Hit screening with multivariate robust outlier detection

Hit screening, which involves the identification of compounds or targets capable of modulating disease-relevant processes, is an important step in drug discovery. Some assays, such as image-based high-content screenings, produce complex multivariate readouts. To fully exploit the richness of such data, advanced analytical methods that go beyond the conventional univariate approaches should be employed. In this work, we tackle the problem of hit identification in multivariate assays. As with univariate assays, a hit from a multivariate assay can be defined as a candidate that yields an assay value sufficiently far away in distance from the mean or central value of inactives. Viewed another way, a hit is an outlier from the distribution of inactives. A method was developed for identifying multivariate hit in high-dimensional data sets based on principal components and robust Mahalanobis distance (the multivariate analogue to the Z- or T-statistic). The proposed method, termed mROUT (multivariate robust outlier detection), demonstrates superior performance over other techniques in the literature in terms of maintaining Type I error, false discovery rate and true discovery rate in simulation studies. The performance of mROUT is also illustrated on a CRISPR knockout data set from in-house phenotypic screening programme.