A drug resistance prediction method and system based on contrast learning and multi-modal fusion

Abstract

The application discloses a drug resistance prediction method and system based on contrast learning and multi-modal fusion. The method first generates a molecular graph and a molecular fingerprint based on the SMILES of the target drug, extracts the drug molecular features through a contrast learning model constructed by combining a graph attention network and a graph convolution network. Then, the protein expression, gene expression and metabolic expression data are obtained from the target tissue cells, the modal features are extracted through a deep convolution network, a Transformer encoder and a multi-dimensional attention network respectively, and the adaptive fusion of multi-modal features is realized through a heterogeneous interaction attention mechanism. Finally, the fused multi-modal features and the drug molecular features are jointly input into a multi-layer perceptron to realize high-precision prediction of the drug resistance of the cells. The application effectively improves the representation ability and prediction accuracy of the model by introducing the contrast learning and multi-modal feature fusion mechanism, and can provide efficient and reliable support for drug screening and clinical decision-making.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of drug response prediction, and particularly relates to a drug resistance prediction method and system based on contrast learning and multi-modal fusion. BACKGROUND

[0002] In the drug development process, drug response prediction is an important link for new drug screening and target optimization, and its accuracy and efficiency are directly related to the development cycle and cost control. Traditional drug response prediction mainly relies on biological experimental methods such as in vitro experiments and in vivo animal experiments. This kind of method usually needs complex experimental procedures and high-precision experimental equipment, and the operation process is time-consuming and laborious, and the technical threshold is high. In addition, the collection and processing of experimental data are often limited by sample quantity, experimental conditions, detection platform and other factors, and there is inevitably a certain degree of error and irreproducibility. Therefore, when facing large-scale drug screening tasks, traditional experimental methods are not only low in efficiency, long in cycle and high in economic cost, and it is difficult to meet the actual needs of the current drug development towards high-throughput, fast-paced and low-cost direction.

[0003] In order to improve the efficiency and automation level of drug response prediction, in recent years, data-driven methods based on machine learning and deep learning have gradually developed. These methods train models to learn the potential relationship between drugs and cells, so as to realize the prediction of drug response. Although this kind of method alleviates the cost pressure brought by experimental means to a certain extent, but the current research still has some significant technical bottlenecks. First, the model structure used by a large number of researches is relatively simple, and shallow neural networks or traditional machine learning methods (such as support vector machine, random forest, etc.) are generally used, which have limited modeling ability and feature extraction ability when facing high-dimensional, nonlinear and complex biological data. Shallow model is difficult to describe the deep potential law in the data, which leads to the difficulty of the model prediction result to reach the ideal level in precision and generalization ability. Especially when the diversified response of drugs in different individuals or different tissues needs to be accurately modeled, the expression ability of the existing shallow method cannot meet the higher level of research demand.

[0004] Secondly, in terms of the use of feature data, most current drug response prediction models still have the problem of insufficient information utilization. The actual drug mechanism usually involves the synergistic regulation of multiple biological levels, such as gene expression, protein expression, metabolite concentration, etc., but many existing models are still based on single type of data input in the modeling process. This single-source feature modeling method ignores the multidimensional complexity of the biological system and is difficult to fully reflect the dynamic interaction process between drugs and cells, tissues and even the whole organism. There is a close relationship and potential coupling between multi-modal biological information, such as changes in gene expression levels may trigger the regulation of protein translation, which in turn affects the changes in the metabolic network, and then determines the actual performance of drug response. Ignoring these cross-information will inevitably limit the performance of the model in data learning and feature fusion, thereby affecting the final prediction accuracy.

[0005] In addition, at the data representation level, although many deep learning methods introduce multi-layer neural network structures to improve the non-linear modeling ability of the model, there are still problems such as unstable feature representation and insufficient utilization of context information in feature representation learning. In particular, in the drug response prediction scenario, the differences between different drug structures and the heterogeneity between different biological systems make it difficult for traditional feature representation methods to achieve good transferability and generalizability. In most scenarios, model training relies on sufficient and balanced labeled data, while in practice, labeled data is often scarce or unevenly distributed, further exacerbating the limitations of existing methods in generalizability and robustness.

[0006] In summary, the existing drug response prediction technology still has problems such as low efficiency, high cost, insufficient prediction accuracy, and limited model representation ability and feature fusion ability, no matter in experimental methods or modeling strategies. With the increasing demand for prediction accuracy and speed in drug development, it is urgent to make technological breakthroughs in feature expression, multi-modal fusion, deep modeling, and other dimensions to achieve a deeper analysis of drug response mechanisms and more efficient prediction capabilities. SUMMARY

[0007] To solve the above technical problems, the present application provides a drug resistance prediction method and system based on contrast learning and multi-modal fusion. The technical solution provided by the present application is as follows:

[0008] A drug resistance prediction method based on contrast learning and multi-modal fusion, comprising the steps of:

[0009] S1, generating a corresponding molecular graph and molecular fingerprint based on the SMILES of the target drug;

