Privacy-Preserving Property Prediction for New Drugs with MPNN

Abstract

Message passing neural network (MPNN) is one of the excellent deep learning models for drug discovery and development, drug laboratory usually outsource the MPNN model to cloud servers to save the research and development cost. However, drug-related data privacy has become a noticeable hindrance for outsourcing cooperation in drug discovery. In this paper, we propose a lightweight privacy-preserving message passing neural network framework (SecMPNN) for property prediction in new drugs. To implement SecMPNN, we design multiple protocols to perform the three stages of MPNN, namely message function, update function, and readout function. The above new-designed secure protocols enable SecMPNN to adapt to the different numbers of participating servers and different lengths of encryption requirements. Moreover, the accuracy, efficiency, and security of SecMPNN are demonstrated through comprehensive theoretical analysis and a large number of experiments. The experimental results show the communication efficiency in multiplication and comparison increases 27.78% and 58.75%, the computation error decreases to 4.64%.