Methods, systems and equipment for predicting molecular properties

By converting molecular data into SELFIES sequences and performing adaptive graph structure learning, the problem that existing models cannot learn to conform to physicochemical laws is solved, and high-precision molecular property prediction is achieved.

CN123120118APending Publication Date: 2026-11-17SHANGHAI LANCHANG AUTO TECH CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611415506.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-13
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing pre-trained models cannot adaptively adjust and learn molecular graph structures that better conform to physicochemical laws, resulting in insufficient accuracy in predicting downstream properties.

Method used

The processor transforms molecular data into SELFIES sequences, obtains initial feature vectors, constructs a similarity matrix, performs adaptive graph structure learning, establishes an objective function for predicting molecular properties, and uses a graph convolutional network to capture deep chemical semantics and higher-order dependencies.

Benefits of technology

It significantly improves the accuracy of downstream small sample molecular property prediction and enhances the precision of molecular property prediction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123120118A_ABST
    Figure CN123120118A_ABST
Patent Text Reader

Abstract

This invention belongs to the interdisciplinary field of computer deep learning and cheminformatics, and particularly relates to a molecular property prediction method, prediction system, and prediction device. The molecular property prediction method involves a processor extracting SMILES representations from molecular data and converting them into SELFIES sequences, obtaining initial feature vectors for each molecule; the processor constructs a similarity matrix based on the initial feature vectors and molecular fingerprint structure information; the processor adaptively learns the graph structure based on the similarity matrix; and the processor establishes an objective function for molecular property prediction based on the learned graph representation features, thereby predicting molecular properties. This achieves the capture of deep chemical semantics and high-order dependencies in molecular graph structures based on high-level features extracted from existing deep chemical models, through an adaptive graph structure learning method, thus improving the accuracy of downstream small-sample molecular property prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer deep learning and cheminformatics, and particularly relates to a method, system and device for predicting molecular properties. Background Technology

[0002] Molecular property prediction (such as toxicity, solubility, metabolic stability, and bioactivity) is a core component of modern drug development, materials science, and fine chemicals. Traditional molecular property assessment mainly relies on wet experiments (such as high-throughput screening), but these methods have inherent drawbacks such as high cost, long processing times, and low success rates. In recent years, the use of computer-aided design (CADD) for artificial intelligence-based molecular property prediction has become an important means to accelerate new drug development.

[0003] In early AI-based predictions of molecular properties, researchers often used one-dimensional molecular sequences (such as SMILES) or two-dimensional molecular fingerprints based on feature engineering (such as ECFP4), combined with traditional machine learning algorithms. However, these methods often fragment the complex spatial topology of molecules, resulting in models that are insufficient in capturing complex chemical features and have poor generalization performance.

[0004] With the development of deep learning, Graph Neural Networks (GNNs) have become the mainstream tool for molecular representation learning because they can naturally model molecules as graph structures (where atoms are considered nodes and chemical bonds are considered edges). Traditional methods directly use chemical bond connections as fixed, static graph structure inputs. However, traditional chemical bond connections are not entirely equivalent to the actual interactions within molecules. Existing pre-trained models cannot adaptively adjust and learn graph structures that better conform to physicochemical laws when faced with these static graphs, thus limiting the accuracy of downstream property predictions.

[0005] Therefore, due to the technical problem that existing pre-trained models cannot meet the accuracy requirements for downstream property prediction, it is necessary to design a molecular property prediction method, prediction system, and prediction device.

[0006] It should be noted that the information disclosed in this background section is only for understanding the background technology of the present application concept, and therefore, the above description is not considered to constitute prior art information. Summary of the Invention

[0007] This disclosure provides at least one method, system, and device for predicting molecular properties.

[0008] In a first aspect, embodiments of this disclosure provide a method for predicting molecular properties, including: The processor extracts SMILES representations from the molecular data and then converts them into SELFIES sequences, and obtains the initial feature vectors for each molecule; The processor constructs a similarity matrix based on the initial feature vector and molecular fingerprint structure information; The processor learns adaptively based on the similarity matrix and graph structure. The processor establishes an objective function for predicting molecular properties based on the learned graphical representation features, and then predicts the molecular properties.

[0009] In one alternative implementation, the method by which the processor extracts SMILES representations from molecular data, converts them into SELFIES sequences, and obtains initial feature vectors for each molecule includes: For each molecule represented as SMILES, the processor converts it into a SELFIES sequence, and then fine-tunes it using a pre-trained molecular model SELFIES-TED to obtain the initial feature vector for each molecule. ; Among them, Hi (0) d is the feature vector of the 8th molecule; N is the number of molecules; d is the dimension of the molecule feature vector.

