Mechanism-aware collaborative anticancer drug combination prediction method and system

Abstract

This invention proposes a mechanism-aware synergistic anticancer drug combination prediction method and system, belonging to the field of biomedical technology. It includes: extracting node-level feature representations of the first and second drugs through a drug encoder, and extracting molecular representations of the cell line through a cell line encoder; generating an attention tensor based on a trilinear cross-modal synergistic strategy using a trilinear attention network to explicitly capture the synergistic dependencies of the drug combination in a specific cellular environment, and updating and enhancing the features through context fusion; finally, concatenating the enhanced features and outputting a synergistic score via a multilayer perceptron. This invention can directly model the joint interaction relationship between the first drug, the second drug, and the cell line, thereby accurately predicting the synergistic effect of the anticancer drug combination.

Description

Technical Field

[0001] This invention belongs to the field of biomedical technology, and in particular relates to a method and system for predicting synergistic anticancer drug combinations based on mechanism perception. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In cancer treatment, single-drug therapy often struggles to effectively block multiple pathogenic pathways involved in tumors and is prone to drug resistance and adverse reactions. Combination therapy, by acting simultaneously on multiple molecular targets and key signaling pathways, can enhance efficacy while reducing the risk of drug resistance, and has become one of the important strategies for cancer treatment. Against this backdrop, accurately predicting synergistic drug combinations not only helps in screening potentially highly effective combinations but also significantly improves the efficiency and safety of combination therapy design, providing a scientific basis for personalized precision medicine.

[0004] Currently, numerous computational methods have emerged for predicting synergistic drugs. These existing methods typically involve first modeling the effects of individual drugs separately in a specific cell line, and then fusing the drug pair representations later. This modeling approach inherently ignores the combined interaction between drug A, drug B, and the cell line, making it difficult to characterize the synergistic mechanism produced when two drugs act together in a specific cellular environment. This results in insufficient model representation of synergistic effects and limited predictive performance. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a mechanism-aware synergistic anticancer drug combination prediction method and system, which can directly model the joint interaction relationship between the first drug, the second drug and the cell line, and thus accurately predict the synergistic effect of the anticancer drug combination.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0007] The first aspect of this invention provides a mechanism-aware method for predicting synergistic anticancer drug combinations.

[0008] A mechanism-aware method for predicting synergistic anticancer drug combinations includes:

[0009] Obtain biological information on the drug combination and target cell line to be predicted, wherein the drug combination includes a first drug and a second drug;

[0010] The SMILES strings of the first drug and the second drug are converted into molecular graph representations. The molecular structures of the first drug and the second drug are extracted by a drug encoder to obtain node-level feature representations of the first drug and the second drug.

[0011] The cell line encoder is used to extract features from the multi-omics data of the target cell line to obtain the molecular characterization of the cell line.

[0012] Characterization is performed based on a trilinear cross-modal collaborative strategy: an attention tensor related to the first drug, the second drug, and the target cell line is generated through a trilinear attention network, and the three feature representations are updated and enhanced using the attention tensor through a trilinear context fusion module;

[0013] The enhanced first drug feature, second drug feature, and cell line molecular features are flattened and concatenated into a vector, and the predicted co-scoring is output through a multilayer perceptron.

[0014] Furthermore, feature extraction of the molecular structure of the target drug is performed using a drug encoder, including:

[0015] The initial atomic features are mapped to the embedding space through a linear layer; a graph convolutional network is used to encode the features of the molecular graph to obtain the feature representation of the molecular graph; a Transformer encoder is introduced into the graph convolutional network, taking the node embedding as input, and the interaction between nodes is modeled through a self-attention mechanism without introducing position encoding to maintain the permutation invariance of node ordering; the attention output is stabilized through residual connections and layer normalization, and the feature representation of each position is independently mapped and dimensionally transformed through a feedforward neural network to obtain node-level feature representation.

[0016] Furthermore, feature extraction is performed on the multi-omics data of the target cell lines using a cell line encoder, including:

[0017] Based on a pre-defined gene set, a multi-omics feature matrix is ​​constructed by integrating gene expression, gene mutation, and copy number variation information. A one-dimensional convolutional neural network is used to process the multi-omics feature matrix, and it is mapped to a unified embedding dimension through a linear layer to obtain feature representations. A pre-layer normalized Transformer encoder is introduced to model the global relationships between genes in the feature representations. A multi-head self-attention module is used to model the global dependencies between different features in the sequence, and a feedforward neural network is used to perform nonlinear feature transformation to generate molecular characterization of cell lines.

[0018] Furthermore, the generation of the attention tensor includes:

[0019] The first drug node-level feature representation, the second drug node-level feature representation, and the cell line molecular representation are aligned to the same embedding dimension through a fully connected layer. A trilinear attention network is constructed among the three features, and a trilinear attention mechanism is constructed using Tucker decomposition to decompose the high-dimensional tensor into a factor matrix and a low-dimensional core tensor. The unnormalized interaction tensor is normalized using the Softmax function to obtain the attention tensor.