[0010] S2, using the molecular graph representation model trained based on the contrast learning mechanism to extract the molecular graph features and splice them with the molecular fingerprint to form the drug molecular features;

[0011] S3, obtaining protein expression data, gene expression data and metabolic expression data of the target tissue cell;

[0012] S4, preprocessing the protein expression data, the gene expression data and the metabolic expression data respectively to obtain a protein expression matrix, a gene expression matrix and a metabolic expression matrix;

[0013] S5, extracting protein expression features, gene expression features and metabolic expression features based on the protein expression matrix, the gene expression matrix and the metabolic expression matrix respectively, and interactively fusing the expression features of the three modalities;

[0014] S6, inputting the multi-modal fusion features after interactive fusion and drug molecule features into a drug resistance prediction network to obtain a drug resistance prediction result of the target tissue cell to the target drug.

[0015] Further, in step S2:

[0016] The molecular graph representation model includes a parallel graph attention network (GAT) and a graph convolution network (GCN) for generating two different embedding representations of the molecular graph; the molecular graph representation model adopts a contrast learning mechanism, and a weight sharing layer is added during training and connected with the output layers of the GAT and the GCN, and a similarity output layer is arranged after the weight sharing layer; by using contrast learning, the two embedding representations of the same molecular graph are regarded as positive pairs, while the two embedding representations of different molecular graphs are regarded as negative pairs; by maximizing the similarity of the positive pair representations and minimizing the similarity of the negative pair representations, the key features of the molecular structure are gradually learned; wherein the weight sharing layer fuses the feature representations generated by the GAT and the GCN by sharing the same network parameters, and the contrast learning mechanism uses a contrast loss function to optimize the distance difference between the positive pairs and the negative pairs, and the optimization goal is to minimize the difference between the feature representations of the same molecular graph and maximize the feature difference between different molecular graphs.

[0017] Further, the protein expression data is reduced in dimension by principal component analysis to form a protein expression matrix, the one-dimensional gene expression data is continuously segmented into gene expression segments to form a gene expression matrix, and the one-dimensional metabolic expression data is continuously segmented into metabolic expression segments to form a metabolic expression matrix.

[0018] Further, in step S5:

[0019] The protein expression feature is obtained by using a protein expression feature extraction network, the protein expression feature extraction network comprising a deep convolution module, a channel attention module and a spatial attention module connected in sequence: first, the deep convolution module uses multiple one-dimensional convolution layers to extract features layer by layer from the input protein expression matrix; then, the multi-dimensional original feature vector output by the deep convolution module is input into the channel attention module, the channel attention mechanism is used to generate channel weights, and then the channel weights are multiplied with the original feature vector, thereby realizing the weighting operation of each channel and highlighting the features of important channels in the original feature vector; then, the feature vector output by the channel attention module is input into the spatial attention module, the spatial attention mechanism is used to generate spatial weights, and then the spatial weights are multiplied with the feature vector output by the channel attention module, thereby dynamically adjusting the weights of different regions in the feature vector; finally, the feature vector output by the spatial attention module is taken as the protein expression feature.

[0020] Further, in step S5, the gene expression feature is obtained by using a gene expression feature extraction network, the gene expression feature extraction network adopting a Transformer encoder, the Transformer encoder being composed of multiple repeated encoder blocks, each encoder block containing input embedding, position encoding, multi-head self-attention mechanism and feedforward neural network, wherein the multi-head self-attention mechanism generates new embedding representation by calculating the relationship between different gene expression fragments in the gene expression matrix.

[0021] Further, in step S5, the metabolic expression feature is obtained by using a metabolic expression feature extraction network, the metabolic expression feature extraction network comprising a row attention module, a column attention module and a triangular attention module connected in sequence;

[0022] The row attention module generates a query vector, a key vector and a value vector for each row of the input metabolic expression matrix respectively, then calculates the attention relevance by dot product operation using the query vector and the key vector, generates an attention weight by a Softmax function, the attention weight is used for weighted combination of the value vector, and finally the output of the row attention module is obtained through linear mapping;

[0023] The column attention module generates a query vector, a key vector and a value vector for each column of the feature vector output by the row attention module respectively, then calculates the attention relevance by dot product operation using the query vector and the key vector, generates an attention weight by a Softmax function, the attention weight is used for weighted combination of the value vector, and finally the output of the column attention module is obtained through linear mapping;

[0024] The triangular attention module adopts a triangular attention mechanism to process the feature vector output by the column attention module, and takes the processed feature vector as the metabolic expression feature.

[0025] Further, in step S5, the protein expression features, gene expression features and metabolic expression features are fused by the heterogeneous interaction attention mechanism for heterogeneous interaction and adaptive weighted fusion, and finally the multi-modal fusion features are output; the heterogeneous interaction attention mechanism:

[0026] First, the self-attention mechanism is used to generate query vectors, key vectors and value vectors for each modality feature;

[0027] Then, the generated vectors are used for heterogeneous interaction to obtain the features after interaction:

[0028]

[0029]

[0030]

[0031] wherein, P, G, M are the protein expression features, gene expression features and metabolic expression features after heterogeneous interaction, , , , , , is the interaction weight, is the dimension of the query vector, and Softmax represents a normalization function; , , are the query vectors, key vectors and value vectors of the generated protein expression features, , , are the query vectors, key vectors and value vectors of the gene expression features, , , are the query vectors, key vectors and value vectors of the metabolic expression features;

[0032] Finally, the protein expression features, gene expression features and metabolic expression features after heterogeneous interaction are adaptively weighted and fused to obtain the multi-modal fusion features.

