Drug-target affinity prediction method and system based on Raman spectrum and multisource heterogeneous characteristics

Abstract

The invention discloses a drug-target affinity prediction method and system based on Raman spectrum and multisource heterogeneous characteristics. The method comprises the following steps: acquiring an amino acid sequence data set and a first Raman spectrum matrix of existing proteins, and an SMILES sequence data set and a second Raman spectrum matrix of existing compounds; obtaining an amino acid sequence feature matrix based on the amino acid sequence data set, obtaining an SMILES sequence feature matrix based on the SMILES sequence data set, and obtaining a first Raman spectrum feature and a second Raman spectrum feature based on the first Raman spectrum matrix and the second Raman spectrum matrix; obtaining a first multi-source heterogeneous feature extraction result based on the amino acid sequence feature matrix and the first Raman spectrum feature, and obtaining a second multi-source heterogeneous feature extraction result based on the SMILES sequence feature matrix and the second Raman spectrum feature; and obtaining a drug-target affinity prediction result based on the first multi-source heterogeneous feature extraction result and the second multi-source heterogeneous feature extraction result.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug research and development, and in particular relates to a drug-target affinity prediction method and system based on Raman spectroscopy and multi-source isomerism characteristics. Background Art

[0002] Drug-Target binding Affinity (DTA) prediction technology can effectively improve the success rate and speed of new drug development and drug repositioning. The current DTA prediction method based on machine learning is mainly based on molecular sequence or secondary structure characteristics, lacks consideration of the three-dimensional structure of proteins and drugs, and is prone to losing important information in the biological reaction process of drug-target space. Raman spectroscopy technology is a spectral analysis method based on the principle of Raman scattering. The Raman spectral data of amino acids and compounds collected by Raman spectrometers can provide more relevant information about the three-dimensional structure of molecules for DTA prediction.

[0003] Although the traditional drug development process has high accuracy, it has problems such as high cost, long time consumption and low efficiency. Virtual drug screening based on machine learning can avoid the financial, energy and time consumption caused by a large number of traditional experimental screenings, reduce drug development investment, and is a hot research direction in the current new drug development technology. Existing research methods are mostly based on protein and compound molecular sequence or secondary structure characteristics for DTA prediction. However, this type of method fails to consider the three-dimensional structural information of proteins and drugs, and cannot fully reflect the spatial biological reaction process of drugs-targets.

[0004] Raman spectroscopy is a spectral analysis method based on the principle of Raman scattering. The Raman spectral data of amino acids and compounds collected by Raman spectrometers can provide more relevant information about the three-dimensional structure of molecules for DTA prediction. Raman spectroscopy can provide molecular vibration information of samples. By analyzing Raman spectra, characteristic peaks of different molecules in the sample can be extracted. These characteristic peaks correspond to specific vibration modes and chemical bond information of molecules. Summary of the invention

[0005] The present invention aims to solve the deficiencies of the prior art and provides the following solutions:

[0006] The drug-target affinity prediction method based on Raman spectroscopy and multi-source heterogeneous features includes the following steps:

[0007] Obtaining an amino acid sequence data set and a first Raman spectrum matrix of an existing protein, and a SMILES sequence data set and a second Raman spectrum matrix of an existing compound;

[0008] Performing feature extraction on the amino acid sequence data set to obtain an amino acid sequence feature matrix, performing feature extraction on the SMILES sequence data set to obtain a SMILES sequence feature matrix, and performing feature extraction on the first Raman spectrum matrix and the second Raman spectrum matrix to obtain a first Raman spectrum feature and a second Raman spectrum feature;

[0009] The amino acid sequence feature matrix is ​​weightedly fused with the first Raman spectrum feature to obtain a first multi-source heterogeneous feature extraction result, and the SMILES sequence feature matrix is ​​weightedly fused with the second Raman spectrum feature to obtain a second multi-source heterogeneous feature extraction result;

[0010] The first multi-source heterogeneous feature extraction result and the second multi-source heterogeneous feature extraction result are cross-fused, and the fused structure is input into a fully connected neural network to obtain a drug-target affinity prediction result.