[0020] Furthermore, the generation of the attention tensor also includes: imposing symmetry constraints on the input drug features to reflect the interaction between the first drug and the second drug.

[0021] Furthermore, the three feature representations are updated and enhanced using the attention tensor through the trilinear context fusion module, including:

[0022] Linear projection is performed on the first drug node-level feature representation, the second drug node-level feature representation, and the molecular representation of the cell line. Pairwise interactive representations are constructed using a broadcast mechanism to obtain joint representations of the cell line and the first drug, the cell line and the second drug, and the first drug and the second drug, respectively. The attention tensor is introduced into the subspace to weight and aggregate the context information from all features, resulting in cell environment context matrices, first drug context matrices, and second drug context matrices aggregated from the other two modalities. Residual connections and linear transformations are used to fuse the context information with the original feature representations, yielding enhanced first drug features, second drug features, and cell line molecular features.

[0023] Furthermore, the ReLU activation function is used between the layers of the multilayer perceptron structure, and a Dropout layer is introduced to alleviate overfitting.

[0024] A second aspect of the present invention provides a mechanism-aware synergistic anticancer drug combination prediction system.

[0025] A mechanism-aware synergistic anticancer drug combination prediction system, comprising:

[0026] The data acquisition module is configured to acquire biological information of the drug combination to be predicted and the target cell line, wherein the drug combination includes a first drug and a second drug;

[0027] The drug structure feature construction module is configured to: convert the SMILES strings of the first drug and the second drug into molecular graph representations, and extract features of the molecular structures of the first drug and the second drug respectively through a drug encoder to obtain the node-level feature representations of the first drug and the second drug.

[0028] The cell line molecular feature capture module is configured to: extract features from the multi-omics data of the target cell line through the cell line encoder to obtain the molecular characterization of the cell line;

[0029] The trilinear cross-modal collaborative processing module is configured to perform representation processing based on a trilinear cross-modal collaborative strategy. Specifically, it generates attention tensors related to the first drug, the second drug, and the target cell line through a trilinear attention network, and updates and enhances the three feature representations using the attention tensors through a trilinear context fusion module.

[0030] The drug combination prediction module is configured to flatten and concatenate the enhanced first drug feature, second drug feature, and cell line molecular features into a vector, and output the predicted collaborative score through a multilayer perceptron.

[0031] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a mechanism-aware synergistic anticancer drug combination prediction method as described in the first aspect of the present invention.

[0032] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a mechanism-aware synergistic anticancer drug combination prediction method as described in the first aspect of the present invention.

[0033] The above one or more technical solutions have the following beneficial effects:

[0034] This invention enhances drug molecule characterization by employing a drug encoder that integrates graph convolutional networks and Transformer encoders to learn local atomic environments and long-range dependencies. It also extracts cell line features by using a cell line encoder that combines one-dimensional convolutional neural networks and Transformer encoders to identify nearest-neighbor regulation and long-range interactions of genes. Furthermore, it models the joint interaction relationships between a first drug, a second drug, and the cell line based on a trilinear cross-modal synergistic strategy. In implementation, a three-dimensional attention map is constructed using a trilinear attention network to explicitly capture the synergistic dependencies of drug combinations in specific cellular environments, enabling deep-level mechanism modeling. Simultaneously, this map guides dynamic feature updates, reinforcing key synergistic information during the interaction process. Compared to existing technologies, this invention can directly model the joint interaction relationships between the first drug, the second drug, and the cell line, thereby accurately predicting the synergistic effects of anticancer drug combinations.

[0035] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0036] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0037] Figure 1 This is a flowchart of a mechanism-aware synergistic anticancer drug combination prediction method according to Embodiment 1 of the present invention.

[0038] Figure 2 This is a flowchart illustrating the implementation of DeepDrugs in Embodiment 1 of the present invention.

[0039] Figure 3 This is a schematic diagram of drug feature extraction based on a drug encoder in Embodiment 1 of the present invention.

[0040] Figure 4 This is a schematic diagram of cell line feature extraction based on cell line encoder in Embodiment 1 of the present invention.

[0041] Figure 5 This is a schematic diagram illustrating the operation of the trilinear attention network in Embodiment 1 of the present invention.

[0042] Figure 6 This is a schematic diagram of the fusion process of the three linear contexts in Embodiment 1 of the present invention. Detailed Implementation

[0043] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0044] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0045] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0046] Example 1

[0047] This embodiment discloses a mechanism-aware method for predicting synergistic anticancer drug combinations.

[0048] like Figure 1 As shown, a mechanism-aware synergistic anticancer drug combination prediction method includes:

[0049] Obtain biological information on the drug combination and target cell line to be predicted, wherein the drug combination includes a first drug and a second drug;

[0050] The SMILES strings of the first drug and the second drug are converted into molecular graph representations. The molecular structures of the first drug and the second drug are extracted by a drug encoder to obtain node-level feature representations of the first drug and the second drug.

