A drug response prediction method and system based on cross-view contrast alignment

Abstract

The application discloses a drug response prediction method and system based on cross-view contrast alignment, and the method comprises the following steps: acquiring multi-omics data, drug molecular fingerprint data and drug-cell line response data of a cell line, constructing an attribute view according to the multi-omics data and the drug molecular fingerprint data, and constructing a correlation view according to the drug-cell line response data; encoding the attribute view and the correlation view respectively to obtain attribute view representation and correlation view representation; obtaining aligned attribute view representation and aligned correlation view representation through cross-view contrast alignment according to the attribute view representation and the correlation view representation; adaptively fusing the aligned attribute view representation and the aligned correlation view representation to obtain fused representation, and predicting a response value of a drug to a cell line according to the fused representation.

Description

Technical Field

[0001] This invention belongs to the field of drug response prediction technology, specifically relating to a drug response prediction method and system based on cross-view comparison alignment. Background Technology

[0002] Accurate prediction of anticancer drug response is a core task of precision oncology. The core of drug response prediction lies in predicting the half-maximal inhibitory concentration (IC50) of a drug on tumor cell lines by analyzing their molecular characteristics and drug molecular structures. However, high-throughput omics data are characterized by high dimensionality and heterogeneity, while drug molecular structure data exhibit diverse formats. Effectively integrating these multi-source, heterogeneous data and learning discriminative characterizations from them remains a major challenge in current drug response prediction.

[0003] In existing technologies, drug response prediction methods can be mainly classified into the following categories: The first category of methods is based on traditional machine learning and matrix factorization. Methods like SRMF and CaDRReS decompose the drug-cell line response matrix into low-order latent factors and combine this with similarity regularization for prediction. However, these methods only utilize the first-order interaction relationships of the response matrix, making it difficult to capture complex nonlinear patterns in multi-omics features and drug structures, resulting in limited prediction accuracy.

[0004] The second category of methods is based on deep learning feature fusion, such as DeepCDR and NeRD. These methods utilize neural networks to extract features from multi-omics data and drug molecule structures, and then fuse them through splicing or attention mechanisms. However, their feature extraction process is mainly based on the intrinsic attributes of individual entities (i.e., attribute views), ignoring the rich correlation information inherent in the drug-cell line response relationship itself.

[0005] The third type of method is based on graph neural networks and contrastive learning. Existing works such as GraphCDR have attempted to introduce contrastive learning into drug response prediction, but in existing methods, contrastive learning is usually used as an auxiliary loss term, which only plays a regularization role and fails to fundamentally bridge the distribution differences between the attribute space and the association space.

[0006] In summary, existing technologies still have significant shortcomings in drug response prediction tasks, mainly manifested in the following aspects: (1) the attribute view and the associated view are separated, and there is a lack of cross-view information alignment mechanism; (2) contrastive learning is only used as an auxiliary constraint and fails to effectively bridge the distribution differences of multimodal heterogeneous data; (3) there is a lack of adaptive multi-view fusion mechanism. These shortcomings greatly limit the application effect of existing methods in drug response prediction. Summary of the Invention

[0007] This invention aims to address the shortcomings of existing technologies by proposing a drug response prediction method and system based on cross-view contrast alignment. This method and system can accurately predict drug-cell line response data (IC50 value) by utilizing multi-omics data such as gene expression profiles, copy number variations, DNA methylation, and gene mutations of cell lines, as well as drug molecular fingerprint data, through cross-view contrast learning and adaptive fusion.

[0008] To achieve the above objectives, the present invention provides the following solution: A drug response prediction method based on cross-view contrast alignment includes the following steps: Acquire multi-omics data of cell lines, drug molecule fingerprint data, and drug-cell line response data; construct an attribute view based on the multi-omics data and the drug molecule fingerprint data; and construct an association view based on the drug-cell line response data. The attribute view and the associated view are encoded respectively to obtain the attribute view representation and the associated view representation; Based on the attribute view representation and the associated view representation, the aligned attribute view representation and the aligned associated view representation are obtained through cross-view comparison and alignment. The aligned attribute view representation and the aligned associated view representation are adaptively fused to obtain a fused representation, and the drug response value to the cell line is predicted based on the fused representation.

