Abstract 1146P
Background
While significant progress has been made in developing new therapies for cancer patients, many patients lack treatments that result in favorable outcomes. Existing patient-therapy matching algorithms frequently rely on mutations or other well-studied targets for which limited FDA-approved therapies exist. In contrast, SHEPHERD’s approach, called DELVE, uses computational and mathematical tools informed by transcriptomic data to match therapies with the models, cancers, and specific patients that will be most impacted by drug treatment, regardless of mutational status.
Methods
DELVE leverages bioinformatics, chemoinformatics, proprietary algorithms, deep learning neural networks, random forest classifiers, and other tools to generate transcriptomic-level drug response-resistance signatures. DELVE was deployed to characterize drug response and resistance across thousands of in vivo, ex vivo, and in vitro cancer models and over 75,000 patient samples representing 125 cancers and healthy tissues.
Results
DELVE was able to correctly classify the highest and lowest responding drug-cell line pairs with 96% sensitivity [CI +/- 0.34%] and 88% specificity [CI +/- 1.71%]. Across published in vivo studies related to 20 FDA-approved cancer therapies, drug-model pairs which achieved a high DELVE score showed greater than or equal to 90% tumor growth inhibition in vivo 78% of the time. Drug-model pairs with low DELVE scores showed less than or equal to 60% tumor growth inhibition in vivo 77% of the time. Across 71 FDA-approved drugs, using chemical structure alone, the platform was able to predict at least one approved solid tumor indication 84% of the time. Random chance indication-therapy pairing was correct only 21% of the time. Across 10 failed therapies, the platform was able to predict failure with 82% accuracy.
Conclusions
DELVE is able to predict drug efficacy on cancer models and to correctly select indications for existing therapies, supporting its utility in predicting new indications for cancer therapies. Because the platform can operate on any single transcriptome, it is possible to use this tool to match patients with therapies from which they may benefit without reliance on a previously classified target.
Clinical trial identification
Editorial acknowledgement
Legal entity responsible for the study
Shepherd Therapeutics.
Funding
Shepherd Therapeutics.
Disclosure
M. Grushko, J. Goldstein, Z. ElSeht, A. Alarcon, M. Samizadeh, Y. Zhu, J. Kaplan, K. Arline: Financial Interests, Personal, Ownership Interest: Shepherd Therapeutics. N. Jones: Financial Interests, Personal, Full or part-time Employment: Shepherd Therapeutics.