[0011] Preferably, the method of obtaining the first Raman spectrum matrix and the second Raman spectrum matrix includes:

[0012] Using a Raman spectrometer to collect data on a sample of the existing protein / the existing compound to obtain Raman spectrum data of the protein / compound;

[0013] Draw a protein / compound Raman spectrum graph based on the protein / compound Raman spectrum data and perform averaging processing;

[0014] The first Raman spectrum matrix and the second Raman spectrum matrix are constructed using the protein / compound Raman spectrum map.

[0015] Preferably, the method for obtaining the amino acid sequence feature matrix comprises:

[0016] Performing embedding coding and position coding on the amino acid sequence data set to obtain a coded amino acid matrix;

[0017] Inputting initial parameters into the convolution kernel of the S4 model, the initial parameters include: angular frequency ω, the encoded amino acid matrix A, initial state matrices Λ, B, C, low-rank decomposition matrices P, Q, and step size Δ;

[0018] Solve using the encoded amino acid matrix A and the initial state matrix C

[0019]

[0020] in, represents a specific matrix in the state space model, I represents the identity matrix, It represents the matrix after the encoded amino acid matrix A passes through the generating function, Represents the matrix after the initial state matrix C is normalized;

[0021] based on Compute the black-box Cauchy kernel:

[0022]

[0023] Among them, k ab (ω) represents the state of the amino acid node at the angular frequency ω, a = 0, 1, b = 0, 1;

[0024] Substitute the black box Cauchy kernel into the Woodbury formula to obtain an intermediate variable matrix, and calculate the state space generating function based on the intermediate variable matrix

[0025]

[0026]

[0027] in, represents the intermediate variable matrix, j represents the complex unit, k represents the time step, and L represents the length of the truncated state space generating function;

[0028] The state space generating function Perform fast Fourier transform to obtain the amino acid sequence feature matrix:

[0029]

[0030]

[0031] Where t represents continuous time, Represents the amino acid sequence feature matrix.

[0032] Preferably, the method for obtaining the SMILES sequence feature matrix includes:

[0033] The SMILES sequence data set is embedded and positionally encoded to obtain the encoded SMILES sequence X D ;

[0034] The encoded SMILES sequence X D Perform linear transformation to obtain the compound two-dimensional matrices Q, K and V, and use the linear projection matrix and Linearly project the compound two-dimensional matrices Q, K and V to obtain the linearly transformed two-dimensional matrix Q D , K D and VD , and calculate the output of the attention layer:

[0035] QW f Q =Q D ,

[0036]

[0037]

[0038] Where, f = 1, 2, ..., m, f represents the number of layers of linear projection, m represents the number of linear projections, d represents the number of linear projections, k represents the dimension of the key vector, and T represents the transposition of the matrix;

[0039] Based on the output of the attention layer, the ReLU activation function is used to calculate the SMILES sequence feature matrix B1:

[0040] B1=ReLU(concat(head f , ..., head m )W D )

[0041] Among them, W D Represents a linear projection matrix.

[0042] Preferably, the method for obtaining the first Raman spectrum feature and the second Raman spectrum feature comprises:

[0043] Performing batch normalization on the first Raman spectrum matrix and the second Raman spectrum matrix, and aggregating information from adjacent nodes to obtain a first output matrix;

[0044] Inputting the first output matrix into a graph convolutional network to obtain a three-dimensional matrix output feature and a two-dimensional matrix output feature;

[0045] Performing node set processing on the three-dimensional matrix output features to obtain three-dimensional hidden features, and performing global pooling processing on the three-dimensional hidden features to obtain pooling features;

[0046] The pooling feature and the two-dimensional matrix output feature are connected and input into a fully connected neural network to obtain the first Raman spectrum feature A2 and the second Raman spectrum feature B2.