[0033] Preferably, in step S6, the drug resistance prediction network adopts a multi-perceptron, and the multi-layer perceptron includes an input layer, a plurality of hidden layers and an output layer, each hidden layer contains a plurality of neurons, and each neuron maps the input to the output through an activation function; the output layer outputs a drug resistance probability value, and the greater the value, the stronger the drug resistance of the target tissue cells to the target drug.

[0034] A drug resistance prediction system based on the above method, comprising a drug molecule feature acquisition module, a protein expression feature acquisition module, a gene expression feature extraction module, a metabolic expression feature extraction module, a multi-modal feature fusion module and a drug resistance prediction module;

[0035] The drug molecule feature acquisition module acquires molecular graph features based on a molecular graph generated by a target drug SMILES, and splices the molecular graph features with molecular fingerprints, and the splicing result is used as the drug molecule features;

[0036] The protein expression feature acquisition module, the gene expression feature extraction module and the metabolic expression feature extraction module respectively generate corresponding protein expression features, gene expression features and metabolic expression features based on protein expression data, gene expression data and metabolic expression data of a target tissue cell;

[0037] The multi-modal feature fusion module performs heterogeneous interaction and adaptive weighted fusion on the generated protein expression features, gene expression features and metabolic expression features through a heterogeneous interaction attention mechanism, and finally outputs multi-modal fusion features;

[0038] The drug resistance prediction module is used to output a drug resistance prediction result of a target tissue cell to a target drug according to the input drug molecule features and multi-modal fusion features.

[0039] Compared with the prior art, the present application has achieved significant improvement in many aspects:

[0040] Firstly, the traditional drug response prediction method relies on single-modal data and shallow learning model, and it is difficult to fully capture the multi-level biological information in the drug action process, resulting in insufficient prediction accuracy and generalization ability. While the present application innovatively introduces three key omics data of protein expression, gene expression and metabolic expression, constructs a heterogeneous feature extraction network, uses deep convolution network combined with attention mechanism, Transformer encoder and multi-dimensional attention module to extract core features in each modality, and further uses heterogeneous interaction attention mechanism to complete the fusion of multi-modal information, effectively improving the diversity and integrity of feature representation, and enhancing the modeling ability of complex biological response mechanism.

[0041] In addition, the application combines graph convolution network and graph attention network in drug molecule modeling, and enhances the recognition ability of the model to structural differences through a contrast learning mechanism, so that the drug molecule features are more stable and have higher discrimination. The contrast learning mechanism maximizes the similarity between multi-view representations of the same molecule and minimizes the similarity between representations of different molecules, thereby improving the robustness and generalization ability of drug structure representation. In the model integration stage, a multilayer perceptron is used to cooperatively model the integrated omics features and drug molecule features, which can not only capture deep interaction patterns between drugs and cells, but also output drug resistance prediction results with biological interpretation.

[0042] Through the integration of the above multi-dimensional structure design and learning mechanism, the application can achieve high-precision prediction of drug resistance in a scene with high sample heterogeneity and complex data dimension, which not only improves the explainability and adaptability of the algorithm, but also significantly reduces the development cost of relying on expensive experimental resources, effectively speeds up the drug screening and preclinical evaluation process, and has important clinical application value and industrial promotion prospects. BRIEF DESCRIPTION OF DRAWINGS

[0043] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, which together with the embodiments of the application, are used to explain the application, and do not constitute a limitation on the application.

[0044] Figure 1 is a schematic diagram of a drug resistance prediction technology framework provided by an embodiment of the application;

[0045] Figure 2 is a schematic diagram of drug molecule feature extraction provided by an embodiment of the application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0047] Embodiment one

[0048] The embodiment provides a drug resistance prediction method and system based on contrast learning and multi-modal fusion, comprising the following steps:

[0049] 1. Data collection

[0050] PDX (patient-derived xenograft) models are established by extracting tumor tissues, and the efficacy of various drugs is tested. Sequencing analysis of transcriptomics, proteomics, and metabolomics is performed to obtain gene expression, protein expression, and metabolic expression data of tumor tissues before drug addition. The SMILES (Simplified Molecular-Input Line-Entry System) of each drug molecule is also obtained, and finally, the drug resistance data of various drugs is obtained to construct a comprehensive dataset combining multidimensional biological information and drug resistance characterization.

