Deep Boltzmann Machine for Inverse Molecular Design
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing molecular design mechanisms rely on evolutionary optimization and supervised shallow machine learning, requiring large datasets and often suggest infeasible molecules, while lacking an efficient method for inverse molecular design with target properties.
Innovation Solution
A method and system using a deep learning Bayesian framework that represents molecular structures with SMILES, converts them to binary form, pre-trains Restricted Boltzmann Machines, constructs a Deep Boltzmann Machine, and combines it with Bayesian inference to generate molecules with target properties using limited molecular property data from Density Functional Theory.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised shallow machine learning is used for molecular design, then structure-property correlation can be obtained, but large datasets are required and accuracy is insufficient
Solution Approach 1:
The patent transitions from shallow machine learning to deep learning architecture, adding depth dimension to the model. The deep neural network with multiple hidden layers enables the system to learn complex molecular structure-property relationships from limited data by creating hierarchical feature representations across multiple levels of abstraction.
Solution Approach 2:
The patent changes the architectural parameters of the machine learning model by introducing depth through multiple hidden layers, changing activation functions, and adjusting network configuration. These parameter changes enable the model to achieve higher accuracy with smaller datasets by learning more efficient feature representations.
2Reliability
If evolutionary optimization methods are used for molecular design, then expert information can be incorporated, but infeasible molecules may be suggested
Solution Approach 1:
The patent replaces evolutionary optimization methods with a deep learning-based generative model. Instead of using mechanical evolutionary algorithms that iteratively optimize molecular structures, the system uses a trained neural network to directly generate feasible molecular structures that satisfy target properties, eliminating the need for complex evolutionary search mechanisms.
Solution Approach 2:
The deep learning model learns to generate feasible molecular structures by training on existing molecular data, enabling the system to serve itself by incorporating implicit expert knowledge about molecular feasibility directly into the generative process, without requiring explicit evolutionary rules or constraints.
3Adaptability or versatility
If forward property prediction is used, then structure-property correlation can be obtained, but inverse molecular design with target properties cannot be performed
Solution Approach 1:
The patent inverts the traditional forward property prediction approach by implementing inverse molecular design. Instead of predicting properties from given structures, the system generates molecular structures that satisfy target property specifications by inverting the prediction function and using it as a generative model with property constraints.
Solution Approach 2:
The deep learning model serves dual functions: it can perform both forward property prediction and inverse molecular design. The same neural network architecture is used to predict properties from structures and to generate structures with target properties, making the system universally applicable to both prediction and design tasks.
Data Source
AI summary
The embodiments herein disclose a method and system for designing molecules by using a machine learning algorithm. The method includes representing molecular structures included in a dataset by using a Simplified Molecular Input Line Entry System (SMILES), where the SMILES uses a series of characters, converting a SMILES representation of the molecular structures into a binary representation, pre-training a stack of Restricted Boltzmann Machines (RBMs) by using the binary representation of the molecular structures, constructing a Deep Boltzmann Machine (DBM) by using the stack of the RBMs, determining limited molecular property data for a subset of the molecule structures in the dataset, training the DBM with the limited molecular property data, combining the pre-trained stack of the RBMs and the trained DBM in a Bayesian inference framework, and generating a sample of molecules with target properties by using the Bayesian inference framework.