A wave of new drug candidates is coming, many discovered with AI, and each one will hit the same bottleneck in the clinical development pipeline: finding, enrolling, and retaining patients.

That was the premise of "More Drugs, Faster: Can Clinical Development Keep Up?", the webinar Unlearn hosted on September 28.

Our Chief Scientific Officer, Jon Walsh, was joined by Dr. Tala Fakhouri, Chief AI and Regulatory Strategy Officer at Parexel, who previously led the development of AI policy for FDA's Center for Drug Evaluation and Research, and Dr. Barbara Bierer, Faculty Director of the MRCT Center of Brigham and Women's Hospital and Harvard.

They covered where AI is already making trials faster and better, what regulators want to see, and why, in Tala’s words, the fear of engaging with regulators is a people problem and a culture problem.

The full recording, including the audience Q&A, is available on demand. [Watch the recording]

1. More drug candidates, the same bottleneck: patients

Tala described what she hears from sponsors: as AI-driven discovery gets potentially more accurate at predicting which candidates should advance, more drug candidates hit the same bottleneck. In rare disease, there are only so many patients for a given indication, and the field already struggles to find enough patients to enroll.
‍
Her conclusion: there simply won't be enough patients to run all of these trials. The field will have to rely more on modeling and simulation, and get better at the science of digital twins, external control arms, and synthetic controls.

Barbara widened the lens to the trial infrastructure itself. Hear which parts she says were never optimized for efficiency or quality, and Tala’s advice on data quality.

2. Don’t be scared of the regulators

Tala said regulators are always encouraging everyone to come talk to the FDA, but from the industry side this rarely happens. Sponsors wait until they have something concrete. Legal and quality teams say no because they are not sure the regulator will accept it. Her diagnosis of where the fear comes from is what she called PTSD: one negative engagement with the regulator that taints every interaction that follows.

Her practical advice covered:

• There is more than one shot on goal: FDA’s draft guidance on AI in regulatory decision-making ends with a table of engagement pathways, and the right door depends on the context of use.
• Why to ask for specific experts to be in the room.
• What separates a useful engagement from a poor one.
• Why publishing use cases, successes and failures alike, matters. Tala and Jon recently posted a preprint applying FDA's draft AI guidance to a digital twin use case.
  ‍

Jon described how Unlearn’s own interactions with FDA have gone since one of its first, a Critical Path Innovation Meeting about six years ago. Barbara raised a counterpoint: insiders know whom to ask for, and many in the extramural community do not. Tala had an answer for that.

3. Evaluate the human-plus-AI team, not AI versus the old way

Tala said she is not currently seeing any autonomous AI in clinical research, and does not know when she will, because the ecosystem is risk-averse for good reasons. What she does see is confusion caused by a false choice: the old way or the AI way. The right question is how the process performs with the humans, AI, and guardrails together, compared with the traditional standard.

She gave two examples of a better method being held back by fear of how regulators will react, and said how she thinks regulators would actually respond.

Jon added that human-in-the-loop may be a long, possibly permanent part of this field. Models and data improve decision-making, but people still make the decisions. 

4. There are two applications the panel singled out

1. Informed consent: Barbara called this a softball because current practice leaves so much room for improvement. Forty-page consents get "shortened" by reducing the font size.
   
   
2. Monitoring as data accumulates: Tala said many in clinical research are already using AI here, detecting anomalies and struggling sites before traditional methods would.
   
   Learn what the MRCT Center is building to reimagine consent, how Tala expects trial data to reach regulators five years from now, where Barbara draws the line between monitoring data quality and judging efficacy or safety, and where Jon says digital twins help.
   

5. Single-arm trials: fit for purpose, and not only for rare disease

Jon framed the question: as drugs become more effective, equipoise erodes, and it gets harder to justify giving patients standard of care. So how do you measure efficacy without a control arm?

Tala pointed out that in rare disease and oncology, regulators are more open to innovative designs. So why not in more prevalent conditions, where she says the modeling can probably be done better?

Barbara laid out when single-arm designs are easiest and when they are hardest. Watch the recording for Tala's answer, her advice to sponsors working in prevalent conditions, and the one difference between the U.S. and European systems that Barbara says complicates it.

Audience questions the panel took on:

• Misalignment within FDA. How often do FDA divisions give differing advice, and what should a sponsor do about it?
• The skills gap. Which skills are most often missing when trial teams adopt AI? Tala has a name for the approach that does not work: trickle-down AI.
• Data sharing. Can the FDA push the industry to share more control data?
  ‍

The answers, and the audience Q&A, are in the full recording.

If the conversation raised questions about a study you are planning or running, reach out.