Planning chemical syntheses with deep neural networks and symbolic AI

Abstract

To plan the syntheses of small organic molecules, chemists use retrosynthesis, a problem-solving technique in which target molecules are recursively transformed into increasingly simpler precursors. Computer-aided retrosynthesis would be a valuable tool but at present it is slow and provides results of unsatisfactory quality. Here we use Monte Carlo tree search and symbolic artificial intelligence (AI) to discover retrosynthetic routes. We combined Monte Carlo tree search with an expansion policy network that guides the search, and a filter network to pre-select the most promising retrosynthetic steps. These deep neural networks were trained on essentially all reactions ever published in organic chemistry. Our system solves for almost twice as many molecules, thirty times faster than the traditional computer-aided search method, which is based on extracted rules and hand-designed heuristics. In a double-blind AB test, chemists on average considered our computer-generated routes to be equivalent to reported literature routes.

Toward Design of Novel Materials for Organic Electronics

“…Recent years have shown the emergence of a novel class of methods that are not based on physical models but on learning systematic relations in large datasets and on methods of statistical inference. [189][190][191][192][193] Examples include the use of regression and classification models such as neural networks for the prediction of molecular or materials properties [194,195] and for synthesis planning [146,149,150] as well as the use of generative models such as variational autoencoders and generative adversarial networks for inverse molecular design. Mater.…”

“…Recent developments include fingerprint based representations of the molecular graph [204,205] or the 3D structure. [144,149,151,152] To date, machine learning based systems already outperform traditional methods in chemistry including examination by human experts and database search in several fields of work. [144,149,151,152] To date, machine learning based systems already outperform traditional methods in chemistry including examination by human experts and database search in several fields of work.…”

“…Template-based approaches consist of two steps including the automatic generation of rules from the available reaction data set and the training of a neural network, which is then used for the reaction decision making via ranking of the most likely occurring reaction template. [149] Monte Carlo tree search with an expansion policy network guides in a second step the search for the most suitable chemical route. [146][147][148] A recently published method for reaction prediction by neural networks in combination with symbolic artificial intelligence (AI) highlights the impact of the current developments and the scope of machine learning methods: A model to retrosynthetic analysis via training of a neural network with a dataset of more than 12.5 million reactions was developed to select the most promising retrosynthesis step.…”

Compounding Meta‐Atoms into Metamolecules with Hybrid Artificial Intelligence Techniques

“…However, photonic media with identical meta-atoms may offer only inadequate www.advmat.de www.advancedsciencenews.com optimizing and designing nanophotonic devices. [24][25][26] Once the dataset is constructed, the machine learning model trained based on the dataset can be employed for various designs in an expeditious way, multiple times. In addition to adjoint methods, genetic algorithms and related variations also play important roles in the design of photonic structures.…”

Machine-learning approach to the design of OSDAs for zeolite beta

“…Examples include algorithms for quantum chemistry (24), retrosynthetic chemistry (25), and de novo design (26, 27). Once properly trained, an ML algorithm is very fast to produce an output from new input.…”