[0047] The present invention also provides a drug-target affinity prediction system based on Raman spectroscopy and multi-source heterogeneous features, wherein the prediction system applies any of the above-mentioned prediction methods and comprises: a data collection module, a feature extraction module, a weighted fusion module and a prediction module;

[0048] The data collection module is used to obtain an amino acid sequence data set and a first Raman spectrum matrix of an existing protein, and a SMILES sequence data set and a second Raman spectrum matrix of an existing compound;

[0049] The feature extraction module is used to perform feature extraction on the amino acid sequence data set to obtain an amino acid sequence feature matrix, perform feature extraction on the SMILES sequence data set to obtain a SMILES sequence feature matrix, and perform feature extraction on the first Raman spectrum matrix and the second Raman spectrum matrix to obtain a first Raman spectrum feature and a second Raman spectrum feature;

[0050] The weighted fusion module is used to perform weighted fusion on the amino acid sequence feature matrix and the first Raman spectrum feature to obtain a first multi-source heterogeneous feature extraction result, and to perform weighted fusion on the SMILES sequence feature matrix and the second Raman spectrum feature to obtain a second multi-source heterogeneous feature extraction result;

[0051] The prediction module is used to perform feature cross-fusion on the first multi-source heterogeneous feature extraction result and the second multi-source heterogeneous feature extraction result, and input the fused structure into a fully connected neural network to obtain a drug-target affinity prediction result.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] The present invention uses machine learning algorithms to perform in-depth mining of the three-dimensional Raman spectrum data and two-dimensional sequence data of drugs and targets, analyzes the affinity between drugs and targets and calculates their intensity, providing certain data references for virtual drug screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0055] Figure 1 A schematic diagram of a method flow of an embodiment of the present invention;

[0056] Figure 2 is a flow chart of the Transformer algorithm of an embodiment of the present invention;

[0057] Figure 3 This is a flow chart of feature extraction of Raman spectroscopy data based on KC-GCN according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Embodiment 1

[0061] In this embodiment, if Figure 1 As shown, the drug-target affinity prediction method based on Raman spectroscopy and multi-source heterogeneous features includes the following steps:

[0062] S1. Obtain an amino acid sequence dataset and a first Raman spectrum matrix of an existing protein, as well as a SMILES sequence dataset and a second Raman spectrum matrix of an existing compound.

[0063] In this embodiment, the method for obtaining an amino acid sequence data set and a SMILES sequence data set includes: collecting and downloading data such as the name, category, use, sequence, drug-target interaction (DTI) and drug-target affinity (DTA) of drugs and targets from databases such as PubChem, Davis and KIBA, and collating and constructing a target protein amino acid FASTA sequence and a drug compound SMILES sequence data set, a total of 671 protein amino acid FASTA sequence data, 2179 compound SMILES sequence data, and 41910 groups of drug-target interaction data.

[0064] The method for obtaining the first Raman spectrum matrix and the second Raman spectrum matrix includes: using a Raman spectrometer to collect data from samples of existing proteins / existing compounds to obtain protein / compound Raman spectrum data; drawing a protein / compound Raman spectrum graph based on the protein / compound Raman spectrum data and performing averaging processing; and constructing the first Raman spectrum matrix and the second Raman spectrum matrix using the protein / compound Raman spectrum graph. The first Raman spectrum matrix and the second Raman spectrum matrix are both composed of a Raman spectrum three-dimensional matrix and a Raman spectrum two-dimensional matrix.

[0065] In this embodiment, in order to correspond to the sequence data downloaded from the above public database, Raman spectrometer was used to collect Raman spectral data of 671 proteins and 2179 compounds corresponding to the above data set. 100 sets of data were collected for each protein and compound molecule.

[0066] (1) Raman spectral data acquisition: Raman spectrometer is used to collect Raman spectral data of amino acids and compounds respectively. Taking compounds as an example, the specific data acquisition and processing steps are as follows: ① Prepare samples: Place the compound to be tested on the sample holder and make it as flat and smooth as possible to ensure the stability of the spectral data; ② Determine instrument parameters: According to the characteristics of the compound, select appropriate laser wavelength, power and detector parameters to ensure the clarity and stability of the spectral signal; ③ Perform background scan: Before acquisition, a background scan is required to eliminate the interference of the instrument background signal on the sample signal; ④ Collect samples: Place the compound on the sample holder for acquisition to obtain accurate spectral data; ⑤ Data processing: Use the spectral instrument software to process and analyze the spectral data obtained from the test, including the measurement and analysis of indicators such as peak position, peak shape, and peak area; ⑥ Interpretation of results: Interpret and evaluate the processed and analyzed spectral data to determine the structure and composition of the compound, as well as characterize the physicochemical properties and performance of the sample.

