Artificial intelligence is moving from a peripheral tool in clinical development into the substrate of how sponsors make decisions. In this interview, Aaron Smith, founder and head of AI at Unlearn, describes how the company is building what he calls a scientific intelligence layer — a platform that connects digital twin models, patient-level data, and historical literature into the decisions clinical teams face across the trial lifecycle.
Smith explains that Unlearn’s machine-learning models — built on harmonized clinical trial data, registries, and real-world data — were originally designed to provide counterfactual predictions of how patients or cohorts would progress under control or standard-of-care. But he argues the larger story is how those models and technology can drive key decisions in clinical development during trial planning, monitoring and analysis.
He points to recent work that illustrates the approach. A new Unlearn preprint in oncology shows how recent summary statistics from the literature can be folded into digital twin models, combining historical precedent with newer results for which patient-level data isn’t available. The same simulation models can be repurposed for blinded data monitoring, flagging clinical sites whose incoming data deviates from the patterns the model expects — a quality check that runs without unblinding the trial. And precedent packages built from published literature can feed directly into computational tools that estimate the probability of technical success, power, and sample size.
Smith also discusses the regulatory work that underwrites all of this. Unlearn’s PROCOVA method — prognostic covariate adjustment, which brings digital twin predictions into the analysis of randomized controlled trials, not just the planning — has received EMA qualification and aligns with FDA guidance, and Smith says Phase 3 trials using digital twin–based analysis are imminent. He frames credibility with regulators as a function of three things: a clear context of use, a controlled understanding of risk, and a validation plan that proves model outputs hold up under scrutiny. His advice to sponsors mirrors FDA’s own: “Take the regulators seriously, and interact with them early and often.”
The conversation closes on synthetic controls, where Smith sees real pressure from ethics and feasibility in oncology and rare disease, but pushes back on overreach. A fully synthetic control “oracle” is a long-term scientific possibility, he says, but today’s reality is hybrid: Unlearn has used digital twin–based synthetic comparators in CNS trials, primarily to sharpen internal decision-making, and pairs them with methods like doubly robust ML to maintain bias control when historical information enters a regulated analysis.