NEW YORK—To avoid the tedium of working out how a specific compound is made in a drug or organic molecule manually, Elsevier is collaborating with Pending.AI (PAI), a startup focused on developing artificial intelligence (AI) solutions for drug discovery, to develop a predictive retrosynthesis tool.
Based on deep learning to support innovation in synthetic and medicinal chemistry, the tool was developed via Elsevier’s R&D Collaboration Network. It is being integrated into Elsevier’s flagship chemistry solution, Reaxys, combining Reaxys’ content with cutting-edge AI and machine learning technologies developed by PAI, potentially cutting the time and cost of bringing a new drug to the market drastically.
Synthesizing new small molecules typically involves multiple cycles of manual planning and execution. Chemists need to assess the synthesizability of proposed target molecules, which could mean evaluating hundreds or possibly thousands of molecules. Elsevier and PAI’s new AI-driven chemistry retrosynthesis tool AI tool, which is based on a deep learning algorithm, can automatically derive more than 400,000 reaction rules from the Reaxys source data of more than 15 million single-step organic reactions, so chemists do not need to rely on hand-encoded rules.
According to Dr. Mark Waller, director at PAI, “AI is becoming essential as scientific data grows in abundance. Our mission is to develop pragmatic solutions using AI and machine learning to empower scientists to advance drug discovery and development of other chemical compounds. We are proud to be working with Elsevier to meet this goal. The Reaxys-PAI Predictive Retrosynthesis tool will complement the knowledge of scientists and teams and help them to rapidly make more informed decisions.”
Originally developed through Elsevier’s R&D Collaboration Network, the Reaxys-PAI Predictive Retrosynthesis solution uses a model that incorporates deep neural networks trained on Reaxys data. The results are found using a Monte Carlo tree search approach to quickly discover promising candidate routes.
Tested thoroughly by the world’s leading pharmaceutical and chemical companies, the tool has shown that it can provide scientifically robust, diverse and innovative synthetic route suggestions. It is described by Waller as “easy and intuitive to use.” He added that it “supports the needs of the business and researchers by being a very good assistant and idea generator.”
The predictive retrosynthesis solution has been trained on both positive and negative reaction data and solves synthesis design questions for novel molecules with direct links to experimental reactions available in Reaxys. The predictive model training and creation is fast, allowing it to “self-learn” from ever-growing chemistry knowledge. Reaxys-PAI Predictive Retrosynthesis can be further augmented by training on proprietary chemistry reaction data, including a customer’s own reaction dataset and building block library. This solution is now available as an add-on module for Reaxys customers.
According to Dr. Ivan Krstic, Elsevier’s director of product management, life science solutions, “AI is set to revolutionize the domain of chemical design and synthesis of small molecules. Over the past decade, the exponential growth in chemistry data [and] the ability to curate and harmonize data, coupled with advancements in computational and digital technologies such as deep learning, has provided ideal grounds for addressing the problem of computer-aided synthesis design.”
He added, “We are very happy that this innovative work is enabled by a partnership between Elsevier and PAI to provide a best-in-class predictive retrosynthesis solution which combines high-quality Reaxys reaction data with industry-leading predictive algorithms developed by PAI. We have strong evidence that the addition of AI-based retrosynthesis to Reaxys can help drive innovation, save researchers considerable time, and radically change how we approach chemical synthesis, but I also want to share with my fellow chemists our strong belief that AI won’t replace chemists. Instead, it will support chemists and their decision-making by paving the way in a more and more complex landscape of data.”