Development and in vivo validation of a machine learning platform to predict ototoxic potential of pharmaceuticals

ABSTRACT Over 150 drugs are linked to significant hearing loss, leading to hearing impairment in over one million people per year. For many drugs, ototoxicity is primarily based on individual case studies, so the true ototoxic burden is largely unknown. Further, medication use is heavily skewed towards elderly patients who often have pre-existing hearing loss. We are unlikely to detect a specific ototoxicity in these patients, nor should we use patients as a testbed for harmful side-effects. We propose a rapid, low-cost solution to detect ototoxicity during pre-clinical drug development. We developed an innovative machine learning (ML) model that uses proprietary algorithms to compare chemical and molecular features of new drugs to a curated database of known ototoxins and non-ototoxins to predict the relative ototoxic potential of these drugs. Our strong pilot data show that this model categorizes ototoxins vs. non-ototoxins with 87% accuracy, within the range of similar models that predict toxicity in other organs. The long-term goal is to develop an online portal with a user-friendly interface for customers to incorporate our ototoxicity prediction model into their early-stage drug development pipelines before they invest substantial time and capital in a lead compound. The goal of this Phase 1 STTR proposal is to improve and experimentally test our ML ototoxicity prediction technology. We will accomplish this goal with two Specific Aims: 1) Optimize and apply our ototoxicity prediction model, and 2) Experimentally validate model predictions in the zebrafish lateral line. For Aim 1 we will use semi-supervised learning, unsupervised learning, and transfer learning approaches to further optimize our model to achieve 90% accuracy, then apply our model to a database of FDA-approved drugs. For Aim 2 we will determine the accuracy of model output using the zebrafish system, which is an ideal platform for rapid in vivo ototoxicity validation. We plan to build upon our zebrafish experience to build a Contract Research Organization (CRO) as a value-add for customers who use our ML service. The FDA recognizes that the lack of preclinical ototoxicity testing presents a danger to patients and that there is a need for cost- effective, sensitive nonclinical ototoxicity assessment options, providing an exciting opportunity for our technology. Further, the new NIDCD Strategic Plan calls for application of ML algorithms to prevent auditory disorders. Our unique combination of experts in ototoxicity, preclinical drug development, machine learning, and commercialization of AI-based technologies ideally positions us for success in this Phase 1 project.