Unlearn and Merck collaborate on medical twins trials

Unlearn and Merck KGaA have announced a partnership to accelerate drug trials using medical digital twins of patients. Unlearn uses recent developments from deep learning to create digital twins of patients in clinical trials. The new technique allows drug researchers to reduce the size of control arms by 30% or more and generate reliable clinical evidence in less time. Merck plans to focus on late-stage clinical trials for immunology drugs initially. 

“Using digital twins in the design and analysis of a clinical trial maintains the highest standard of evidence while reducing the number of required patients and making the trial more attractive to potential patients because they are more likely to receive the experimental drug instead of the control,” Unlearn CEO and founder Charles Fisher told Venturebeat. “Therefore, we believe that digital twins will be included in all clinical trials in the future, which is a multi-billion-dollar opportunity.”

The company’s significant innovation is a novel statistics and analytics workflow to minimize bias and errors in predictions. These new techniques could increase efficiency in the drug industry and also improve digital twin simulations in other industries as well. 

Improving research efficiency

The drug industry spends far more on research and development than any other industry to ensure drugs are both safe and effective. The International Federation of Pharmaceutical Manufacturers & Associations estimates drug companies spent $196 billion on R&D in 2021, expected to grow to $213 billion in 2024. Clinical trials make up a significant component of this cost. A study by Presence Research expects the cost of clinical trials to grow from $51 billion in 2021 to $84 billion in 2030.

Regulators have shown an interest in the promise of synthetic data techniques to generate a synthetic control arm that simulates the likely progression of diseases for patients with or without an existing treatment. In the best case, it would be great to give all study participants the new drug and then use the synthetic data as a control. But this can introduce bias, since the synthetic data generated controls are only as good as the fidelity and depth of the actual patient data. 

One recent survey of synthetic and external controls in clinical trials observed that synthetic controls might suffer bias even if the employed external sources appear similar and unbiased. There may be differences with respect to factors that either cannot or are rarely recorded in data. For example, palliative care practices, such as nutrition, massage, or even spiritual interventions, may vary by physician or care center and not be captured in traditional drug research studies. 

How it works

Unlearn’s approach is unique because, while it also enables smaller, more efficient trials, it does not introduce these types of biases and is suitable for most clinical trials under current regulations. Its innovation is a statistical process for simulating the likely progression of disease across experimental and control groups and then verifying the results after the fact for both.  

Unlearn invented two types of technologies. First, it has developed deep learning methods to create digital twins of patients in trials. Second, it has developed methods for using digital twins to design and analyze randomized controlled trials. 

Researchers in a trial first train the digital twin model to create realistic simulations of disease progression using the database of historical patient data. Then, they collect medical records that describe the current state of a particular patient and input the data into our model. This model is designed to simulate how the patient’s disease will progress in the future. 

A clinical trial compares how a patient’s disease would progress on a new drug to how it would progress on an existing drug or placebo (the control). A patient’s digital twin provides information about how their disease would progress on the control, allowing one to determine the difference between the effect of the new drug and the control more precisely.

The process of designing and analyzing trials includes digital twins of all trial participants, whether they are randomly assigned to receive the experimental drug or the control. Digital twins are created for all of these patients at the beginning of the study before their treatment is assigned. The more accurately the digital twins predict patients’ disease progression on the control, the smaller the control group in the trial needs to be. 

Fisher said, “Computer models like digital twins can’t be perfect. Whether the digital twins are created using deep learning or traditional mathematical modeling, the model will make mistakes.”

Unlearn has created a way to incorporate digital twins into clinical trials that are immune to potential mistakes made by the model to overcome this problem. “Similar strategies could be applied in other industries, allowing digital twins to be used for critical applications by preventing inevitable errors in the digital twins from propagating downstream,” Fisher said. 

Unlearn plays in a relatively new market for AI tools to improve statistical analysis. Competitors working on tools for synthetic data in clinical trials include Cytel, Dassault Systèmes Medidata, and FlatIron Health.