Similarity metric learning on perturbational datasets improves functional identification of perturbations

Analysis of high-throughput perturbational datasets, including the Next Generation Connectivity Map (L1000) and the Cell Painting projects, uses similarity metrics to identify perturbations or disease states that induce similar changes in the biological feature space. Similarities among perturbations are then used to identify drug mechanisms of action, to nominate therapeutics for a particular disease, and to construct bio-logical networks among perturbations and genes. Standard similarity metrics include correlations, cosine distance and gene set enrichment methods, but these methods operate on the measured features without refinement by transforming the measurement space. We introduce Perturbational Metric Learning (PeML), a weakly supervised similarity metric learning method to learn a data-driven similarity function that maximizes discrimination of replicate signatures by transforming the biological measurements into an intrinsic, dataset-specific basis. The learned similarity functions show substantial improvement for recovering known biological relationships, like mechanism of action identification. In addition to capturing a more meaningful notion of similarity, data in the transformed basis can be used for other analysis tasks, such as classification and clustering. Similarity metric learning is a powerful tool for the analysis of large biological datasets.