Two years ago, our conversations with leading sponsors about AI-generated digital twins of patients started the same way. 

Is the method real? 

Will an agency accept it?

Who wants to risk their study’s success on a new technology? 

We’ve moved past those questions, mostly because AI is commonly used in day-to-day work and is being intentionally pursued to fix the sluggish pace of clinical development. But the questions being decided now are harder: who builds the long-term infrastructure and capabilities, what is the data foundation, and how do you build both to hold up when the FDA or the EMA asks about them five years from now? We have witnessed some of the larger sponsors in the industry attempt to build that capability in-house: “digital twin” models built on data they have access to, including their own clinical trials, that their own teams could run. This is the right move; internal teams that know an indication and their molecule’s pharmacology better than anyone should be the ones training and using the models that inform their decisions. However, getting adoption of these internally built models remains the biggest hurdle.

Those internal teams ask all too familiar questions:

Is the method real? 

Will an agency accept it?

Who wants to risk their study’s success on a new technology? 

What these sponsors need is a trusted partner. One that is incorporating digital twins in dozens of prospective studies today. One that has worked hand in hand with regulators to shape how digital twin models are used in studies. One that has built the data and technology infrastructure from the ground up over nearly a decade. And not one that is using digital twins as a bolt-on marketing tactic. 

These sponsors need Unlearn.

The model, the mandate, and the data underneath

Internal builds run into the same three issues.

The first is the model. Much of what the industry now calls a digital twin is a single-study model with a better name, and some of it is a matched historical cohort with a marketing budget. Clinical programs often run over a decade. Throughout this time, endpoints, visit schedules, and target patient populations evolve, and the model should be flexible to account for these changes. A model trained for one specific study measuring specific endpoint(s) in a specific patient population answers a narrow set of questions. It cannot follow the program when the strategy evolves and the next study changes the target population, study duration, or endpoint, and it must be rebuilt and revalidated before it can be defended in front of a regulatory agency. 

What the industry should be focusing on are disease models that predict outcomes across a range of patient populations so the model is ready to be used across many studies in different clinical programs with variable characteristics without major rework and revalidation.

The second is getting clinical development teams to use them. A digital twin that never reaches a protocol changes nothing, no matter how good the model behind it. Between a mandate from R&D leadership and a study that prospectively uses digital twins, there’s clinical development, biostatistics, regulatory, and governance teams.  Each team has its own expertise that informs their questions on how digital twins can add value to a study.  This leads to a lot of internal debate and swirl on how best to move forward.  The industry needs a partner that has spent years building trust cross-functionally and knows how to get teams aligned to be excited and comfortable about incorporating digital twins into their studies.

The third is data. Even with the first two points solved, the foundation of making digital twin adoption a reality is the data. In one recent engagement, we harmonized a sponsor's clinical data from more than twenty of its own studies and provided the sponsor with the ability to view the data in aggregate for the first time.  When we showed them the dashboard, they immediately asked if more users could get logins - a clear “aha moment” for the sponsor. This situation is not unusual. We’ve seen the same pattern at many large enterprises, including those with large data lakes and established repositories, and we’ve seen it across neuroscience, obesity, oncology, and other therapeutic areas. Storing the data is rarely the problem. Harmonizing it so it is useful for clinical insight and modeling is, and that problem likely exists at every sponsor that has run more than one study in an indication. Harmonizing data so it is useful for digital twin modeling is the foundation, and nothing else works without it.

Harmonized data first: what we’ve learned working with leading sponsors

In many engagements, Unlearn has been asked to harmonize a sponsor's trial data into one governed, analysis-ready layer the sponsor keeps. The end goal of having this data is for Unlearn to build a digital twin model to support a sponsor’s clinical development program. Unlearn builds and validates a disease model on that data. Once the model is built, it’s deployed in the sponsor's environment. This is the base capability, but it's insufficient to make the model useful for the long haul of a clinical program. Alongside the model are credibility assessments built on the FDA’s seven-step framework, validation documentation that is paramount for regulatory acceptance, carefully thought-through text for statistical analysis plans, and the myriad of traceability and provenance processes needed for a submission. 

These models are used when a study is being designed, studies are in progress, and after study readout. “In design” means simulating candidate designs and sizing the study on the sponsor's own evidence, enhancing whatever insights can be compiled from historical studies and literature. In progress, it means an interim read sharper than the raw data allows. After readout, it means a more precise estimate of the treatment effect and a model already better suited for the next study. The field guide walks through the methods behind each of these decision points.

That is how our largest engagements are set up today, one study at a time. The next step is carrying the same model and data across clinical programs and therapeutic areas. We also want to empower the sponsor’s technical users to work directly with digital twins. 

Building digital twins of patients before there was a market for them 

Unlearn has been working on these problems since 2017, when building machine learning models of disease progression was an eccentric thing to base a company on, and nobody was opening a chat window to ask an AI anything. The principles Jon Walsh wrote down earlier this month came out of that work: build large, relevant datasets, build general models and fine-tune them for the use case, and validate for the decision the model will inform. Digital twins are those principles applied to patients. Most of the field is still learning them, and a few are learning them from our papers.

The regulatory foundation for incorporating digital twins into clinical studies came from the same discipline. PROCOVA went through EMA qualification in 2022, sits inside the FDA's 2023 guidance on covariate adjustment, and is an example FDA staff have used when they describe a low-risk use of AI in drug development. More than a dozen agency interactions since have covered sample size reduction, power gains, and model credibility. As Luis Olmos wrote earlier this month, the innovations that last are the ones that advance alongside regulatory science. Unlearn continues to build on this foundation and brings this expertise when helping partners build digital twin capabilities.

A model that gets better with every study 

Today, sponsor-built digital twin models are intended for one study and set aside when that study reads out. The right unit for that model is the disease. A disease model built on an organization's own harmonized data addresses questions across populations and endpoints that no single-study model was trained to answer.  This supports running simulations to decide on program-level and study-level strategy, covariate adjustment in one study, an external control arm in the next, and it stays with the program with its provenance intact. Within a few years, building a new model for every study will look like running power calculations by hand. The sponsors who get there first will design studies their competitors cannot, using data they already own.

What Unlearn is building now puts that in sponsors' hands: the tools it built to do this work for itself, so a sponsor's own data science and biostatistics teams can harmonize, build, validate, and govern their models with Unlearn's scientists alongside them. Each study adds to what the model knows, and by the time the next protocol is drafted, the team is simulating against everything the organization has already learned.

If your team is building this now, or has tried and stalled, we would like to compare notes.