[0009] Preferably, the method for constructing the attribute view includes: For each omics modality of the multi-omics data and the drug molecule fingerprint type in the drug molecule fingerprint data, a K-nearest neighbor graph is constructed based on cosine similarity, where nodes are cell lines or drug entities and edges are the similarity relationships between nodes in the corresponding modality feature space. The attribute view is constructed based on the K-nearest neighbor graph.

[0010] Preferably, the method for constructing the associated view includes: Based on the drug-cell line response matrix, drug-cell line pairs with half-maximal inhibitory concentration values ​​below a threshold are marked as sensitive, and a binary relationship edge between the drug-cell line response is constructed based on the marking results. A heterogeneous graph is constructed based on the binary relation edges of the responses, and singular value decomposition is performed on the drug-cell line response matrix to obtain the preferred embeddings as initial features of the nodes: in, This represents the drug-cell line response matrix. express The left singular vector matrix, Represents a singular value diagonal matrix. express The right singular vector matrix, express The transpose of the matrix, Indicates by and Initial node features obtained by concatenating based on entity dimensions express The square root matrix; The associated view is constructed based on the initial features of the nodes.

[0011] Preferably, the method for constructing the attribute view representation includes: encoding the node features in the attribute view through a multi-layer graph attention network to obtain multimodal features. in, Represents a node v In the l The representation of layer +1, l This indicates the layer number of the graph attention network. K This indicates the number of attention heads in the graph. k Indicates the attention head number, σ Represents a non-linear activation function. Represents a node v The set of neighboring nodes, u This represents the neighboring nodes. Indicates the first k Nodes in each attention head u For nodes v Attention weights Indicates the first k The learnable weight matrix for each attention head. Represents a node u In the l Layer representation; The multimodal features are fused into a unified embedding using a semantic attention mechanism to obtain the attribute view representation: in, Represents the attribute view representation. m This indicates the type number of the omics modality or drug molecule fingerprint. Indicates the first m Semantic attention weights for each modality Indicates the first m The feature representation of each modality, softmax represents the normalization exponential function, and the semantic attention formula contains... v Represents the semantic attention vector. Represents the learnable weight matrix. This represents the bias vector.

[0012] Preferably, the method for constructing the associated view representation includes: propagating the preferred embedding on the heterogeneous graph through a graph attention network to generate the associated view representation.

[0013] Preferably, the method for obtaining the aligned attribute view representation and the aligned associated view representation includes: For the same entity, the attribute view representation and the associated view representation are taken as positive sample pairs; For other different entities, the attribute view representation and the association view representation are used as negative sample pairs; Based on the positive sample pairs and the negative sample pairs, InfoNCE contrastive loss is used for optimization to semantically align the attribute view representation and the associated view representation, thereby obtaining the aligned attribute view representation and the aligned associated view representation.

[0014] Preferably, the method for obtaining the fusion characterization includes: Based on the aligned attribute view representation and the aligned associated view representation, the fusion weight is calculated through a gating mechanism; Based on the fusion weight, the aligned attribute view representation and the aligned associated view representation are weighted and summed to obtain the fusion representation.

