Improving Few- and Zero-Shot Reaction Template Prediction Using Modern Hopfield Networks.

Philipp Seidl,Paulo Neves,Sepp Hochreiter,Günter Klambauer,Philipp Renz,Natalia Dyubankova,Marwin Segler,Jonas Verhoeven,Jörg K Wegner

doi:10.1021/acs.jcim.1c01065

Abstract

Finding synthesis routes for molecules of interest is essential in the discovery of new drugs and materials. To find such routes, computer-assisted synthesis planning (CASP) methods are employed, which rely on a single-step model of chemical reactivity. In this study, we introduce a template-based single-step retrosynthesis model based on Modern Hopfield Networks, which learn an encoding of both molecules and reaction templates in order to predict the relevance of templates for a given molecule. The template representation allows generalization across different reactions and significantly improves the performance of template relevance prediction, especially for templates with few or zero training examples. With inference speed up to orders of magnitude faster than baseline methods, we improve or match the state-of-the-art performance for top-k exact match accuracy for k ≥ 3 in the retrosynthesis benchmark USPTO-50k. Code to reproduce the results is available at github.com/ml-jku/mhn-react.

Highlights

The design of a new molecule starts with an initial idea of a chemical structure with hypothesized desired properties.[1]
Pubs.acs.org/jcim this allows a more fine-grained analysis of the template ranking obtained by the models, because it ignores errors stemming from multiple potential application locations
A fully connected network with a softmax output in which each output unit corresponds to a reaction template, conceptually similar to the model introduced in ref 11

Summary

Introduction

The design of a new molecule starts with an initial idea of a chemical structure with hypothesized desired properties.[1] Desired properties might be the inhibition of a disease or a virus in drug discovery or thermal stability in material science.[2,3] From the design idea of the molecule, a virtual molecule is constructed, the properties of which can be predicted by means of computational methods.[4,5] to test its properties and to make use of it, the molecule must be made physically available through chemical synthesis. To aid in finding synthesis routes, chemists have resorted to computer-assisted synthesis planning (CASP) methods.[6,8] Chemical synthesis planning is often viewed in the retrosynthesis setting in which a molecule of interest is recursively decomposed into less complex molecules until only readily available precursor molecules remain.[9] Such an approach relies on a single-step retrosynthesis model, which, given a product, tries to propose sets of reactants from which it can be synthesized. These methods, require extensive manual curation.[9−11] Recently, there have been increased efforts to model chemical reactivity from reaction databases using machine learning methods.[9,12−15]

Methods

Results

Conclusion