[0067] (2) Raman spectral data processing: Based on the collected protein and compound Raman spectral data, protein Raman spectra and compound Raman spectra are plotted respectively and averaged. Through the Raman spectra, the differences in wavenumbers and peak values ​​between proteins and compounds can be determined. Using the difference in wavenumbers, we can determine the chemical bond vibrations and specific chemical groups of proteins and compounds; using the difference in peaks, we can determine the three-dimensional structural information of proteins and compounds. Next, the Raman spectra are used to construct a three-dimensional Raman spectrum matrix and a two-dimensional Raman spectrum matrix of proteins and compounds. These matrices contain rich structural information about the target proteins and compounds, preparing for subsequent feature extraction. In this embodiment, the first Raman spectrum matrix is ​​taken as an example to introduce the process. The specific process is as follows: ① Generation of the three-dimensional Raman spectrum matrix: First, the number of occurrences of the k-mers fragment of the target amino acid is calculated. Next, the combinations of all three residues are listed to form a combination set P a Then, regardless of the order of the residues, the occurrences of the k-mers segment in the amino acids are counted and the results are mapped to the above combination set P a Finally, a three-dimensional matrix of Raman spectra was established Where L is P a ② Generation of Raman spectrum two-dimensional matrix: First, all possible combinations of two residues in the protein Raman spectrum sequence are represented, so that the protein characteristics can be captured more accurately. Then, the Cartesian product operation is used to process the set P of all residues in the three-dimensional matrix M obtained above. a, use the Cartesian product operation with itself to process the set of all residues of the target sequence to obtain the intermediate variable set Y. All the combinations of remainders appearing in set Y are counted in the order of remainders and further mapped to P a Finally, a two-dimensional matrix of Raman spectra was established.

[0068] S2. Perform feature extraction on the amino acid sequence data set to obtain an amino acid sequence feature matrix, perform feature extraction on the SMILES sequence data set to obtain a SMILES sequence feature matrix, perform feature extraction on the first Raman spectrum matrix and the second Raman spectrum matrix to obtain the first Raman spectrum feature and the second Raman spectrum feature.

[0069] Most amino acid sequences are long sequences. When extracting features from long sequences, problems such as inaccurate feature extraction results often occur because the maximum length of the amino acid sequence exceeds the maximum length that the algorithm can calculate. The S4 (Structured State Spaces) algorithm is a type of algorithm in natural language processing algorithms that is suitable for processing long sequence problems. This embodiment considers using the S4 algorithm to extract features from long amino acid sequences. In this embodiment, the method for obtaining an amino acid sequence feature matrix includes: embedding and position encoding the amino acid sequence data set, the embedding encoding is a linear conversion process, and the embedding layer has trainable weights. Position encoding encodes the relative position information, absolute position information, and relative position information of each amino acid or compound molecule. During position encoding, the dimension is set to the same dimension as the embedded coding, and then the encoded amino acid matrix is ​​calculated using the following formula:

[0070]

[0071]

[0072] Among them, i represents the position of the amino acid, q represents the dimension, d represents the size of the amino acid position code, PE T Indicates the position encoding of the i-th amino acid in the protein sequence; input the initial parameters into the convolution kernel of the S4 model (structured state space sequence model), the initial parameters include: angular frequency ω, encoded amino acid matrix A, initial state matrix Λ, B, C, low rank decomposition matrix P, Q, step size Δ, among which:

[0073] A=A-PQ *

[0074] Solve using the encoded amino acid matrix A and the initial state matrix C

[0075]

