A Method and System for Predicting Drug Synergistic Effects Based on Cell Semantic-Guided Dynamic Biological Pathway Weights

Abstract

This invention discloses a method and system for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights. This method integrates semantic priors from a large language model, dynamic biological pathway gating, and a triplet self-attention mechanism. It dynamically regulates gene-pathway mapping relationships through cellular semantic features to form soft constraints, and applies biological hard constraints to network weights through a binary mask matrix constructed from gene-pathway subordination relationships. The combination of these two approaches, while adhering to the underlying facts of molecular biology, introduces semantic soft gating to adapt to the heterogeneity of the cellular microenvironment. On this basis, the complex interactions between drugs and cells are captured through a sequenced triplet self-attention mechanism, achieving accurate identification of drug synergy and deep coupling of cross-modal information, thereby improving the biological interpretability and generalization ability of the prediction.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and bioinformatics, and in particular relates to a method and system for predicting drug synergistic effects based on dynamic biological pathway weights guided by cellular semantics. Background Technology

[0002] In the field of artificial intelligence applications in drug development and precision medicine, drug prediction technologies mainly include disease-based drug prediction, drug interaction prediction, and drug synergy prediction. Among them, drug interaction prediction aims to avoid the risks of combined drug use, while drug synergy prediction aims to screen highly effective therapeutic drug combinations.

[0003] Common methods for predicting drug synergy include three main categories: graph neural network-based methods, network pharmacology-based methods, and auxiliary task enhancement methods.

[0004] Graph neural network-based methods, such as SDDSYnergy, model microscopic graph structures using graph neural networks to capture substructural information of drug molecules. Combined with cellular features, the prediction head outputs drug synergy predictions. HANSynergy and KGANSynergy encode drugs and cells as graph structures of biological entities, using deep neural networks to complete intra-graph information transfer and feature fusion, and then using the fused representation to predict downstream drug synergies. However, these methods, due to their static network structure modeling, cannot adapt to weight variations in different cell line backgrounds and are prone to oversmoothing, leading to a loss of discriminative power in the extracted features and affecting prediction performance.

[0005] Network pharmacology-based methods, such as SyndrumNET and RareDDIE, construct sub-networks of diseases and drugs, embedding drugs into gene networks using network proximity and meta-learning techniques, respectively. They then combine these with gene regulatory networks to mine the associations between drugs and biological functions, learning the mechanisms of action to predict drug synergistic effects. However, these methods rely on simple network proximity calculations, failing to simulate the complex feedback regulation and compensatory mechanisms after drugs enter the human body, and cannot intuitively represent the synergistic mechanisms of drugs at the functional pathway level, resulting in poor model interpretability.

[0006] Methods based on auxiliary task enhancement, such as MultiComb and ComboNet, introduce auxiliary tasks like target prediction and drug sensitivity. These methods simultaneously complete multiple downstream auxiliary tasks while performing the main task of predicting drug synergy, aiming to enhance the predictive ability of synergistic effects and compensate for the limitations of single feature representation. However, these methods rely heavily on additional labeled data such as drug sensitivity and target prediction. Furthermore, if the datasets for the auxiliary tasks are of poor quality, the noise they contain can significantly interfere with the training process of the main task. Moreover, these methods simply concatenate the output features of the main task and each auxiliary task dimensionally, without exploring the inherent semantic connections and feature complementarity between the main task and auxiliary tasks, or among the auxiliary tasks themselves, by understanding the biological mechanisms of drug synergy. They lack deep semantic alignment and fusion design across tasks, which can easily lead to information redundancy or feature conflicts in the feature space, making it difficult to leverage the enhancing effect of auxiliary tasks on the main task. Summary of the Invention

[0007] The purpose of this invention is to address the problems existing in the prior art by proposing a method and system for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights. This method integrates semantic priors from a large language model, dynamic biological pathway gating, and a triplet self-attention mechanism. It utilizes cellular semantic features generated by the large language model to dynamically regulate biologically meaningful gene-pathway mapping relationships to form soft constraints, while applying biological hard constraints to network weights through a binary mask matrix constructed from gene-pathway subordination relationships. This combination, while adhering to the underlying facts of molecular biology, introduces semantic soft gating to adapt to the heterogeneity of the cellular microenvironment. Based on this, a sequential triplet self-attention mechanism captures the complex interactions between drugs and cells, enabling the identification of drug synergies. Furthermore, through deep coupling of cross-modal information, it improves the biological interpretability and generalization ability of the prediction.