[0051] Gene expression, protein expression, and metabolic expression are important biological indicators in tumor research, reflecting the activity state of genes, proteins, and metabolites in cells or tissues, which helps to comprehensively understand the biological characteristics of tumors and their response to drugs. SMILES is a chemical language used to describe molecular structure, which represents the atoms, bonds, and cyclic structures of molecules in linear strings, enabling computers to efficiently process and analyze molecular information.

[0052] 2. Data preprocessing

[0053] Protein expression data is usually presented in the form of a multidimensional matrix. Since protein expression data usually has high dimensions and contains a lot of redundant or irrelevant information, this embodiment uses principal component analysis (PCA) to reduce the dimension of the protein expression matrix. PCA is a commonly used unsupervised dimensionality reduction method that retains the main information in the data, removes noise and redundant features, and reduces the dimension of the data. After PCA dimensionality reduction, a relatively simple protein expression matrix containing main features can be obtained, which facilitates subsequent processing.

[0054] Gene expression data is usually presented in the form of one-dimensional data, with each gene's expression value as a feature. However, these one-dimensional data cannot directly reflect the complex interaction between genes. To solve this problem, this embodiment uses a fragmentation processing method to divide the one-dimensional gene expression data into multiple continuous fragments to form a two-dimensional matrix, with each fragment representing the local change of gene expression within a certain range. Specifically, assuming that the original sample length of gene expression data is L and the length of each fragment is P, N (N = L / P) sample fragments can be obtained, and the two-dimensional gene expression matrix composed of these N sample fragments has a dimension size of N × P .

[0055] Similar to gene expression data, metabolic expression data is also typically presented as one-dimensional data, with the expression value of each metabolite serving as a feature. To effectively uncover the complex relationships between metabolites, this embodiment also employs a fragmentation approach, dividing the one-dimensional metabolic expression data into multiple continuous fragments to form a two-dimensional matrix, thus better capturing the interrelationships between metabolites. Assuming the original sample length of the metabolic expression data is L, and the length of each fragment is P, then N (N = L / P) sample fragments can be obtained. The dimension of the two-dimensional metabolic expression matrix composed of these N sample fragments is... N × P .

[0056] 3. Multimodal feature extraction and fusion

[0057] (1) Construct a protein expression feature extraction network to extract protein expression features.

[0058] like Figure 1 As shown, the protein expression feature extraction network consists of a deep convolutional module, a channel attention module, and a spatial attention module connected in sequence.

[0059] The deep convolutional module utilizes multiple convolutional layers to extract features layer by layer from the input protein expression matrix. One-dimensional convolution is employed within these layers, effectively capturing local dependencies in sequence data. This is crucial for extracting potential patterns and important features from the protein expression matrix. By applying one-dimensional convolutional kernels to the protein expression matrix, local variation patterns in protein expression can be learned, which is highly useful for drug response prediction. To further enhance the network's feature representation capabilities, residual connections are also introduced into the deep convolutional module. Residual connections are a technique in deep neural networks that allow signals to "skip connections" within the network, avoiding the gradient vanishing or information loss problems common in deep networks. Through residual connections, the network can be more easily optimized, especially when dealing with complex data, helping to avoid overfitting and accelerating training.

[0060] In this embodiment, the feature vector extracted directly from the protein expression matrix by the additional branch convolutional layer is added element-wise to the feature vector extracted through multiple convolutional layers to achieve residual connections. Alternatively, the feature vector extracted by the previous convolutional layer is added element-wise to the feature vector extracted by the middle convolutional layer, and then used as the input vector for the subsequent network layer.

[0061] Channel attention mechanism can automatically assign weights to different feature channels of multi-dimensional feature vectors, focusing on more important feature channels for the task, thereby improving the selectivity of features and the expressive power of the model. Spatial attention mechanism focuses on the spatial position of features, and by weighting each position of the input, it can adjust the importance of features in the spatial dimension. The combination of these two attention mechanisms can help the model better understand which parts of the input data are important for the final prediction, thereby improving the accuracy and robustness of feature learning.

[0062] In this embodiment, the multi-dimensional original feature vector output by the deep convolution module is first input into the channel attention module, and the channel weight is generated through the channel attention mechanism, and then the channel weight is multiplied with the original feature vector, thereby realizing the weighting operation of each channel, and the important channel features in the original feature vector are highlighted. Then, the channel weighted feature vector output by the channel attention module is input into the spatial attention module, and the spatial weight is generated through the spatial attention mechanism, and then the spatial weight is multiplied with the input feature vector, thereby dynamically adjusting the weight of different regions in the feature vector, so that the network can pay attention to the important spatial position in the feature vector.

[0063] Specifically, the channel attention mechanism mainly includes the following steps:

[0064] Global pooling operation: compress the input feature vector through global average pooling or global maximum pooling to obtain a feature compression vector with a channel number of 1;

[0065] Fully connected processing: passing the feature compression vector through a fully connected layer to generate a channel weight vector, usually using a Sigmoid function to limit the weight value to between 0 and 1;

[0066] Weighted fusion: multiplying the generated channel weight vector with the original input feature vector.

