A method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion

Abstract

This disclosure proposes a compound blood-brain barrier permeability prediction method based on multimodal fusion. The specific implementation steps are as follows: (1) Collect SMILES expressions of compounds with blood-brain barrier permeability tags, perform data preprocessing on all compounds to obtain an enhanced dataset; (2) Utilize the SMILES expression of each compound in the dataset to generate molecular three-dimensional structure features (3DF), molecular text features (MNF), molecular descriptor features (MDF), and molecular graph features (MGF). Normalize MNF, 3DF, MDF, and MGF to obtain a normalized molecular three-dimensional structure feature. (2) Construct the feature P-3DF, normalized text feature P-MNF, normalized molecular descriptor feature P-MDF, and normalized molecular graph feature P-MGF; (3) Construct the feature fusion network BBBNet to construct a multimodal fusion feature vector dataset; (4) Divide the dataset into a training set and a test set, and input the training set into the Kolmogorov-Arnold model KAN for model training; after training, use the test set to evaluate the performance of the trained KAN model; (5) Use the trained KAN model to predict the blood-brain barrier permeability of the compound.

Description

Technical Field

[0001] This disclosure pertains to the field of bioinformatics processing, and further relates to a method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion in the field of deep learning technology. This invention can be used to predict the blood-brain barrier permeability of drugs under development. Background Technology

[0002] In the development of compounds targeting the central nervous system (CNS), the permeability of the blood-brain barrier (BBB) ​​is a crucial indicator. As a barrier protecting the brain and surrounding neurons, the BBB selectively filters and restricts the entry and exit of substances into the brain, preventing harmful substances, pathogens, or foreign substances from entering while maintaining the homeostasis of brain cells. Therefore, assessing the BBB permeability of candidate drugs is of great significance in early drug discovery and development, especially in optimizing the efficacy of CNS drugs. However, while traditional clinical trial methods are accurate, they are often costly and time-consuming.

[0003] In recent years, deep learning and machine learning methods have been widely used in predicting the permeability of compounds along the blood-brain barrier (BBB), achieving significant progress in improving prediction efficiency and accuracy. However, existing methods still have room for performance improvement. Most studies focus on molecular textual features (SMILES expressions) or molecular graph features, neglecting the role of molecular three-dimensional structure. In fact, the three-dimensional structure of molecules plays a crucial role in assessing their physicochemical properties and interactions with biological targets. Therefore, the lack of full utilization of three-dimensional structure limits the comprehensiveness and accuracy of the models.

[0004] Furthermore, data imbalance remains a challenge. In current compound datasets, there are relatively more compounds that cross the blood-brain barrier (BBB+) and fewer that do not, leading to poor prediction accuracy for the minority class. Therefore, existing models struggle to effectively address this issue when dealing with real-world compound development. Training deep learning models typically requires a large amount of sample data; insufficient data can prevent the model from learning adequately, resulting in incomplete training and affecting the final blood-brain barrier permeability prediction. Moreover, current multimodal feature fusion methods have shortcomings in feature alignment and fusion mechanisms, failing to fully exploit the potential correlations between different features, further limiting the model's blood-brain barrier permeability prediction and generalization capabilities. Therefore, effectively addressing issues such as data imbalance, insufficient data volume, and multimodal feature fusion remains a significant challenge in current compound blood-brain barrier permeability prediction. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of this section, the abstract, and the title of the invention; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] The purpose of this invention is to address the problems existing in the above-mentioned background technology by proposing a method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion.

[0007] Terminology Explanation:

[0008] Multimodal: A modality refers to different types or forms of data, with different modalities representing different data sources or perception channels. Multimodal processing involves integrating and processing data from multiple modalities to achieve a more comprehensive understanding and reasoning.

[0009] Blood-brain barrier permeability label: A marker indicating whether a compound can cross the blood-brain barrier. If a compound can cross the blood-brain barrier, its blood-brain barrier permeability label is BBB+; otherwise, it is BBB-.

[0010] Positive samples: Compound samples that can cross the blood-brain barrier. The blood-brain barrier permeability label is BBB+, also known as BBB+ samples.

[0011] Negative samples: Compound samples that cannot cross the blood-brain barrier. The blood-brain barrier permeability label is BBB-, also known as BBB- samples.

[0012] Minority class samples: In classification tasks, minority class samples refer to samples of a certain class in the dataset that are significantly fewer than samples of other classes, and usually account for a small proportion of the overall distribution.

[0013] ADASYN: An oversampling technique for handling class imbalance. It generates new minority class samples, especially more samples in the boundary regions, making the classifier pay more attention to minority class samples and improving the model's ability to identify minority classes.

[0014] Simplified Molecular Linear Input Standard: This is a standard that uses ASCII characters to describe molecular structures. Its main feature is that it uses strings to represent the atoms and their connections in compounds, making the input, storage, and computer processing of molecular structures more convenient.

[0015] SMILES expressions are a simplified linear canonical representation of molecules, commonly used to represent chemical structures. They consist of a series of atomic symbols and connectors, accurately describing the molecular structure for computer processing and storage. SMILES expressions can be imported into most molecular editing software and converted into two-dimensional graphics or three-dimensional molecular models. Based on the fundamental principle of cheminformatics that similar molecules have similar properties, they can therefore be used to predict biochemical characteristics.

[0016] Three-dimensional molecular structure: The three-dimensional molecular structure refers to the specific form of a molecule in three-dimensional space. It is determined by the spatial arrangement of atoms, stereochemical factors, and intermolecular interactions, and has an important influence on the physical, chemical, and biological activities of molecules.

[0017] The Kolmogorov-Arnold model (KAN) is designed based on the Kolmogorov-Arnold representation theorem. It can decompose multidimensional input features into single-variable combinations layer by layer to approximate complex multidimensional functions.

[0018] Chemical toolkits: A set of computer programs used to process chemical information and molecular data. These tools can be used for tasks such as molecular modeling, computational chemical property prediction, and molecular structure analysis. Some commonly used chemical toolkits include RDKit, Open Babel, and ChemAxon. These toolkits provide rich functionality and algorithms that can be used to generate molecular fingerprints and descriptors for applications such as molecular similarity calculations, compound screening, and quantitative structure-activity relationship (QSAR) modeling.

[0019] Node Features: The feature matrix represents the attributes of each node (i.e., atom) in the graph. Each row represents a node, and each column corresponds to a feature of the node, such as atom type, valence, number of bonds, etc.

[0020] Layer Normalization: Standardizes the output of each layer to improve the training speed and stability of the model.

[0021] Query vector space: In attention mechanisms, the query vector space is a multi-dimensional space used to represent the query vector. In this space, the query vector is matched with the key vector to calculate attention weights, thereby determining the importance of each part in the information aggregation process.

[0022] Key vector space: In the attention mechanism, the key vector space is used to represent the multidimensional space in which the key vectors reside. The key vectors and query vectors are compared in this space to measure the relevance of different information fragments to the current query, thereby determining the weights for attention allocation.

[0023] Value vector space: In the attention mechanism, the value vector space is used to represent the multidimensional space in which value vectors reside. Each value vector corresponds to a piece of information, and its relationship with the query vector and key vector is adjusted through attention weights.

[0024] Self-Attention Score: This is a value that measures the correlation or similarity between different elements. In the self-attention mechanism, the self-attention score is used to determine the strength of the association between an element and other elements, thereby determining how that element should be weighted in the final output.

[0025] n-gram model: The n-gram model is a probabilistic language model widely used in text modeling in natural language processing. It estimates the probability of a word or character in a sequence by analyzing the frequency of occurrence of a continuous substring of length n (n-gram).

[0026] The n-gram vocabulary refers to the set of all possible n-grams included in an n-gram model. An n-gram is a subsequence of n consecutive words or characters used to represent local contextual relationships within a text. The size of the vocabulary depends on the number of all possible n-grams in the text and is a crucial foundation for building an n-gram model.

[0027] A substring of length 2 refers to all possible subsequences of two consecutive characters extracted from a string in character order.

