Relatively small and dispersed data sets and privacy issues represent two challenges in applying machine learning in drug development. Enter predictive analytics firm Owkin, which aims to overcome them through transfer and federated learning. It has secured $11 million in series A to do just that.
With the $11 million, plus the $2.1 million from seed financing, Owkin will improve its artificial intelligence platform called Owkin Socrates, cultivate partnerships with leading healthcare organizations and fuel internal growth, the company said in a statement.
Owkin Socrates builds machine-learning models and algorithms that analyze molecular and medical imaging libraries as well as patient profiles to uncover complex biomarker patterns behind diseases. Designed to assist drug researchers in preclinical and clinical studies and clinicians in healthcare settings, it can build models for automated diagnostics, prediction of treatment outcomes or clinical trial optimization.
Unlike many other AI-based drug development platforms, Owkin Socrates utilizes two machine learning tools—transfer learning and federated learning. Gilles Wainrib, PhD., co-founder and CSO of Owkin, explained to FierceBiotech via email how they work.
Transfer learning improves learning capacities on one data set or project through knowledge gathered from another related task. This allows machine learning to be applicable in smaller data sets—a few hundreds of patients from clinical trials, for instance. In drug development, deep learning applied to pathology image analysis can provide insight to stratify patients or to predict drug response efficiently.
“Owkin’s platform takes advantage of its unique access to our network of partners, such as large cancer centers, to build a library of various pre-trained models that power the Owkin Socrates machine learning engine,” said Wainrib.
Federated learning, a technique made popular by Google, takes out the requirement of storing and training data in one place (usually in the cloud). Because healthcare data are extremely private, sharing them with outsiders is one of the major bottlenecks for applying AI in healthcare, Wainrib explained.
“Our technology provides a way to circumvent this problem by training large-scale AI algorithms on multicentric databases, while the data itself never leaves the hospitals,” he said. “Only algorithms travel.”
In other words, each data gatherer can train a model on their own, and only needs to contribute updates—without specific data—to improve the shared model. The method facilitates cooperation between healthcare providers and biopharma companies without the concern of data privacy.
“Since there is no centralized repository for computing or data storage, not only can our service ensure that participants retain full control over their privacy and property, but it also provides a truly scalable solution for machine learning on healthcare data,” said Wainrib.
New York-based Owkin has already inked deals with several biopharma companies, including with Amgen and Actelion on applications of its technology to clinical trial optimization and real-world data analysis. It is also working with leading cancer centers in Europe, including Institut Curie and Centre Léon Bérard, to build predictive models of disease evolution and drug response. Wainrib said partnerships with U.S. cancer centers are also coming soon.
The series A round was led by Otium Venture, and digital startup-focused Cathay Innovation, Plug and Play and NJF Capital also participated.