[0015] The present invention also discloses a drug response prediction system based on cross-view comparison alignment, wherein the system applies the above-mentioned method and includes: a view construction module, a feature encoding module, a feature alignment module and a fusion prediction module; The view building module is used to acquire multi-omics data of cell lines, drug molecule fingerprint data, and drug-cell line response data, construct an attribute view based on the multi-omics data and the drug molecule fingerprint data, and construct an association view based on the drug-cell line response data; The feature encoding module encodes the attribute view and the associated view respectively to obtain the attribute view representation and the associated view representation; The feature alignment module obtains the aligned attribute view representation and the aligned associated view representation by cross-view comparison alignment based on the attribute view representation and the associated view representation; The fusion prediction module is used to adaptively fuse the aligned attribute view representation and the aligned association view representation to obtain a fused representation, and predict the drug response value to the cell line based on the fused representation.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) Cross-view contrast alignment mechanism bridges the semantic gap between attributes and associations: The cross-view contrast alignment module proposed in this invention is the core innovation that distinguishes it from existing technologies. Existing methods usually use contrast learning as an auxiliary loss term, which only plays a regularization role. This invention innovatively performs end-to-end semantic alignment between attribute views and association views through InfoNCE contrast loss, fundamentally bridging the distribution differences between attribute space and association space. Experiments show that after removing this module, all indicators decrease (PCC decreases from 0.8312 to 0.7771, R² decreases from 0.6898 to 0.6034), fully verifying the effectiveness of this module.

[0017] (2) Adaptive gating fusion enables dynamic integration of multi-view information: This invention uses a gating mechanism to dynamically learn the fusion weights of attribute views and associated views for each entity. Ablation experiments show that after removing adaptive gating fusion, PCC decreases from 0.8312 to 0.8252, and R² decreases from 0.6898 to 0.6803, verifying the contribution of this module.

[0018] (3) Enhanced response signal based on SVD-based association priority embedding: In the construction of association view, the present invention introduces the priority embedding extracted by SVD decomposition as the initial feature of the node, which directly encodes the global structural information of the drug-cell line response matrix, enabling the association view to capture the potential semantics of response relationship more effectively.

[0019] (4) Multi-omics semantic attention enables effective fusion of heterogeneous features: This invention introduces a semantic attention mechanism into the attribute view encoding to automatically learn the relative importance of each omics modality. The full-modality configuration achieves optimal performance (PCC 0.8312, R² 0.6898, MSE 0.1950), and the lack of any single modality leads to a decrease in performance.

[0020] (5) Strong generalization ability in cold start scenarios: The present invention achieves a PCC of 0.7972 and an R² of 0.6318 in the cell line cold start setting, which is superior to the existing methods in all indicators. This is due to the synergistic effect of cross-view contrast alignment and adaptive gating fusion. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2The diagram illustrates the ablation experiment and omics data contribution analysis in this embodiment of the invention. (a) shows the ablation experiment results of the cross-view comparison alignment module and the adaptive gating fusion module, and (b) shows the contribution analysis results of gene expression profiles, copy number variations, gene mutations and DNA methylation data. Detailed Implementation

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

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1 In this embodiment, as Figure 1 As shown, a drug response prediction method based on cross-view contrast alignment includes the following steps: S1. Acquire multi-omics data of cell lines, drug molecule fingerprint data, and drug-cell line response data. Construct an attribute view based on the multi-omics data and drug molecule fingerprint data, and construct an association view based on the drug-cell line response data.

[0026] In this embodiment, firstly, multi-omics data of cell lines (including gene expression profiles, copy number variation data, DNA methylation data, and gene mutation data), drug molecular fingerprint data (including various molecular fingerprint encoding methods such as MACCS, PubChem, and CDK), and drug-cell line response data (characterized by the half-maximal inhibitory concentration (IC50) value) are collected from publicly available databases. Gene expression profiles reflect the transcriptional activity status of the cell lines, copy number variation and mutation data reflect genomic-level abnormalities, drug molecular fingerprints encode the chemical structure and properties of the drug, and the IC50 value directly measures the inhibitory effect of the drug on the cell lines.

[0027] In this embodiment, the method for constructing the attribute view includes: for each omics modality of multi-omics data and the drug molecule fingerprint type in drug molecule fingerprint data, constructing K-nearest neighbor graphs based on cosine similarity, where nodes are cell lines or drug entities, and edges represent the similarity relationship between nodes in the corresponding modality feature space; and constructing the attribute view based on the K-nearest neighbor graphs. Specifically, for each omics modality (including gene expression profiles, copy number variations, DNA methylation, gene mutations, etc.) and drug molecule fingerprint type, constructing K-nearest neighbor graphs (K equals 10) based on cosine similarity, where nodes are cell lines or drug entities, and edges represent the similarity relationship between nodes in the corresponding modality feature space.

