Retrosynthetic Planning with Dual Value Networks

Retrosynthesis Zero: Self-Improving Global Synthesis Planning Using Reinforcement Learning

“…Furthermore, most of these methods utilize significantly larger rule data sets than ours. For example, Retro*+, EG-MCTS, and PDVN all employed the 318k rule data set. In contrast, ReSynZ, as presented in this work, operated with a considerably smaller data set of up to 13k rules and used limited computing resources, yet still achieves commendable performance.…”

“…Their method is dedicated to goal-driven synthesis planning but achieves decent performance in general planning as well. The PDVN algorithm trains one policy network and two value networks using data collected from route planning with the MCTS algorithm. PDVN also improves the three neural networks in an iterative fashion using reinforcement learning, and the policy network is continuously updated throughout the learning process, training with data collected from successful synthesis routes of target molecules.…”

“…To ensure data quality, a forward reaction prediction model was used to filter out implausible hypothetical reactions. Later, EG-MCTS and PDVN adopted an RL framework similar to AlphaGo. The methods improved the retrosynthetic policy and learned a value function to guide the route search simultaneously in an iterative fashion.…”

Models Matter: the impact of single-step retrosynthesis on synthesis planning

“…22,23 Finally, self-play approaches, motivated by their success in Go, 24 learn to guide the tree search by leveraging information gathered from prior runs of synthesis planning. [25][26][27] Single-step retrosynthesis prediction and multi-step synthesis planning are inherently intertwined where the single-step method denes the maximum searchable reaction network, and the search algorithm tries to efficiently traverse this network by repeatedly applying the chemical information that is stored in the single-step model. However, this connection is not reected in contemporary research, with only few novel single-step models testing their approaches within a multi-step synthesis planning framework.…”

“…These methods are typically only tested on reactant prediction and not within multi-step search algorithms, therefore their usability for synthesis planning is not assessed. Similarly, multi-step search algorithms benchmark the route-nding capabilities of their method using a single single-step model, oen based on the template-based NeuralSym model, 2,[18][19][20]27 and evaluate the success rate of nding potential synthesis routes for molecules of interest. However, multi-step approaches do not consider the impact of alternative single-step models, a vital aspect of the search, as the route planning algorithm uses the reaction information stored in the single-step model to nd synthesis routes and create alternate reaction pathways within the reaction network.…”