Predicting protein targets for drug-like compounds using transcriptomics

The development of an expanded chemical space for screening is an essential step in the challenge of identifying chemical probes for new, genomic-era protein targets. However, the difficulty of identifying targets for novel compounds leads to the prioritization of synthesis linked to known active scaffolds that bind familiar protein families, slowing the exploration of available chemical space. To change this paradigm, we validated a new pipeline capable of identifying compound-protein interactions even for compounds with no similarity to known drugs. Based on differential mRNA profiles from drug treatments and gene knockdowns across multiple cell types, we show that drugs cause gene regulatory network effects that correlate with those produced by silencing their target protein-coding gene. Applying supervised machine learning to exploit compound-knockdown signature correlations and enriching our predictions using an orthogonal structure-based screen, we achieved top-10/top-100 target prediction accuracies of 26%/41%, respectively, on a validation set 152 FDA-approved drugs and 3104 potential targets. We further predicted targets for 1680 compounds and validated a total of seven novel interactions with four difficult targets, including non-covalent modulators of HRAS and KRAS. We found that drug-target interactions manifest as gene expression correlations between drug treatment and both target gene knockdown and up/down-stream knockdowns. These correlations provide biologically relevant insight on the cell-level impact of disrupting protein interactions, highlighting the complex genetic phenotypes of drug treatments. Our pipeline can accelerate the identification and development of novel chemistries with potential to become drugs by screening for compound-target interactions in the full human interactome.