[0028] In this embodiment, the method for constructing the associated view includes: based on the drug-cell line response matrix, marking drug-cell line pairs with a half-maximal inhibitory concentration (WMC) value below a threshold as sensitive, and constructing binary relation edges for drug-cell line responses based on the marking results; constructing a heterogeneous graph based on the binary relation edges, and performing singular value decomposition on the drug-cell line response matrix to obtain preferred embeddings as initial features for nodes. in, This represents the drug-cell line response matrix. express The left singular vector matrix, Represents a singular value diagonal matrix. express The right singular vector matrix, express The transpose of the matrix, Indicates by and Initial node features obtained by concatenating based on entity dimensions express The square root matrix.

[0029] Specifically, based on the drug-cell line response matrix (where matrix elements represent the IC50 values ​​of each drug for each cell line, with rows corresponding to cell lines and columns to drugs), drug-cell line pairs with IC50 values ​​below a threshold (e.g., 1 μmol / L) are marked as sensitive. Singular value decomposition (SVD) is performed on the response matrix to obtain 128-dimensional preferred embeddings as initial node features for the associated view, and the associated view is constructed based on these initial node features.

[0030] S2. Encode the attribute view and the associated view respectively to obtain the attribute view representation and the associated view representation.

[0031] In this embodiment, the method for constructing the attribute view representation includes: encoding the node features in the attribute view through a multi-layer graph attention network to obtain multimodal features. in, Represents a node v In the l The representation of layer +1, l This indicates the layer number of the graph attention network. K This indicates the number of attention heads in the graph. k Indicates the attention head number, σ Represents a non-linear activation function. Represents a node v The set of neighboring nodes, u This represents the neighboring nodes. Indicates the first k Nodes in each attention head u For nodes v Attention weights Indicates the first k The learnable weight matrix for each attention head. Represents a node u In the l Layer representation.

[0032] Multimodal features are fused into a unified embedding using a semantic attention mechanism to obtain an attribute view representation: in, Represents the attribute view representation. m This indicates the type number of the omics modality or drug molecule fingerprint. Indicates the first m Semantic attention weights for each modality Indicates the first m The feature representation of each modality, softmax represents the normalization exponential function, and the semantic attention formula contains... v Represents the semantic attention vector. Represents the learnable weight matrix. This represents the bias vector.

[0033] In this embodiment, the method for constructing an association view representation includes: propagating the preferential embedding on a heterogeneous graph through a graph attention network to generate an association view representation that incorporates global signals of drug-cell line response relationships.

[0034] S3. Based on the attribute view representation and the associated view representation, obtain the aligned attribute view representation and the aligned associated view representation by cross-view comparison alignment.

[0035] The method for obtaining aligned attribute view representations and aligned related view representations includes: for the same entity, the attribute view representation and the related view representation are used as positive sample pairs; for other different entities, the attribute view representation and the related view representation are used as negative sample pairs; based on the positive sample pairs and the negative sample pairs, InfoNCE contrastive loss is used for optimization so that the attribute view representation and the related view representation are semantically aligned, thereby obtaining aligned attribute view representations and aligned related view representations.

[0036] In this embodiment, for each cell line or drug entity, its attribute view representation obtained from gene expression profiles and drug molecule fingerprint encoding, and its association view representation obtained from drug-cell line response matrix SVD decomposition encoding, constitute positive sample pairs, while the representations of other entities in the same mini-batch are considered negative samples. Alignment is performed using InfoNCE contrastive loss. in, Let represent the cross-view comparison alignment loss, where log represents the natural logarithm function, exp represents the exponential function, and sim represents the cosine similarity. i Indicates the current entity number. j This indicates the entity number that is being compared within the same small batch. Representing entities i The attribute view representation, Representing the same entity i The associated view representation, This represents the associated view representation of entity j. τ This represents the temperature scaling parameter. By maximizing the similarity between positive sample pairs (attribute view representations and association view representations of the same entity) and minimizing the similarity between negative sample pairs (representations of different entities), the representation spaces of attribute information such as gene expression profiles and drug molecule fingerprints are semantically aligned with the drug-cell line response association information, thus effectively bridging the distribution differences between attribute information and association information. During model prediction, for the cell line-drug pair to be predicted, its gene expression profile and drug molecule fingerprint are input into the attribute view encoder. Through the contrast-aligned representation space mapping, the IC50 predicted value, which is semantically consistent with the association view, can be output.

