INTRODUCTION
There is great interest in harnessing the power and versatility of artificial intelligence (AI) methods in clinical pharmacology and pharmacometrics to facilitate modeling and simulation, accelerate drug development, and improve patient outcomes in drug therapy.
Participants in clinical trials overall are not fully representative of the population of patients 1-5. In 2011, African Americans and Hispanics comprised 12% and 16% of the United States population, respectively, but only 5% and 1% of clinical trial participants 2. The Food and Drug Administration (FDA) recently issued a guidance to address representation in clinical trials5 that recommended broadening study eligibility criteria and addressing other study design and recruitment logistic factors to improve the participant pools. There is clearly an unmet need and knowledge gap to enable modeling of treatment effects and safety in diverse populations and in under-represented groups.
We posit that when successful, the use of AI technologies in the context of already available “big data” could be a transformative computational strategy to mitigate the impact of under-representation of race/ethnicity groups in pharmacometrics and drug development. The clinical pharmacology and pharmacometrics research community has not meaningfully leveraged potentially promising AI-enhanced approaches to address treatment variability, group-level outcome disparities, and real world clinical applications 6. The increasing availability of public health databases, de-identified health records and pharmacogenomics data provides new opportunities for forecasting of biomarker profiles and drug outcomes in diverse and in under-represented populations 7.
Generative adversarial networks (GANs) are a powerful, deep learning (DL)-based, AI technology that enables realistic simulations of complex patterns 8. GANs utilize two neural networks called the generator and discriminator that learn from data to generate complex high-dimensional pattern distributions. The generator neural network synthesizes instances of plausible samples that conform to the population distribution of interest. The discriminator network is a classifier that attempts to categorize whether each instance presented as input belongs to the training data or has been synthesized by the generator. The adversarial training process results in a generator that produces samples that mimic the training data.
We hypothesized that GANs could be an effective approach for generating realistic biomarker profiles for clinical trial and pharmacometrics simulations. The objective was to design and assess GANs capable of generating disease-relevant biomarker profiles by learning from observations in real-world, population-based studies. An additional goal was to extend the GAN approach for generating biomarker joint distributions for under-represented race/ethnicity groups.