Using Invertible Plugins in Autoencoders for Fast and Customizable Post-training Optimization

Abstract

One of the main motivations for modern drug research is the production of new compounds that act as drugs, however developing a new drug is an excessively time and resource intensive process. Deep generative neural networks might provide a solution. With their help, we may be able to search in a continuous latent space to find drug molecules that are not yet known but have suitable chemical and structural properties (e.g. solubility, interaction with a given target protein). In this paper, we propose a model which can generate novel drug candidates, that are suitable for a pre-specified objective function of arbitrary properties. The model consists of a generative network and a predictor. The former is an autoencoder which utilizes attention to handle the textual representation of molecules, while the latter uses matrix factorization to predict drug-target interactions (DTI). With a genetic algorithm we can generate novel compounds from the continuous latent space, but if there are changes in the objective function, we may need to train the whole model again. This problem is typical of conditional generative models, to address it, we separated the predictor from the pretrained autoencoder thus forming the plugin. In addition to getting a flexible architecture without any deterioration in the so far achieved results, our model can also be used in a distributed setup by concatenating the plugins. In this way, the objective function can be broken down to smaller subtasks, which can be solved by different plugins without sharing any data.