Prediction method of dntf co-crystal ligand molecules based on multi-layer information cross fusion

Abstract

The application provides a DNTF co-crystal ligand molecule prediction method based on multi-layer information cross fusion, obtains current DNTF crystal data to be detected, and pre-processes the DNTF crystal data to obtain pre-processed DNTF crystal data, and then inputs the pre-processed DNTF crystal data into a trained DMSP network to obtain a prediction result of the DNTF crystal; wherein the trained DMSP network comprises a dynamic sparse projection layer, a multi-order interaction module, a structure perception module, a multi-layer attention fusion module and a classifier; since the DMSP network used in the application can extract co-crystal feature information that frequently appears and has clear patterns in crystal data through an explicit interaction layer in the multi-order interaction module, learn potential feature information of the co-crystal through a neural network of an implicit interaction layer, and then fuse information of two perspectives, and combine co-crystal structure features extracted in the structure perception module, the application truly realizes all-around capture of co-crystal information, and thus improves the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of DNTF eutectic materials, and specifically relates to a prediction method for DNTF eutectic ligand molecules based on multi-layer information cross-fusion. Background Art

[0002] The design, synthesis, and application of high-energy-density materials are crucial for improving the energy properties of propellants and explosives. Currently, CL-20 (hexanitrohexaazaisowurtzitane) is the most widely used material in the materials field, but its high cost restricts its widespread application. Recent research has revealed that DNTF (3,4-dinitrofurazanylfurazan oxide) offers high energy density, high detonation velocity, and excellent stability. Most importantly, DNTF's performance is similar to CL-20's, but at only half the cost. Therefore, DNTF holds high research value and holds great potential for development as a high-performance oxidizer for solid propellants.

[0003] Research has revealed that DNTF still presents numerous challenges, such as poor thermal stability, poor mechanical properties, crystallization, and poor compatibility, all of which can be addressed through eutectics. In recent years, eutectics have emerged as an effective approach to improving material properties. Energetic eutectic technology can modify the molecular composition (molecular arrangement) and crystal structure (crystal packing coefficient, density, etc.) of energetic materials without destroying the original molecular structure of the energetic compound. This reduces the free volume of the crystals, resulting in dense packing and further reducing the probability of hotspot formation. Consequently, physical properties such as density, melting point, sensitivity, and explosiveness change, achieving a balance between energy and safety. However, due to the significant experimental costs and safety concerns associated with eutectic technology, while eutectic materials have gradually demonstrated their potential in various application areas over the past few decades, such as electronics, semiconductor engineering, and chemical engineering, they have not yet been widely adopted in the field of energetic materials due to the complexity of the eutectic system and the relative limitations of eutectic structure research.

[0004] The field of artificial intelligence (AI) and machine learning is rapidly developing. Within the field of deep learning within AI, as the performance of deep learning models in prediction and classification improves, deep learning methods are beginning to be applied to materials science. Compared to traditional methods that extract simple low-level features through data mining, deep learning-based methods can learn from large amounts of data a wealth of high-level feature representations that are indiscernible to the naked eye, thereby improving predictive capabilities in complex data environments. Furthermore, deep learning methods, through large-scale data training, are more robust and adaptive, adapting to different data formats and avoiding the problem of widely varying prediction results associated with traditional methods due to the complexity of materials data.

[0005] However, searching for cocrystal ligand molecules for DNTF energetic materials through manual experiments is undoubtedly time-consuming and inefficient. Several issues remain to be addressed in the research on deep learning methods for predicting DNTF cocrystal ligands: First, the ratio of positive to negative samples in existing databases is unbalanced, and the different dimensions hinder the neural network's attention and weight adjustment during model training. Second, there is currently no deep learning framework established specifically for DNTF cocrystal ligand molecules, so deep learning cannot be fully utilized, and a corresponding deep learning model must be rebuilt for the data model. Third, many current models ignore the iterative nature of the cocrystal research process during prediction, resulting in poor prediction results. Summary of the Invention

