Unlearn.AI

The models are the engine of the core offering, and the company now sells a partnership around them. Digital Twin Generators, disease specific machine learning models trained on historical patient level data, forecast each participant's outcomes under control from baseline data, and the twin-adjusted analysis (PROCOVA and its Bayesian and repeated measures variants) is what a sponsor buys.

The solutions pages describe the rest of what is sold: Unlearn scientists who draft the statistical analysis plan and deliver the analysis, an AI-powered shared workspace, harmonized historical data, design scenario simulation and precedent search, and regulatory strategy. PROCOVA itself is a statistical method that works with any prognostic score. Remove the generators and the design, precedent and statistical consulting remain, which is the model as the engine of a core module.

The structure is described for randomized trials and is missing its limits for the newer uses. In a randomized trial the twin supplies a prognostic covariate, randomization is preserved, the analysis is pre-specified before unblinding and delivered by Unlearn scientists, and the method's error control holds even when the model is wrong.

The company now also markets digital twins as the comparator in single-arm studies and the design of externally controlled studies where randomization is not an option (solutions page; arXiv 2026, Digital Twins as Synthetic Controls in Single-Arm Trials), and a blog post says not every use of AI in trial analysis carries the same risk. What is not published is where the company stops: the conditions under which a twin-based external comparator is not appropriate, how bias is bounded there, and who decides.

The approach and model families are published; version and update practice are not. The research list names the model work behind the generators (probabilistic neural networks, 2019; conditional restricted Boltzmann machines for Alzheimer's disease, Nature 2019; Neural Boltzmann Machines, 2023; Digital Twin Generators for Disease Modeling, arXiv 2024) and the statistical methods (PROCOVA, De Gruyter 2020; Bayesian and repeated measures variants), and the data acknowledgments page names training sources such as ADNI.

The technology page lists generators for 26 conditions and says specification sheets on each model's capabilities and underlying data are available on request. No version identifier, release history or statement of how a generator is updated and how sponsors are told is published.

The model layer is the company's own, and its data sources are partly named. The generators are built in house, and the data acknowledgments page names contributors such as the Alzheimer's Disease Neuroimaging Initiative for the Alzheimer's generator; custom generators can run entirely inside a sponsor's environment, where the transfer does not occur.

The privacy notice names website and application analytics providers (FullStory, Sentry, Auth0, Google Analytics) but no hosting provider or subprocessor list for sponsor trial data, and nothing states retention or use terms for data once it reaches the hosted application.

Published research short of independent validation, measuring statistical efficiency. The research list includes a Nature (2019) paper on forecasting Alzheimer's progression, Alzheimer's Association presentations evaluating digital twins on completed phase 2 trial data (2022, 2024), and arXiv papers on the generators and PROCOVA variants.

The solutions pages report re-analyses of completed trials: an 18 percent variance reduction and 92 fewer participants for a phase 3 ALS trial, 17 percent fewer participants for an AbbVie phase 2 Alzheimer's study, and power at an interim rising from 60 to 70 percent with PROCOVA and 78 percent with Bayesian PROCOVA in a completed phase 2 study. These are retrospective re-analyses and simulations run by the company and its partners, not prospective use as the primary analysis of a completed registrational trial, and several are preprints.

General assurances and one route-specific commitment, without terms for sponsor data. The technology page says sponsors can build custom generators from their proprietary data entirely within their own environments and under their control. The privacy notice (effective 1 April 2026) keeps information for as long as reasonably necessary and says its AI supports trial design and analysis rather than decisions about individuals.

For sponsors using the hosted application, nothing states how trial datasets are retained, returned or segregated, or whether data from one sponsor's trials can train generators used for another's. The generators are trained on historical data, including research consortium data acknowledged on the data acknowledgments page (ADNI for Alzheimer's disease). Nothing is published on safety engineering.

Business associate status is stated conditionally, with a substantive privacy notice behind it and no agreement published. The privacy notice (effective 1 April 2026) has a section on healthcare data and HIPAA saying that where the company receives protected health information from a covered entity or business associate and signs a business associate agreement, it commits to HIPAA's safeguards.

Most of its work uses historical and sponsor trial data under sponsor agreements and consent, where no covered entity may be involved, and the notice does not say which engagements fall under which regime. No agreement template or subcontractor terms are published.