[0028] The purpose of this invention is to address the following problems:

[0029] (1) In terms of feature extraction, existing methods only focus on molecular graph features and molecular text features, and do not make full use of molecular three-dimensional information, resulting in the inability to fully capture the spatial structural features of molecules, and thus have low accuracy in the blood-brain barrier permeability prediction task.

[0030] (2) Existing methods have the following shortcomings in multimodal feature fusion: On the one hand, they do not consider the alignment problem between features of different modalities. Different modal data have significant differences in feature scale and distribution, and traditional weighted average methods are difficult to achieve effective cross-modal semantic alignment. On the other hand, fixed-weight feature fusion strategies cannot adapt to the differences in modality of different molecular samples, resulting in insufficient sensitivity to key modal information;

[0031] (3) Models trained on blood-brain barrier datasets with insufficient data have the following drawbacks: On the one hand, insufficient data prevents the model from learning the complex relationships between molecules, resulting in insufficient generalization ability; on the other hand, when the amount of data is small, the gradient of error backpropagation becomes very small, causing the weight updates of the first few layers of the network to be extremely slow, eventually leading to the gradient vanishing problem. In addition, most blood-brain barrier datasets also suffer from data imbalance, i.e., there are far more BBB+ samples than BBB- samples, which makes the existing methods less accurate in predicting compounds that cannot cross the blood-brain barrier.

[0032] To address the aforementioned issues, this invention proposes a multimodal fusion-based method for predicting the blood-brain barrier permeability of compounds. This method extracts molecular graph features, molecular text features, molecular descriptor features, and molecular 3D features, and then fuses these four types of features. This method can predict molecular properties, understand chemical behavior, and increase the accuracy of blood-brain barrier permeability prediction.

[0033] The compound blood-brain barrier permeability prediction method based on multimodal fusion includes the following steps:

[0034] S1: Represent each collected compound as a binary tuple. ,in For compounds The SMILES expression, for Blood-brain barrier permeability tags were obtained; all collected compounds were preprocessed to obtain an enhanced dataset. ;

[0035] S2: For any compound The following steps are taken: 3D molecular structure features (3DF), molecular text features (MNF), molecular descriptor features (MDF), and molecular graph features (MGF) are generated. 3DF, MNF, MDF, and MGF are then normalized to obtain normalized molecular 3D structure features (P-3DF), normalized text features (P-MNF), normalized molecular descriptor features (P-MDF), and normalized molecular graph features (P-MGF).

[0036] S3: Construct the feature fusion network BBBNet to build a multimodal fusion feature vector dataset. ;

[0037] S4: Will The model is divided into a training set and a test set. The training set is input into the Kolmogorov-Arnold model (KAN) for model training. After training, the performance of the trained KAN model is evaluated using the test set.

[0038] S5: Use the trained KAN model to predict the blood-brain barrier permeability of compounds.

[0039] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of data preprocessing in S1 are as follows:

[0040] S11: Using resampling technology to... The specific steps for performing data balancing are as follows: For any minority class sample binary... ,right Use the ADASYN method to generate a new SMILES expression. Then construct the binary pair Representing new minority class samples and adding them to middle;

[0041] S12: Yes Each compound in the sample is subjected to the SMILES enhancement operation:

[0042] (1) For any compound with an asymmetric structure ,Will The rotated SMILES expression is obtained by rotating the central atom along an axis perpendicular to the molecular plane. Construct binary pairs and will join in The rotation operation is represented as follows:

[0043]

[0044] in, Indicates to Functions that perform rotation operations. The rotation angle is randomly selected;

[0045] (2) For compounds containing two or more closed rings The following adjustments are made to all closed rings in the compound: the connection order of all atoms on the closed rings is rearranged randomly, and the SMILES expression is generated based on the adjusted atomic connection order. Construct binary pairs and will join in ;

[0046] (3) For compounds containing double bonds Randomly select a double bond and swap its stereo markings. If the original stereo markings were cis, adjust them to trans; otherwise, adjust them to cis. The adjusted markings generate the SMILES expression. Construct binary pairs and will join in .

[0047] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of S2 are as follows:

[0048] S21: For any compound The normalized molecular 3D structure features P-3DF are generated using the RDKit library. The specific steps are as follows:

[0049] (1) Normalized Euclidean distance between any two atoms in a compound , The calculation formula is as follows:

[0050] (1)

[0051] in, , Represents any two atoms in a compound; yes The three-dimensional coordinates yes 3D coordinates; This indicates a normalization operation;

[0052] (2) Any atom in a compound Normalized initial three-dimensional eigenvectors , The calculation formula is as follows:

[0053] (2)

[0054] in, It is a One-Hot encoded vector, representing Atom type; This is a vector concatenation operation; yes The three-dimensional coordinate vector represents Position in three-dimensional space;

[0055] (3) Any edge in the compound Normalized initial edge eigenvectors , The calculation formula is as follows:

[0056] (3)

[0057] in It is a One-Hot encoded vector representing an atom. and atoms The chemical bonds between them, of which , It is a binary value, representing the first... chemical bonds, and ; This represents the number of all chemical bond types. express and The Euclidean distance between them;

[0058] (4) Normalized angular eigenvectors in compounds , The calculation process is as follows:

[0059] (4)

[0060] Where, vector Indicates from atoms To the atom Direction and distance, vector Indicates from To the atom Direction and distance; The function is used to calculate two vectors. and The angle corresponding to the cosine of the included angle; It is the edge and The resulting normalized angular eigenvector.

[0061] S22: For any compound The RDKit library is used to extract information such as the physicochemical properties and topological structure of compounds to generate molecular descriptors; at the same time, a molecular graph representation of the corresponding compound is constructed; the SMILES string of each compound is concatenated with the generated molecular descriptor to construct a molecular descriptor feature set MDF; at the same time, the SMILES expression is concatenated with the corresponding molecular graph to construct a molecular graph feature set MGF.

[0062] S23: For any compound The SMILES expressions of all compounds are extracted to construct the text feature set MNF.

[0063] S24: Normalize MNF, MDF and MGF to obtain normalized text features P-MNF, normalized molecular descriptor features P-MDF and normalized molecular graph features P-MGF.

[0064] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of S3 are as follows:

[0065] S31: Construct the MolTransNet molecular transformation module, input the P-MGF of each compound into MolTransNet, and generate the molecular graph feature vector of the compound. ;

[0066] S32: Construct a molecular 3D encoder module GEMM, input the P-3DF of each compound into GEMM, and generate the molecular 3D feature vector of the compound. ;

[0067] S33: Input the P-MNF of each compound into the Transformer module to generate the text feature vector of the compound. ;

[0068] S34: Convert the P-MDF of each compound into a molecular descriptor vector. The compound , 、 , The features are input together into the attention feature fusion module Afusion for feature fusion, resulting in the feature fusion vector of the compound. According to the SMILES expression of the compound. Blood-brain barrier permeability tags and Construct a triple ;Will All compounds in Construct a multimodal fusion feature vector dataset ,in for The number of triples in the array.