[0006] To address the above-mentioned problems in the prior art, the present invention provides a method for predicting DNTF cocrystal ligand molecules based on multi-layer information cross-fusion. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] A prediction method for DNTF cocrystal ligand molecules based on multi-layer information cross-fusion, comprising:

[0008] S100, obtaining the DNTF crystal data to be tested, and preprocessing it to obtain preprocessed DNTF crystal data;

[0009] S200, input the preprocessed DNTF crystal data into the trained DMSP (dynamic multi-order structure perception) network to obtain the prediction results of eutectic or non-eutectic between DNTF crystals; wherein, the trained DMSP network includes a dynamic sparse projection layer, a multi-order interaction module, a structure perception module, a multi-level attention fusion module and a classifier; the dynamic sparse projection layer is used to filter the preprocessed DNTF crystal data; the multi-order interaction module is used to perform adaptive interaction between features through a dynamic routing mechanism; the structure perception module is used to capture the time, space and graph structure features in the preprocessed DNTF crystal data, and fuse the three; the multi-level attention fusion module is used to fuse the features output by the multi-order interaction module and the structure perception module through the attention mechanism and the gating function, and the classifier is used to perform prediction classification based on the fusion features output by the multi-level attention fusion module to output the prediction results.

[0010] Beneficial effects:

[0011] The present invention provides a prediction method for DNTF eutectic ligand molecules based on multi-layer information cross-fusion, which obtains the current DNTF crystal data to be detected, preprocesses it to obtain preprocessed DNTF crystal data, and then inputs it into a trained DMSP network to obtain a prediction result of eutectic or non-eutectic between DNTF crystals; wherein the trained DMSP network includes a dynamic sparse projection layer, a multi-order interaction module, a structure perception module, a multi-level attention fusion module and a classifier; the dynamic sparse projection layer is used to filter the preprocessed DNTF crystal data; the multi-order interaction module is used to perform adaptive interaction between features through a dynamic routing mechanism; the structure perception module is used to capture the time, space and graph structure features in the preprocessed DNTF crystal data, and fuse the three; the multi-level attention fusion module is used to fuse the features output by the multi-level interaction module and the structure perception module through an attention mechanism and a gating function, and the classifier is used to perform prediction and classification according to the fusion features output by the multi-level attention fusion module to output a prediction result. The DMSP network adopted in the present invention can extract the eutectic feature information that frequently appears and has clear patterns in the crystal data through the explicit interaction layer in the multi-order interaction module, learn the potential feature information of the eutectic through the neural network of the implicit interaction layer, and finally fuse the information from the two perspectives and combine it with the eutectic structural features extracted in the structure perception module to truly realize the all-round capture of the eutectic information, thereby improving the accuracy of the prediction.

[0012] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of a prediction method for DNTF cocrystal ligand molecules based on multi-layer information cross-fusion provided by the present invention.

[0014] Figure 2 It is a structural diagram of the DMSP network provided by the present invention.

[0015] Figure 3 This is a schematic diagram of the lightweight CNN residual network provided by the present invention.

[0016] Figure 4 Schematic diagram of the structure perception module provided by the present invention.

[0017] Figure 5 Schematic diagram of the multi-level attention fusion module provided by the present invention. DETAILED DESCRIPTION

[0018] The present invention will be further described in detail below with reference to specific examples, but the embodiments of the present invention are not limited thereto.

[0019] like Figure 1 As shown, the present invention provides a prediction method for DNTF cocrystal ligand molecules based on multi-layer information cross-fusion, including:

[0020] S100 , obtaining DNTF crystal data to be detected, and preprocessing the data to obtain preprocessed DNTF crystal data.

[0021] Specifically, S100 includes:

[0022] S110, obtaining a character string corresponding to the DNTF crystal data to be tested;

[0023] S120 , performing morgan encoding on the character string corresponding to the DNTF crystal data to be detected, to obtain pre-processed DNTF crystal data.

