BioCentury
ARTICLE | Translation in Brief

Private sharing

MIT AI models drug-target interactions from confidential pooled data sets

November 1, 2018 7:18 PM UTC

A new cryptographic computational protocol from the Massachusetts Institute of Technology will allow drug companies to pool data for predictive modeling of drug-target interactions without revealing the underlying drugs, targets or observed interactions. The method has attracted interest from companies such as Pfizer Inc. (NYSE:PFE) and Biogen Inc. (NASDAQ:BIIB) that wish to boost pharmacological collaboration without risking the privacy of their data.

The approach, described in a paper in Science on Oct. 19, relies on machine learning to train convolutional neural networks to predict drug-target interactions while hiding the raw data through a secure multiparty computation framework. Using publicly available drug-target interaction databases to train the network, the MIT team showed that the protocol could be used on large-scale data sets including millions of drug-target interactions, and generated more accurate predictions than other techniques including matrix factorization, network diffusion and heterogeneous data integration (see Distillery)...