[0069] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of S31 are as follows:

[0070] S311: For any atom in a compound Solve One-hop neighbor node set and the set of two-hop neighbor nodes Then calculate atomic eigenvectors , The following iterative equations (5) to (14) were used to calculate the following:

[0071] (5)

[0072] (6)

[0073] (7)

[0074] (8)

[0075] (9)

[0076] (10)

[0077] (11)

[0078] (12)

[0079] (13)

[0080] (14)

[0081] in, yes One-hop neighbor node; yes One-hop neighbor nodes, among which express The set of one-hop neighbor nodes; yes The two-hop neighbor nodes, among which express The set of two-hop neighbor nodes; , , These are One-Hot encoded vectors, representing respectively , , Atom type; , , These are One-Hot encoded vectors, representing respectively , , The degree; , , These are One-Hot encoded vectors, representing respectively... , , The number of hydrogen atoms connected; It is a dimension Atomic feature vectors; Yes conduct The atomic feature vectors obtained after rounds of iteration; It is a dimension Atomic feature vectors; It is a dimension Atomic feature vectors; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration; , , All are of size The weight matrix; It is the activation function ReLU; This represents the dot product operation, used to calculate the inner product of two vectors or matrices;

[0082] yes The two-hop neighbor nodes; yes One-hop neighbor nodes, among which express The set of one-hop neighbor nodes; yes The two-hop neighbor nodes, among which express The set of two-hop neighbor nodes; , , These are One-Hot encoded vectors, representing respectively , , Atom type; , , These are One-Hot encoded vectors, representing respectively , , The degree; , , These are One-Hot encoded vectors, representing respectively... , , The number of hydrogen atoms connected; It is a dimension ; It is a dimension Atomic feature vectors; It is a dimension ; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration;

[0083] It is a dimension Atomic feature vectors; It is a One-Hot encoded vector, representing The degree; It is a One-Hot encoded vector, representing the... The number of hydrogen atoms connected; Yes conduct The atomic feature vectors obtained after rounds of iteration;

[0084] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, each compound atom is generated in step S311. One-jump neighbor set and two-hop neighbor set The specific steps are as follows:

[0085] S3111: Find all atoms in the compound that are related to atoms Atoms directly connected by chemical bonds constitute all the above atoms. ;

[0086] S3112: For Each atom in Searching directly through chemical bonds All atoms connected ,if Or not with A direct connection will... join in .

[0087] S312: Based on Attention features for calculating the correlation between atoms within a compound The specific calculation formula is as follows:

[0088] (15)

[0089] in, It is a weight matrix generated by random initialization, with a size of ; Indicates that, except for Any atom other than the specified atom; query matrix express Mapping of atomic feature vectors in the query vector space; key matrix express Mapping in the key vector space; value matrix express Mapping in the value space, used to store ; The dimension size is The attention weight matrix represents and The relationship between atomic eigenvectors;

[0090] S313: Use and Calculate transformation features The specific calculation formula is as follows:

[0091] (16)

[0092] (17)

[0093] in, This indicates a normalization operation; Indicates a residual join operation; It is an intermediate feature; It is a feedforward neural network, used for... Perform nonlinear transformation;

[0094] S314: [This refers to] all atoms of the compound... Perform average pooling The operation yields the molecular graph feature vector of the compound. , The calculation formula is as follows:

[0095] (18)

[0096] in, It indicates the number of atoms in a compound.

[0097] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of S32 are as follows:

[0098] S321: For any atom in a compound Extraction from P-3DF of the compound Normalized initial three-dimensional eigenvectors and edge Normalized initial edge eigenvectors ,in yes Any adjacent point;

[0099] S322: Calculation 3D feature vectors , The following iterative equations (19) to (22) were used to calculate the following:

[0100] (19)

[0101] (20)

[0102] (twenty one)

[0103] (twenty two)

[0104] in, express and The Euclidean distance between them; It is a One-Hot encoded vector, representing and Chemical bonds between them; Yes conduct The edge feature vector obtained after round iteration; yes The number of one-hop neighbor nodes; express The set of one-hop neighbor nodes; Representing an edge and The resulting normalized angular feature vector; yes One-hop neighbor node;

[0105] It is a One-Hot encoded vector, representing Atom type; yes The three-dimensional coordinate vector represents Position in three-dimensional space Yes conduct The three-dimensional feature vector after round iteration; Yes conduct Edge feature vectors after round iteration; Yes conduct The three-dimensional feature vector after round iteration;

[0106] It is the size of Angle weighting matrix; It is the size of The node weighted matrix; It is a vector-level addition operation, used to add the elements of two vectors at corresponding positions; Yes conduct The three-dimensional feature vector after round iteration;

[0107] S323: Through The operation fuses the three-dimensional features of all atoms in the compound into a three-dimensional molecular vector. , The calculation formula is as follows:

[0108] (twenty three).

[0109] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of S33 are as follows:

[0110] S331: For any atom in a compound Build using n-gram model bag-of-words vectors The specific construction formula is as follows:

[0111] (twenty four)

[0112] in, Indicates the first Frequency of n-gram occurrences in SMILES expressions It is the size of the n-gram vocabulary;

[0113] S332: Generation Position encoding vector ,use generate Position-enhanced bag-of-words vectors The specific generation formula is as follows:

[0114] (25)

[0115] (26)

[0116] (27)

[0117] (28)

[0118] in, express Index in the SMILES expression; yes Dimension index, when When it is even, use generate ;when When it is an odd number, use generate ;

[0119] S333: Use Calculate text attention features The specific calculation formula is as follows:

[0120] (29)

[0121] in, It is a randomly initialized text weight matrix with a size of ; Indicates that, except for Any atoms other than those in the text query matrix; express Position-enhanced bag-of-words vector mapping in the query vector space; text key matrix express Mapping in the key vector space; text value matrix express Mapping in the value space, used to store ; and besides Used to calculate atoms and atoms The correlation of textual features between them; The dimension size is The text attention weight matrix represents and The position enhances the relationship between bag-of-words vectors; Text attention features representing the correlations between atoms within a compound;

[0122] S334: Use and Calculate text transformation features The specific calculation formula is as follows:

[0123] (30)

[0124] (31)

[0125] in, This indicates a normalization operation; It is an intermediate feature of the text; It is a feedforward neural network, used for... Perform nonlinear transformation;

[0126] S335: [This refers to] all atoms of the compound... conduct This operation yields the molecular text feature vector of the compound. , The calculation formula is as follows:

[0127] (32).

[0128] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of S34 are as follows:

[0129] S341: Will , , , The fusion process yields a vector that integrates molecular graph features, 3D molecular features, textual molecular features, and molecular descriptor features. , The specific calculation formula is as follows:

[0130] (33)

[0131] S342: Use Calculate the molecular query vector Molecular bond vector , Molecular value vector The specific formula is as follows:

[0132] (34)

[0133] (35)

[0134] (36)

[0135] in, It is the size of Modal query weight matrix; express Mapping in the query vector space; It is the size of The modal bond weight matrix; express Mapping in the key vector space; It is the size of The modal bond weight matrix; express Mapping in the value vector space; and Used for calculation Similarity between different features;

[0136] S343: Use and Calculate self-attention weights , The specific generation formula is as follows:

[0137] (37)

[0138] in, It is the self-attention score; This represents the matrix transpose operation; It is a scaling factor used to prevent The value is too large; It is a normalization function used to normalize... Convert to a probability distribution; Indicates the self-attention weights;

[0139] S344: Use and The feature fusion vector of the compound was calculated. The specific calculation formula is as follows:

[0140] (38).

[0141] As a preferred embodiment of the compound blood-brain barrier permeability prediction method based on multimodal fusion described in this invention, the specific steps of S4 are as follows:

[0142] S41: Will Divided into training set and test set ,in Used for training the KAN model. Used to evaluate the performance of the KAN model;

[0143] S42: Will any The input is fed into the KAN model for training, and the result is obtained. Blood-brain barrier permeability label prediction values ​​of one compound The specific calculation formula is as follows:

[0144] (39)

[0145] (40)

[0146] (41)

[0147] in, It is the first One B-spline basis function; It is the first A superposition function of B-spline elements, by indivual Summation yields the result; It is the first Bias terms of a KAN mapping layer; It is a linear function, representing the first... One KAN mapping layer; It is a random initial weight, representing Initial weights of the KAN mapping layer; Yes conduct The weights obtained after the first iteration represent the weights of the first iteration. The weight coefficients of the KAN mapping layer; for Quantity; yes Number of triplet pairs; The range of values ​​for is {BBB-, BBB+};

[0148] S43: Configure the cross-entropy loss function The specific formula is as follows:

[0149] (42)

[0150] in, Indicates the first The crossover loss function for each compound;

[0151] S44: Use gradient descent method to... Perform iterations: Set the change in the loss function Convergence threshold is ; calculate in each iteration When satisfied or Stop iteration and output when... This yields the trained KAN model; The following iterative equations (43)~(44) were used to calculate the following:

[0152] (43)

[0153] (44)

[0154] in, It is a fixed learning rate; Yes conduct The weights obtained after rounds of iteration; Indicates the first Blood-brain barrier permeability tags for the compounds; Indicates the first Predicted values ​​of blood-brain barrier permeability tags for each compound; express right The gradient; It is a dynamically adjustable value, set according to the size of the data.