[0051] The cell line encoder is used to extract features from the multi-omics data of the target cell line to obtain the molecular characterization of the cell line.

[0052] Characterization is performed based on a trilinear cross-modal collaborative strategy: an attention tensor related to the first drug, the second drug, and the target cell line is generated through a trilinear attention network, and the three feature representations are updated and enhanced using the attention tensor through a trilinear context fusion module;

[0053] The enhanced first drug feature, second drug feature, and cell line molecular features are flattened and concatenated into a vector, and the predicted co-scoring is output through a multilayer perceptron.

[0054] Based on the above process, this invention can directly model the combined interaction between the first drug, the second drug, and the cell line, thereby accurately predicting the synergistic effect of the anticancer drug combination. To facilitate understanding of the technical solution of this invention, the specific implementation methods of this invention will be further explained and described below.

[0055] like Figure 2 As shown, the implementation process of the mechanism-aware synergistic anticancer drug combination prediction method (hereinafter referred to as DeepDrugs) in this invention is as follows: The SMILES strings of two drugs and multi-omics data (gene expression, gene mutation, copy number variation) of the cell line are input into the drug encoder and cell line encoder for processing, respectively. A trilinear cross-modal synergistic module captures the interaction relationships between drug A (i.e., the first drug), drug B (i.e., the second drug), and the cell line, generating enhanced features. Subsequently, the three feature matrices are flattened and concatenated into a vector, and finally, the predicted synergistic score is output through a multilayer perceptron (MLP).

[0056] like Figure 3 As shown, the drug encoder uses a graph convolutional network to generate node-level representations to capture molecular graph information, while using a Transformer encoder to model long-range interactions between atoms.

[0057] like Figure 4As shown, during the processing of the cell line encoder, the cell line is characterized using multi-omics data. In this embodiment, this specifically includes gene expression, mutation, and copy number variation information for 1,061 key genes. The cell line encoder first captures short-range dependencies between genes using a one-dimensional convolutional neural network, and then introduces a Transformer encoder to enhance the representation ability of distant site interactions using an attention mechanism.

[0058] like Figure 5 As shown, the trilinear attention network is used to model the cooperative effects between drug pairs in a specific cellular environment and generate attention tensors. Figure 6 As shown, the trilinear context fusion module further enhances the feature representation by combining the attention tensor with the input drug A, drug B, and cell line features.

[0059] Furthermore, the main technical features of the present invention will be explained in further detail below:

[0060] Step S1: Construct the structural features of the drug based on the drug encoder.

[0061] To construct the structural features of the drugs, the RDKit tool was used to convert the drug's SMILES strings into corresponding molecular graph representations. Specifically, for each drug molecule, it was represented as an undirected graph. Among them, the atoms in the molecule constitute a set of nodes. any node For each atom, the set of edges It represents the chemical bond between atoms.

[0062] Based on the molecular graph described above, an adjacency matrix is ​​constructed. ;in, This indicates the number of atoms in a molecule. When atoms... With atoms When chemical bonds exist, ;otherwise .

[0063] Regarding node feature construction, for each node... The corresponding atom extracts its corresponding 74-dimensional feature vector. This feature vector encodes atom type, atomic degree, number of implicit hydrogen atoms, formal charge, number of free radical electrons, hybridization mode, total number of hydrogen atoms, and aromaticity information. From this, an initial molecular representation of the drug can be obtained. ,in This indicates the number of atoms in a drug molecule.

[0064] To further enhance feature representation capabilities, the drug encoder first maps the initial atomic features to a linear layer. The molecular graph is then embedded in a 3D embedding space, and subsequently feature-encoded using a graph convolutional network (GCN). The GCN... The node representation update rules for the layer are as follows:

[0065] ;

[0066] in, For the first The learnable weight matrix of a layer GCN This represents the adjacency matrix after adding self-connections; It is a diagonal matrix whose diagonal elements correspond to the adjacency matrix. The degree of each node in the equation; It is the first The node feature representation of the layer, where express Feature representation obtained through linear transformation mapping; This represents the ReLU activation function.

[0067] The molecular graph feature representation obtained from GCN is denoted as... ,in, For feature dimension, Indicates the number of layers in the GCN; Indicates after the first The molecular graph node feature matrix is ​​obtained after layer graph convolution. To capture the global dependencies between arbitrary nodes, a Transformer encoder is introduced on GCN, using node embeddings as input and a self-attention mechanism to model the interactions between nodes, without introducing positional encoding to maintain permutation invariance of node ordering. The calculation method of the Transformer encoder and the core attention mechanism are defined as follows:

[0068] ;

[0069] ;

[0070] ;

[0071] ;

[0072] in, The learnable parameter matrix; These represent the query vector, key vector, and value vector matrices extracted from the node features, respectively. Key vector matrix The feature dimension. In multi-head attention mechanisms, the first... The output representations of each attention head and the final multi-head attention representation are as follows:

[0073] ;

[0074] ;

[0075] in, , The weight matrix is ​​a learnable weight matrix; Indicates the number of attention heads. Indicates the first The features are calculated for each position. The attention output is then stabilized using residual connections and layer normalization. Building upon this, the encoder further introduces a feedforward neural network (FFN) to perform independent nonlinear mapping and dimensionality transformation on the feature representation at each location, and then integrates them using the same residual connections and layer normalization.