A SOC 2 Type 2 attestation is named in the company's own material without the report. The technology and solutions pages say the technology is built to meet GxP, 21 CFR Part 11 and SOC 2 Type 2 requirements. No report, auditor, period or trust center is published, and no ISO/IEC 27001 certification or penetration testing cadence is stated. What the company holds is commercially sensitive: unpublished protocols, endpoint strategies and interim data for sponsors' programs.

A named regulatory position outside the device framework, from a non-US regulator. The European Medicines Agency's CHMP issued a qualification opinion (2022) that PROCOVA is acceptable for primary analysis of phase 2 and 3 trials with continuous outcomes and can increase power or reduce sample size; the company describes this as EMA qualification and reports positive feedback from FDA, quoting FDA's guidance on covariate adjustment rather than any FDA decision on its product.

The method and the generators are trial analysis tools used by sponsors, not devices, and no device claim is made. The qualification is narrow. It covers a statistical method for continuous endpoints rather than a specific generator, and the company's newer single arm and external comparator uses sit outside it.

Governance is expressed through a regulator reviewed method and material written for sponsors' review, with no subgroup evaluation of the generators. The European Medicines Agency's CHMP qualified PROCOVA for primary analysis of phase 2 and 3 trials with continuous outcomes, and the company publishes whitepapers on a risk-based approach to AI in clinical trials and on positioning digital twin analyses with regulators, which a sponsor's review committee can use.

The method is built so that, in a randomized trial, a poor prognostic model costs power rather than introducing bias. What is not published is any evaluation of the generators across patient groups: models trained on historical trial populations inherit their composition, and forecast accuracy by age, sex, race or disease stage is not reported on the site or in the research list.

Mechanisms make the output checkable, and nothing stands behind it. The twin-adjusted analysis is pre-specified in the statistical analysis plan before unblinding, the method is published (EMA qualification opinion; papers on PROCOVA, Bayesian PROCOVA and PROCOVA-MMRM), and in a randomized trial the method's type I error control does not depend on the generator being accurate, so a regulator or the sponsor's statisticians can audit it.

The company reports what the analyses delivered in re-analyses of completed trials (an 18 percent variance reduction in a phase 3 ALS trial, 17 percent fewer participants in an AbbVie phase 2 Alzheimer's study), which are efficiency figures rather than an error rate for the forecasts. The terms of service disclaim responsibility for the accuracy and reliability of the services, and no warranty, indemnity or remediation commitment attaches.

There is no clinical record integration; the product works on trial datasets in the sponsor's statistical environment. The company says generators can run as a web application or inside the sponsor's cloud, and that twins are generated from baseline data already collected.

It names no electronic data capture system, clinical data repository, CDISC standard or statistical platform it connects to, and does not describe how prognostic scores reach the analysis in a form the sponsor's statisticians or a regulator can reproduce.

Deployment options are stated, with the location of the hosted option left open. The technology and solutions pages say generators are available as a web-based application or as a secure on-premises solution within the sponsor's cloud infrastructure, and that custom generators can be built and deployed from the sponsor's own data entirely within the sponsor's environment.

For the hosted application no provider, region or retention schedule for sponsor trial data is published; the privacy notice covers transfers of personal data out of the EEA for the company's own processing.

No price is published, and how the work is sold is discoverable. Engagements are partnerships with Unlearn scientists across a single study, a program or a portfolio, with generators delivered through a web application or deployed in the sponsor's cloud. The company states value rather than price: about 20 percent sample size reduction locked into design, about $2 million saved on a phase 1/2 study, and about $14 million less for the next ALS trial in one re-analysis. No fee, unit of charge or contract basis is published.

Coverage is named with validation behind part of it. The technology page lists generators for 26 conditions across neurology, oncology, immunology, respiratory, metabolic and psychiatric disease, and the company sells into trial design, in-progress analysis and post-readout work for phase 1 to 3 studies.

Published or reported validation covers Alzheimer's disease and ALS in re-analyses of completed trials, and the EMA qualification is limited to phase 2 and 3 trials with continuous outcomes. For most listed conditions no validation is published, and binary and time to event endpoints sit outside the qualification.