Predictive Modeling of Clinical Trial Outcomes for Novel Drugs using Digital Twin Patient Cohorts and GenerativeAI

ABSTRACT

There is a problem of clinical trial failure, as each new drug should surpass the effectiveness of existing treatment regimens, which becomes increasingly challenging over time. Another significant issue is treating patients who have developed resistance to the current therapies.

Essentially, the use of drug combinations or off-label drug use, where the indication does not match the diagnosis, is akin to an experiment, as there is insufficient data on which drug or combination to use.

This work proposes an approach utilizing computer modeling of patients using gene expression and clinical data. Deep learning and generative adversarial networks are employed as modeling tools. The training data for the algorithms were sourced from publicly available databases such as TCGA and Drugbank.

The modeling is based on the hypothesis of similarity between patients, similarity between drugs, as well as the similarity between individual organs and patient tissues with cell lines, with similarity being computed mathematically. As a result, a patient model is created, where the input consists of drugs and their combinations, and the output provides survival probability values. These model data can be generated in any required quantity with generative adversarial networks (GAN) technology to create observation and control groups. Consequently, it becomes possible to simulate clinical trials, forecasting their outcomes, and, most importantly, optimizing the trial parameters to maximize the likelihood of success.