Generative Neural Network Using Global-Shape Representations for 3D Molecular Modeling

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing generative learning models struggle to accurately approximate the distribution of equilibrium 3D structures due to inefficient and non-robust use of string and 2D graph representations, which lack the necessary robustness and efficiency in capturing 3D domain information.

Innovation Solution

The use of global-shape input representations, such as persistence images, in conjunction with local point-level and node-level characteristics, within a generative neural network to create a more effective generative model of 3D domains, enhancing the model's ability to represent and reproduce the functionality of 3D shapes and geometries.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If string and 2D graph representations are used to represent 3D molecular structures, then the model can process input data, but the model fails to accurately capture 3D domain information and approximate the distribution of equilibrium 3D structures

Engineering Contradiction:

Improve3D domain informationVSAvoidaccuracy of approximating 3D structure distribution

Core Design Contradiction:

Loss of informationVSReliability

Solution Approach 1:

The patent transitions from 2D graph representations to 3D point cloud representations, adding spatial dimensionality to capture true 3D molecular geometry. This enables the model to represent equilibrium 3D structures accurately by processing actual 3D coordinates rather than flattened 2D projections, directly resolving the information loss in conventional representations.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The patent changes the representation parameters from string/2D graph formats to 3D point cloud coordinates with associated features. This parameter transformation allows the neural network to process genuine 3D spatial information, improving both the completeness of 3D domain information capture and the reliability of 3D structure distribution approximation.

Inventive Principle:

Principle #35Parameter changes

2Reliability

If conventional point-level representations are used, then the model can be trained, but the model lacks robustness and efficiency in capturing 3D domain information

Engineering Contradiction:

Improverobustness in capturing 3D informationVSAvoidefficiency in capturing 3D domain information

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The patent merges local point-level features with global shape representations into a unified 3D point cloud framework. By combining both local atomic properties and global molecular geometry information within a single 3D representation system, the model achieves both robustness and efficiency in capturing 3D domain information simultaneously.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The 3D point cloud representation serves multiple functions: it provides local geometric information, captures global molecular shape, and enables both training and inference operations. This multi-functional representation eliminates the need for separate processing systems, improving both robustness and computational efficiency.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS20230395202A1Using global-shape representations to generate a deep generative model

Publication Date: 2023.12.07 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

Embodiments of the invention provide a computer-implemented method that includes applying input representations of a three-dimensional (3D) domain to a generative neural network (GNN); and using the GNN to form a generative model of the 3D domain based at least in part on the input representations. The input representations include a global-shape input representation of the 3D domain.