Neural Network Retrosynthesis for Faster Reactant Set Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for retrosynthesis are inefficient and computationally costly, with large search spaces and inaccuracies in predicting optimal reactant sets for synthesizing target compounds.

Innovation Solution

A neural network system utilizing a three-subnetwork architecture processes input data to generate predicted probabilities for reactant sets, employing a dual loss function to ensure consistency between forward and backward reaction directions, thereby reducing computational costs and improving accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If existing methods for retrosynthesis are used, then the process can determine reactant sets for target compounds, but the computational cost and time consumption are excessively high due to large search spaces

Engineering Contradiction:
Improveretrosynthesis efficiencyVSAvoidcomputational time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent segments the retrosynthesis problem into two distinct neural network models: a forward reaction prediction model and a backward retrosynthesis model. This segmentation allows each model to be optimized for its specific direction, improving overall efficiency by avoiding exhaustive search through the entire reaction space. The forward model learns from known reactions while the backward model leverages this learned knowledge to efficiently predict reactants.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs preliminary action by using the forward reaction prediction model to pre-process and learn from extensive reaction data before the backward model performs retrosynthesis. This preliminary training phase allows the system to capture reaction patterns and constraints in advance, so that when retrosynthesis is performed, the search space is already constrained by learned knowledge, significantly reducing computational time.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If existing graph-based models are used for retrosynthesis, then predictions can be generated, but accuracy is insufficient due to inaccuracies in predicting optimal reactant sets

Engineering Contradiction:
Improveprediction accuracyVSAvoidpredictive reliability
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent implements feedback by establishing a consistent training framework where the forward reaction prediction model provides feedback to the backward retrosynthesis model. The forward model's predictions and loss calculations inform the backward model's training, creating a feedback loop that continuously improves prediction accuracy. This feedback mechanism ensures that the backward model learns from accurate forward predictions rather than training in isolation.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent applies parameter changes by using different loss functions and training parameters for the forward and backward models. The forward model uses a loss function optimized for reaction prediction while the backward model uses a loss function tailored for retrosynthesis. This parameter differentiation allows each model to achieve optimal accuracy for its specific task, improving overall predictive reliability.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4118658B1Retrosynthesis using neural networks
Publication Date: 2025.08.06 GOOGLE LLC
  • EP4118658B1 patent drawingFigure 1
  • EP4118658B1 patent drawingFigure 2
  • EP4118658B1 patent drawingFigure 3A

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for performing retrosynthesis using a neural network. One of the methods includes generating a prediction of a set of a plurality of predicted reactants that are combinable to generate a target compound, the generating comprising processing, for each of a plurality of candidate sets of reactants, a network input characterizing the candidate set using a neural network, determining, for each candidate set of the plurality of candidate sets, a score using the generated probabilities; and selecting a particular candidate set of one or more reactants using the determined scores.