Unlearn

Unlearn is an enterprise clinical development and biostatistics software company that started out in Cambridge, Boston in twenty seventeen and does only clinical trial design and analysis. It squeezes a patient digital twin into a single "prognostic covariate" and pushes it into an analysis of covariance, cutting sample size by cutting variance — it is not a large language model, and there is no self-serve sign-up.

Core capabilities

• TrialPioneer design bench: Scout searches the literature and regulatory precedents, Hindsight checks historical trial benchmarks, SimLab compares sample-size scenarios, and Monitor watches site-level data anomalies.
• Patient digital twins: baseline data is used to project the longitudinal outcome trajectory a subject would have followed in the control arm; the official site lists twenty-six indications, and the model can be deployed inside a sponsor's own cloud.
• PROCOVA covariate adjustment: the twin's predicted value enters the ANCOVA as a single covariate, lowering residual variance to raise power — the only method for which the company holds a regulatory qualification opinion.
• Conditional restricted Boltzmann machine: the foundation is an unsupervised probabilistic model plus statistical refinements, and the company calls it "a special case of ANCOVA"; searching nine front-end bundles turns up no trace of a third-party large model.

Where it fits Built for pharma companies and clinical-stage biotechs that have Phase II and Phase III trials with continuous endpoints in neurodegenerative or immunometabolic areas, plus an in-house biostatistics team able to talk to regulators. Teams that want self-serve sign-up, or want AI to replace the control arm for a registration filing, are not the audience here.

Distinct advantages There are plenty of players in this lane, but Unlearn is the only one holding a first-hand regulatory qualification document: the EMA opinion from twenty twenty-two is so far the only formal document addressing "machine learning that reduces sample size in confirmatory trials", and competitors will struggle to replicate it in the near term.

Editorial assessment Three gaps. First, the regulators' position is more conservative than the website: that opinion recognises only a prognostic-score covariate in randomised trials with continuous endpoints, and never endorses "a twin replacing the control arm", yet the website writes it up as digital twins as the comparator; the letter of intent submitted to the FDA ISTAND pilot was refused on the grounds that the method is already covered by existing guidance, so "dual certification" is a misstatement. Second, the company's AI lead himself describes fully synthetic controls as merely a "long-term possibility"; the reality today is a hybrid design serving internal decisions only. Third, the reduction in the control arm ranges from eighteen percent to thirty-three percent across channels, all self-reported; there is no public audit report for SOC 2 or GxP, and G2, Capterra and Trustpilot carry no profile at all.