Insilico Medicine ("Insilico"), a clinical-stage end-to-end generative artificial intelligence (AI) drug discovery company, has demonstrated that it can predict the outcome of Phase II to Phase III clinical trial success using its proprietary transformer-based AI clinical trial prediction tool called inClinico with a high degree of accuracy. The research has been published in Clinical Pharmacology and Therapeutics, an authoritative cross-disciplinary journal in experimental and clinical medicine.
The AI engines used in the study are integrated into Insilico's inClinico system designed to predict the outcomes of clinical trials and is a part of the Medicine42 clinical trials analysis and planning platform.
The inClinico system is now available for use in pilots, collaboration programs, as well as for use by qualifying industry analysts, hedge funds, and banks interested in comparing human analytical performance with the performance of multiple AI algorithms. The research paper included three types of validation of AI engines trained to predict the probability of success of Phase II trials including retrospective, quasi-prospective, and prospective validation.
The transformer-based AI software platform inClinico combines various engines leveraging generative AI and multimodal data (including text, omics, clinical trial design, and small molecule properties), and was trained on over 55,600 unique Phase II clinical trials over the last 7 years. The subsequent model for clinical trial probability of success developed by Insilico researchers demonstrated 79% accuracy on the outcomes of real-world trials in the prospective validation set where those outcomes were able to be measured.
Around 90% of drug development fails at the clinical stage for reasons including inability to show efficacy, safety concerns, and the complexity of diseases and data, resulting in the loss of trillions of dollars and decades of work.
"Clinical trial failures are complex problems that AI is uniquely positioned to solve," says Alex Aliper, PhD, president of Insilico Medicine and one of the paper's authors. "With this tool, we can help companies determine which programs to prioritize and give investors critical insights into the drug discovery programs that are most likely to succeed."
The inClinico platform was validated in retrospective, quasi-prospective, and prospective validation studies internally and with pharmaceutical companies and financial institutions. The platform achieved 0.88 ROC AUC in predicting the Phase II to Phase III transition on a quasi-prospective validation dataset, a measure of performance in machine learning indicating a high level of discrimination capability.
The first prospective predictions were made and placed on date-stamped preprint servers in 2016. In addition to forecasting outcomes for several Phase II clinical trials and achieving 79% accuracy for the recently completed trials in the prospective validation set, inClinico also demonstrated the platform's usefulness to investors – using a date stamped virtual trading portfolio demonstrating 35% 9-month return on investment (ROI). The paper notes that investment portfolios' performance is dependent on numerous factors, many of which are very hard to foresee or model, and that these findings are intended to instead serve as proof-of-concept based on empirical evidence.
"Currently, more than half of Phase II trials fail, resulting in the loss of ten of millions of dollars and decades of effort. Accurate prediction of the likelihood of success of Phase II to Phase III transitions could be a game changer, giving biotech and pharma companies the opportunity to steer clinical trials toward successful outcomes earlier in the drug discovery process, and providing investors with valuable insights about which drugs in development are most likely to succeed," says Alex Zhavoronkov, PhD, founder and CEO of Insilico Medicine and the study's co-author.
The findings indicate that target choice is much more likely to impact clinical trial outcome prediction than trial design, underscoring that lack of efficacy is the primary driver of clinical trial failures. And the tool's successful prediction included that of LNP023, a first-in-class factor B inhibitor for the rare, life-threatening blood disease paroxysmal nocturnal hemoglobinuria, indicating that inClinico could be useful even without prior information on the clinical relevance of the mechanism of the drug's action in the disease.
The findings also demonstrate that inClinico could provide technical due diligence insights for investors, as well as help pharma companies prioritize their drug development programs.
This promising outcome gives us much to build on. We plan to continue to refine this tool by identifying even more granular components of clinical trial protocol affecting clinical success as well as developing generative AI models that can create optimal clinical protocols from scratch and generate the most relevant criteria for patient selection."
Petrina Kamya, PhD, Head of AI Platforms, President of Insilico Medicine Canada and co-author
Aliper, A., et al. (2023) Prediction of clinical trials outcomes based on target choice and clinical trial design with multi-modal artificial intelligence. Clinical Pharmacology & Therapeutics. doi.org/10.1002/cpt.3008.