[0037] S4. Adaptively fuse the aligned attribute view representation and the aligned associated view representation to obtain the fused representation, and predict the drug response value of the cell line based on the fused representation.

[0038] The methods for obtaining the fusion representation include: calculating the fusion weight based on the aligned attribute view representation and the aligned associated view representation through a gating mechanism; and obtaining the fusion representation by weighted summation of the aligned attribute view representation and the aligned associated view representation based on the fusion weight.

[0039] In this embodiment, complementary information from the attribute view (features extracted from gene expression profiles, drug molecule fingerprints, etc.) and the association view (features extracted from the drug-cell line response matrix) is adaptively integrated. For each cell line or drug entity, a fusion weight is calculated through a gating mechanism: in, Represents the gate vector, This represents the Sigmoid activation function. The learnable weight matrix representing the gating mechanism. This represents the aligned attribute view representation. This represents the aligned associated view representation. This represents the bias vector. This represents the fused representation obtained by gating and weighting the aligned attribute view representation and the aligned associated view representation. This represents element-wise multiplication. Each dimension of the gating vector ranges from 0 to 1, adaptively controlling the contribution ratio of the attribute view (derived from multi-omics data such as gene expression profiles and drug molecular fingerprints) and the association view (derived from the drug-cell line IC50 response matrix) to the entity. The fused representations are concatenated into a cell line-drug pair representation, which is input into the prediction network and outputs the predicted IC50 value of the drug for that cell line. The smaller the predicted IC50 value, the stronger the inhibitory effect of the drug on the cell line. The prediction results can be directly compared with the true IC50 value to evaluate the model accuracy.

[0040] Subsequently, the IC50 prediction value of the drug response is output based on the fused characterization (obtained through adaptive fusion of multi-omics attribute views such as gene expression profiling and drug molecular fingerprinting, and drug-cell line response correlation views). The overall training objective consists of two parts: in, Indicates the overall training objective; The predicted loss is represented by the difference between the predicted IC50 value and the actual IC50 value (from drug-cell line response data) calculated using mean squared error (MSE). This represents the cross-view contrast alignment loss, ensuring that the attribute view representations encoded from multi-omics data such as gene expression profiles and drug molecular fingerprints are semantically aligned with the association view representations encoded from the IC50 response matrix. This represents the balancing hyperparameter, adjusting the weights between the two losses. Through joint optimization, the model learns to accurately predict IC50 values ​​based on input data such as gene expression profiles and drug molecular fingerprints, while maintaining semantic consistency between the attribute view and the association view.

[0041] Example 2 This embodiment uses actual data to illustrate the process of the present invention.

[0042] 1. Dataset Preparation This embodiment uses the GDSC (Genomics of Drug Sensitivity in Cancer) database as the experimental data source. The GDSC database contains response profile data for 732 cancer cell lines and 174 drugs, covering a variety of cancer types and drug categories.

[0043] Multi-omics data include gene expression profiles, copy number variations (CNVs), DNA methylation, and gene mutations. Drug molecular characteristics are encoded using molecular fingerprints (MACCS, PubChem, CDK). Preprocessing steps: (1) Standardize gene expression data; (2) Encode molecular fingerprints in binary form; (3) Construct a drug-cell line response matrix based on IC50 values, with IC50 < 1 μM as the sensitivity threshold.