[0076] ;

[0077] ;

[0078] ;

[0079] in, For learnable weight matrix, For bias terms, It is the ReLU activation function. Through the above encoding process, drug A ( ) and drug B ( The node-level features of the molecule preserve the differences in importance between different atoms within the molecule, providing a foundation for more fine-grained interaction modeling in subsequent modules.

[0080] Step S2: Capture the molecular features of cell lines based on the cell encoder.

[0081] To capture the molecular characteristics of cell lines, a gene screening and multi-omics integration strategy was employed. Based on 978 marker genes from the L1000 platform, and combined with cancer-related target genes from the Therapeutic Target Database (TTD), after deduplication and deletion removal, 963 marker genes and 98 target genes were retained as a gene set. Based on this gene set, gene expression, gene mutation, and copy number variation (CNV) information were integrated from the Cancer Dependency Map (DepMap) database to construct multi-omics features. Expression and CNV were treated as continuous variables, mutations were one-hot encoded, and ordered by chromosomal position to maintain genomic structural consistency. The final multi-omics feature matrix provided a comprehensive molecular characterization. The final characterization of the cell lines is represented as follows: ,in It refers to the number of genes.

[0082] The cell line encoder first uses a one-dimensional convolutional neural network (i.e., 1D-CNNs) to process the cell line feature matrix. Each convolutional layer is followed by a max-pooling operation to progressively shorten the sequence length and increase the number of channels. The transformation of a convolutional layer can be represented as:

[0083] ;

[0084] in, , For learnable convolutional kernels; Represents the ReLU activation function. This represents the max pooling operation in the hidden layer. The number of input channels for the three convolutional layers are [3, 8, 16], and the number of output channels are [8, 16, 32]. The kernel size for all layers is fixed at 8, and the corresponding max pooling window size is [2, 2, 2]. 1D-CNNs output ( The total number of layers is mapped to a unified embedding dimension through linear layers. ), to obtain feature representation .

[0085] Subsequently, a pre-layer normalized Transformer encoder was introduced to... The global relationships between genes are modeled, and the gradient problem is alleviated by input layer normalization. The calculation process is as follows:

[0086] ;

[0087] in, This indicates that the input feature matrix The result after layer normalization.

[0088] Subsequently, the normalized features are input into the multi-head self-attention module to model the global dependencies between different features in the sequence. The computation process of multi-head self-attention and the feedforward neural network is as follows:

[0089] ;

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] ;

[0096] ;

[0097] in, , , , For learnable parameter matrix, , For bias terms; Key vector matrix The dimension; This represents a fully connected layer in a feedforward neural network.

[0098] After completing the multi-head self-attention computation, the attention output is added to the original input through residual connections, followed by layer normalization. The result is then fed into a feedforward neural network for nonlinear feature transformation, ultimately generating cell line representations. .

[0099] Step S3: Execute the trilinear cross-modal cooperative strategy based on the trilinear cross-modal cooperative module.

[0100] The trilinear cross-modal collaboration module consists of a trilinear attention network and a trilinear context fusion module. The trilinear attention network generates attention tensors associated with drug A, drug B, and cell line, while the trilinear context fusion module uses these attention tensors to update and enhance the three feature representations.

[0101] 1) Trilinear attention network.

[0102] The parameters input to the trilinear cross-modal cooperative module include drug A features. and characteristics of drug B (Generated by the drug encoder), and cell line characteristics. (Generated by cell line encoder), where and These represent the number of atoms in drug A and drug B, respectively. All inputs are aligned to the same embedding dimension using a fully connected layer.

[0103] ;

[0104] ;

[0105] ;

[0106] in, , , For learnable projection matrices, This represents the ReLU activation function.

[0107] Subsequently, a trilinear attention network is constructed among these three features:

[0108] ;

[0109] in, It is an unnormalized interaction tensor Learnable tensors. Operators express 1st-order modal product.

[0110] Learnable Tensors The high dimensionality of the third-order tensor makes the learning process difficult. To reduce the parameter and computational overhead of third-order tensor modeling, a Tucker decomposition mechanism is used to construct an efficient trilinear attention mechanism, which decomposes the high-dimensional tensor into a factor matrix and a low-dimensional core tensor, making the third-order tensor more efficient. This can be concisely represented as:

[0111] ;

[0112] in, For the factor matrix, The core tensor forms the decomposition process, which significantly improves computational efficiency while reducing parameter dimensionality. Furthermore, the Tucker decomposition introduces low-rank structure constraints, enabling the model to learn more generalizable cross-modal interaction patterns, thereby mitigating the risk of overfitting.

