Molecule Synthesizability Scoring with Two-Stage Retrosynthesis ML

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for predicting the synthesizability of molecules are computationally expensive and require large amounts of labeled training data, which is scarce, making them inefficient and inaccurate.

Innovation Solution

A two-stage machine learning approach involving a retrosynthesis featurization model to generate a feature array and a synthesizability prediction model to produce a synthesizability score, leveraging trained models with smaller architectures and efficient training data generation from programmatically assembled retrosynthesis trees.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing methods for predicting synthesizability are used, then prediction accuracy can be maintained, but computational cost increases and training data requirements increase

Engineering Contradiction:
Improvesynthesizability prediction accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The prediction system is divided into two independent models: a retrosynthesis featurization model that processes molecular structures and generates feature arrays, and a synthesizability prediction model that consumes these features to predict synthesizability scores. This segmentation allows each model to be optimized independently, reducing overall computational complexity while maintaining prediction accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system extracts essential chemical features from molecular structures through the featurization model, transforming complex molecular data into condensed feature arrays. This extraction process separates the essential information needed for prediction from the raw molecular complexity, enabling more efficient processing by the prediction model.

Inventive Principle:
Principle #2Taking out (Extraction)

2Reliability

If existing methods for predicting synthesizability are used, then comprehensive analysis can be performed, but training data requirements increase

Engineering Contradiction:
Improveprediction reliabilityVSAvoidtraining data quantity
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The system employs self-supervised learning mechanisms where the retrosynthesis featurization model learns to generate meaningful features without requiring extensive labeled synthesizability data. The model serves itself by creating structured feature representations that capture essential chemical relationships, reducing dependence on large annotated datasets while maintaining prediction reliability.

Inventive Principle:
Principle #25Self-service

3Measurement precision

If complex models are used for synthesizability prediction, then prediction accuracy improves, but training time and computational resources increase

Engineering Contradiction:
Improvesynthesizability prediction accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

By dividing the prediction task into two specialized models, each model can be trained more efficiently on its specific subtask. The featurization model learns molecular representation patterns, while the prediction model learns synthesizability patterns from features, reducing the overall training time compared to a single complex model learning all patterns simultaneously.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The retrosynthesis featurization model performs preliminary processing of molecular structures, transforming raw molecular data into meaningful feature arrays before the prediction model processes them. This preliminary action prepares the data in an optimized format, reducing the computational burden and training time for the final prediction model.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4687146A1Predicting molecule synthesizability using machine learning
Publication Date: 2026.02.04 ISOMORPHIC LABS LTD
  • EP4687146A1 patent drawingFigure 1
  • EP4687146A1 patent drawingFigure 2
  • EP4687146A1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for predicting synthesis complexity. In one aspect, a method comprises: receiving data characterizing a molecule; processing a model input that comprises the data characterizing the molecule, using a retrosynthesis featurization machine learning model and in accordance with values of a set of retrosynthesis featurization machine learning model parameters, to generate a feature array comprising a plurality of features of a retrosynthesis tree for the molecule; and processing a model input that comprises the feature array comprising the plurality of features of the retrosynthesis tree for the molecule, using a synthesizability prediction machine learning model and in accordance with values of a set of synthesizability prediction machine learning model parameters, to generate a synthesizability score that characterizes a synthesis complexity of the molecule.