[0076] in, represents a specific matrix in the state space model, I represents the identity matrix, It represents the matrix after the encoded amino acid matrix A passes through the generating function, Represents the matrix after the initial state matrix C is standardized; based on Compute the black-box Cauchy kernel:

[0077]

[0078] Among them, k ab (ω) represents the state of the amino acid node at the angular frequency ω, a=0, 1, b=0, 1; the black box Cauchy kernel is substituted into the Woodbury formula to obtain the intermediate variable matrix, and the state space generating function is calculated based on the intermediate variable matrix

[0079]

[0080]

[0081] in, represents the intermediate variable matrix, j represents the complex unit, k represents the time step, and L represents the length of the truncated state space generating function; the state space generating function Perform fast Fourier transform to obtain the amino acid sequence feature matrix:

[0082]

[0083]

[0084] Where t represents continuous time, Represents the amino acid sequence feature matrix.

[0085] The SMILES sequences of the compounds are all short sequences. If the S4 algorithm suitable for long sequence feature extraction is continued to be used, the complexity of the model will increase. Therefore, this embodiment uses the Transformer algorithm in the natural language processing algorithm to extract features. The Transformer algorithm is mainly based on the attention mechanism, such as Figure 2 In this embodiment, the method for obtaining the SMILES sequence feature matrix includes: performing embedding coding and position coding on the SMILES sequence data set to obtain the encoded SMILES sequence X D ; For the encoded SMILES sequence X D Perform linear transformation to obtain the compound two-dimensional matrices Q, K and V, and use the linear projection matrix and Perform m linear projections on the compound two-dimensional matrices Q, K and V to obtain the linearly transformed two-dimensional matrix Q D , K D and V D , and calculate the output of the attention layer:

[0086]

[0087]

[0088]

[0089]

[0090] Where, f = 1, 2, ..., m, f represents the number of layers of linear projection, m represents the number of linear projections, d represents the number of linear projections, k represents the dimension of the key vector, d v Represents the dimension of the value vector, d k =d v , T represents matrix transpose, R represents the total Raman spectrum data set, and n represents the maximum length of the compound; based on the output of the attention layer, the ReLU activation function is used to calculate the SMILES sequence feature matrix B1:

[0091] B1=ReLU(concat(head f , ..., head m )W D )

[0092] Among them, W D Represents a linear projection matrix.

[0093] The method for obtaining the first Raman spectral feature and the second Raman spectral feature includes: performing batch normalization operations on the first Raman spectral matrix and the second Raman spectral matrix, and aggregating information from adjacent nodes to obtain a first output matrix; inputting the first output matrix into a graph convolutional network to obtain a three-dimensional matrix output feature and a two-dimensional matrix output feature; performing node set processing on the three-dimensional matrix output feature to obtain a three-dimensional hidden feature, and performing global pooling processing on the three-dimensional hidden feature to obtain a pooled feature; connecting the pooled feature and the two-dimensional matrix output feature, and inputting and outputting them into a fully connected neural network to obtain a first Raman spectral feature A2 and a second Raman spectral feature B2.

[0094] In this embodiment, the Raman spectral data of amino acids and compounds have three-dimensional structural characteristics, and basic algorithms are difficult to fully extract their features. Therefore, this embodiment considers using a graph neural network to extract features. Ordinary graph neural networks cannot quickly extract the relationship between nodes, so this embodiment considers using the graph neural network algorithm KC-GCN for improved data processing. The feature extraction process of amino acid Raman spectral data and the feature extraction process of compound Raman spectral data both use the KC-GCN algorithm, and the steps are the same. Next, the feature extraction method of amino acid Raman spectral data is introduced as an example. According to the results of the DTA prediction process, the compound two-dimensional structure database is queried, the corresponding compound two-dimensional structure data is obtained, and it is sent into the graph neural network space to simulate the two-dimensional space biological reaction process of compounds and proteins. The simulation process includes:

[0095] Due to the two ways of processing protein data, two graph neural networks are used for protein feature extraction. Therefore, the whole model includes three inputs and three corresponding neural networks, two of which are used for protein feature mining and the other one is used for mining between molecular node features. Then, the three amino acid hidden features obtained are connected to output the amino acid feature extraction results based on Raman spectroscopy data, such as Figure 3 The specific steps are as follows. The graph convolutional network (GCN) is a neural network model designed specifically for processing graph structure data. The k-mers three-dimensional matrix M and two-dimensional matrix N of the amino acid Raman spectra obtained in the data processing stage are sent to Figure 3 In the model shown, the data input and graph neural network calculation process of the three-dimensional matrix and the two-dimensional matrix are the same. Here, three-dimensional data is used as an example:

[0096] ① Data input: batch normalize the k-mers three-dimensional matrix M of amino acid Raman spectra, and obtain the corresponding output by aggregating information from adjacent nodes, expressed as matrix H; ② Graph neural network calculation: define the number of GCN layers as l+1. First, input the information matrix H to layer 0 0 GCN learns the node feature calculation and outputs the l+1-layer amino acid node feature matrix H after multiple update iterations. (l+1) :

[0097]

[0098] Among them, σ represents the activation function, D -1 / 2 represents the diagonal matrix of the graph, represents the graph frontage matrix with self-loops added, H l represents the protein node feature matrix of layer l, W l Represents the weight matrix of layer l; then, the amino acid node feature matrix H (l+1)To represent the node, its spatial position is represented as x and y. Using the weight matrix W X Perform a linear transformation on the nodes and calculate the attention coefficient e between each input node J and its first-order neighbors in the graph xy , to characterize the importance of node y relative to node x:

[0099]

[0100] Among them, α represents the attention function; then, the attention coefficient e is adjusted using the sofimax function. xy Normalize and get the normalized attention coefficient α xy :

[0101]

[0102] Among them, s represents the adjacent node; then according to the obtained attention coefficient α xy Calculate the node output feature h xy

[0103]

[0104] Among them, N (x) Represents the set of adjacent nodes of node x; the node output feature of the three-dimensional matrix is ​​represented as h, and the node output feature of the two-dimensional matrix can be obtained by the same method, that is, the hidden feature of the two-dimensional matrix, represented as h2; ③ Process the node set of the node output feature h of the three-dimensional matrix. First, perform GIN (Generalized Inverted Index) calculation to obtain the hidden feature h1, as shown in the following formula:

[0105]

[0106] in, represents the feature vector of node v in layer p, N (v) represents the set of adjacent nodes of node v, MLP represents multi-layer perceptron, ε (p) is a learnable bias vector. Because different molecules have different numbers of atoms, molecular graphs can be constructed with different numbers of nodes. Therefore, after the GIN calculation, the hidden feature h1 of the three-dimensional matrix is ​​globally pooled to ensure that its pooling result h3 can have the same dimension as the hidden feature h2 of the two-dimensional matrix. Finally, the hidden features h3 and h2 of the three-dimensional matrix and the two-dimensional matrix are connected and sent to the fully connected neural network to obtain the first Raman spectral feature A2 of the amino acid Raman spectral data. Using the same method, the second Raman spectral feature B2 of the compound Raman spectral data can be obtained.

[0107] S3. Perform weighted fusion of the amino acid sequence feature matrix and the first Raman spectrum feature to obtain a first multi-source heterogeneous feature extraction result, and perform weighted fusion of the SMILES sequence feature matrix and the second Raman spectrum feature to obtain a second multi-source heterogeneous feature extraction result.

[0108] In this embodiment, first, the internal feature extraction results of amino acids and compounds are cross-fused. The amino acid sequence feature matrix A1 and the first Raman spectrum feature A2 of the amino acid Raman spectrum data are converted to the same space and weighted to obtain the multi-source isomer feature extraction result of amino acids, which is expressed as V D The SMILES sequence feature matrix B1 of the compound sequence and the second Raman spectrum feature B2 of the compound Raman spectrum data are converted to the same space and weighted to obtain the multi-source heterogeneous feature extraction result of the compound, which is expressed as V T .