[0010] In one optional implementation, the method by which the processor constructs a similarity matrix based on the initial feature vector and molecular fingerprint structure information includes: The processor uses Morgan fingerprint feature vectors to calculate the similarity between molecules: ; in, f represents the similarity between the i-th and j-th molecules; i f j For the Morgan fingerprint information of the i,j molecule; , representing the molecular fingerprint dimension; Construct a molecular similarity diagram: ; ; in, This is the intermolecular similarity matrix. The threshold value is used.

[0011] In one alternative implementation, the method by which the processor learns the adaptive graph structure based on the similarity matrix includes: Processor-adaptive graph similarity estimation maps nodes to M learnable subspaces to obtain their representation vectors in those subspaces. : ; in, For the node embedding in the (t-1)th iteration; Let be the parameter matrix to be learned.

[0012] In one alternative implementation, the processor acquires molecules. Similarity in the m-th space: ; Where T is the vector transpose; The constructed adaptive graph similarity matrix is: ; in, ; Multi-view iterative update fusion: ; in, Let be the initial similarity matrix, the similarity matrix obtained in the first iteration, and the similarity matrix obtained in the t-th iteration; These are weight parameters; For the convolutional representation of the l-layer graph at step t, the feature update is as follows: ; in, ; , It is a diagonal matrix. I is the identity matrix; The parameter matrix to be learned; It is a non-linear activation function; The embedding for the t-th iteration is: .

[0013] In one optional implementation, the processor establishes an objective function for predicting molecular properties based on the learned graphical representation features, and the method for predicting molecular properties includes: The objective function for predicting molecular properties is: ; in, These are weight parameters used to describe the importance of each part of the objective function; The graph-related loss function is constructed as follows: Construct the target binary similarity matrix ; ; The loss function is then: ; The processor predicts molecular properties using an objective function for molecular property prediction.

[0014] Secondly, embodiments of this disclosure also provide a molecular property prediction system, comprising: The initial feature vector module is configured to extract SMILES representations from molecular data, convert them into SELFIES sequences, and obtain the initial feature vectors for each molecule. The similarity matrix module is configured to construct a similarity matrix based on the initial feature vector and molecular fingerprint structure information; The graph structure learning module is configured to adaptively learn graph structures based on the similarity matrix; The prediction module is configured to establish an objective function for predicting molecular properties based on the learned graphical representation features, thereby predicting molecular properties.

[0015] Thirdly, embodiments of this disclosure also provide a molecular property prediction device, comprising: A processor configured to predict molecular properties using the aforementioned molecular property prediction method.

[0016] Fourthly, embodiments of this disclosure also provide a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described molecular property prediction method.

[0017] Fifthly, embodiments of this disclosure also provide a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described molecular property prediction method.

[0018] The beneficial effects of this invention are that the molecular property prediction method extracts SMILES representations from molecular data and converts them into SELFIES sequences, and obtains the initial feature vectors of each molecule; the processor constructs a similarity matrix based on the initial feature vectors and molecular fingerprint structure information; the processor adaptively learns the graph structure based on the similarity matrix; the processor establishes an objective function for molecular property prediction based on the learned graph representation features, and then predicts the molecular properties. This achieves the capture of deep chemical semantics and high-order dependencies in molecular graph structures based on high-level features extracted from existing deep chemical models, thereby improving the accuracy of downstream small-sample molecular property prediction.

[0019] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0021] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 A flowchart of a molecular property prediction method provided in an embodiment of this disclosure. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] As used herein, the phrases “in one embodiment,” “according to one embodiment,” “in some embodiments,” etc., generally refer to the fact that a particular feature, structure, or characteristic following the phrase can be included in at least one embodiment of this disclosure. Therefore, a particular feature, structure, or characteristic can be included in more than one embodiment of this disclosure, such that these phrases do not necessarily refer to the same embodiment. As used herein, the terms “example,” “exemplary,” etc., are used to “serve as an example, instance, or illustration.” Any implementation, aspect, or design described herein as “example” or “exemplary” is not necessarily to be construed as preferred or superior to other implementations, aspects, or designs. Rather, the use of the terms “example,” “exemplary,” etc., is intended to present concepts in a specific manner.

[0025] Traditional methods directly input chemical bonds as a fixed, static graph structure. However, conventional chemical bonds do not fully represent actual interactions within molecules. Existing pre-trained models, when faced with these static graphs, cannot adaptively adjust and learn graph structures that better conform to physicochemical laws, thus limiting the accuracy of downstream property predictions.