[0155] S45: Use The performance of the trained KAN model is evaluated.

[0156] Compared with the prior art: (1) The present invention designs a data preprocessing method, which first uses resampling technology to process the data. Data balancing was performed to resolve the data imbalance problem. Subsequently, the SMILES augmentation operation was used to effectively address the vanishing gradient and underfitting issues that occurred when the amount of data was insufficient, further improving the training performance of the model.

[0157] (2) The present invention designs a feature fusion network BBBNet to extract four different features: molecular text features, molecular graph features, molecular three-dimensional features and molecular descriptor features, which solves the problem of not making full use of molecular information in traditional methods and improves the comprehensive characterization of complex properties of compounds.

[0158] (3) The present invention also designed a feature fusion module Afusion based on attention mechanism; this module can efficiently fuse features of four different modalities and enable the model to learn the weights of each modality through attention mechanism, thereby solving the current problem of insufficient multimodal feature fusion and improving the performance of the model in the blood-brain barrier permeability prediction task. Attached Figure Description

[0159] Figure 1 This is an overall flowchart of the present invention.

[0160] Figure 2 Flowchart for generating SMILES expression features;

[0161] Figure 3 Here is a structural diagram of the feature fusion network BBBNet;

[0162] Figure 4 A flowchart for predicting blood-brain barrier permeability tags. Detailed Implementation

[0163] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0164] Example

[0165] This application provides a method that can effectively solve the problems mentioned above. The following will describe in detail how to implement a compound blood-brain barrier permeability prediction method based on multimodal fusion, using several embodiments.

[0166] Figure 1 A flowchart illustrating a compound blood-brain barrier permeability prediction method based on multimodal fusion is shown, including:

[0167] S1: Represent each collected compound as a binary tuple. ,in The compound The SMILES expression, for Blood-brain barrier permeability tags; data preprocessing of all compounds to obtain an enhanced dataset. ;

[0168] In this embodiment of the application, in order to realize the compound blood-brain barrier permeability prediction method based on multimodal fusion, a data preprocessing flow is constructed, and the execution flow of the data preprocessing is as follows:

[0169] S11: Using resampling technology to... The specific steps for performing data balancing are as follows: For any minority class sample binary... ,right Use the ADASYN method to generate a new SMILES expression. Then construct the binary pair Representing new minority class samples and adding them to middle;

[0170] Use the ADASYN method to generate a new minority class SMILES expression. The specific process is as follows:

[0171] S111: First calculate the minority class samples Importance Then according to Calculate the number of new samples The formula is as follows:

[0172]

[0173]

[0174] in, express The number of most recent majority class samples, The maximum number of neighbors in the minority class samples. This represents the number of minority samples generated.

[0175] S112: For each minority class sample Utilizing its neighboring majority class samples Generate new minority class samples The generation process is as follows:

[0176]

[0177] in, It is in Random numbers in an interval are used to... New minority class samples are generated between the minority class and its neighboring majority class samples. This process is performed iteratively until the required number of minority class samples are generated.

[0178] For example, this invention uses the B3DB dataset from McMaster University in Canada, which contains 7807 samples, including 4956 BBB+ samples and 2851 BBB- samples. The minority class samples in this dataset refer to BBB- samples. Assume a BBB- sample... Using its neighboring BBB+ samples Assuming the number of recent BBB+ samples Maximum number of neighbors ,Depend on Total number of BBB samples generated Set as First, calculate the importance of the BBB-samples. ;get Number of generated BBB samples ; It is a value in A random number within a range, taken in the current example. ;Depend on Generate new BBB-samples The process is iterated continuously using the method described above until the first generation is generated. There are 10 BBB-compound samples, bringing the total number of BBB-compound samples to 4900.

[0179] S12: Apply the SMILES enhancement operation to each compound in BBBD:

[0180] (1) For any compound with an asymmetric structure ,Will The rotated SMILES expression is obtained by rotating the central atom along an axis perpendicular to the molecular plane. Construct binary pairs and will join in The rotation operation is represented as follows:

[0181]

[0182] in, Indicates to Functions that perform rotation operations. The rotation angle is randomly selected;

[0183] For example, the SMILES expression for 1,2-dichloroethylene is: It is a compound with an asymmetric structure. The SMILES-enhanced operation was applied to 1,2-dichloroethylene, specifically by selecting an axis perpendicular to the plane of 1,2-dichloroethylene as the rotation axis and choosing a rotation angle... Perform a rotation operation. After rotation, the expression for the rotated SMILES is: .

[0184] (2) For compounds containing two or more closed rings The following adjustments are made to all closed rings in the compound: the connection order of all atoms on the closed rings is rearranged randomly, and the SMILES expression is generated based on the adjusted atomic connection order. Construct binary pairs and will join in ;

[0185] For example, the expression for naphthalene's SMILES is: naphthalene is a compound with two closed ring structures. Applying the SMILES enhancement operation to naphthalene involves adjusting the two closed rings by rearranging the connection order of all atoms on the closed rings. The resulting SMILES expression is: .

[0186] (3) For compounds containing double bonds Randomly select a double bond and swap its stereo notation. If the original stereo notation was cis, adjust it to trans; if the original stereo notation was trans, adjust it to cis. The adjusted notation generates the SMILES expression. Construct binary pairs and will join in

[0187] For example, the SMILES expression for olefins is: It is a compound containing a double bond. The SMILES enhancement operation is applied to alkenes, specifically: randomly selecting stereolabels for double bond exchange; the first... Indicates that the original solid mark is in the cis form; the second The stereo notation is trans; swapping the stereo notation of the double bonds generates new SMILES. The expression is: The cis and trans arrangements reflect the relative positions of the substituents on both sides of the double bond in three-dimensional space.

[0188] S2: For any compound The following steps are taken: 3D molecular structure features (3DF), molecular text features (MNF), molecular descriptor features (MDF), and molecular graph features (MGF) are generated. 3DF, MNF, MDF, and MGF are then normalized to obtain normalized molecular 3D structure features (P-3DF), normalized text features (P-MNF), normalized molecular descriptor features (P-MDF), and normalized molecular graph features (P-MGF).

[0189] In this embodiment of the application, in order to realize the compound blood-brain barrier permeability prediction method based on multimodal fusion, a flowchart for generating SMILES expression features was constructed; Figure 2 This paper describes the steps involved in generating SMILES expression features, mainly including: generating molecular 3D structural features (3DF), molecular text features (MNF), molecular descriptor features (MDF), and molecular graph features (MGF); and normalizing 3DF, MNF, MDF, and MGF. The specific steps of feature generation are as follows:

[0190] S21: For any compound The RDKit library is used to generate molecular 3D structural features, which are then normalized to obtain P-3DF. The specific steps are as follows:

[0191] (1) Normalized Euclidean distance between any two atoms in a compound , The calculation formula is as follows:

[0192] (1)

[0193] in, , Represents any two atoms in a compound; yes The three-dimensional coordinates yes 3D coordinates; This indicates a normalization operation;

[0194] (2) Any atom in a compound Normalized initial three-dimensional eigenvectors , The calculation formula is as follows:

[0195] (2)

[0196] in, It is a One-Hot encoded vector, representing Atom type; This is a vector concatenation operation; yes The three-dimensional coordinate vector represents Position in three-dimensional space;

[0197] (3) Any edge in the compound Normalized initial edge eigenvectors , The calculation formula is as follows:

[0198] (3)

[0199] in It is a One-Hot encoded vector representing an atom. and atoms The chemical bonds between them, of which , It is a binary value, representing the first... chemical bonds, and ; This represents the number of all chemical bond types. express and The Euclidean distance between them;

[0200] (4) Normalized angular eigenvectors in compounds , The calculation process is as follows:

[0201] (4)

[0202] Where, vector Indicates from atoms To the atom Direction and distance, vector Indicates from To the atom Direction and distance; The function is used to calculate two vectors. and The angle corresponding to the cosine of the included angle; It is the edge and The resulting normalized angular eigenvector.