[0067] Similarly, the spatial attention mechanism mainly includes the following steps:

[0068] Convolution operation: performing convolution operation on the input feature vector to extract spatial features;

[0069] Activation processing: compressing the convolution result to between 0 and 1 through a Sigmoid function to generate a spatial weight vector;

[0070] Weighted fusion: multiplying the generated spatial weight vector with the original input feature vector.

[0071] Through the above steps, especially the hybrid mechanism of channel attention and spatial attention, we can extract more effective and representative feature representations from the reduced protein expression matrix, which will serve as the input of the subsequent prediction model.

[0072] (2) Constructing a gene expression feature extraction network to extract gene expression features

[0073] The gene expression feature extraction network adopts a Transformer encoder to extract gene expression features by performing data projection, encoding, and position embedding on gene fragments in the gene expression matrix.

[0074] The Transformer encoder is composed of multiple repeated encoder blocks, each block containing input embedding, position embedding, multi-head self-attention mechanism (Multi-head Self-Attention), and feedforward neural network. The multi-head self-attention mechanism generates new representations by calculating the relationships between different parts (gene expression fragments divided during preprocessing) in the input sequence. Specifically, the input sequence is processed with position encoding to preserve the order of the sequence, and then mapped into query vectors (Query), key vectors (Key), and value vectors (Value). These vectors calculate similarity through dot product and use the softmax function for normalization to generate attention weights. Finally, the attention weights are multiplied with the value vectors and weighted summation is performed to obtain the final output. The feedforward neural network performs nonlinear transformation on the output of the self-attention mechanism to further extract features.

[0075] Transformer has an advantage in processing sequence data, especially in capturing long-term dependencies, far superior to traditional convolutional neural networks (CNN). Through the self-attention mechanism, Transformer can assign different weights to each gene expression fragment, focusing on key fragments related to drug response. The self-attention mechanism can also effectively reveal long-distance dependencies between genes, automatically focusing on gene fragments that have a greater impact on the prediction result, thereby improving the effectiveness of feature extraction. Transformer not only can extract local features, but also can capture global relationships between genes. This allows the model to consider the role of each gene in different fragments and understand the complex interactions between genes. By combining fragmentation processing with the Transformer model, it can more accurately extract potential patterns in the gene expression matrix, and thus provide more rich and accurate feature representations for drug resistance prediction.

[0076] (3) Constructing a metabolic expression feature extraction network to extract metabolic expression features

[0077] As shown in Figure 1 , the metabolic expression feature extraction network includes a row attention module (RowAttention), a column attention module (Column Attention), and a triangle attention module (TriangleAttention) connected in sequence.

[0078] 1) Row attention mechanism is a component of multi-head self-attention mechanism, which is specifically designed to process the rows of input data, and its main purpose is to capture the correlation between features within the row by focusing on the rows of input data. In this embodiment, the row attention module generates query vectors, key vectors and value vectors for each row of the input metabolic expression matrix respectively, calculates the attention correlation by dot product operation using the query vectors and key vectors, and generates attention weights by Softmax function, these weights are used for weighted combination of value vectors, and finally the final output is obtained through linear mapping.

[0079] 2) Column attention mechanism is similar to row attention mechanism, but it focuses on the relationship between features within the column of input data. In this embodiment, the column attention module generates query vectors, key vectors and value vectors for each column of the feature vectors output by the row attention module, calculates the attention correlation by dot product operation using the query vectors and key vectors, and generates attention weights by Softmax function, these weights are used for weighted combination of value vectors, and finally the final output is obtained through linear mapping.

[0080] 3) Triangular attention mechanism is one of the core innovations of AlphaFold series, which is mainly used to model the geometric constraints and chemical interactions between residues in protein structure, and solves the core challenge of simultaneously satisfying local chemical bonding and global spatial constraints in protein structure prediction through geometry-driven bidirectional path modeling, its design concept has been extended to other biological molecule interaction prediction fields. In this embodiment, the triangular attention mechanism is used to further process the feature vectors output by the column attention module.

[0081] Specifically, assuming that the feature vector output by the column attention module is , represents the vector element of the i th row and the j th column, first, project each position to generate three groups of vectors respectively:

[0082]

[0083] wherein , and are the query vectors, key vectors and value vectors corresponding to the corresponding position respectively, , and are the corresponding projection weights, , and are bias terms.

[0084] ① Start node mode calculation

[0085] For each position in the first i row, compute the dot product similarity and scale:

[0086]

[0087] Along k the dimension, perform Softmax normalization:

[0088] Update the representation of all positions in the current row:

[0089] ② Termination node mode calculation

[0090] Directly reuse the weights of the starting node mode, but transpose the row and column indices:

[0091]

[0092] Update the representation of all positions in the current column:

[0093] ③ Merge and output

[0094] Concatenate the outputs of the two modes and fuse through a linear transformation layer to obtain the final output.

[0095] The above is the main calculation process of the triangular attention module. This process implements the triangular attention mechanism in the AlphaFold series, and its core lies in focusing dynamically through rows and columns, which shows the triangular inequality constraint.

