Expert Q&A: Maya Petersen and Mark Van der Laan on how their work is improving clinical drug trials

UC Berkeley School of Public Health: What are the possibilities you envision for real-world data?

Maya Peterson (MP): We work on the use of real-world data, either alone or in combination with clinical trial data with data-fusion methods—to generate evidence for what types of drugs work and for whom; our methods can be used across the drug-development pipeline, from early stage screening through post-market safety surveillance and the discovery of new clinical uses for existing drugs.

How are you helping the FDA create its framework for adding more real-world evidence to its regulatory approval and drug safety monitoring process?

MP: Mark and I have had lots of engagement with the FDA at multiple levels. The nice thing about working on regulatory approvals is that it’s really where the rubber meets the road, in terms of the standards. They are extremely high and they should be extremely high. If you develop a new methodology, you have to be able to rigorously show that it works before it is used. This is an ideal setting for the methodology that we at the Center for Targeted Machine Learning and Causal Inference have always worked on.

The long-standing premise of our Center is that statistics and the process of learning from data should not be an art. It should be a science. We need methods that can take messy health data and actually deliver rigorous inferential answers to real world questions, along with accurate and understandable quantification of how uncertain those answers are.

How does JICI work?

MP: We develop and test new methods, and then we convene stakeholders around methodology, ours and others’. We’re not the only people in the world, obviously, to say, ‘Okay, this is a space where openly sharing knowledge about best practices, and jointly working to improve these best practices, benefits everybody.’

In this regulatory space, whether you’re talking about designing and analyzing a clinical trial, analyzing real-world data, or putting the two together, if there’s no consensus across what the best practices are, how to implement them and how to document them, then any evidence you generate will have a much more limited impact.

Tell me about the upcoming gathering you have planned for spring 2026.

MP: We’re convening our fourth annual gathering this spring, the Forum on the Integration of Observational and Randomized Data. We call it FIORD. It’s an invitation-only forum that we do in the Washington, DC, area. We bring together folks from the FDA who are methodology-oriented; academic professors, and representatives from the private sector, including Pharma. It’s been great, because it’s just been a really, really healthy culture; very focused on science and consensus building; respectful and productive.

Mark, How did you get started looking at clinical trials?

Mark Van der Laan (MV): I’m a biostatistician, so I’ve always been working with excellent people in causal inference.

The challenge is, how do you really analyze data and learn from it in an objective way? At first the FDA was more interested in using advanced approaches, such as targeted learning, for safety analysis. That’s where my initial collaborations with the FDA started.

On the academic side, I worked with James M. Robins among other top methodologists. We wrote a book together, and eventually my journey in understanding statistical inference from complex observational data converged into an approach called targeted learning. The FDA was already interested in it early on, which was kind of remarkable, and they made it a demonstration project. This interest evolved over time in many more collaborative projects, workshops, and working groups involving Targeted Learning. Now, within pharma companies where they initially were interested in these approaches for post-market safety analysis, they are now also really starting to do it for designing prospective studies for drug approvals or new indications for approved drugs.

Can you mention any specific diseases you’re working on?

MV: There is a real desperate need for these modern approaches, for example, for rare diseases. We have a collaboration with the FDA people and UC Berkeley, and a startup company I’m involved in, called Targeted ML Solutions, as well. In rare diseases the sample size is often small, say 50 people, so it’s very challenging and the kind of approaches we developed can be used for these problems. We meet every few weeks talking about these statistical methods in targeted learning for the purpose of analyzing rare disease trials.

MP: There really isn’t wiggle room in regulatory science. That’s what’s great. There are real benchmarks, and you have to meet them. When you tackle challenging real world data problems, you have to bring in modern machine learning and advanced statistical methods to succeed.

For example, recently we have been working on how to develop high-quality surrogate endpoints that let you screen candidate drugs faster for potential effectiveness. Or, we could also take a single drug and figure out faster if it might work for new things. You might be able to get faster to the first stage —not as the final evidence. This kind of approach can help you find drugs that work faster.

Is the pharmaceutical industry supportive of these efforts?

MP: Yes, absolutely. I’d never worked with drug companies before. I think what makes this such an interesting area of work and such a good thing to be honed academically, is that the incentives are truly aligned.

As academics we want to develop better methods, and we want these methods to be used and have impact. Regulatory bodies want to get better answers sooner, while maintaining the highest standards of rigor. Pharma also wants ways to get better answers sooner,, but they would not like to waste a bunch of money and time pursuing some promising but novel methodology approach and then have the FDA turn around and say, ‘Wait, but this actually doesn’t count as good evidence.’

Jointly advancing both the quality of methodology and consensus around its best practice use is good for everyone.

Do your students get to work on these drug development methods?

MP: Absolutely, students play a key role in our work, both while they are here at Berkeley and after they graduate. CTML sits at the school of public health and at the University of California, and in that sense, a key piece of our mission is to train the next generation of leaders. Our students graduate and go on to advance the process of developing drugs and deploying them in a way that improves lives and health. Our graduates are professors at top universities, and also leaders at the FDA and Pharma companies and at healthcare organizations such as Kaiser. Students are vital to the success of what we do. We wouldn’t be able to do it without them.

Are you getting most of your students from mathematics?

MV: No, we have a very diverse group. We have students who are very interested in math, but a lot of them are just really interested in the translation, and doing work in collaborative groups, working with the FDA, working with Pharma, working in healthcare, in precision medicine. They don’t necessarily have to do the proofs, but they have to understand it. Maya is very good at teaching high-level math in a way that makes sense to people who care about practice and are not necessarily great mathematicians.