[0109] S4. Perform feature cross-fusion on the first multi-source heterogeneous feature extraction result and the second multi-source heterogeneous feature extraction result, and input the fused structure into a fully connected neural network to obtain a drug-target affinity prediction result.

[0110] In this embodiment, the feature cross-fusion between amino acids and compounds is performed. The feature extraction results V of amino acids and compounds obtained in the previous step are D and V T After the connection, it is fed into a fully connected neural network. Finally, an adaptive average pooling layer is used to adjust the feature size of the attention block output and output the drug-target affinity prediction result.

[0111] Embodiment 2

[0112] In this embodiment, a drug-target affinity prediction system based on Raman spectroscopy and multi-source heterogeneous features includes: a data collection module, a feature extraction module, a weighted fusion module and a prediction module.

[0113] The data collection module is used to obtain the amino acid sequence data set and the first Raman spectrum matrix of the existing protein, and the SMILES sequence data set and the second Raman spectrum matrix of the existing compound.

[0114] The feature extraction module is used to extract features from the amino acid sequence data set to obtain an amino acid sequence feature matrix, to extract features from the SMILES sequence data set to obtain a SMILES sequence feature matrix, and to extract features from the first Raman spectrum matrix and the second Raman spectrum matrix to obtain first Raman spectrum features and second Raman spectrum features.

[0115] The weighted fusion module is used to weightedly fuse the amino acid sequence feature matrix with the first Raman spectrum feature to obtain the first multi-source heterogeneous feature extraction result, and to weightedly fuse the SMILES sequence feature matrix with the second Raman spectrum feature to obtain the second multi-source heterogeneous feature extraction result.

[0116] The prediction module is used to perform feature cross-fusion on the first multi-source heterogeneous feature extraction result and the second multi-source heterogeneous feature extraction result, and input the fused structure into a fully connected neural network to obtain a drug-target affinity prediction result.

[0117] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A drug-target affinity prediction method based on Raman spectroscopy and multi-source isomerism features, characterized in that: The following steps are involved: Obtaining an amino acid sequence data set and a first Raman spectrum matrix of an existing protein, and a SMILES sequence data set and a second Raman spectrum matrix of an existing compound; Performing feature extraction on the amino acid sequence data set to obtain an amino acid sequence feature matrix, performing feature extraction on the SMILES sequence data set to obtain a SMILES sequence feature matrix, and performing feature extraction on the first Raman spectrum matrix and the second Raman spectrum matrix to obtain a first Raman spectrum feature and a second Raman spectrum feature; The amino acid sequence feature matrix is ​​weightedly fused with the first Raman spectrum feature to obtain a first multi-source heterogeneous feature extraction result, and the SMILES sequence feature matrix is ​​weightedly fused with the second Raman spectrum feature to obtain a second multi-source heterogeneous feature extraction result; The first multi-source heterogeneous feature extraction result and the second multi-source heterogeneous feature extraction result are cross-fused, and the fused structure is input into a fully connected neural network to obtain a drug-target affinity prediction result.

2. The drug-target affinity prediction method based on Raman spectroscopy and multi-source isomerism characteristics according to claim 1, characterized in that: The method of obtaining the first Raman spectrum matrix and the second Raman spectrum matrix includes: Using a Raman spectrometer to collect data on a sample of the existing protein / the existing compound to obtain Raman spectrum data of the protein / compound; Draw a protein / compound Raman spectrum graph based on the protein / compound Raman spectrum data and perform averaging processing; The first Raman spectrum matrix and the second Raman spectrum matrix are constructed using the protein / compound Raman spectrum map.