[0024] The string corresponding to the DNTF crystal data to be tested is preprocessed and converted into morgan code, and the data format is a binary feature vector .

[0025] S200, inputting the pre-processed DNTF crystal data into the trained DMSP network to obtain a prediction result of eutectic or non-eutectic between the DNTF crystals.

[0026] In a specific embodiment of the present invention, the training process of the trained DMSP network includes:

[0027] S000 uses crystallographic data from the Cambridge Crystallographic Library to construct positive and negative sample pairs of energetic molecules and their ligands.

[0028] S001, multiple positive and negative sample pairs are combined into a data set, where the positive and negative sample pairs are character strings corresponding to the crystals of two eutectics.

[0029] S002, converting the data set into a binary data structure through Morgan coding to obtain a preprocessed data set.

[0030] The present invention constructs sample pairs of energetic molecules and their ligands using crystal data from the Cambridge Crystallographic Data Library (CSD). After querying and extracting a large number of sample pairs from the database, a dataset of energetic material cocrystals is constructed. The positive sample pairs in the constructed dataset are the smiles (Simplified Molecular Input Line EntrySystem) strings of two cocrystal crystals, which can be used to encode the atomic connection order, bond type, and stereochemical information of the molecule, and the corresponding label is 1. The negative sample pairs are the smiles strings of non-cocrystal crystals, and the corresponding label is 0. After the dataset is created, the dataset is converted into a [0, 1] feature structure through Morgan encoding in the data preprocessing part, while the corresponding labels remain unchanged.

[0031] S003, iteratively training the preset DMSP network using the preprocessed data set to obtain a trained DMSP network.

[0032] refer to Figure 2 As shown in the figure, the trained DMSP network includes a dynamic sparse projection layer, a multi-order interaction module, a structure perception module, a multi-level attention fusion module and a classifier; the dynamic sparse projection layer is used to filter the preprocessed DNTF crystal data; the multi-order interaction module is used to perform adaptive interaction between features through a dynamic routing mechanism; the structure perception module is used to capture the temporal, spatial and graph structure features in the preprocessed DNTF (bis(triaza)tetrafluoroethylene) crystal data and fuse the three; the multi-level attention fusion module is used to fuse the features output by the multi-order interaction module and the structure perception module through the attention mechanism and the gating function, and the classifier is used to perform prediction and classification based on the fusion features output by the multi-level attention fusion module to output the prediction results.

[0033] The dynamic sparse projection layer of the present invention is implemented using an embedding layer. The role of the embedding layer is to perform dynamic sparse projection on the input data, with the purpose of reducing the embedding dimension, adaptively filtering unimportant features, and alleviating the curse of dimensionality. The embedding layer is expressed as:

[0034] (1);

[0035] Where, Represents the embedding layer The output result of dimension, for The corresponding sparse gating coefficients, express activation function, represents the weight matrix, Indicates the binary vector corresponding to the DNTF crystal data to be tested.

[0036] Combine Figure 2 and Figure 3 ,The multi-order interaction module of the present invention includes an explicit low-order interaction module and an implicit high-order interaction module, and the low-order interaction module and the high-order interaction module perform adaptive interaction through a dynamic routing mechanism;

[0037] The low-level interaction module is expressed mathematically as:

[0038] (2);

[0039] Where, Represents the output of the low-level interaction module, Represents a vector The corresponding weight, Represents a vector No. dimensional vector, Represents the vector dimension, represents the vector dimension, Represents a vector With vector The weight corresponding to the second-order crossover, Represents a vector No. The value of the dimension, Represents a vector ,vector With vector The weight corresponding to the third-order crossover, Represents a vector No. dimensional vector.

[0040] The low-order interaction module is designed based on the principle of factorization machine and supports the third-order intersection of vectors.