[0044] 2. Model Training This embodiment uses the CCLDRP model proposed in this invention, and the main steps are as follows: (1) Graph construction stage: The attribute view constructs a K-nearest neighbor graph (K=10) based on cosine similarity, and the association view constructs a heterogeneous graph based on the SVD decomposition of the response matrix. The attribute view uses multi-layer GAT encoding (8 heads), and the association view uses GAT encoding (128 dimensions).

[0045] (2) Contrast Alignment Stage: Temperature Parameters Lost in InfoNCE Contrast Set to 0.5, compare alignment weights Set it to 0.1.

[0046] (3) Gated fusion and prediction stage: The cell line and drug characterization inputs are spliced ​​into the prediction network, and the MSE loss is used for regression training. The Adam optimizer is used with a learning rate of 10⁻³ for 200 epochs.

[0047] 3. Experimental Results and Analysis (1) Overall performance comparison: Compared with 6 existing methods, this embodiment achieves the best performance in all indicators: PCC 0.8312, R2 0.6898, MSE 0.1950. As shown in Table 1.

[0048] Table 1 .

[0049] (2) Cold start performance: PCC 0.7972 and R2 0.6318 under cold start of cell line.

[0050] (3) Ablation experiment: After removing cross-view contrast alignment, PCC decreased from 0.8312 to 0.7771 (a decrease of 6.5%), and R2 decreased from 0.6898 to 0.6034 (a decrease of 12.5%); after removing adaptive gating fusion, PCC decreased to 0.8252; after removing both modules, PCC decreased to 0.8077.

[0051] like Figure 2 As shown in Figure (a), the ablation experiment results of the cross-view comparison alignment module and the adaptive gating fusion module are presented. The normalized values ​​of all indicators in the complete configuration are 100%; after removing any module or removing both modules at the same time, all indicators change, which is consistent with the above ablation experiment results.

[0052] (4) Data ablation: Full-modal configuration is optimal; the lack of any single learning mode will lead to performance degradation.

[0053] like Figure 2 As shown in Figure (b), the contribution analysis results of gene expression profiles, copy number variation, gene mutation, and DNA methylation data are presented. All indicators in the full modality configuration were normalized to 100%; after removing any single omics modality, all indicators changed, indicating that all four types of omics data contributed to the drug response prediction results.

[0054] Example 3 In this embodiment, a drug response prediction system based on cross-view comparison alignment includes: a view construction module, a feature encoding module, a feature alignment module, and a fusion prediction module.

[0055] The view construction module acquires multi-omics data of cell lines, drug molecule fingerprint data, and drug-cell line response data. It constructs attribute views based on the multi-omics and drug molecule fingerprint data, and association views based on the drug-cell line response data. The feature encoding module encodes the attribute views and association views respectively, obtaining attribute view representations and association view representations. The feature alignment module aligns the attribute view representations and association view representations through cross-view comparison, obtaining aligned attribute view representations and aligned association view representations. The fusion prediction module adaptively fuses the aligned attribute view representations and aligned association view representations to obtain a fused representation, and predicts the drug's response value to the cell line based on the fused representation.

[0056] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A drug response prediction method based on cross-view contrast alignment, characterized in that, Includes the following steps: Acquire multi-omics data of cell lines, drug molecule fingerprint data, and drug-cell line response data; construct an attribute view based on the multi-omics data and the drug molecule fingerprint data; and construct an association view based on the drug-cell line response data. The attribute view and the associated view are encoded respectively to obtain the attribute view representation and the associated view representation; Based on the attribute view representation and the associated view representation, the aligned attribute view representation and the aligned associated view representation are obtained through cross-view comparison and alignment. The aligned attribute view representation and the aligned associated view representation are adaptively fused to obtain a fused representation, and the drug response value to the cell line is predicted based on the fused representation.

2. The drug response prediction method based on cross-view comparison alignment according to claim 1, characterized in that, The method for constructing the property view includes: For each omics modality of the multi-omics data and the drug molecule fingerprint type in the drug molecule fingerprint data, a K-nearest neighbor graph is constructed based on cosine similarity, where nodes are cell lines or drug entities and edges are the similarity relationships between nodes in the corresponding modality feature space. The attribute view is constructed based on the K-nearest neighbor graph.

