Chemical Structure Embeddings for Accurate Similarity Search

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing chemical information databases struggle to accurately identify structurally similar chemicals and account for nuanced functional relationships due to the limitations of rule-based algorithms and fingerprint methods, which are often slow and inefficient.

Innovation Solution

Utilizing deep learning models, specifically graph neural networks, to generate embeddings of chemical structures that incorporate both structural and property information, forming a vector database for efficient chemical similarity searches.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If rule-based algorithms and fingerprint methods are used for chemical similarity search, then the search process is simple to implement, but the accuracy in identifying structurally similar chemicals and capturing functional relationships deteriorates

Engineering Contradiction:
Improveease of implementationVSAvoidaccuracy of chemical similarity identification
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent replaces traditional rule-based algorithms and fingerprint methods (mechanical/systematic approaches) with a neural network-based embedding system. The neural network learns to represent chemical structures as high-dimensional vectors that capture both structural and functional relationships, substituting the old mechanical comparison methods with a learned representation system that achieves superior accuracy in identifying chemically similar compounds.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Ease of manufacture

If rule-based algorithms and fingerprint methods are used for chemical similarity search, then the implementation is straightforward, but the search efficiency and speed deteriorate

Engineering Contradiction:
Improveease of implementationVSAvoidsearch efficiency
Core Design Contradiction:
Ease of manufactureVSProductivity

Solution Approach 1:

The patent applies preliminary action by pre-computing embeddings for all chemicals in the database and storing them in advance. When a similarity search is performed, the system only needs to compute the embedding for the query molecule and compare it against pre-computed embeddings, significantly improving search efficiency. This pre-processing step transforms the search process from computationally intensive on-the-fly calculations to rapid comparisons of pre-generated representations.

Inventive Principle:
Principle #10Preliminary action

3Device complexity

If traditional chemical search methods are used, then the system complexity is low, but the ability to capture complex molecular characteristics and relationships deteriorates

Engineering Contradiction:
Improvesystem complexityVSAvoidcapture of molecular characteristics
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent transitions from low-dimensional fingerprint representations to high-dimensional embedding vectors. This dimensional expansion allows the neural network to capture complex molecular characteristics and functional relationships that cannot be represented in traditional low-dimensional spaces. The high-dimensional embedding space provides sufficient capacity to encode nuanced structural and functional information while maintaining manageable system complexity through the use of standard neural network architectures.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4708304A1Chemical similarity search using fine-tuned neural network
Publication Date: 2026.03.11 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4708304A1 patent drawingFigure 1
  • EP4708304A1 patent drawingFigure 2
  • EP4708304A1 patent drawingFigure 3

AI summary

Examples are disclosed that relate to forming embeddings comprising vector representations of chemical structures, and performing chemical similarity searches. One example provides a method of forming a vector database using a pre-trained neural network, the method comprising inputting (402) labeled training data into the pre-trained neural network configured to form embeddings of chemical structures, the labeled training data comprising structural information and a value of a property for each chemical object in a first set of chemical objects. The method further comprises fine-tuning (408) the pre-trained neural network and forming (414) the vector database by inputting a reference dataset into the fine-tuned neural network to generate embeddings of chemical structures, the reference dataset comprising structural information for each chemical object in a second set of chemical objects. Each embedding stored in the vector database comprises a vector representation of a chemical structure and embedded information for the property for the chemical structure.