Developing new compounds with desirable drug-like properties such as increased target activity while maintaining good ADME is a challenging feat. During the cycle of novel chemical design, chemists are tasked with creation of new analogs from a target molecule. Machine learning (ML) is often integrated into this cycle, providing a way to predict activity and score new compounds according to learned data representation. These traditional ML models, however, lack the ability to generate new compounds from learned data, leaving the task to chemistry experts.

Recently, ML  models for generating de novo libraries of compounds have been introduced into the literature, including Recurrent Neural Networks (RNNs). RNNs are a neural network architecture which learn the structure of sequential data, keeping track of the most salient information at every step and every previously seen step, affecting the interpretation of the current step. 

RNNs have been used as generative models, successfully learning SMILES representation of chemicals and are capable of generating new, synthetically reasonable compounds with desirable drug-like properties. We have implemented several different RNNs which can benefit from the thousands of datasets/models we have curated in MegaTox, MegaTrans, MegaPredict, and other datasets. 

Our RNN implementation called MegaSyn, utilizes a state-of-the-art multi-objective optimization algorithm to optimize multiple parameters simultaneously during the RNN training. Our in-house automated analog designer scores these generated molecules on their synthetic feasibility, allowing us to select optimized generated compounds that are synthetically feasible.