[0026] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0027] The following detailed description of some embodiments of the present invention is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0028] like Figure 1 As shown, at least one disclosed embodiment provides a molecular property prediction method, including: a processor extracting SMILES representations from molecular data and converting them into SELFIES sequences, and obtaining initial feature vectors for each molecule; the processor constructing a similarity matrix based on the initial feature vectors and molecular fingerprint structure information; the processor adaptively learning a graph structure based on the similarity matrix; and the processor establishing an objective function for molecular property prediction based on the learned graph representation features, thereby predicting molecular properties. This achieves the capture of deep chemical semantics and high-order dependencies in molecular graph structures based on high-level features extracted from existing deep chemical models, through an adaptive graph structure learning method, thereby improving the accuracy of downstream small-sample molecular property prediction.

[0029] In this embodiment, adaptive graph structure learning is performed on high-level semantic features using graph convolutional networks, which can significantly improve prediction accuracy compared to other methods.

[0030] In this embodiment, the molecular property prediction method can be used, but is not limited to, for predicting the properties of water molecules.

[0031] In this embodiment, the molecular property prediction method using adaptive graph structure learning can fully capture the molecular graph structure of deep chemical semantics and high-order dependencies. Experimental results show that the boundary ambiguity of this method is significantly reduced compared with other methods.

[0032] In this embodiment, the initial feature vectors of each molecule are first obtained using the pre-trained molecular model SELFIES-TED, fingerprint representation vectors are constructed to calculate the molecular similarity graph, and then molecular similarity is obtained through graph structure learning, thereby realizing molecular property prediction.

[0033] In one optional implementation, the method by which the processor extracts SMILES representations from molecular data and then converts them into SELFIES sequences, and obtains the initial feature vectors for each molecule includes: the processor converts each molecule represented as SMILES into a SELFIES sequence, and then obtains the initial feature vectors for each molecule by fine-tuning a pre-trained molecular model SELFIES-TED. ; Among them, Hi (0) d is the feature vector of the 8th molecule; N is the number of molecules; d is the dimension of the molecule feature vector.

[0034] In one optional implementation, the method by which the processor constructs a similarity matrix based on the initial feature vector and molecular fingerprint structure information includes: the processor using the Morgan fingerprint feature vector to calculate the similarity between molecules. ; in, f represents the similarity between the i-th and j-th molecules; i f j For the Morgan fingerprint information of the i,j molecule; , representing the molecular fingerprint dimension; Construct a molecular similarity diagram: ; ; in, This is the intermolecular similarity matrix; The threshold value is used.

[0035] In one alternative implementation, the method by which the processor learns the adaptive graph structure based on the similarity matrix includes: Processor-adaptive graph similarity estimation maps nodes to M learnable subspaces to obtain their representation vectors in those subspaces. : ; in, For the node embedding in the (t-1)th iteration; Let be the parameter matrix to be learned.

[0036] In one alternative implementation, the processor acquires molecules. The similarity e(t)ij,m in the m-th space: ; Where T is the vector transpose; The constructed adaptive graph similarity matrix is: ; in, ; Multi-view iterative update fusion: ; in, Let be the initial similarity matrix, the similarity matrix obtained in the first iteration, and the similarity matrix obtained in the t-th iteration; These are weight parameters; For the convolutional representation of the l-layer graph at step t, the feature update is as follows: ; in, I is the identity matrix; The parameter matrix to be learned; It is a non-linear activation function; The embedding for the t-th iteration is: .

[0037] In one optional implementation, the processor establishes an objective function for predicting molecular properties based on the learned graph representation features, and the method for predicting molecular properties includes: the objective function for predicting molecular properties is: ; in, These are weight parameters used to describe the importance of each part of the objective function; Where the prediction loss function Typically, cross-entropy loss or mean loss function is chosen. The graph-related loss function is constructed as follows: Construct the target binary similarity matrix ; ; The loss function is then: ; The processor predicts molecular properties using an objective function for molecular property prediction.

[0038] At least one other disclosed embodiment also provides a molecular property prediction system, comprising: an initial feature vector module configured to extract SMILES representations from molecular data and convert them into SELFIES sequences, and obtain initial feature vectors for each molecule; a similarity matrix module configured to construct a similarity matrix based on the initial feature vectors and molecular fingerprint structure information; a graph structure learning module configured to adaptively learn graph structures based on the similarity matrix; and a prediction module configured to establish an objective function for molecular property prediction based on the learned graph representation features, and then predict molecular properties.

[0039] In this embodiment, the functions of each module can be integrated into the processor.

[0040] At least one other disclosed embodiment also provides a molecular property prediction device, comprising: a processor configured to predict molecular properties using the above-described molecular property prediction method.