[0041] The high-order interaction module uses a lightweight CNN (deep neural network) residual network; the CNN residual network uses residual links and bottleneck structure (to reduce the number of parameters) to learn nonlinear combinations. Figure 3 As shown in the figure, the lightweight CNN residual network includes a first convolutional layer, a first batch normalization layer, a residual connection layer, a pooling layer and a first linear mapping layer connected in sequence; the residual connection layer includes a first weight layer, a first ReLU activation layer and a second weight layer connected in sequence, and the fusion feature obtained by fusing the output of the second weight layer with the input of the first weight layer is the output of the residual connection layer.

[0042] The output characteristics of the multi-order interaction module are expressed mathematically as follows:

[0043] (3);

[0044] Where, represents the output features of the multi-stage interaction module, represents the learnable weights, Represents a deep neural network function.

[0045] The structure perception module of the present invention includes a sequence perception branch, a spatial perception branch, and a graph structure perception branch. The three perception branches are responsible for capturing three hierarchical structural features respectively. In the sequence perception branch, a lightweight TCN network is used to capture temporal dependencies, in the spatial perception branch, a binary CNN is used to extract local spatial patterns, and in the graph structure perception branch, a lightweight GNN is used to aggregate the adjacency features of the morgan code. Figure 4 ,exist Figure 4 From left to right in the figure, they are the sequence perception branch, graph perception branch, and spatial perception branch. The sequence perception branch includes the dilated causal convolution layer, the second batch normalization layer, the second ReLU activation layer, and the second linear mapping layer; the spatial perception branch includes the adjacency matrix aggregation layer, the third linear mapping layer, and the third ReLU activation layer; the graph structure perception branch includes the binarization approximation layer, the second convolution layer, the third batch normalization layer, and the third linear mapping layer; the output features of the sequence perception branch, the spatial perception branch, and the graph structure perception branch are fused and then used as the output features of the structure perception module.

[0046] The output features of the structure perception module are expressed as follows:

[0047] (4);

[0048] Where, - Represents the weights corresponding to the sequence perception branch, space perception branch, and graph structure perception branch, represents the sequence-aware branching function, represents the spatial perception branch function, Represents graph structure-aware branching function.

[0049] The multi-level attention fusion module achieves dynamic weight calibration during feature fusion through dual-path gating coordination and structural constraint reinforcement. Its specific implementation is as follows:

[0050] (5);

[0051] Where, represents the fusion features output by the multi-level attention fusion module, Represents the two-way differentiated attention function corresponding to the multi-level attention fusion module.

[0052] refer to Figure 5As shown, the multi-level attention fusion module includes a structural space gating branch, a multi-order channel gating branch and a dynamic weight allocation layer. The structural space gating branch includes a structural space gating function layer and a first feature cross layer connected in sequence. The multi-order channel gating branch includes a multi-order channel gating function layer and a second feature cross layer connected in sequence. The output of the structural space gating function layer is connected to the input of the second feature cross layer, and the output of the multi-order channel gating function layer is connected to the input of the first feature cross layer. The first feature cross layer and the second feature cross layer undergo cross calculation and output the calculation results to the dynamic weight allocation layer. The dynamic weight allocation layer assigns weights to the calculation results of the first feature cross layer and the second feature cross layer and calculates and outputs fusion features.

[0053] Among them, the calculation process of the first feature cross layer and the second feature cross layer is the same due to the cross calculation.

[0054] The structure space gating function layer adopts the structure space gating function, which is expressed as follows:

[0055] (6);

[0056] The multi-order channel gating function layer adopts a multi-order channel gating function, which is expressed as follows:

[0057] (7);

[0058] Where, represents the structure space gating function, represents a multi-order channel gating function, represents the sigmoid function, represents the one-dimensional convolution function, represents the feature concatenation function, represents the average pooling function, represents the output features of the multi-stage interaction module, Represents the output result of the multi-order interaction module calculated by the maximum pooling function, represents the output features of the structure perception module, Represents feature splicing.

[0059] The cross calculation process of the first feature cross layer and the second feature cross layer is expressed as follows:

[0060] (8);

[0061] Where, represents element-wise multiplication, represents a learnable parameter.

