NPIF Phenotypic virtual screening - exploiting machine learning techniques to derive predictive models from high content screening datasets

尽管应用了更多的资源和更现代的技术，但识别和验证新治疗靶点的传统发现方法的有效性一直在急剧下降。一流治疗学的发展来自于表型筛选工作，也许能够在细胞反应的最终读数中捕捉复杂的系统生物学和网络效应，并仅引起人们对有效化合物的注意[1]。爱丁堡大学的奥尔和卡拉格实验室处于独特的地位，可以将机器学习技术应用于现有的高含量表型数据集，创建一个新的虚拟筛查领域。传统上，有两个分支：基于结构的，通过对感兴趣的蛋白质目标进行建模，并将小分子对接到它以预测结合，从而预测酶抑制潜力。基于配基的虚拟筛选旨在利用小分子结合剂和抑制剂的知识来预测类似活性的小分子。我们提出了第三种方法：基于表型的虚拟筛选，根据观察到的分子与表型诱导的生物活性分子的相似性来评估分子。这个项目的成功候选者将把机器学习技术应用到现有的高含量成像数据的大型数据集(卡拉格实验室，目前是专利的，未发表)，捕捉八个细胞系对14,000种不同小分子的表型反应。细胞表型反应由每个细胞340个特征编码[2]。将教授机器学习技术将表型反应与小分子特征相关联，训练有素的人工智能软件将能够将一系列可用的分子指纹中的特征与表型反应相关联[3]。这项工作的最终用途在于将这些预测器应用于新的、看不见的分子，以便预测能够引起所需的、预先定义的、可开发的、与治疗相关的有效反应的分子。

Traditional discovery methods to identify and validate new therapeutic targets has been undergoing a drastic decline in effectiveness, despite greater resources and more modern techniques being applied. A surprising rise in the development of first in class therapeutics has come from phenotypic screening efforts, perhaps able to capture complex systems biology and network effects in the final readout of cellular response, and drawing attention only to efficacious compounds [1].The Auer and Carragher labs at the University of Edinburgh are in a unique position to apply machine learning techniques to an existing high content phenotypic dataset for the creation of a new field of virtual screening. Traditionally, there have been two branches; structure- based, whereby a protein target of interest is modelled and small molecules docked to it to predict binding and therefore enzymatic inhibition potential. Ligand-based virtual screening aims to use knowledge of small molecule binders and inhibitors to predict similarly active small molecules. We propose a third approach; phenotype-basedvirtual screening whereby molecules are assessed based on their observed similarityto phenotype inducing biologically active molecules.The successful candidate for this project will apply machine learning techniques to an existing, large dataset of high content imaging data (Carragher lab, currently proprietary, unpublished), capturing the phenotypic response of eight cell lines to 14,000 diverse small molecules. Cellular phenotypic response is encoded in 340 features per cell [2]. Machine learning techniques will be taught to associate phenotypic responses with small molecule features, Trained AI software will be able to associate features present in a battery of available molecular fingerprints and associate these with a phenotypic response [3]. The ultimate utility in this work lays in the application of these predictors to new, unseen molecules, in order to predict molecules capable of eliciting a desired, predefined, exploitable, therapeutically relevant, efficacious response.