3. The drug response prediction method based on cross-view comparison alignment according to claim 1, characterized in that, The method for constructing the associated view includes: Based on the drug-cell line response matrix, drug-cell line pairs with half-maximal inhibitory concentration values ​​below a threshold are marked as sensitive, and a binary relationship edge between the drug-cell line response is constructed based on the marking results. A heterogeneous graph is constructed based on the binary relation edges of the responses, and singular value decomposition is performed on the drug-cell line response matrix to obtain the preferred embeddings as initial features of the nodes: in, This represents the drug-cell line response matrix. express The left singular vector matrix, Represents a singular value diagonal matrix. express The right singular vector matrix, express The transpose of the matrix, Indicates by and Initial node features obtained by concatenating based on entity dimensions express The square root matrix; The associated view is constructed based on the initial features of the nodes.

4. The drug response prediction method based on cross-view comparison alignment according to claim 3, characterized in that, The method for constructing the attribute view representation includes: encoding the node features in the attribute view through a multi-layer graph attention network to obtain multimodal features. in, Represents a node v In the l The representation of layer +1, l This indicates the layer number of the graph attention network. K This indicates the number of attention heads in the graph. k Indicates the attention head number, σ Represents a non-linear activation function. Represents a node v The set of neighboring nodes, u This represents the neighboring nodes. Indicates the first k Nodes in each attention head u For nodes v Attention weights Indicates the first k The learnable weight matrix for each attention head. Represents a node u In the l Layer representation; The multimodal features are fused into a unified embedding using a semantic attention mechanism to obtain the attribute view representation: in, Represents the attribute view representation. m This indicates the type number of the omics modality or drug molecule fingerprint. Indicates the first m Semantic attention weights for each modality Indicates the first m The feature representation of each modality, softmax represents the normalization exponential function, and the semantic attention formula contains... v Represents the semantic attention vector. Represents the learnable weight matrix. This represents the bias vector.

5. The drug response prediction method based on cross-view contrast alignment according to claim 4, characterized in that, The method for constructing the associated view representation includes: propagating the preferred embedding on the heterogeneous graph through a graph attention network to generate the associated view representation.

6. The drug response prediction method based on cross-view contrast alignment according to claim 1, characterized in that, The methods for obtaining the aligned attribute view representation and the aligned associated view representation include: For the same entity, the attribute view representation and the associated view representation are taken as positive sample pairs; For other different entities, the attribute view representation and the association view representation are used as negative sample pairs; Based on the positive sample pairs and the negative sample pairs, InfoNCE contrastive loss is used for optimization to semantically align the attribute view representation and the associated view representation, thereby obtaining the aligned attribute view representation and the aligned associated view representation.

7. The drug response prediction method based on cross-view contrast alignment according to claim 1, characterized in that, The methods for obtaining the fusion characterization include: Based on the aligned attribute view representation and the aligned associated view representation, the fusion weight is calculated through a gating mechanism; Based on the fusion weight, the aligned attribute view representation and the aligned associated view representation are weighted and summed to obtain the fusion representation.

8. A drug response prediction system based on cross-view contrast alignment, wherein the system applies the method described in any one of claims 1-7, characterized in that, include: The module includes a view construction module, a feature encoding module, a feature alignment module, and a fusion prediction module. The view building module is used to acquire multi-omics data of cell lines, drug molecule fingerprint data, and drug-cell line response data, construct an attribute view based on the multi-omics data and the drug molecule fingerprint data, and construct an association view based on the drug-cell line response data; The feature encoding module encodes the attribute view and the associated view respectively to obtain the attribute view representation and the associated view representation; The feature alignment module obtains the aligned attribute view representation and the aligned associated view representation by cross-view comparison alignment based on the attribute view representation and the associated view representation; The fusion prediction module is used to adaptively fuse the aligned attribute view representation and the aligned association view representation to obtain a fused representation, and predict the drug response value to the cell line based on the fused representation.

Images