Public Release: 

Accelerating drug discovery by crowdsourcing confidential data

American Association for the Advancement of Science

Leveraging modern cryptographic and machine learning tools, researchers seeking to accelerate drug discovery have developed a way for multiple pharmaceutical companies and laboratories to collaborate without revealing confidential data. According to the related report, the shared experimental datasets improve the ability of predictive models designed to identify drug-target interactions (DTI), to predict new therapeutic candidates. Using this approach, drug candidates could be identified at a rate and scale far greater than current state-of-the-art methods allow, the authors say. Developing a new drug takes years of research and a large amount of resources. To address this, pharmaceutical companies sometimes collaborate, sharing knowledge and resources. While such efforts have been shown to be fruitful in some cases, they are often limited in scope due to concerns about intellectual property and competing financial interests. What's more, the sharing of data between multiple entities is restricted by a need to maintain confidentiality. Secure multiparty computation (MPC) protocols offer a modern cryptographic solution for facilitating collaboration while ensuring data privacy. However, existing MPC frameworks lack the ability to perform complex algorithms over the large datasets required to predict new therapeutic drugs. To address this need, Brian Hie and colleagues developed a computational protocol for collaborative DTI prediction based on secure MPC, which blinds sensitive data and divides computational tasks across collaborating groups. Together, the secured combined dataset - comprising millions of chemical compounds and drug target interactions - was used to train a neural network model for DTI prediction. This represents one of the first such demonstrations of secure neural networks training on large-scale real-world data, the authors say. According to their results, the protocol allowed the model to produce accurate results in under four days of training. While the method demonstrates a promising solution for pharmaceutical collaboration, the authors suggest it could also be used in other areas hindered by a lack of collaboration due to privacy concerns.


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.