[0041] At least one other disclosed embodiment also provides a computer-readable storage medium having a computer program / instructions stored thereon that, when executed by a processor, implements the steps of the above-described molecular property prediction method.

[0042] At least one other disclosed embodiment also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described molecular property prediction method.

[0043] In summary, this molecular property prediction method extracts SMILES representations from molecular data and converts them into SELFIES sequences, obtaining initial feature vectors for each molecule. The processor then constructs a similarity matrix based on the initial feature vectors and molecular fingerprint structure information. The processor then performs adaptive graph structure learning based on the similarity matrix. Finally, the processor establishes an objective function for molecular property prediction based on the learned graph representation features, thereby predicting molecular properties. This method achieves the goal of capturing deep chemical semantics and high-order dependencies in molecular graph structures based on high-level features extracted from existing deep chemical models, through an adaptive graph structure learning method, thus improving the accuracy of downstream small-sample molecular property prediction.

[0044] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for predicting molecular properties, characterized in that, include: The processor extracts SMILES representations from the molecular data and then converts them into SELFIES sequences, and obtains the initial feature vectors for each molecule; The processor constructs a similarity matrix based on the initial feature vector and molecular fingerprint structure information; The processor learns adaptively based on the similarity matrix and graph structure. The processor establishes an objective function for predicting molecular properties based on the learned graphical representation features, and then predicts the molecular properties.

2. The molecular property prediction method as described in claim 1, characterized in that: The processor extracts SMILES representations from molecular data, converts them into SELFIES sequences, and obtains the initial feature vectors for each molecule using the following methods: The processor converts each molecule represented as SMILES into a SELFIES sequence, and then fine-tunes the pre-trained molecular model SELFIES-TED to obtain the initial feature vector for each molecule. : ; Among them, Hi (0) d is the feature vector of the 8th molecule; N is the number of molecules; d is the dimension of the molecule feature vector.

3. The molecular property prediction method as described in claim 2, characterized in that: The method by which the processor constructs a similarity matrix based on the initial feature vector and molecular fingerprint structure information includes: The processor uses Morgan fingerprint feature vectors to calculate the similarity between molecules: ; in, f represents the similarity between the i-th and j-th molecules; i f j For the Morgan fingerprint information of the i,j molecule; , representing the molecular fingerprint dimension; Construct a molecular similarity diagram: ; ; in, This is the intermolecular similarity matrix. The threshold value is used.

4. The molecular property prediction method as described in claim 3, characterized in that: The processor's method for adaptive graph structure learning based on the similarity matrix includes: Processor-adaptive graph similarity estimation maps nodes to M learnable subspaces to obtain their representation vectors in those subspaces. : ; in, For the node embedding in the (t-1)th iteration; Let be the parameter matrix to be learned.

5. The molecular property prediction method as described in claim 4, characterized in that: Processor acquires molecules Similarity in the m-th space: ; Where T is the vector transpose; The constructed adaptive graph similarity matrix is: ; in, ; Multi-view iterative update fusion: ; in, Let be the initial similarity matrix, the similarity matrix obtained in the first iteration, and the similarity matrix obtained in the t-th iteration; These are weight parameters; For the convolutional representation of the l-layer graph at step t, the feature update is as follows: ; in, ; , It is a diagonal matrix. I is the identity matrix; The parameter matrix to be learned; It is a non-linear activation function; The embedding for the t-th iteration is: 。 6. The molecular property prediction method as described in claim 5, characterized in that: The processor establishes an objective function for predicting molecular properties based on the learned graphical representation features, and the methods for predicting molecular properties include: The objective function for predicting molecular properties is: ; in, These are weight parameters used to describe the importance of each part of the objective function; The graph-related loss function is constructed as follows: Construct the target binary similarity matrix ; ; The loss function is then: ; The processor predicts molecular properties using an objective function for molecular property prediction.

7. A molecular property prediction system, characterized in that, include: The initial feature vector module is configured to extract SMILES representations from molecular data, convert them into SELFIES sequences, and obtain the initial feature vectors for each molecule. The similarity matrix module is configured to construct a similarity matrix based on the initial feature vector and molecular fingerprint structure information; The graph structure learning module is configured to adaptively learn graph structures based on the similarity matrix; The prediction module is configured to establish an objective function for predicting molecular properties based on the learned graphical representation features, thereby predicting molecular properties.

8. A molecular property prediction device, characterized in that, include: A processor configured to predict molecular properties using the molecular property prediction method as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When executed by a processor, the computer program / instructions implement the steps of the molecular property prediction method according to any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When executed by a processor, the computer program / instructions implement the steps of the molecular property prediction method according to any one of claims 1-6.