[0113] For learning tensors The calculations using the above decomposition method are as follows:

[0114] ;

[0115] in, For slicing parameters; Represents the learnable Tucker core tensor. , It is a learnable factor matrix.

[0116] Apply symmetry constraints to the input drug characteristics to reflect drug A ( ) and drug B ( The interaction between them. The learning tensor of Tucker decomposition. Included In the expression, the unnormalized interaction tensor It can be rewritten as:

[0117] ;

[0118] Attention Tensor The calculation formula is as follows:

[0119] ;

[0120] in, This represents a vectorization operator used to flatten a tensor into a one-dimensional vector. This indicates that the vector will be reconstructed into a form with size . The tensor.

[0121] 2) Trilinear context fusion.

[0122] To enable the model to capture the effects of drug combinations in a cell-specific context, trilinear attention weights are introduced to update and enhance drug and cell features. First, a linear projection is performed on each input feature:

[0123] ;

[0124] ;

[0125] ;

[0126] in, , , The projected matrix, , , This is a learnable weight matrix.

[0127] By constructing pairwise interactive representations, the joint representations of cell line and drug A were obtained. Joint representation of cell line and drug B The combined representation of drug A and drug B :

[0128] ;

[0129] ;

[0130] ;

[0131] in, , , The above addition operation is implemented through a broadcast mechanism. Specifically, it will... and Extend to the same three-dimensional tensor shape This allows each cell line characteristic to be paired element-wise with all characteristics of drug A. Subsequently, element-wise addition is used to obtain... This is used to encode the combined characteristics of all cell lines in combination with drug A. Similarly, (Depend on and (Construction) and (Depend on and The same broadcast mechanism is also used for computation in the construction.

[0132] Based on this, an attention tensor is introduced within each subspace. We perform weighted aggregation of contextual information from all features:

[0133] ;

[0134] ;

[0135] ;

[0136] in, Indicates cell line characteristic index. This represents the feature index of drug A. Indicates the characteristic index of drug B. Indicates the th in the cell line feature matrix The feature vector, the feature matrix of drug A The feature vector and the feature matrix of drug B Attention weights under a combination of feature vectors; ; , , These represent the context matrices of the cell environment, drug A, and drug B obtained from the aggregation of the other two modalities, respectively. Representation matrix The Middle Line 1 The elements of the column.

[0137] To fuse contextual information with the original feature representation, residual connections and linear transformations are used to update the aggregated features:

[0138] ;

[0139] ;

[0140] ;

[0141] in, , , The learnable parameter matrix; , , These represent cell line characteristics enhanced by drug A characteristics and drug B characteristics, drug A characteristics enhanced by cell line characteristics and drug B characteristics, and drug B characteristics enhanced by cell line characteristics and drug A characteristics, respectively. This represents a random deactivation operation, where some elements of the input features are randomly masked during training. Through these steps, each modality's features can fuse information from the other two modalities, thereby achieving feature updates and enhancements.

[0142] 3) Regression layer.

[0143] Cell line characteristics obtained from the trilinear cross-modal collaborative module Drug A characteristics Drug B characteristics After flattening and stitching, the data is input into a multilayer perceptron (MLP) for drug synergistic scoring prediction. To improve learning efficiency, the ReLU activation function is used between layers, and a Dropout layer is introduced to alleviate overfitting and enhance model robustness.

[0144] ;

[0145] in, This represents the dimension transformation operation that reconstructs a multidimensional feature tensor into a one-dimensional vector; It is an operation that concatenates multiple input vectors along a specified dimension to form a unified representation; This represents a multilayer perceptron with the following structure: linear layer → ReLU activation function → Dropout → linear layer → ReLU activation function → Dropout → linear layer. This structure progressively reduces the feature dimension and does not use an activation function in the last layer, directly outputting a continuous collaborative scoring scalar.

[0146] Step S4: Effect evaluation.

[0147] This invention mainly evaluates the model from the following three aspects: 1) comparing the performance of this invention with other algorithms on two datasets; 2) comparing the performance of this invention with other algorithms on leave-one-out validation experiments; 3) conducting ablation experiments on the key modules of this invention.

[0148] Predicting the synergistic score of anticancer drug combinations is a regression problem; therefore, this invention employs mean squared error (MSE). ), root mean square error ( Mean absolute error ( ), coefficient of determination ( Pearson correlation coefficient ( Spearman correlation coefficient ( As evaluation indicators, each indicator is calculated as follows:

[0149] ;

[0150] ;

[0151] ;

[0152] ;

[0153] ;

[0154] ;

[0155] in, and They represent the first The true and predicted collaborative scores for each sample Represents the total number of samples. and represent the mean of the actual collaborative score and the predicted collaborative score, respectively. Indicates the first The ranking difference of each sample.

