的表达匹配良好的基因更有可能是BCL6靶点。使用这种方法，我们能够高精度地识别BCL6靶点，使我们成为该挑战赛的并列最佳表现者。
