Pharma R&D Today
Ideas and Insight supporting all stages of Drug Discovery & Development
At the leading edge of an innovation hotspot: predictive reaction modeling
Posted on September 8th, 2021 by Louise Springthorpe in Chemistry
Indisputably, a surge in predictive machine learning models to support synthetic chemistry has galvanized a movement to make these tools effective R&D solutions. Reaxys has been part of this innovation journey from the beginning. Today, Reaxys includes an award-winning Predictive Retrosynthesis solution that merges deep neural networks trained on Reaxys data with a Monte Carlo tree search to quickly discover promising candidate routes. Users explore these routes and new synthetic spaces via an intuitive interface that links to Reaxys content on commercial availability and accelerates design-make-test-analyze (DMTA) cycles.
Decisions about what that solution is today and how it will evolve build on a dialog with users. In a series of Expert Forums, Reaxys users are invited to share ideas, soundboard new features and provide constructive feedback. The Predictive Retrosynthesis solution was the topic of discussion in the most recent forum. A broad audience of chemists joined the two-hour event, featuring talks by scientists on three areas of innovation: improved model training, novel reaction representations and predictive tool adoption.
The importance of negative reaction data
Dr. Martin Villalba from Bayer presented results from a research project examining the effect of negative reaction data on a model Bayer uses to predict the viability of a novel reaction. Through a systematic comparison of different data constructs that included real or synthetic negative reaction data as a training set, Villalba was able to demonstrate the importance of a diverse training set. He encouraged viewers to use the negative reaction data in electronic lab notebooks to improve the accuracy of models.
Reducing complexity with Condensed Graph of Reactions
Representing chemical reactions for predictive models poses an interesting challenge, explained Dr. Alexandre Varnek from the University of Strasbourg. They involve different types of molecules, often proceed in multiple steps and success is highly dependent on conditions. Varnek and collaborators have reduced that complexity into a single molecule graph called Condensed Graph of Reaction, or CGR. In his presentation, Varnek showed how CGR can be used to shorten SMILES, correct atom-to-atom mapping, classify reactions, and assess optimal reaction conditions or protective group reactivity based on CGR similarity comparisons.
Retrosynthesis tools in the hands of synthetic chemists
No retrosynthesis tool is of value if not adopted by the intended user. So, Dr. Jessica Herrick from Corteva asked the question, “How do retrosynthesis tools match up in a head-to-head comparison?” Using known transformations from the literature, her team evaluated various solutions. Herrick listed key features of a valuable solution, including that it delivers a multitude of candidate routes, identifies commercially available starting materials and easily integrates into the daily routine of users. She also recommended strategies to promote adoption, like finding early adopters, offering support through subject matter experts and sharing success stories.
Based on user input, the Reaxys Predictive Retrosynthesis tool has a clear upgrade roadmap for this year. Customers will soon be able to define bonds to be made or broken, as well as include or exclude intermediates. An enhanced model will handle protective group strategies and integrate the roughly 60 million commercially available substances in Reaxys. The user experience will improve further with expandable synthesis plans.
With sights on the future, the Reaxys team surveyed the forum audience for thoughts on further improvements. Predicting reaction conditions and synthetic accessibility emerged as top priorities. The audience was also asked about challenges to data customization to better inform Reaxys efforts to offer integration of in-house data. Most respondents pointed to issues with data readiness, security and costs as important hurdles.
Such are the outcomes of Reaxys Expert Forums – ideas and important requirements for an evolved, powerful tool designed with and for users. Additional forums will take place to capture user opinions. If you’d like to be part of one of these events, please join the Reaxys User Group on LinkedIn for updates. The team looks forward to collaborating with you.
R&D Solutions for Pharma & Life SciencesWe're happy to discuss your needs and show you how Elsevier's Solution can help.
Manager, Corporate Markets Marketing, Elsevier
- Can predictive retrosynthesis become a valuable part of a chemist’s toolkit?
- AI-driven innovation in life sciences: Unlocking the page – 20+ years of digital innovation at Elsevier
- AI-driven innovation in life sciences: Data as data should be
- AI-driven innovation in Life Sciences: Data Science & AI amplifies your innovative drug R&D capabilities
- Accelerating synthetic chemistry with a predictive retrosynthesis solution