[0156] Based on the above methods, compared with the prior art, the present invention achieves the following technical breakthroughs:

[0157] 1) DeepDrugs outperforms existing methods on different datasets.

[0158] This embodiment evaluates the performance on the DrugComb and O'Neil datasets. The DrugComb dataset contains 330,917 samples, involving 387 drugs and 170 cancer cell lines. The synergistic effect is quantified using the S-score, and the training, validation, and test sets are divided in an 8:1:1 ratio. The O'Neil dataset contains 18,929 samples, consisting of 38 drugs and 32 cell lines, and uses the Loewe score as the synergistic effect indicator, with the training and test sets divided in a 4:1 ratio. This embodiment compares DeepDrugs with five current state-of-the-art methods: JointSyn, SynergyX, MGAE-DC, DeepSynergy, and DFFNDDS. Model performance is measured using the aforementioned six regression evaluation metrics to comprehensively reflect the model's predictive capabilities across different aspects.

[0159] As shown in Table 1, on the DrugComb dataset, DeepDrugs achieved the lowest MSE (78.96), a reduction of 1.1% to 54.2% compared to other methods; simultaneously, its RMSE also reached the lowest value (8.89), indicating that this model has a significant advantage in controlling large and systematic errors. More importantly, DeepDrugs performed best in correlation metrics, with a coefficient of determination (... The correlation coefficient was 0.848, and the Pearson correlation coefficient was ( The correlation coefficient was 0.921, and the Spearman correlation coefficient was ( The score is 0.853, which is better than existing methods in capturing the linear trend of collaborative scoring and ranking consistency.

[0160] Table 1. Performance comparison of DeepDrugs (in this invention) with baseline methods on the DrugComb dataset.

[0161]

[0162] As shown in Table 2, DeepDrugs also exhibits the best performance on the O'Neil dataset. It is 127.93, compared to JointSyn ( = 131.45, p<0.01) represents a 2.7% improvement, compared to SynergyX ( =137.26, p<0.001) improved by 5.7%, and compared to DFFNDDS ( = 277.36) achieved an error reduction of over 50%. Furthermore, DeepDrugs in (0.662) (0.815) and The (0.786) indicator achieved the highest value, further verifying its robustness under more limited combination space conditions.

[0163] Table 2 Performance comparison of DeepDrugs (in this invention) with baseline methods on the O'neil dataset.

[0164]

[0165] 2) In leave-one-out validation experiments, DeepDrugs showed stronger generalization ability compared to existing methods.

[0166] In real-world synergistic drug combination discovery, models need to be able to predict previously untested drug combinations, chemical entities newly entering clinical trials, and synergistic effects in patient-derived cell lines not observed during training. To evaluate the practical application value of DeepDrugs in these clinically significant scenarios, this embodiment designs three rigorous validation strategies based on the O'Neil dataset to simulate different real-world application challenges:

[0167] a. Leave out drug combination validation: Rigorous 10-fold cross-validation was performed on 583 drug combinations to simulate the ability to predict untested combinations in a known drug library;

[0168] b. Validation with one cell line: Use all samples from each cell line as the test set and the remaining samples as the training set to evaluate the model’s ability to generalize to cell environments from which no patients were found.

[0169] c. Leave-one-out drug validation: For each drug, all relevant samples are used as the test set, and the remaining samples are used as the training set to evaluate the model's ability to predict new chemical entities from scratch.

[0170] In all validation strategies, to avoid data leakage, all models were trained from scratch in each round of experiments, without using any pre-trained weights or fine-tuning strategies. Model performance passed... , , , , as well as An evaluation was conducted, and the results were directly compared with five baseline methods under the same conditions.

[0171] As shown in Table 3, DeepDrugs achieved [results] in the validation of drug combinations. = 140.16、 =11.78、 =8.26、 =0.625、 =0.792 and The optimal result of 0.778 represents a 1.6% reduction in error compared to JointSyn and a more than 36% reduction compared to DFFNDDS and DeepSynergy. This result demonstrates that the model possesses strong combinatorial reasoning capabilities, enabling reliable ranking of a large number of unexplored drug combinations within the approved or investigational drug set.

[0172] Table 3. Performance comparison of DeepDrugs and baseline methods in retention drug combination validation.

[0173]

[0174] As shown in Table 4, in the cell line validation with one cell line, this embodiment sequentially used 32 cell lines as the test set to evaluate the model's generalization ability in unknown cellular environments. DeepDrugs once again achieved optimal performance ( = 220.49、 =14.46、 =10.58、 =0.404、 =0.667、 = 0.651), compared to the suboptimal model JointSyn, the MSE decreased by 3.9%. This represents a 4.9% improvement. This indicates that DeepDrugs can effectively transfer learned synergistic patterns to entirely new genomic contexts, supporting its personalized screening applications in patient-derived models.

[0175] Table 4. Performance comparison of the DeepDrugs method of this invention with the baseline method in leave-one-cell line validation.