[0062] The fusion feature is expressed as follows:

[0063] (9);

[0064] Where, Represents prior knowledge, which is used to dynamically adjust the fusion weights. It means to find the Hadamard product of two elements.

[0065] The multi-level attention fusion module of the present invention not only preserves the cross-synergistic effect of features, but also prevents excessive gating inhibition.

[0066] In the downstream tasks of the model, the DNTF crystal molecule is used as one of the sample pairs, and it is respectively combined with a variety of ligand molecules to produce a DNTF test data set. The DNTF test data set is also in the form of a smiles string, and then converted into learnable binary data through morgan encoding. The binary data is input into the trained DMSP network to obtain the final prediction result. Based on the prediction results, all crystal sample pairs predicted to be co-crystals are screened out, and further manual experiments are carried out for verification. The verification results show that the present invention can improve the prediction accuracy.

[0067] It is worth noting that the terms "first" and "second" in this disclosure are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, features defined as "first" or "second" may explicitly or implicitly include one or more of such features. In the description of this disclosure, "plurality" means two or more, unless otherwise specifically defined.

[0068] The above is a further detailed description of the present invention in conjunction with specific preferred embodiments, and the specific implementation of the present invention should not be considered to be limited to these descriptions. For those skilled in the art of the present invention, without departing from the concept of the present invention, several simple deductions or substitutions can be made, which should be considered to fall within the scope of protection of the present invention.

Claims

1. A prediction method for DNTF cocrystal ligand molecules based on multi-layer information cross-fusion, characterized in that: include: S100, obtaining the DNTF crystal data to be tested, and preprocessing it to obtain preprocessed DNTF crystal data; S200, inputting the preprocessed DNTF crystal data into the trained DMSP network to obtain a prediction result of eutectic or non-eutectic between DNTF crystals; wherein the trained DMSP network includes a dynamic sparse projection layer, a multi-order interaction module, a structure perception module, a multi-level attention fusion module and a classifier; the dynamic sparse projection layer is used to filter the preprocessed DNTF crystal data; the multi-order interaction module is used to perform adaptive interaction between features through a dynamic routing mechanism; the structure perception module is used to capture the temporal, spatial and graph structure features in the preprocessed DNTF crystal data and fuse the three; the multi-level attention fusion module is used to fuse the features output by the multi-order interaction module and the structure perception module through an attention mechanism and a gating function, and the classifier is used to perform prediction and classification based on the fusion features output by the multi-level attention fusion module to output a prediction result; The multi-order interaction module includes an explicit low-order interaction module and an implicit high-order interaction module, and the low-order interaction module and the high-order interaction module perform adaptive interaction through a dynamic routing mechanism; the high-order interaction module adopts a lightweight CNN residual network; the lightweight CNN residual network includes a first convolutional layer, a first batch normalization layer, a residual connection layer, a pooling layer and a first linear mapping layer connected in sequence; the residual connection layer includes a first weight layer, a first ReLU activation layer and a second weight layer connected in sequence, and the fusion feature obtained by fusing the output of the second weight layer with the input of the first weight layer is the output of the residual connection layer; The structure perception module includes a sequence perception branch, a space perception branch and a graph structure perception branch; The multi-level attention fusion module includes a structure space gating branch, a multi-order channel gating branch and a dynamic weight distribution layer. The structure space gating branch includes a structure space gating function layer and a first feature cross layer connected in sequence. The multi-order channel gating branch includes a multi-order channel gating function layer and a second feature cross layer connected in sequence. The output of the structure space gating function layer is connected to the input of the second feature cross layer, and the output of the multi-order channel gating function layer is connected to the input of the first feature cross layer. The first feature cross layer and the second feature cross layer undergo cross calculation and output the calculation results to the dynamic weight distribution layer. The dynamic weight distribution layer assigns weights to the calculation results of the first feature cross layer and the second feature cross layer and calculates and outputs fusion features.

