Oncology clinical trials fail frequently due to not fully characterizing the clinical candidates to identify the true target, off-target effects, or mechanism of action. Clinical candidates generally are characterized with limited tools based on the target and the known target indications. These biases have contributed to failed clinical trials. Using cutting-edge novel technologies, like large-scale PRISM profiling of 900 cell lines that have been genomically profiled in the Dependency Map, with machine learning analysis will provide a more thorough and agnostic understanding of oncology drugs. We show that systematic characterization of our Oncology Reference (OncRef) library of over 200 well-known oncology drugs, of which 64% have never been tested before with large-scale cell line profiling, enables better understanding of on-target and off-target effects.

Over 200 drugs, including encorafenib, dabrafenib, pevonedistat, palbociclib, ribociclib, abemaciclib, docetaxel and paclitaxel were screened across ~900 PRISM cell lines at 8 point dose in triplicate. Mixtures of 20-25 cell lines were plated in 384 well plates in RPMI 1640 with 10%-20% FBS. After 5 days, cells were lysed and mRNA was amplified by RT-PCR prior to detection with Luminex FlexMap scanners.

Of the over 200 drugs that were screened, 41% of 192 inhibitory molecules had their target as the top CRISPR or shRNA correlation. This confirmed that PRISM can be used to generate high-quality data in a short amount of time across a large panel of cell lines. Two interesting findings include the reported target of pevonedistat, of NEDD8 activating enzyme (NAE) inhibitor, and the specificity of the CDK4/6 inhibitors palbociclib, ribociclib and abemaciclib. Pevonedistat, a drug that failed in phase III clinical trials in AML was found to be broadly toxic and had no gene expression or functional genomic correlations with the target. If pevonedistat had been tested with a large-scale cell line panel, they would have further optimized the compound prior to clinical trials. For the CDK4/6 inhibitors that are approved in HR+/HER2- breast cancer, palbociclib only had CRISPR and shRNA correlations with CDK6 and ribociclib and abemaciclib only had CRISPR and shRNA correlations with CDK4.

Our findings establish the OncRef dataset and PRISM as a benchmark for evaluating current and emerging cancer therapies. Using large-scale systematic PRISM cell line profiling, genomic characterization and machine learning analysis enables better characterization of oncology drugs and will potentially reduce clinical trials failures.