[0096] After obtaining the metabolic expression matrix, the metabolic expression feature extraction network extracts the relationship information between metabolites by calculating row, column and triangular attention. Row attention mainly focuses on the global information of metabolites in different fragments, column attention focuses on the relationship between different metabolites, and triangular attention further focuses on the deeper interactions between metabolites. This multi-dimensional attention mechanism can help the model capture complex patterns in metabolic data, thereby improving the understanding of metabolic networks and the prediction ability of drug reactions. In the metabolic expression feature extraction network, this embodiment also introduces residual connection, which helps to alleviate the gradient vanishing problem and improve the training efficiency and expression ability of the model by directly passing the input information to the subsequent layer. By combining row, column and triangular attention with residual connection, the metabolic expression feature extraction network can fully exploit the details and deep features in metabolic data, providing more accurate support for drug resistance prediction and other tasks.

[0097] (4) Multi-modal feature interaction fusion

[0098] Through the above steps, the features of three different modalities of protein expression, gene expression and metabolic expression are obtained. Gene expression, protein expression and metabolite expression each carry different biological information, and by reasonably interacting and fusing these features, the potential rules of drug resistance can be more accurately captured.

[0099] As shown in the formula (1), the embodiment fuses the features of the three different modalities through a heterogeneous interaction attention module (Heterogeneous Attention) to perform heterogeneous interaction and adaptive weighted fusion, and finally outputs a multi-modal fusion feature. Figure 1

[0100] Specifically, first, the self-attention mechanism is used to generate query vectors, key vectors and value vectors for each modality feature. Assuming that the query vectors, key vectors and value vectors of the generated protein expression features are , , , the query vectors, key vectors and value vectors of the gene expression features are , , , and the query vectors, key vectors and value vectors of the metabolic expression features are , , .

[0101] Then, the generated vectors are used for heterogeneous interaction to obtain the features after interaction:

[0102]

[0103]

[0104]

[0105] wherein P, G, M are the protein expression features, gene expression features and metabolic expression features after heterogeneous interaction, , , , , , are the interaction weights, is the dimension of the query vector, and Softmax represents a normalization function.

[0106] Finally, the protein expression features, gene expression features and metabolic expression features after heterogeneous interaction are adaptively weighted and fused to obtain a multi-modal fusion feature:

[0107]

[0108] ​, and are adaptive weights of corresponding modalities, F is the multi-modal fusion feature.

[0109] The protein, gene and metabolic expression are respectively extracted by a single network model, which can better capture the specific information of each modality, and separate processing can reduce the noise and interference generated by direct fusion of multi-modal data. After using different networks to extract features and then fusing, it can ensure that the key information of each modality will not be covered by the data of other modalities. After independent extraction of features and then interactive fusion, the synergistic effect of multi-modal can be realized. The fusion feature can capture the complex interaction between gene expression, protein and metabolite, thereby improving the understanding and prediction accuracy of the biological system.

[0110] 4. Drug molecule feature extraction

[0111] First, based on the drug SMILES, a molecular graph (Molecular Graph) and a molecular fingerprint (Molecular Fingerprints) are generated to represent the structure of the drug molecule. The molecular graph represents the drug molecule as a graph structure, where atoms are nodes and chemical bonds are edges, which can more intuitively reflect the connection relationship between molecules. Molecular fingerprints, such as ECFP fingerprints (Extended Connectivity Fingerprints), can effectively capture the ring structure, branch structure and connection relationship between atoms in the drug molecule. By recursively considering the structure information of each atom and its neighborhood, ECFP fingerprints can describe the topological structure of the molecule in the form of a fixed-length binary vector. In this embodiment, the molecular graph and molecular fingerprint are generated directly by calling the relevant functions of the RDKit library, which is a commonly used function library in the field, and the specific details are not repeated here.

[0112] Then, the molecular graph feature is extracted by a molecular graph representation model, and the extracted molecular graph feature is spliced with the molecular fingerprint to obtain the final drug molecule feature.

[0113] For example, Figure 2As shown, the molecular graph representation model in this embodiment adopts a contrastive learning mechanism, including two different graph neural networks: GAT (Graph Attention Network) and GCN (Graph Convolutional Network), for generating two different embedding representations for each molecular graph. The model is trained using data from the ZINC database (a drug molecule database created and maintained by UCSF). During model training, a weight sharing layer is added to connect the output layers of GAT and GCN, and a similarity output layer is set after the weight sharing layer. By using contrastive learning, the two embedding representations of the same molecular graph are considered as positive pairs, while the two embedding representations of different molecular graphs are considered as negative pairs. By maximizing the similarity of positive pair representations and minimizing the similarity of negative pair representations, the model gradually learns the key features of molecular structures. The weight sharing layer fuses the feature representations generated by GAT and GCN by sharing the same network parameters. The contrastive learning mechanism uses a contrastive loss function to optimize the distance difference between positive and negative pairs. The optimization goal is to minimize the difference between the feature representations of the same molecular graph and maximize the feature difference between different molecular graphs.

