Unlearn.ai, a company that designs software tools for clinical research, today announced that it secured $12 million in equity financing. Unlearn’s “digital twin” approach to trials, in which digital models are used in place of real test subjects, could reduce the number of people required to run a trial without sacrificing standards of evidence.
Unlearn’s technology could also help to solve the systemic reproducibility problem in clinical research, which a pair of surveys by Bayer and Amgen recently brought into sharp relief. Bayer reported successfully replicating just 25% of published preclinical studies it analyzed, while Amgen confirmed findings in just 6 of 53 landmark cancer studies (11%).
Unlearn was cofounded in 2017 by physicists Charles Fisher, Aaron Smith, and Jon Walsh, who initially built the company’s platform atop an AI architecture called restricted Boltzmann machines (RBMs). RBMs are inspired by statistical mechanics and can model a person’s characteristics while remaining robust in the face of missing data, but they poorly model data from different groups, producing blended rather than distinct distributions of, for example, patients.
To address such shortcomings, the team architected an open source package called Paysage, which implemented unsupervised learning algorithms (meaning they use data that hasn’t been classified or labeled) including a hybrid of an RBM and generative adversarial networks: a Boltzmann Encoded Adversarial Machine (BEAM). GANs are two-part AI models consisting of a generator that creates samples and a discriminator that attempts to differentiate between the generated samples and real-world samples, and this unique arrangement enables them to achieve impressive feats of media synthesis.
Unlearn’s DiGenesis platform is built upon this hybrid model. It processes historical clinical trial data sets from thousands of patients to build the disease-specific machine learning models, which are used to create digital twins and their corresponding virtual medical records. Digital twin records are longitudinal and include demographic information, common lab tests, and endpoints and/or biomarkers that look identical to actual patient records in a clinical trial.
In a case study published last year, Unlearn applied its system to predict Alzheimer’s disease progression, in essence projecting the symptoms that individual patients will experience at any point in the future. It simultaneously computed predictions and confidence intervals for multiple characteristics of a patient at once using a BEAM, which was trained and tested on the Coalition Against Major Diseases (CAMD) Online Data Repository for Alzheimer’s Disease. The data set consisted of 5,000 patients measured over a period of 18 months covering 50 variables, including the individual components of ADAS-Cog (a widely used cognitive subscale) and Mini-Mental State Examination, a questionnaire used to measure cognitive impairment in clinical and research settings.
In the course of the study, Unlearn leveraged the trained model to generate “virtual patients” and their associated cognitive exam scores, laboratory tests, and clinical data. Simulations were run for individual patients to project their disease progression in areas such as word recall, orientation, and naming, which were in turn used to compute the overall ADAS-Cog score.
The result: The unsupervised model was able to make accurate ADAS-Cog predictions out to at least 18 months.
Unlearn says that undisclosed pharmaceutical companies have expressed interest in DiGenesis — which isn’t surprising. It takes on average over $2 billion and 10 years to develop and sell a new medicine, and much of the costs arise in the trial phases, during which around 90% of candidate treatments are proven ineffective or unsafe.
“Patients who volunteer for clinical trials take some risk; they could receive a treatment that doesn’t work, or experience serious side-effects. Therefore, it’s really important that we run these trials as efficiently as possible while providing reliable evidence to further medical science,” Fisher told VentureBeat via email. “We believe that our [platform] will have a profound impact on this problem, and are excited to partner with 8VC to realize a shared vision to use technology to improve the lives of patients.”
Unlearn’s aspirational goal is to develop a digital twin for every patient, which it envisions will help physicians evaluate the risks each patient faces and develop the best course of treatment for that patient. In the near term, Unlearn intends to focus on neurological diseases, starting with Alzheimer’s disease and multiple sclerosis.
This financing round — a series A — was led by 8VC with participation from all of Unlearn’s existing investors including DCVC, DCVC Bio, and Mubadala Capital Ventures. (It brings Unlearn’s total raised to date to over $17 million.) Through its investment, 8VC principal Francisco Gimenez joined the company’s board of directors.
According to LinkedIn data, San Francisco-based Unlearn has 16 employees.