[0203] S22: For any compound The RDKit library is used to extract information such as the physicochemical properties and topological structure of compounds to generate molecular descriptors; at the same time, a molecular graph representation of the corresponding compound is constructed; the SMILES string of each compound is concatenated with the generated molecular descriptor to construct a molecular descriptor feature set MDF; at the same time, the SMILES expression is concatenated with the corresponding molecular graph to construct a molecular graph feature set MGF.

[0204] S23: For any compound The SMILES expressions of all compounds are extracted to construct the text feature set MNF.

[0205] S24: Normalize MNF, MDF and MGF to obtain normalized text features P-MNF, normalized molecular descriptor features P-MDF and normalized molecular graph features P-MGF.

[0206] For example, the standard deviation of the molecular descriptor feature set MDF is normalized using a normalization method, as shown in the following formula:

[0207]

[0208] in, For a sample in the molecular descriptor feature set MDF. This represents the mean of the feature set of molecular descriptors. The standard deviation of the feature set of molecular descriptors. This is the normalized value.

[0209] S3: Construct the feature fusion network BBBNet to build a multimodal fusion feature vector dataset. ;

[0210] In the embodiments of this application, in order to realize a multimodal fusion method for predicting the blood-brain barrier permeability of compounds, a feature fusion network BBBNet was designed. The execution flow of this model is as follows: Figure 3 As shown, the BBBNet building blocks perform feature extraction, and the attention-based feature fusion module Afusion is designed to perform feature fusion, finally resulting in a multimodal fused feature vector dataset. ;

[0211] The feature fusion network BBBNet consists of four modules: MolTransNet, GEMM, Transformer, and an attention-based feature fusion module called Afusion. The MolTransNet module is used to extract molecular graph feature vectors of compounds. The GEMM module is used to extract the three-dimensional feature vectors of compounds. The Transformer module is used to extract textual feature vectors of compounds. The attention feature fusion module Afusion is used to generate feature fusion vectors for compounds. ;

[0212] S31: Construct the MolTransNet molecular transformation module, input the P-MGF of each compound into the MolTransNet module, and generate molecular graph feature vectors. ;

[0213] The construction process of the MolTransNet module is as follows:

[0214] Using P-MGF as input to the MolTransNet module, the module first extracts the initial feature vector for each atom from P-MGF. Then, through multiple rounds of information passing and neighbor feature aggregation, the atomic feature vector of each atom is progressively updated. The module further processes these features using Transformer to finally obtain the molecular graph feature vector of the compound. ;

[0215] Finally, the molecular graph feature vector of the compound is obtained. As output.

[0216] S311: For any atom in a compound Solve One-hop neighbor node set and the set of two-hop neighbor nodes Then calculate atomic eigenvectors , The following iterative equations (5) to (14) were used to calculate the following:

[0217] (5)

[0218] (6)

[0219] (7)

[0220] (8)

[0221] (9)

[0222] (10)

[0223] (11)

[0224] (12)

[0225] (13)

[0226] (14)

[0227] in, yes One-hop neighbor node; yes One-hop neighbor nodes, among which express The set of one-hop neighbor nodes; yes The two-hop neighbor nodes, among which express The set of two-hop neighbor nodes; , , These are One-Hot encoded vectors, representing respectively , , Atom type; , , These are One-Hot encoded vectors, representing respectively , , The degree; , , These are One-Hot encoded vectors, representing respectively... , , The number of hydrogen atoms connected; It is a dimension Atomic feature vectors; Yes conduct The atomic feature vectors obtained after rounds of iteration; It is a dimension Atomic feature vectors; It is a dimension Atomic feature vectors; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration; , , All are of size The weight matrix; It is the activation function ReLU; This represents the dot product operation, used to calculate the inner product of two vectors or matrices;

[0228] yes The two-hop neighbor nodes; yes One-hop neighbor nodes, among which express The set of one-hop neighbor nodes; yes The two-hop neighbor nodes, among which express The set of two-hop neighbor nodes; , , These are One-Hot encoded vectors, representing respectively , , Atom type; , , These are One-Hot encoded vectors, representing respectively , , The degree; , , These are One-Hot encoded vectors, representing respectively... , , The number of hydrogen atoms connected; It is a dimension ; It is a dimension Atomic feature vectors; It is a dimension ; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration;

[0229] It is a dimension Atomic feature vectors; It is a One-Hot encoded vector, representing The degree; It is a One-Hot encoded vector, representing the... The number of hydrogen atoms connected; Yes conduct The atomic feature vectors obtained after rounds of iteration;

[0230] For example, The size is , The number of atomic nodes in each compound; The 66-bit information includes: 44 bits representing a high-dimensional vector generated by One-Hot encoding mapping of the atom type, 13 bits representing the atom degree, and 10 bits representing the number of hydrogen atoms connected to the atom.

[0231] calculate Need to be done In each update round, the feature vectors of all atoms in the compound are aggregated with the feature vectors of all their one-hop neighbors and two-hop neighbors to form a complete update round; the specific steps are as follows:

[0232] First, calculate the number of atoms in the compound. , , Next, feature aggregation is performed: and and Perform feature aggregation, and aggregate into ;at the same time and and Aggregate into Similarly, and and Aggregate into Once all atoms in a compound have completed one characteristic polymerization cycle, an update cycle ends; then, the above characteristic polymerization process is repeated. In each round, the update is based on the feature vector from the previous round, resulting in... .

[0233] One jump to the neighbor collection and two-hop neighbor set The specific construction process is as follows:

[0234] S3111: Find all atoms in the compound that are related to atoms Atoms directly connected by chemical bonds constitute all the above atoms. ;

[0235] S3112: For Each atom in Searching directly through chemical bonds All atoms connected ,if Or not with A direct connection will... join in .

[0236] For example, suppose the SMILES expression of a compound is CC(=O)O, and the compound contains the following atoms: the first carbon atom The second carbon atom oxygen atom in carbonyl group and the oxygen atom in the hydroxyl group. These atoms are connected by chemical bonds, forming an undirected graph, where each atom is a node and the chemical bonds are edges in the graph. , , , , , , , The specific calculation process is as follows:

[0237] First, calculate the set of one-hop neighbor nodes for each atomic node. and Directly connected, therefore . Directly connected via chemical bonds , and ,therefore .for It is directly related to Connected, therefore .for It is also directly related to Connected, therefore .

[0238] Next, we calculate the set of two-hop neighbors for each atom, that is, the set of other atoms that can reach that atom through two chemical bonds. For Its set of one-hop neighbor nodes is ;exist A collection of neighbors in a jump Remove those that already belong or with After directly connecting atoms, we get ;for Its set of one-hop neighbor nodes is In the direct connection atoms of each neighboring node, remove those that already belong to... or with After directly connecting atoms, the result is an empty set, therefore .for One of its neighbors is ;exist In the set of neighbors of the jump Remove those that already belong or with After directly connecting atoms, we get Similarly, for Its set of one-hop neighbor nodes is After calculation, we get .

[0239] Finally, the set of neighbor nodes for each atom is obtained as follows: ; ; ; ; ; ; ; .

[0240] S312: Based on Attention features for calculating the correlation between atoms within a compound The specific calculation formula is as follows:

[0241] (15)

[0242] in, It is a weight matrix generated by random initialization, with a size of ; Indicates that, except for Any atom other than the specified atom; query matrix express Mapping of atomic feature vectors in the query vector space; key matrix express Mapping in the key vector space; value matrix express Mapping in the value space, used to store ; The dimension size is The attention weight matrix represents and The relationship between atomic eigenvectors;

[0243] S313: Use and Calculate transformation features The specific calculation formula is as follows:

[0244] (16)

[0245] (17)

[0246] in, This indicates a normalization operation; Indicates a residual join operation; It is an intermediate feature; It is a feedforward neural network, used for... Perform nonlinear transformation;

[0247] S314: [This refers to] all atoms of the compound... Perform average pooling The operation yields the molecular graph feature vector of the compound. , The calculation formula is as follows:

[0248] (18)

[0249] in, It indicates the number of atoms in a compound.