[0114] 5. Drug resistance prediction

[0115] As shown, a Multilayer Perceptron (MLP) is constructed to input the multi-modal fusion features based on protein expression, gene expression, and metabolic expression, as well as drug molecule features into the MLP, and output drug resistance probability values. The larger the value, the stronger the drug resistance of the tumor tissue to the drug. Figure 1

[0116] MLP is a feedforward artificial neural network, including an input layer, multiple hidden layers, and an output layer. Each hidden layer contains several neurons, and each neuron maps the input to the output through an activation function. This network is widely used in classification, regression, and pattern recognition tasks, and its training process usually uses the backpropagation algorithm to minimize the prediction error. The MLP used to output drug resistance probability values in this embodiment is fully trained and verified in a previously constructed dataset, which can ensure the accuracy of the prediction results to a certain extent. Regarding the construction of the dataset, as previously mentioned, a large number of drug experiments were conducted by establishing PDX models of tumor tissues, and a comprehensive dataset combining multi-dimensional biological information and drug resistance representation was constructed.

[0117] Example Two

[0118] ​Based on the above method, the embodiment provides a drug resistance prediction system, which mainly comprises a drug molecule feature acquisition module, a protein expression feature acquisition module, a gene expression feature extraction module, a metabolic expression feature extraction module, a multi-modal feature fusion module and a drug resistance prediction module.

[0119] The drug molecule feature acquisition module acquires molecular graph features based on a molecular graph generated by a target drug SMILES, and splices the molecular graph features with molecular fingerprints, and the splicing result is taken as the drug molecule features.

[0120] The protein expression feature acquisition module, the gene expression feature extraction module and the metabolic expression feature extraction module respectively generate corresponding protein expression features, gene expression features and metabolic expression features based on protein expression data, gene expression data and metabolic expression data of a target tissue cell.

[0121] The multi-modal feature fusion module performs heterogeneous interaction and adaptive weighted fusion on the generated protein expression features, gene expression features and metabolic expression features through a heterogeneous interaction attention mechanism, and finally outputs multi-modal fusion features.

[0122] The drug resistance prediction module is used for outputting a drug resistance prediction result of the target tissue cell to the target drug according to the input drug molecule features and multi-modal fusion features.

[0123] The above system can execute the drug resistance prediction method based on contrast learning and multi-modal fusion described in embodiment one, has the corresponding functional modules and beneficial effects of the method, and the technical details not described in detail in this embodiment can be referred to the drug resistance prediction method based on contrast learning and multi-modal fusion provided in embodiment one of the present application.

[0124] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software plus a general hardware platform, and of course, it can also be realized by hardware. Based on such understanding, the above technical solutions or the part that contributes to the related art can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0125] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, but not limited to them; under the idea of the present application, the technical features of the above examples or different examples can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of the present application as described above, which are not provided in details for simplicity; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting drug resistance based on contrastive learning and multimodal fusion, characterized in that, Including the following steps: S1. Generate corresponding molecular maps and molecular fingerprints based on the SMILES of the target drug; S2. Using the molecular graph representation model trained based on the contrastive learning mechanism, extract molecular graph features and splice them with molecular fingerprints to form drug molecule features; S3. Obtain protein expression data, gene expression data, and metabolic expression data of the target tissue cells; S4. Preprocess the protein expression data, gene expression data, and metabolic expression data respectively to obtain the protein expression matrix, gene expression matrix, and metabolic expression matrix. S5. Extract protein expression features, gene expression features, and metabolic expression features based on the protein expression matrix, gene expression matrix, and metabolic expression matrix, respectively, and perform interactive fusion of the expression features of the three modalities; S6. Input the multimodal fusion features after interactive fusion and the drug molecule features into the drug resistance prediction network to obtain the drug resistance prediction results of the target tissue cells to the target drug.

2. The drug resistance prediction method as described in claim 1, characterized in that, In step S2: The molecular graph representation model comprises a parallel Graph Attention Network (GAT) and a Graph Convolutional Network (GCN) to generate two distinct embedding representations of the molecular graph. The model employs a contrastive learning mechanism, adding a weight-sharing layer during training, connected to the output layers of both GAT and GCN. A similarity output layer is then added after the weight-sharing layer. Through contrastive learning, two embedding representations of the same molecular graph are considered positive pairs, while two embedding representations of different molecular graphs are considered negative pairs. By maximizing the similarity of positive pairs while minimizing the similarity of negative pairs, key features of the molecular structure are learned progressively. The weight-sharing layer fuses the feature representations generated by GAT and GCN by sharing the same network parameters. The contrastive learning mechanism uses a contrastive loss function to optimize the distance difference between positive and negative pairs. The optimization objective is to minimize the difference between feature representations of the same molecular graph and maximize the feature difference between different molecular graphs.

3. The drug resistance prediction method as described in claim 1, characterized in that, In step S4: Principal component analysis is used to reduce the dimensionality of protein expression data to form a protein expression matrix. One-dimensional gene expression data is continuously segmented into several gene expression fragments to form a gene expression matrix. One-dimensional metabolic expression data is continuously segmented into several metabolic expression fragments to form a metabolic expression matrix.

