AI-Driven Synthetic Route Design Incorporated with Retrosynthesis Knowledge

Abstract

Computer-aided synthesis planning (CASP) aims to assist chemists in performing retrosynthetic analysis for which they utilize their experiments, intuition, and knowledge. Recent breakthroughs in machine learning (ML) techniques, including deep neural networks, have significantly improved data-driven synthetic route designs without human intervention. However, learning chemical knowledge by ML for practical synthesis planning has not yet been adequately achieved and remains a challenging problem. In this study, we developed a data-driven CASP application integrated with various portions of retrosynthesis knowledge called “ReTReK” that introduces the knowledge as adjustable parameters into the evaluation of promising search directions. The experimental results showed that ReTReK successfully searched synthetic routes based on the specified retrosynthesis knowledge, indicating that the synthetic routes searched with the knowledge were preferred to those without the knowledge. The concept of integrating retrosynthesis knowledge as adjustable parameters into a data-driven CASP application is expected to enhance the performance of both existing data-driven CASP applications and those under development.

Recent advances in deep learning for retrosynthesis

“…Ishida et al 87 proposed a promising research direction to integrate various portions of retrosynthesis knowledge into multi‐step retrosynthesis planning. On the basis of MCTS architecture proposed by Segler et al, 20 they made two major improvements: (1) replaced the highway network with a more advanced GCN‐based policy network to make better single‐step predictions; (2) defined a series of retrosynthesis scores to evaluate promising search directions, including a convergent disconnection score (CDScore), an available substances score (ASScore), a ring disconnection score (RDScore), and a selective transformation score (STScore).…”

Predictive chemistry: machine learning for reaction deployment, reaction development, and reaction discovery

“…88 Another common strategy is to better estimate the value function for any node without the expensive rollout. Injection of chemical heuristics in selection can be as simple as using a combination of reaction likelihood and complexity assessment score like SCScore , 89 as was done in Schwaller et al 22 Similarly, ReTReK 90 defines four heuristic scores to guide MCTS towards convergent synthesis, ring-forming reactions, and reactants with fewer reaction centers, harkening back to the early days of formalizing retrosynthesis where “x-oriented” (starting material-, stereochemistry-, topology-, etc. ) strategies were proposed.…”

Recent advances in artificial intelligence for retrosynthesis

“…More specifically, by modifying the attention masks, they forced different attention heads in self-attention to pay attention to atom pairs at a different distance, and tried to make the cross-attention mask as close as possible to the atom-mapping mask. The first application of artificial intelligence to multi-step retrosynthesis Expert DFPN-E [74] Balance the number of AND and OR nodes Expert Monte Carlo Tree Search 3N-MCTS [24] Evaluate the pathway quality via double-blind AB tests Neuralsym AutoSyn [42] Combine MCTS with the template-free model Transformer Transformer ReTReK [75] Introduce retrosynthesis knowledge to guide the search direction GCN A* Search Retro* [76] Consider both the informed cost and estimated future value of nodes Neuralsym RetroGraph [77] Represent the search process as a directed graph rather than a search tree Neuralsym Others Self-Improved [78] A end-to-end framework for improve single-step solver Neuralsym Hyper-graph [79] Apply forward model to filter suggestions Transformer SimulatedExp [80] Apply reinforcement learning to find a optimal search policy Neuralsym…”

“…Ishida et al [75] proposed a promising research direction to integrate various portions of retrosynthesis knowledge into multi-step retrosynthesis planning. On the basis of MCTS architecture proposed by Segler et al [24], they made two major improvements: 1) replaced the highway network with a more advanced GCN-based policy network to make better single-step predictions; 2) defined a series of retrosynthesis scores to evaluate promising search directions, including a convergent disconnection score (CDScore), an available substances score (ASScore), a ring disconnection score (RDScore), and a selective transformation score (STScore).…”