[0250] S32: Construct a molecular 3D encoder module GEMM, input the P-3DF of each compound into GEMM, and generate the molecular 3D feature vector of the compound. ;

[0251] In the embodiments of this application, in order to realize a multimodal fusion method for predicting the blood-brain barrier permeability of compounds, a molecular three-dimensional encoder module (GEMM) was designed to generate molecular three-dimensional feature vectors of compounds. The specific process of constructing GEMM is as follows:

[0252] Using P-3DF as the input to the molecular 3D encoder module GEMM, the 3D feature vectors and edge feature vectors of each atom are first extracted from the P-3DF. Then, through multiple rounds of information passing and neighbor feature aggregation, the 3D feature vectors of each atom are progressively optimized. Finally, pooling operations are used to aggregate the 3D features of all atoms in the compound to obtain the molecular 3D vector. ;

[0253] S321: For any atom in a compound Extraction from P-3DF of the compound Normalized initial three-dimensional eigenvectors and edge Normalized initial edge eigenvectors ,in yes Any adjacent point;

[0254] S322: Calculation 3D feature vectors , The following iterative equations (19) to (22) were used to calculate the following:

[0255] (19)

[0256] (20)

[0257] (twenty one)

[0258] (twenty two)

[0259] in, express and The Euclidean distance between them; It is a One-Hot encoded vector, representing and Chemical bonds between them; Yes conduct The edge feature vector obtained after round iteration; yes The number of one-hop neighbor nodes; express The set of one-hop neighbor nodes; Representing an edge and The resulting normalized angular feature vector; yes One-hop neighbor node;

[0260] It is a One-Hot encoded vector, representing Atom type; yes The three-dimensional coordinate vector represents Position in three-dimensional space Yes conduct The three-dimensional feature vector after round iteration; Yes conduct Edge feature vectors after round iteration; Yes conduct The three-dimensional feature vector after round iteration;

[0261] It is the size of Angle weighting matrix; It is the size of The node weighted matrix; It is a vector-level addition operation, used to add the elements of two vectors at corresponding positions; Yes conduct The three-dimensional feature vector after round iteration;

[0262] One iteration consists of two steps, as follows:

[0263] (1) Use calculate After the edge feature vectors of all edges in the compound have been updated, proceed to the next stage.

[0264] (2) Use and besides calculate After the three-dimensional feature vectors of all atoms in the compound are updated, one iteration is completed.

[0265] (3) Repeat (1) and (2) iteratively. After the round, I received .

[0266] S323: Through The operation fuses the three-dimensional features of all atoms in the compound into a three-dimensional molecular vector. , The calculation formula is as follows:

[0267] (twenty three).

[0268] S33: Input the P-MNF of each compound into the Transformer to generate the text feature vector of the compound. ;

[0269] The specific steps for inputting P-MNF into the Transformer for each compound are as follows:

[0270] S331: For any atom in a compound Build using n-gram model bag-of-words vectors The specific construction formula is as follows:

[0271] (twenty four)

[0272] in, Indicates the first Frequency of n-gram occurrences in SMILES expressions It is the size of the n-gram vocabulary;

[0273] S332: Generation Position encoding vector ,use generate Position-enhanced bag-of-words vectors The specific generation formula is as follows:

[0274] (25)

[0275] (26)

[0276] (27)

[0277] (28)

[0278] in, express Index in the SMILES expression; yes Dimension index, when When it is even, use generate ;when When it is an odd number, use generate ;

[0279] S333: Use Calculate text attention features The specific calculation formula is as follows:

[0280] (29)

[0281] in, It is a randomly initialized text weight matrix with a size of ; Indicates that, except for Any atoms other than those in the text query matrix; express Position-enhanced bag-of-words vector mapping in the query vector space; text key matrix express Mapping in the key vector space; text value matrix express Mapping in the value space, used to store ; and besides Used to calculate atoms and atoms The correlation of textual features between them; The dimension size is The text attention weight matrix represents and The position enhances the relationship between bag-of-words vectors; Text attention features representing the correlations between atoms within a compound;

[0282] S334: Use and Calculate text transformation features The specific calculation formula is as follows:

[0283] (30)

[0284] (31)

[0285] in, This indicates a normalization operation; It is an intermediate feature of the text; It is a feedforward neural network, used for... Perform nonlinear transformation;

[0286] S335: [This refers to] all atoms of the compound... conduct This operation yields the molecular text feature vector of the compound. , The calculation formula is as follows:

[0287] (32).

[0288] S34: Convert the P-MDF of each compound into a molecular descriptor vector. The compound , 、 , The features are input together into the attention feature fusion module Afusion for feature fusion, resulting in the feature fusion vector of the compound. According to the SMILES expression of the compound. Blood-brain barrier permeability tags and Construct a triple ;Will All compounds in Construct a multimodal fusion feature vector dataset ,in for The number of triples in the array.

[0289] For example, suppose P-MDF contains the following characteristics: molecular weight (MW), number of hydrogen bond donors, number of hydrogen bond acceptors, LogP, and number of rings. The specific properties of this compound are: molecular weight of... The number of hydrogen bond donors is The number of hydrogen bond acceptors is LogP is The number of rings is The P-MDF of this compound was converted to .

[0290] This invention designs an attention-based feature fusion module, Afusion, for feature fusion. Afusion learns the relative importance of different features through an attention mechanism, thereby effectively fusing information from various modalities and improving the accuracy of blood-brain barrier permeability prediction. The construction process of the attention-based feature fusion module Afusion is as follows:

[0291] S341: Will , , , The fusion process yields a vector that integrates molecular graph features, 3D molecular features, textual molecular features, and molecular descriptor features. , The specific calculation formula is as follows:

[0292] (33)

[0293] S342: Use Calculate the molecular query vector Molecular bond vector , Molecular value vector The specific formula is as follows:

[0294] (34)

[0295] (35)

[0296] (36)

[0297] in, It is the size of Modal query weight matrix; express Mapping in the query vector space; It is the size of The modal bond weight matrix; express Mapping in the key vector space; It is the size of The modal bond weight matrix; express Mapping in the value vector space; and Used for calculation Similarity between different features;

[0298] For example, , , The dimension is set to ,and The value of each matrix element follows the mean. Standard deviation is It follows a normal distribution.

[0299] S343: Use and Calculate self-attention weights , The specific generation formula is as follows:

[0300] (37)

[0301] in, It is the self-attention score; This represents the matrix transpose operation; It is a scaling factor used to prevent The value is too large; It is a normalization function used to normalize... Convert to a probability distribution; Indicates the self-attention weights;

[0302] For example, the scaling factor Set as Used to prevent Gradient explosion or numerical instability caused by excessively large values.

[0303] S344: Use and The feature fusion vector of the compound was calculated. The specific calculation formula is as follows:

[0304] (38).

[0305] S4: Will The model is divided into a training set and a test set. The training set is input into the Kolmogorov-Arnold model (KAN) for model training. After training, the performance of the trained KAN model is evaluated using the test set.

[0306] Blood-brain barrier permeability label prediction using the Kolmogorov-Arnold (KAN) model; the prediction process for blood-brain barrier permeability labels is as follows: Figure 4 As shown. The specific process of KAN's model training and performance evaluation is as follows:

[0307] S41: Will Divided into training set and test set ,in Used for training the KAN model. Used to evaluate the performance of the KAN model;

[0308] For example, The number of samples containing BBB+ and BBB- are respectively and ;Will Divided into 70%:30% , Training set There is Number of samples, including the number of triples ,and It contains 7,064 BBB+ samples and One BBB-sample; test set Include 1 sample, containing One BBB+ sample and One BBB-sample.

