How Sanofi uses digital twins in modern clinical trials

Technology-forward pharmaceutical companies are increasingly experimenting with digital twins to streamline the clinical trial process.

Digital twins are virtual copies of organs and human bodies companies use to test drugs, and the tech can help predict disease progression and optimize study design. Over the last few years, French pharma company Sanofi has built digital twins to simulate the behaviors of its drugs as well as patient outcomes in dozens of disease areas.

Although artificial intelligence (AI) is currently the industry’s favorite buzzword, Paris-based Sanofi’s digital twins do not rely solely on input from AI models, Matt Truppo, global head of research platforms and computational R&D at Sanofi, told Healthcare Brew. The company’s models also incorporate the quantitative systems pharmacology (QSP) approach that combines math and biology modeling based on differential equations.

AI models are often referred to as “black boxes” because how they derive an output from a given input can be unclear, and this is troublesome if researchers need to explain why a model made a mistake. QSP is “more deterministic,” Truppo added, and it helps make the models more “explainable.”

Informed decisions

To determine which drug candidates will perform best in the market, the model simulates the behavior and performance of a drug asset and its competitor products in a digital environment all the way from cellular protein level through to the organ, the patient, and the group of patients in a clinical trial, Truppo explained.

Data on the patient’s biology and corresponding drug responses are fed into “target ID engines,” a suite of AI and machine learning methods for multimodal data analysis coupled with large language models for data interrogation, which help Sanofi researchers model the progression of a specific disease and find proteins or molecules that could be targeted with a novel therapeutic. The ID engine can also help pinpoint which patient population would benefit most from a given drug candidate.

The researchers then use the QSP approach on available experimental data to predict the possible effects the drug could have on the human body.

Sanofi has built these models to simulate human biology across dozens of conditions in immunology, oncology, and rare disease. “Each of these models enable us to build a digital twin concept for the patient to ultimately run these digital clinical trials,” Truppo said.

The pharma company, for example, used digital twins in the last two years to perform comparative modeling for its asthma drug candidate lunsekimig, which is progressing toward a Phase 3 trial. The predicted clinical endpoints matched the actual outcomes, and the modeling supported data from real patients.

“As we generate more data, it can reinform those models, and it can refine them to behave better and give data that’s more amenable to making decisions,” Truppo added.

As they were building the digital twins for the asthma patients in the lunsekimig study, he said, Sanofi researchers withheld some data from the ongoing trials to see if the models would predict the same outcomes observed in real life—and it did, Truppo added.

Accurately predicting clinical responses means drug makers can shorten the time it takes to run studies—an especially huge benefit for research on rare diseases, which have a limited pool of patients. For one rare disease called acid sphingomyelinase deficiency, or Niemann-Pick disease, Sanofi used digital twin modeling to get its drug candidate Xenpozyme in front of regulators without having to run as large of a study on children.

Proceed with caution

Generative AI technology suffers from well-known problems around transparency and hallucinations, which is when AI models produce inaccurate or nonsensical information.

One of the major cons of AI models is that they tend to be “brittle,” Jun Deng, professor of therapeutic radiology at Yale School of Medicine, told Healthcare Brew. And part of the issue is that the training set is often a very “narrow field,” meaning most of the data comes from a single institution and doesn’t extrapolate to a more diverse population.

Sanofi is working to combat these problems by “ensuring that the demographics of the disease are represented” in its clinical trials, Truppo said.

Better and more diverse data helps models make more accurate predictions, and these AI predictive models tend to work best with diseases when there’s an abundance of clinical data. “We need this data to train, test, and validate digital twins,” Deng noted.