Scientific Expert Sees AI Used for Specific Use Cases
By Rebecca Stauffer
“I believe AI is a new tool in the toolbox we call model-informed drug development (MIDD),” Certara Senior Vice President and Head of Quantitative Systems Pharmacology (QSP) Piet van der Graaf, Pharm.D., Ph.D., said in his talk during the closing session of the 2024 PharmSci 360.
Van der Graaf’s talk offered a peek at what he called “grounded applications” of AI in pharmacometrics, PK/PD, physiological-based pharmacokinetics (PBPK), PK modeling, and QSP.
AI and Modeling Similarities
When it comes to AI, “it’s really a way to simulate human behavior to automate and accelerate tasks and delivery,” he explained. “This is based on several concepts that are very similar to other modeling approaches. We have a model, and we have an algorithm, and that’s fed by data, and we can ask questions.
“What’s probably generated most excitement, certainly, in recent years, is the whole area of natural language processing, large language models, and these magical kinds of things that are called ‘GPTs.’”
Yet while off-the-shelf GPTs (Generative Trained Processors), such as ChatGPT or CoPilot, offer benefits, they are not perfect solutions. As van der Graaf reminded the audience, “the big issue in our discipline is that we make decisions on therapeutics that are literally about life and death.”
In early 2023, van der Graaf co-wrote an editorial for an AI-themed issue of Clinical Pharmacology & Therapeutics. This editorial identified three main areas of opportunity for AI in pharmaceutical research:
- Drug Discovery: AI applied to target identification, drug design, drug repurposing, and compound screening
- Drug Development: Applications include trial design, patient recruitment, biomarker/endpoint assessment, dose optimization, etc.
- Patient Care: Applications include disease diagnosis/prognosis, treatment personalization, patient monitoring, and patient adherence
“Currently, the main applications are drug discovery and patient care,” van der Graaf said. “That is where we have large datasets. So, it’s very easy to generate lots of data from thousands, hundreds of thousands, or millions of compounds in early discovery. You have massive datasets and AI loves that.”
Drug development does not offer the same massive data troves to analyze. “In actual drug development where we run clinical trials, the datasets tend to be relatively small in the context of what you need for meaningful AI and machine learning (ML),” van der Graaf said.
Returning to the special issue of Clinical Pharmacology & Therapeutics, he discussed an article from authors at the International Consortium for Innovation and Quality in Pharmaceutical Development. Within this IQ consortium, an AI/ML working group consisting of representatives from 14 large pharmaceutical companies drafted this white paper, describing how the field of AI is currently used in drug discovery and development.
“The purpose of this paper was to compare and contrast different methodologies to establish good practices, to encourage working together, and to generate tools and datasets that people can start to leverage,” van der Graaf said.
The paper also outlined how AI applications for drug discovery and drug development fit under quantitative modeling. Van der Graaf drew from much of this paper for the rest of his talk as he examined AI’s potential in both drug discovery and drug development.
What Really Matters in Pharma R&D?
25 years ago, van der Graaf was also a co-author on a paper examining pharmacokinetics predictions using what was then termed “multivariate quantitative structure-PK relationships.”
“Now, you would call that machine learning, and what we showed is that you could predict the pharmacokinetics clearance, volume, and fraction unbound of adenosine A1 receptor agonists pretty well,” van der Graaf said. “Recently, several people have published similar approaches. The main difference now is that the datasets are much, much bigger.”
The next steps in this area, he said, involve integrating AI and machine learning with physiological-based pharmacokinetics (PBPK) for high-throughput pharmacokinetics. This involves mechanistically analyzing the data of large numbers of compounds.
Virtual compounds are the next phase of this evolution. “Do we actually need data on real compounds or can we kind of apply this idea and generate lots and lots of virtual compounds?”
Recently, researchers at Leiden University in The Netherlands created DrugEx, an open-source software library for de novo design of small molecules using deep learning generative models. This tool enables researchers to determine if compounds have favorable pharmacokinetics using an algorithm called QSPRpred. This allows researchers to study multiple properties simultaneously.
“For example, we can look at compounds and examine them simultaneously for binding affinity of the receptor and various pharmacokinetics,” he said. His team actually applied the algorithm to an Adenosine A2A antagonist. Ultimately, they found that simultaneously optimizing for pharmacokinetics and binding affinity remains very challenging.
Van der Graaf turned to research from the University of Oxford, Relation Therapeutics, and the Francis Crick Institute that asks different questions.
“They said that what we really need to do is look at the biological targets because that’s where the big money is, and that’s actually where very little effort is being focused.”
He reminded the audience that the biggest challenge and opportunity to increase probability of R&D success lies in phase II.
“If you could improve that one, it would cause massive improvement in pharmaceutical R&D,” he said. “It’s really what we need to focus on.”
QSP Models Next Stage in Model-Informed Drug Development
The last portion of van der Graaf’s talk examined quantitative systems pharmacology (QSP) applications—the next phase of model-informed drug development as he sees it.
At Certara, there is a QSP model for inflammatory bowel disease covering all the different cell types involved, cell mediators, etc. With QSP, they can generate hundreds of biomarkers. His team is analyzing hundreds of virtual biomarkers to determine clinical outcomes. Currently, they are using this model to identify combination therapies for inflammatory bowel disease therapeutics.
“The model predicted improvement in clinical outcome for these untested novel combination therapies. Here, we can use a QSP model combined with AI and machine learning to run virtual trials to predict actual clinical outcomes for untested scenarios.”
In another case study, this one involving a consortium led by Certara with eight companies, used QSP for immuno-oncology (IO). Virtual trials were run on an IO platform. The model was used to predict clinical response from two compounds used in a combination therapy, ultimately finding a stronger response for the combination therapy as opposed to a therapeutic using just one compound.
In addition, QSP models can be used to select biomarkers by running virtual clinical trials that do not take 12 months like regular clinical trials.
Van der Graaf sees the future of AI in pharmaceutical and development being used strategically in specific areas.
“Let’s not fall into the trap thinking that AI will solve all of Pharma’s challenges in the near term,” he concluded. “Let’s focus on grounded applications by considering AI as another, integrated extension of MIDD.”