[0309] S42: Will any The input is fed into the KAN model for training, and the result is obtained. Blood-brain barrier permeability label prediction values ​​of one compound The specific calculation formula is as follows:

[0310] (39)

[0311] (40)

[0312] (41)

[0313] in, It is the first One B-spline basis function; It is the first A superposition function of B-spline elements, by indivual Summation yields the result; It is the first Bias terms of a KAN mapping layer; It is a linear function, representing the first... One KAN mapping layer; It is a random initial weight, representing Initial weights of the KAN mapping layer; Yes conduct The weights obtained after the first iteration represent the weights of the first iteration. The weight coefficients of the KAN mapping layer; for Quantity; yes Number of triplet pairs; The range of values ​​for is {BBB-, BBB+};

[0314] S43: Configure the cross-entropy loss function The specific formula is as follows:

[0315] (42)

[0316] in, Indicates the first The crossover loss function for each compound;

[0317] S44: Use gradient descent method to... Perform iterations: Set the change in the loss function Convergence threshold is ; calculate in each iteration When satisfied or Stop iteration and output when... This yields the trained KAN model; The following iterative equations (43)~(44) were used to calculate the following:

[0318] (43)

[0319] (44)

[0320] in, It is a fixed learning rate; Yes conduct The weights obtained after rounds of iteration; Indicates the first Blood-brain barrier permeability tags for the compounds; Indicates the first Predicted values ​​of blood-brain barrier permeability tags for each compound; express right The gradient; It is a dynamically adjustable value, set according to the size of the data.

[0321] S45: Use The performance of the trained KAN model is evaluated.

[0322] For example, the initial weights of the KAN model Set it to 0.5; then, put The input is fed into the KAN model for training; a fixed learning rate is set to ensure model convergence. And set the change in the loss function. Convergence threshold The value is 0.01. During training, gradient descent is used to adjust the weight parameters of the KAN mapping layer. Perform iterative updates, calculating in each iteration When satisfied Or reaching the maximum number of iterations. Stop the iteration and output the final optimized version. .

[0323] S5: Use the trained KAN model to predict the blood-brain barrier permeability of compounds.

[0324] The KAN model takes any real-world compound as input and outputs a predicted value for the compound's blood-brain barrier permeability. .

[0325] Based on the above, this invention designs a multimodal fusion-based method for predicting the blood-brain barrier permeability of compounds, aiming to improve the accuracy and generalization ability of the model in predicting the blood-brain barrier permeability of compounds. To address the feature extraction problem of compounds, this invention further designs a feature fusion network, BBBNet, which extracts four types of features and uses the Afusion module for feature fusion to construct a multimodal fusion feature vector dataset. This invention addresses the problem of incomplete molecular characterization in traditional methods. It utilizes the KAN model to predict blood-brain barrier permeability, obtaining predicted labels. This method effectively extracts and fuses multiple modal features, significantly improving the performance of the blood-brain barrier permeability prediction task.

[0326] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion, characterized in that... To consider the impact of class imbalance on prediction results and improve the accuracy of predicting blood-brain barrier permeability using a feature fusion network, this method includes the following steps: S1: Represent each collected compound as a binary tuple. ,in The compound The SMILES expression, for Blood-brain barrier permeability tags were obtained; all collected compounds were preprocessed to obtain an enhanced dataset. ; S2: For any compound The following steps are taken: 3D molecular structure features (3DF), molecular text features (MNF), molecular descriptor features (MDF), and molecular graph features (MGF) are generated. 3DF, MNF, MDF, and MGF are then normalized to obtain normalized molecular 3D structure features (P-3DF), normalized text features (P-MNF), normalized molecular descriptor features (P-MDF), and normalized molecular graph features (P-MGF). The normalized molecular three-dimensional structural feature P-3DF includes the following information: (1) Normalized Euclidean distance between any two atoms in a compound , The calculation formula is as follows: (1) in, , Represents any two atoms in a compound; yes The three-dimensional coordinates yes 3D coordinates; This indicates a normalization operation; (2) Any atom in a compound Normalized initial three-dimensional eigenvectors , The calculation formula is as follows: (2) in, It is a One-Hot encoded vector, representing Atom type; This is a vector concatenation operation; yes The three-dimensional coordinate vector represents Position in three-dimensional space; (3) Any edge in the compound Normalized initial edge eigenvectors , The calculation formula is as follows: (3) in It is a One-Hot encoded vector representing an atom. and atoms The chemical bonds between them, of which , It is a binary value, representing the first... chemical bonds, and ; This represents the number of all chemical bond types. express and The Euclidean distance between them; (4) Normalized angular eigenvectors in compounds , The calculation process is as follows: (4) Where, vector Indicates from atoms To the atom Direction and distance, vector Indicates from To the atom Direction and distance; The function is used to calculate two vectors. and The angle corresponding to the cosine of the included angle; It is the edge and The resulting normalized angular feature vector; S3: Construct the feature fusion network BBBNet to build a multimodal fusion feature vector dataset. ; The specific steps for constructing the feature fusion network BBBNet include: S31: Construct the MolTransNet molecular transformation module, input the P-MGF of each compound into MolTransNet, and generate the molecular graph feature vector of the compound. ; S32: Construct a molecular 3D encoder module GEMM, input the P-3DF of each compound into GEMM, and generate the molecular 3D feature vector of the compound. ; S33: Input the P-MNF of each compound into the Transformer module to generate the text feature vector of the compound. ; S34: Convert the P-MDF of each compound into a molecular descriptor vector. The compound , 、 , The features are input together into the attention feature fusion module Afusion for feature fusion, resulting in the feature fusion vector of the compound. According to the SMILES expression of the compound. Blood-brain barrier permeability tags and Construct a triple ;Will All compounds in Construct a multimodal fusion feature vector dataset ,in for The number of triples in the data; S4: Will The model is divided into a training set and a test set. The training set is input into the Kolmogorov-Arnold model (KAN) for model training. After training, the performance of the trained KAN model is evaluated using the test set. S5: Use the trained KAN model to predict the blood-brain barrier permeability of compounds.

2. The method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion according to claim 1, characterized in that, The data preprocessing in S1 includes: S11: Using resampling technology to... The specific steps for performing data balancing are as follows: For any minority class sample binary... ,right Use the ADASYN method to generate a new SMILES expression. Then construct the binary pair Representing new minority class samples and adding them to middle; S12: Apply the SMILES enhancement operation to each compound in BBBD: (1) For any compound with an asymmetric structure ,Will The rotated SMILES expression is obtained by rotating the central atom along an axis perpendicular to the molecular plane. Construct binary pairs and will join in The rotation operation is represented as follows: in, Indicates to Functions that perform rotation operations. The rotation angle is randomly selected; (2) For compounds containing two or more closed rings The following adjustments are made to all closed rings in the compound: the connection order of all atoms on the closed rings is rearranged, and the SMILES expression is generated based on the adjusted atomic connection order. Construct binary pairs and will join in ; (3) For compounds containing double bonds Randomly select a double bond and swap its stereo notation; if the original stereo notation was cis, adjust it to trans; if the original stereo notation was trans, adjust it to cis; the adjusted notation generates the SMILES expression. Construct binary pairs and will join in .