2. The prediction method of DNTF cocrystal ligand molecules based on multi-layer information cross-fusion according to claim 1, characterized in that: S100 includes: S110, obtaining a character string corresponding to the DNTF crystal data to be tested; S120 , performing morgan encoding on the character string corresponding to the DNTF crystal data to be detected to obtain pre-processed DNTF crystal data.

3. The prediction method of DNTF cocrystal ligand molecules based on multi-layer information cross-fusion according to claim 1, characterized in that: The dynamic sparse projection layer is implemented using an embedding layer, which is expressed as follows: ； Where, Represents the embedding layer The output result of dimension, for The corresponding sparse gating coefficients, ()express activation function, represents the weight matrix, Indicates the binary vector corresponding to the DNTF crystal data to be tested.

4. The prediction method of DNTF cocrystal ligand molecules based on multi-layer information cross-fusion according to claim 1, characterized in that: The low-order interaction module is expressed as follows: ； Where, Represents the output of the low-level interaction module, Represents a vector The corresponding weight, Represents a vector No. dimensional vector, Represents the vector dimension, represents the vector dimension, Represents a vector With vector The weight corresponding to the second-order crossover, Represents a vector No. The value of the dimension, Represents a vector ,vector With vector The weight corresponding to the third-order crossover, Represents a vector No. dimensional vector; The output characteristics of the multi-stage interaction module are expressed as follows: ； Where, represents the output features of the multi-stage interaction module, represents the learnable weights, Represents a deep neural network function.

5. The prediction method of DNTF cocrystal ligand molecules based on multi-layer information cross-fusion according to claim 1, characterized in that: The sequence perception branch includes an expanded causal convolution layer, a second batch normalization layer, a second ReLU activation layer, and a second linear mapping layer; the spatial perception branch includes an adjacency matrix aggregation layer, a third linear mapping layer, and a third ReLU activation layer; The graph structure perception branch includes a binarization approximation layer, a second convolutional layer, a third batch normalization layer, and a third linear mapping layer; the output features of the sequence perception branch, the spatial perception branch, and the graph structure perception branch are fused and then used as the output features of the structure perception module.

6. The method for predicting DNTF cocrystal ligand molecules based on multi-layer information cross-fusion according to claim 5, characterized in that: The output features of the structure perception module are expressed as follows: ； Where, represents the weight of the sequence perception branch, represents the weight of the spatial perception branch, represents the weight of the graph structure perception branch, represents the sequence-aware branching function, represents the spatial perception branch function, Represents graph structure-aware branching function.

7. The method for predicting DNTF cocrystal ligand molecules by multi-layer information cross-fusion according to claim 1, characterized in that: The structure space gating function layer adopts the structure space gating function, and the structure space gating function is expressed by the formula: ； The multi-stage channel gating function layer adopts a multi-stage channel gating function, which is expressed as follows: ； Where, represents the structure space gating function, represents a multi-order channel gating function, represents the sigmoid function, represents the one-dimensional convolution function, represents the feature concatenation function, represents the average pooling function, represents the output features of the multi-stage interaction module, Represents the output result of the multi-order interaction module calculated by the maximum pooling function, represents the output features of the structure perception module, Represents feature splicing; The cross calculation process of the first feature cross layer and the second feature cross layer is expressed as follows: ； Where, represents element-wise multiplication, represents a learnable parameter; The fusion feature is expressed as follows: ； Where, Represents prior knowledge, which is used to dynamically adjust the fusion weights. It means to find the Hadamard product of two elements.

8. The method for predicting DNTF cocrystal ligand molecules based on multi-layer information cross-fusion according to claim 1, characterized in that: The training process of the trained DMSP network includes: S000, uses crystallographic data from the Cambridge Crystallographic Library to construct positive and negative sample pairs of energetic molecules and their ligands; S001, forming a data set from a plurality of positive and negative sample pairs, wherein the positive and negative sample pairs are character strings corresponding to two eutectic crystals; S002, converting the data set into a binary data structure through Morgan coding to obtain a preprocessed data set; S003, iteratively training the preset DMSP network using the preprocessed data set to obtain a trained DMSP network.

Images