Advancing the Use of Artificial Intelligence In Harmony with Drug Regulation

Recently the FDA released a publication, “Using Artificial Intelligence & Machine Learning in the Development of Drug and Biological Products,” and requested comments on the framework outlined within it from subject matter experts in the industry. VeriSIM Life, which provides a computational drug development platform that uses AI and ML to streamline the drug development process, was eager to contribute our thoughts and feedback to the conversation. The full text of our response can be found at the Regulations.gov website. 

We fully endorse the FDA's commendable initiative to establish standardized protocols for the application of artificial intelligence and machine learning (AI/ML) technologies in drug and biologics research. VeriSIM Life stands firmly behind these efforts, recognizing the transformative potential they hold. VeriSIM Life agrees with FDA that advances in predictions and simulations have greatly facilitated the integration of various factors, including the physicochemical properties of the active pharmaceutical ingredient (API) and the drug's pharmacokinetic and pharmacodynamic properties, leading to more efficient studies while mitigating the risk of each phase for the development of safer and more effective therapeutics for patients.

Furthermore, AI-algorithms are capable of recognizing specific target-related pathways and identify the most efficacious drug target engagement for the potential therapies regardless of dataset complexity (Bello, 2023). This allows for a new level of versatility that is not achieved by traditional methods of in silico model-informed drug development (MIDD). 

That being said, there are many considerations when taking into account how to use AI and ML safely, ethically and responsibly in a drug development context. For instance, in the context of data collection and curation, it is crucial to emphasize that raw data used for model training and result validation must undergo meticulous manual checks and curation. This step is essential to prevent unavoidable discrepancies and outliers. While AI/ML can aid in detecting certain anomalies, “human-in-the-loop” manual review should be considered a mandatory and indispensable practice. Additionally, evaluation of the training data’s integrity and completeness is paramount to ensure the validity and robustness of the AI/ML-driven insights generated during the drug development process. 

Integrating AI with deterministic systems also mitigates issues of transparency and utility. Mechanistic PBPK-PD models have been combined with AI, resulting in “hybrid” models capable of predicting drug PK and PD properties. These hybrid models demonstrate enhanced accuracy, attributed to their transparent simulation and prediction processes, as well as clearly defined result assessment procedures. This approach prevents the indiscriminate application of AI models and enables informed development. (Antontsev, Bundey et al., 2021.) One emerging and highly promising domain for the application of AI lies in the realm of predicting drug toxicity.

Another critical area to pay attention to when implementing an AI-informed approach to drug development is bias. Developers are using various processes to identify and manage bias in AI/ML, such as incorporating diverse and representative data sets during training and validation to ensure the model does not favor certain outcomes. We at VeriSIM Life ensure that BIOiSIM’s datasets include a wide range of populations, demographics, and characteristics to reduce potential for bias, while the AI models can provide more accurate and equitable results. Transparency in model development is crucial for identifying and addressing bias. As a developer, we document the entire AI development process, including data sources, pre-processing steps, and model architecture. This transparency allows regulators and researchers to understand potential sources of bias and take appropriate measures to mitigate them.

At VeriSIM Life, we evaluate and validate AI models for potential bias before deployment. This evaluation process involves assessing model performance across different subgroups and populations to identify any disparities or discrepancies. If bias is detected, it is carefully examined and corrected before the model is put into practical use. And, as bias management is an ongoing process, we perform continuous monitoring to detect and address any bias that may emerge after the model is deployed. This helps ensure that the AI system remains safe and unbiased over time.

In our response to the FDA, we also emphasize that one emerging and highly promising domain for the application of AI lies in the realm of predicting drug toxicity. This is achieved through the integration of hybrid AI-driven PBPK models. “Integrating AI with PBPK-PD models has resulted in the development of hybrid models capable of simulating drug PK and PD properties. These hybrid models demonstrate enhanced accuracy, attributed to their transparent simulation and prediction processes, as well as clearly defined result assessment procedures” (Antontsev, Bundey et al., 2021.)

‍Learn more about how VeriSIM Life has focused on all these areas into its development of the computational platform BIOiSIM, or reach out to connect directly with a member of our team.