3. The method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion according to claim 1, characterized in that, The specific process of constructing the molecular transformation module MolTransNet in S31 includes: S311: For any atom in a compound Solve One-hop neighbor node set and the set of two-hop neighbor nodes Then calculate atomic eigenvectors , The following iterative equations (5) to (14) were used to calculate the following: (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) in, yes One-hop neighbor node; yes One-hop neighbor nodes, among which express The set of one-hop neighbor nodes; yes The two-hop neighbor nodes, among which express The set of two-hop neighbor nodes; , , These are One-Hot encoded vectors, representing respectively , , Atom type; , , These are One-Hot encoded vectors, representing respectively , , The degree; , , These are One-Hot encoded vectors, representing respectively... , , The number of hydrogen atoms connected; It is a dimension Atomic feature vectors; Yes conduct The atomic feature vectors obtained after rounds of iteration; It is a dimension Atomic feature vectors; It is a dimension Atomic feature vectors; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration; , , All are of size The weight matrix; It is the activation function ReLU; This represents the dot product operation, used to calculate the inner product of two vectors or matrices; yes The two-hop neighbor nodes; yes One-hop neighbor nodes, among which express The set of one-hop neighbor nodes; yes The two-hop neighbor nodes, among which express The set of two-hop neighbor nodes; , , These are One-Hot encoded vectors, representing respectively , , Atom type; , , These are One-Hot encoded vectors, representing respectively , , The degree; , , These are One-Hot encoded vectors, representing respectively... , , The number of hydrogen atoms connected; It is a dimension ; It is a dimension Atomic feature vectors; It is a dimension ; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration; Yes conduct The atomic feature vectors obtained after rounds of iteration; It is a dimension Atomic feature vectors; It is a One-Hot encoded vector, representing The degree; It is a One-Hot encoded vector, representing the... The number of hydrogen atoms connected; Yes conduct The atomic feature vectors obtained after rounds of iteration; S312: Based on Attention features for calculating the correlation between atoms within a compound The specific calculation formula is as follows: (15) in, It is a weight matrix generated by random initialization, with a size of ; Indicates that, except for Any atom other than the specified atom; query matrix express Mapping of atomic feature vectors in the query vector space; key matrix express Mapping in the key vector space; value matrix express Mapping in the value space, used to store ; The dimension size is The attention weight matrix represents and The relationship between atomic eigenvectors; S313: Use and Calculate transformation features The specific calculation formula is as follows: (16) (17) in, This indicates a normalization operation; Indicates a residual join operation; It is an intermediate feature; It is a feedforward neural network, used for... Perform nonlinear transformation; S314: [This refers to] all atoms of the compound... Perform average pooling The operation yields the molecular graph feature vector of the compound. , The calculation formula is as follows: (18) in, It indicates the number of atoms in a compound.

4. The method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion according to claim 3, characterized in that, Any atom in the compound generated in S311 One-hop neighbor node set and the set of two-hop neighbor nodes The specific process includes: S3111: Find all atoms in the compound that are related to atoms Atoms directly connected by chemical bonds constitute all the above atoms. ; S3112: For Each atom in Searching directly through chemical bonds All atoms connected ,if Or not with A direct connection will... join in .

5. The method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion according to claim 1, characterized in that, The specific process of constructing the molecular three-dimensional encoder module GEMM in S32 includes: S321: For any atom in a compound Extraction from P-3DF of the compound Normalized initial three-dimensional eigenvectors and edge Normalized initial edge eigenvectors ,in yes Any adjacent point; S322: Calculation 3D feature vectors , The following iterative equations (19) to (22) were used to calculate the following: (19) (20) (21) (22) in, express and The Euclidean distance between them; It is a One-Hot encoded vector, representing and Chemical bonds between them; Yes conduct The edge feature vector obtained after round iteration; yes The number of one-hop neighbor nodes; express The set of one-hop neighbor nodes; Representing an edge and The resulting normalized angular feature vector; yes One-hop neighbor node; It is a One-Hot encoded vector, representing Atom type; yes The three-dimensional coordinate vector represents Position in three-dimensional space Yes conduct The three-dimensional feature vector after round iteration; Yes conduct Edge feature vectors after round iteration; Yes conduct The three-dimensional feature vector after round iteration; It is the size of Angle weighting matrix; It is the size of The node weighted matrix; It is a vector-level addition operation, used to add the elements of two vectors at corresponding positions; Yes conduct The three-dimensional feature vector after round iteration; S323: Through The operation fuses the three-dimensional features of all atoms in the compound into a three-dimensional molecular vector. , The calculation formula is as follows: (23)。 6. The method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion according to claim 1, characterized in that, In step S33, the P-MNF of each compound is input into the Transformer module to generate the text feature vector of the compound. The specific process includes: S331: For any atom in a compound Build using n-gram model bag-of-words vectors The specific construction formula is as follows: (24) in, Indicates the first Frequency of n-gram occurrences in SMILES expressions It is the size of the n-gram vocabulary; S332: Generation Position encoding vector ,use generate Position-enhanced bag-of-words vectors The specific generation formula is as follows: (25) (26) (27) (28) in, express Index in the SMILES expression; yes Dimension index, when When it is even, use generate ;when When it is an odd number, use generate ; S333: Use Calculate text attention features The specific calculation formula is as follows: (29) in, It is a randomly initialized text weight matrix with a size of ; Indicates that, except for Any atoms other than those in the text query matrix; express Position-enhanced bag-of-words vector mapping in the query vector space; text key matrix express Mapping in the key vector space; text value matrix express Mapping in the value space, used to store ; and besides Used to calculate atoms and atoms The correlation of textual features between them; The dimension size is The text attention weight matrix represents and The position enhances the relationship between bag-of-words vectors; Text attention features representing the correlations between atoms within a compound; S334: Use and Calculate text transformation features The specific calculation formula is as follows: (30) (31) in, This indicates a normalization operation; It is an intermediate feature of the text; It is a feedforward neural network, used for... Perform nonlinear transformation; S335: [This refers to] all atoms of the compound... conduct This operation yields the molecular text feature vector of the compound. , The calculation formula is as follows: (32)。 7. The method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion according to claim 1, characterized in that, The specific process of feature fusion performed by the attention feature fusion module Afusion in S34 includes: S341: Will , , , The fusion process yields a vector that integrates molecular graph features, 3D molecular features, textual molecular features, and molecular descriptor features. , The specific calculation formula is as follows: (33) S342: Use Calculate the molecular query vector Molecular bond vector Molecular value vector The specific formula is as follows: (34) (35) (36) in, It is the size of Modal query weight matrix; express Mapping in the query vector space; It is the size of The modal bond weight matrix; express Mapping in the key vector space; It is the size of The modal bond weight matrix; express Mapping in the value vector space; and Used for calculation Similarity between different features; S343: Use and Calculate self-attention weights , The specific generation formula is as follows: (37) in, It is the self-attention score; This represents the matrix transpose operation; It is a scaling factor used to prevent The value is too large; It is a normalization function used to normalize... Convert to a probability distribution; Indicates the self-attention weights; S344: Use and The feature fusion vector of the compound was calculated. The specific calculation formula is as follows: (38)。 8. The method for predicting the blood-brain barrier permeability of compounds based on multimodal fusion according to claim 1, characterized in that, The S4 will The dataset is divided into a training set and a test set. The training set is then input into the Kolmogorov-Arnold model KAN. After training is complete, the performance of the trained KAN model is evaluated using a test set. The specific steps for training and evaluating the performance of a model include: S41: Will Divided into training set and test set ,in Used for training the KAN model. Used to evaluate the performance of the KAN model; S42: Will any The input is fed into the KAN model for training, and the result is obtained. Blood-brain barrier permeability label prediction values ​​of one compound The specific calculation formula is as follows: (39) (40) (41) in, It is the first One B-spline basis function; It is the first A superposition function of B-spline elements, by indivual Summation yields the result; It is the first Bias terms of a KAN mapping layer; It is a linear function, representing the first... One KAN mapping layer; It is a random initial weight, representing Initial weights of the KAN mapping layer; Yes conduct The weights obtained after the first iteration represent the weights of the first iteration. The weight coefficients of the KAN mapping layer; for Quantity; yes Number of triplet pairs; The range of values ​​for is {BBB-, BBB+}; S43: Configure the cross-entropy loss function The specific formula is as follows: (42) in, Indicates the first The crossover loss function for each compound; S44: Use gradient descent method to... Perform iterations: Set the change in the loss function Convergence threshold is ; calculate in each iteration When satisfied or Stop iteration and output when... This yields the trained KAN model; The following iterative equations (43)~(44) were used to calculate the following: (43) (44) in, It is a fixed learning rate; Yes conduct The weights obtained after rounds of iteration; Indicates the first Blood-brain barrier permeability tags for the compounds; Indicates the first Predicted values ​​of blood-brain barrier permeability tags for each compound; express right The gradient; It is a dynamically adjustable value, set according to the size of the data at different scales; S45: Use The performance of the trained KAN model is evaluated.

Images