With new approaches to gradient descent based on a novel branch of computational calculus known as automatic differentiation, deep neural networks have proven in the last half decade to outperform several machine learning methods across multiple divisions of data science. Their utility has already been demonstrated in the development of fully-autonomous vehicle navigation, speech recognition, object detection and that are in production today. With the advent of deep convolutional networks with automatic differentiation, it is now commonplace to build what are known as ‘end-to-end models’. These classes of models map directly from inputs to output (i.e. SMILES string to toxicity, structure to functionality, and inversely, mapping from desired function to possible structure). We discuss the three most popular deep learning models, convolutional neural networks, long-short term memory networks, and generative adversarial networks (GANs) and their utility to several branches of computational chemistry. GAN’s are of great particular interest to computational chemistry; previous models were limited to classification or regression tasks whereas the latest of generative models can generate high resolution multi dimensional (2-dimensional & 3-dimensional) structures and sequences, providing new compound candidates for virtual high throughput screening. Furthermore, these deep networks are transparent in that the model parameters the networks learn can be visualized, interpreted, and most importantly transferred to new applications.

Advances in deep learning and their applied utility toward chemical informatics & drug discovery

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.