4. The drug resistance prediction method as described in claim 1, characterized in that, In step S5: Protein expression features are obtained using a protein expression feature extraction network, which includes a deep convolutional module, a channel attention module, and a spatial attention module connected in sequence. First, the deep convolutional module extracts features from the input protein expression matrix layer by layer using multiple one-dimensional convolutional layers. Next, the multidimensional original feature vector output by the deep convolutional module is input to the channel attention module, where channel weights are generated using a channel attention mechanism. These channel weights are then multiplied by the original feature vector, thus weighting each channel and highlighting the features of important channels in the original feature vector. Next, the feature vector output by the channel attention module is input to the spatial attention module, where spatial weights are generated using a spatial attention mechanism. These spatial weights are then multiplied by the feature vector output by the channel attention module, dynamically adjusting the weights of different regions in the feature vector. Finally, the feature vector output by the spatial attention module is used as the protein expression feature.

5. The drug resistance prediction method as described in claim 1, characterized in that, In step S5, gene expression features are obtained using a gene expression feature extraction network. The gene expression feature extraction network uses a Transformer encoder, which consists of multiple repeating encoder blocks. Each encoder block includes input embedding, position encoding, multi-head self-attention mechanism, and feedforward neural network. The multi-head self-attention mechanism generates new embedding representations by calculating the relationship between different gene expression segments in the gene expression matrix.

6. The drug resistance prediction method as described in claim 1, characterized in that, In step S5, metabolic expression features are obtained using a metabolic expression feature extraction network, which includes a row attention module, a column attention module, and a triangular attention module connected in sequence. The row attention module generates a query vector, a key vector, and a value vector for each row of the input metabolic expression matrix. Then, it calculates the attention relevance using the query vector and the key vector through a dot product operation, and generates attention weights through the Softmax function. The attention weights are used to perform weighted combination of the value vectors, and finally, the output of the row attention module is obtained through a linear mapping. The column attention module generates a query vector, a key vector, and a value vector for each column of the feature vector output by the row attention module. Then, it calculates the attention relevance by using the query vector and the key vector through a dot product operation, and generates attention weights through the Softmax function. The attention weights are used to weightedly combine the value vectors, and finally, the output of the column attention module is obtained through a linear mapping. The triangular attention module uses a triangular attention mechanism to process the feature vectors output by the column attention module, and uses the processed feature vectors as metabolic expression features.

7. The method for predicting drug resistance as described in claim 1, characterized in that, In step S5, a heterogeneous interactive attention mechanism is used to perform heterogeneous interaction and adaptive weighted fusion of protein expression features, gene expression features, and metabolic expression features, ultimately outputting multimodal fusion features; the heterogeneous interactive attention mechanism: First, a query vector, key vector, and value vector are generated for each modality feature using a self-attention mechanism. Next, heterogeneous interactions are performed based on the generated vectors to obtain the interacted features: in, P, G, M These are the protein expression characteristics, gene expression characteristics, and metabolic expression characteristics after heterogeneous interaction, respectively. , , , , , For interactive weights, The dimension of the query vector is Softmax, which represents the normalization function. , , These are the query vector, key vector, and value vector representing the generated protein expression features, respectively. , , These are the query vector, key vector, and value vector representing gene expression characteristics, respectively. , , These are the query vector, key vector, and value vector representing metabolic expression features, respectively. Finally, the protein expression features, gene expression features, and metabolic expression features after heterogeneous interaction are adaptively weighted and fused to obtain multimodal fusion features.

8. The method for predicting drug resistance as described in claim 1, characterized in that, In step S6, the drug resistance prediction network uses a multilayer perceptron, which includes an input layer, multiple hidden layers, and an output layer. Each hidden layer contains several neurons, and each neuron maps the input to the output through an activation function. The output layer outputs a drug resistance probability value, and the larger the value, the stronger the drug resistance of the target tissue cells to the target drug.

9. A drug resistance prediction system based on the method of any one of claims 1 to 8, characterized in that, It includes a drug molecule feature acquisition module, a protein expression feature acquisition module, a gene expression feature extraction module, a metabolic expression feature extraction module, a multimodal feature fusion module, and a drug resistance prediction module; The drug molecule feature acquisition module acquires molecular graph features based on the molecular graph generated by the target drug SMILES, and splices it with the molecular fingerprint. The splicing result is used as the drug molecule feature. The protein expression feature acquisition module, gene expression feature extraction module, and metabolic expression feature extraction module generate corresponding protein expression features, gene expression features, and metabolic expression features based on the protein expression data, gene expression data, and metabolic expression data of the target tissue cells, respectively. The multimodal feature fusion module performs heterogeneous interaction and adaptive weighted fusion of the generated protein expression features, gene expression features and metabolic expression features through a heterogeneous interaction attention mechanism, and finally outputs multimodal fused features; The drug resistance prediction module is used to output the drug resistance prediction results of target tissue cells to the target drug based on the input drug molecule characteristics and multimodal fusion characteristics.

Images