[0176]

[0177] As shown in Table 5, DeepDrugs achieved [results] in the drug validation process. = 266.88、 =15.97、 =11.82 and = 0.164, a 7.0% reduction in error compared to JointSyn and approximately 38.2% reduction compared to MGAE-DC. Despite dealing with unseen molecular graph structures and chemical substructures, the model still maintains good predictive ability.

[0178] Table 5. Performance comparison of the DeepDrugs method of this invention with the baseline method in leave-one-out drug validation.

[0179]

[0180] 3) DeepDrugs conducts ablation experiments on core technologies.

[0181] To investigate the roles of each component module in modeling drug synergistic effects in a specific cellular context, this embodiment conducted targeted ablation experiments on the DrugComb dataset. Four model variants were constructed by systematically removing or simplifying key modules, and all models were trained from scratch under the same settings:

[0182] aw / o TCMSM: Remove the trilinear cross-modal collaboration module and replace the original ternary joint attention mechanism with feature concatenation;

[0183] bw / o DE-TE: Removes the Transformer encoder from the drug encoder and relies only on the atomic structure features extracted by the graph convolutional network;

[0184] cw / o CLE-TE: Removes the Transformer encoder from the cell line encoder, retaining only the short-range gene regulation patterns extracted by the 1D convolutional neural network;

[0185] d.LandmarkGenesOnly: Only 963 signature genes from the L1000 platform were used as input, and 98 additional cancer-related target genes were removed from the TTD database.

[0186] As shown in Table 6, the experimental results indicate that removing any key module leads to a decrease in performance across all evaluation metrics. When the trilinear cross-modal collaborative module used to model the ternary interaction between drug A, drug B, and cell line is removed... It rose to 80.38 (+1.8%). It dropped to 0.846 ( 0.24%), indicating that the apparent trilinear attention mechanism is crucial for capturing synergistic effects. Removing the Transformer encoder in the drug encoder (w / o DE-TE) and the Transformer encoder in the cell line encoder (w / o CLE-TE) respectively... The performance improved to 82.74 (+4.79%) and 82.21 (+4.12%), while the correlation index also decreased significantly, indicating that modeling long-range dependencies in chemical molecular maps and genomic expression data plays an important role. Although using only marker genes resulted in a relatively small performance decrease, it still showed a consistent downward trend. = 79.98, +1.29%; = 0.846, 0.24%; =0.920, 0.11%; = 0.852, (0.12%), indicating that the expanded gene set can provide additional pathway-level information.

[0187] Table 6. Ablation Experiment Study of Key Components of DeepDrugs in this Invention

[0188]

[0189] In summary, the trilinear cross-modal synergistic module, the bi-branch Transformer encoder, and the extended genetic prior information all play irreplaceable roles in the model and function synergistically support accurate and mechanistically interpretable predictions of drug synergistic effects in the cellular environment.

[0190] Example 2

[0191] This embodiment discloses a mechanism-aware synergistic anticancer drug combination prediction system.

[0192] A mechanism-aware synergistic anticancer drug combination prediction system, comprising:

[0193] The data acquisition module is configured to acquire biological information of the drug combination to be predicted and the target cell line, wherein the drug combination includes a first drug and a second drug;

[0194] The drug structure feature construction module is configured to: convert the SMILES strings of the first drug and the second drug into molecular graph representations, and extract features of the molecular structures of the first drug and the second drug respectively through a drug encoder to obtain the node-level feature representations of the first drug and the second drug.

[0195] The cell line molecular feature capture module is configured to: extract features from the multi-omics data of the target cell line through the cell line encoder to obtain the molecular characterization of the cell line;

[0196] The trilinear cross-modal collaborative processing module is configured to perform representation processing based on a trilinear cross-modal collaborative strategy. Specifically, it generates attention tensors related to the first drug, the second drug, and the target cell line through a trilinear attention network, and updates and enhances the three feature representations using the attention tensors through a trilinear context fusion module.

[0197] The drug combination prediction module is configured to flatten and concatenate the enhanced first drug feature, second drug feature, and cell line molecular features into a vector, and output the predicted collaborative score through a multilayer perceptron.

[0198] Example 3

[0199] The purpose of this embodiment is to provide a computer-readable storage medium.

[0200] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of a mechanism-aware synergistic anticancer drug combination prediction method as described in Embodiment 1 of this disclosure.

[0201] Example 4

[0202] The purpose of this embodiment is to provide an electronic device.

[0203] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a mechanism-aware synergistic anticancer drug combination prediction method as described in Embodiment 1 of this disclosure.