3. The drug-target affinity prediction method based on Raman spectroscopy and multi-source isomerism characteristics according to claim 1, characterized in that: The method for obtaining the amino acid sequence feature matrix comprises: Performing embedding coding and position coding on the amino acid sequence data set to obtain a coded amino acid matrix; Inputting initial parameters into the convolution kernel of the S4 model, the initial parameters include: angular frequency ω, the encoded amino acid matrix A, initial state matrices Λ, B, C, low-rank decomposition matrices P, Q, and step size △; Solve using the encoded amino acid matrix A and the initial state matrix C in, represents a specific matrix in the state space model, I represents the identity matrix, It represents the matrix after the encoded amino acid matrix A passes through the generating function, Represents the matrix after the initial state matrix C is standardized; based on Compute the black-box Cauchy kernel: Among them, k ab (ω) represents the state of the amino acid node at the angular frequency ω, a=0,1, b=0,1; Substitute the black box Cauchy kernel into the Woodbury formula to obtain an intermediate variable matrix, and calculate the state space generating function based on the intermediate variable matrix in, represents the intermediate variable matrix, j represents the complex unit, k represents the time step, and L represents the length of the truncated state space generating function; The state space generating function Perform fast Fourier transform to obtain the amino acid sequence feature matrix: Where t represents continuous time, Represents the amino acid sequence feature matrix.

4. The drug-target affinity prediction method based on Raman spectroscopy and multi-source isomerism characteristics according to claim 1, characterized in that: The method for obtaining the SMILES sequence feature matrix includes: The SMILES sequence data set is embedded and positionally encoded to obtain the encoded SMILES sequence X D ; The encoded SMILES sequence X D Perform linear transformation to obtain the compound two-dimensional matrices Q, K and V, and use the linear projection matrix and Linearly project the compound two-dimensional matrices Q, K and V to obtain the linearly transformed two-dimensional matrix Q D , K D and V D , and calculate the output of the attention layer: Where, f = 1, 2, ..., m, f represents the number of layers of linear projection, m represents the number of linear projections, dk represents the dimension of the key vector, and T represents transposing the matrix; Based on the output of the attention layer, the ReLU activation function is used to calculate the SMILES sequence feature matrix B1: B1=ReLU(concat(head f ,...,head m )W D ) Among them, W D Represents a linear projection matrix.

5. The method for predicting drug-target affinity based on Raman spectroscopy and multi-source isomerism characteristics according to claim 1, characterized in that: The method for obtaining the first Raman spectrum feature and the second Raman spectrum feature comprises: Performing batch normalization on the first Raman spectrum matrix and the second Raman spectrum matrix, and aggregating information from adjacent nodes to obtain a first output matrix; Inputting the first output matrix into a graph convolutional network to obtain a three-dimensional matrix output feature and a two-dimensional matrix output feature; Performing node set processing on the three-dimensional matrix output features to obtain three-dimensional hidden features, and performing global pooling processing on the three-dimensional hidden features to obtain pooling features; The pooling feature and the two-dimensional matrix output feature are connected and input into a fully connected neural network to obtain the first Raman spectrum feature A2 and the second Raman spectrum feature B2.

6. A drug-target affinity prediction system based on Raman spectroscopy and multi-source isomerism characteristics, wherein the prediction system applies the prediction method according to any one of claims 1 to 5, characterized in that: include: Data collection module, feature extraction module, weighted fusion module and prediction module; The data collection module is used to obtain an amino acid sequence data set and a first Raman spectrum matrix of an existing protein, and a SMILES sequence data set and a second Raman spectrum matrix of an existing compound; The feature extraction module is used to perform feature extraction on the amino acid sequence data set to obtain an amino acid sequence feature matrix, perform feature extraction on the SMILES sequence data set to obtain a SMILES sequence feature matrix, and perform feature extraction on the first Raman spectrum matrix and the second Raman spectrum matrix to obtain a first Raman spectrum feature and a second Raman spectrum feature; The weighted fusion module is used to perform weighted fusion on the amino acid sequence feature matrix and the first Raman spectrum feature to obtain a first multi-source heterogeneous feature extraction result, and to perform weighted fusion on the SMILES sequence feature matrix and the second Raman spectrum feature to obtain a second multi-source heterogeneous feature extraction result; The prediction module is used to perform feature cross-fusion on the first multi-source heterogeneous feature extraction result and the second multi-source heterogeneous feature extraction result, and input the fused structure into a fully connected neural network to obtain a drug-target affinity prediction result.

Images