Retrosynthetic Planning with Experience-Guided Monte Carlo Tree Search

Abstract

Retrosynthetic planning problem is to analyze a complex molecule and give a synthetic route using simple building blocks. The huge number of chemical reactions leads to a combinatorial explosion of possibilities, and even the experienced chemists often have difficulty to select the most promising transformations. The current approaches rely on human-defined or machine-trained score functions which have limited chemical knowledge or use expensive estimation methods such as rollout to guide the search. In this paper, we propose EG-MCTS, a novel MCTS-based retrosynthetic planning approach, to deal with retrosynthetic planning problem. Instead of exploiting rollout, we build an Experience Guidance Network to learn knowledge from synthetic experiences during the search. Experiments on benchmark USPTO datasets show that, our EG-MCTS gains significant improvement over state-of-the-art approaches both in efficiency and effectiveness. Routes designed by EG-MCTS for real drugs or compounds exhibit the effectiveness of our approach on assisting chemists performing retrosynthetic analysis. Our EG-MCTS system solves almost a quarter more and twice times faster than the traditional computer-aided MCTS search method. In a comparative experiment with the literature, our computer-generated routes were generally viewed to be equivalent to reported literature routes by chemists.

RetroGraph: Retrosynthetic Planning with Graph Search

“…We compare our system with representative baselines on this task and the details are available in Appendix. Briefly, Greedy DFS [17] is a classic planning method that always prioritizes the node with max probability; DFPN-E [22] is a Depth-First Proof-Number search method; MCTS-rollout [17], EG-MCTS [17], and EG-MCTS-0 [17] are MCTS methods; Retro* [5], Retro*-0 [5], Retro*+ and Retro*+-0 [19] are A* search methods.…”

“…Then Kim et al [19] future improve this method with self-training and forward models. Meanwhile, the Monte Carlo tree search (MCTS) based methods have also been applied to this task [17,27,29,35]. Despite their achievements, all of these works are based on tree search and inevitably suffer from intra-target duplication.…”

“…• Greedy DFS [17]: This is a classic planning method that always prioritizes the node with max probability. We use tree-based Greedy DFS with 10 as max search depth following previous works.…”

“…Such a process is usually modeled as a tree-search problem, where the starting point is the target molecule, and the path is the chemical reaction. Research focus [5,17,19] is how to find a feasible pathway with as few trials as possible. Various data-driven search policies have been leveraged / proposed for this task, such as A* search [5], proof number search [22], and Monte Carlo tree search [17].…”

“…Research focus [5,17,19] is how to find a feasible pathway with as few trials as possible. Various data-driven search policies have been leveraged / proposed for this task, such as A* search [5], proof number search [22], and Monte Carlo tree search [17]. They treat the planning progress as a tree expansion problem.…”