[0204] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0205] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0206] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A mechanism-aware predictive method for synergistic anticancer drug combinations, characterized in that, include: Obtain biological information on the drug combination and target cell line to be predicted, wherein the drug combination includes a first drug and a second drug; The SMILES strings of the first drug and the second drug are converted into molecular graph representations. The molecular structures of the first drug and the second drug are extracted by a drug encoder to obtain node-level feature representations of the first drug and the second drug. The cell line encoder is used to extract features from the multi-omics data of the target cell line to obtain the molecular characterization of the cell line. Characterization is performed based on a trilinear cross-modal collaborative strategy: an attention tensor related to the first drug, the second drug, and the target cell line is generated through a trilinear attention network, and the three feature representations are updated and enhanced using the attention tensor through a trilinear context fusion module; The enhanced first drug feature, second drug feature, and cell line molecular features are flattened and concatenated into a vector, and the predicted co-scoring is output through a multilayer perceptron. The molecular structure of the target drug is characterized by a drug encoder, including: Atomic initial features are mapped to the embedding space through linear layers; a graph convolutional network is used to encode the features of the molecular graph to obtain the feature representation of the molecular graph; a Transformer encoder is introduced into the graph convolutional network, taking node embeddings as input, and the interaction between nodes is modeled through a self-attention mechanism without introducing positional encoding to maintain the permutation invariance of node ordering; the attention output is stabilized through residual connections and layer normalization, and a feedforward neural network is used to perform independent nonlinear mapping and dimensional transformation on the feature representation of each position to obtain node-level feature representation; Feature extraction of multi-omics data from target cell lines was performed using a cell line encoder, including: Based on a pre-defined gene set, a multi-omics feature matrix is ​​constructed by integrating gene expression, gene mutation, and copy number variation information. A one-dimensional convolutional neural network is used to process the multi-omics feature matrix, and it is mapped to a unified embedding dimension through a linear layer to obtain feature representations. A pre-layer normalized Transformer encoder is introduced to model the global relationships between genes in the feature representations. A multi-head self-attention module is used to model the global dependencies between different features in the sequence, and a feedforward neural network is used to perform nonlinear feature transformation to generate molecular characterization of cell lines. The generation of the attention tensor includes: The first drug node-level feature representation, the second drug node-level feature representation, and the cell line molecular representation are aligned to the same embedding dimension through a fully connected layer; a trilinear attention network is constructed among the three features, and a trilinear attention mechanism is constructed using Tucker decomposition to decompose the high-dimensional tensor into a factor matrix and a low-dimensional core tensor; the unnormalized interaction tensor is normalized using the Softmax function to obtain the attention tensor. The three feature representations are updated and enhanced using an attention tensor through a trilinear context fusion module, including: Linear projection is performed on the first drug node-level feature representation, the second drug node-level feature representation, and the molecular representation of the cell line. Pairwise interactive representations are constructed using a broadcast mechanism to obtain joint representations of the cell line and the first drug, the cell line and the second drug, and the first drug and the second drug, respectively. The attention tensor is introduced into the subspace to weight and aggregate the context information from all features, resulting in the cell environment context matrix, the first drug context matrix, and the second drug context matrix obtained from the aggregation of the other two modalities. Residual connections and linear transformations are used to fuse the context information with the original feature representations, yielding enhanced first drug features, second drug features, and molecular features of the cell line.

2. The mechanism-aware synergistic anticancer drug combination prediction method as described in claim 1, characterized in that, The generation of the attention tensor further includes: imposing symmetry constraints on the input drug features to reflect the interaction between the first drug and the second drug.

3. The mechanism-aware synergistic anticancer drug combination prediction method as described in claim 1, characterized in that, The multilayer perceptron uses the ReLU activation function between its layers and introduces a Dropout layer to alleviate overfitting.

4. A mechanism-aware synergistic anticancer drug combination prediction system, employing the mechanism-aware synergistic anticancer drug combination prediction method as described in any one of claims 1-3, characterized in that, include: The data acquisition module is configured to acquire biological information of the drug combination to be predicted and the target cell line, wherein the drug combination includes a first drug and a second drug; The drug structure feature construction module is configured to: convert the SMILES strings of the first drug and the second drug into molecular graph representations, and extract features of the molecular structures of the first drug and the second drug respectively through a drug encoder to obtain the node-level feature representations of the first drug and the second drug. The cell line molecular feature capture module is configured to: extract features from the multi-omics data of the target cell line through the cell line encoder to obtain the molecular characterization of the cell line; The trilinear cross-modal collaborative processing module is configured to perform representation processing based on a trilinear cross-modal collaborative strategy. Specifically, it generates attention tensors related to the first drug, the second drug, and the target cell line through a trilinear attention network, and updates and enhances the three feature representations using the attention tensors through a trilinear context fusion module. The drug combination prediction module is configured to flatten and concatenate the enhanced first drug feature, second drug feature, and cell line molecular features into a vector, and output the predicted collaborative score through a multilayer perceptron.

5. A computer-readable storage medium having a program stored thereon, characterized in that, When executed by a processor, the program implements the steps in a mechanism-aware synergistic anticancer drug combination prediction method as described in any one of claims 1-3.

6. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the mechanism-aware synergistic anticancer drug combination prediction method as described in any one of claims 1-3.

Images