[0008] To achieve the above objectives, the present invention adopts the following technical solutions:

[0009] A method for predicting drug synergistic effects based on cellular semantic-guided dynamic biological pathway weights, the method comprising:

[0010] Obtain the semantic features E of the target cell line cell ;

[0011] According to the semantic feature E cell Generate biological pathway activation weights S corresponding to the target cell line. gate ;

[0012] Obtain the gene relationship characteristics X of the first drug A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line. cell ;

[0013] A binary mask matrix is ​​constructed based on the hierarchical relationship between genes and biological pathways, and the binary mask matrix is ​​used to apply hard constraints to the weights of the neural network.

[0014] Based on the weights after hard constraints and the biological pathway activation weights S gate X-ray the genetic relationship characteristics of the first drug and the second drug. A X B and the gene expression characteristics X of the target cell line cell Mapping to the pathway level respectively, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell ;

[0015] The first pathway feature P A Second pathway characteristics P B and cell pathway characteristics P cell Interactive fusion is performed to obtain fusion features;

[0016] Based on the fusion characteristics, the synergistic effect score of the first drug and the second drug on the target cell line is predicted.

[0017] In the aforementioned method for predicting drug synergy based on cell semantic-guided dynamic biological pathway weights, the semantic features E of the target cell line are obtained in the following manner. cell :

[0018] Gene expression data of the target cell line were retrieved from publicly available bioinformatics databases;

[0019] The gene expression data is input into a pre-trained cellular big language model, and the attention mechanism of the cellular big language model is used to capture the semantic associations between genes, outputting the semantic feature E. cell .

[0020] In the aforementioned drug synergy prediction method based on cell semantic-guided dynamic biological pathway weights, biological pathway activation weights S corresponding to the target cell line are generated through semantic pathway gating. gate :

[0021] S gate =σ(f(E cell1 ))

[0022] Where f is a linear mapping function and σ is an activation function.

[0023] In the aforementioned method for predicting drug synergistic effects based on cell semantic-guided dynamic biological pathway weights, the semantic pathway gating is implemented using a two-layer neural network:

[0024] S gate =Sigmoid(W2*GELU(W1*E cell +b1)+b2)

[0025] Where W1 and W2 are learnable weight matrices, b1 and b2 are bias terms, and GELU is the activation function.

[0026] In the aforementioned method for predicting drug synergy based on cell semantic-guided dynamic biological pathway weights, the gene relationship feature X of the first drug is obtained in the following manner. A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line cell :

[0027] Gene relationship data of the first drug, gene relationship data of the second drug, and gene expression data of the target cell line were retrieved from publicly available bioinformatics databases.

[0028] Gene relationship data of the first drug and gene relationship data of the second drug are respectively organized into gene relationship matrices;

[0029] The gene expression data of the target cell line were organized into a gene expression matrix;

[0030] Each matrix is ​​standardized to obtain the gene relationship feature X of the first drug. A The genetic relationship characteristics of the second drug X B and gene expression characteristics of the target cell line X cell .

[0031] In the aforementioned method for predicting drug synergistic effects based on cell semantic-guided dynamic biological pathway weights, each element of the binarized mask matrix is ​​used to indicate whether the corresponding gene belongs to the corresponding biological pathway, and each feature is mapped to the pathway level in the following manner:

[0032] Using the weights after hard constraints, the first drug gene relationship feature X is analyzed. A Second drug gene relationship characteristics X B Gene expression characteristics X cell Linear transformations are performed to map the features from the gene dimension to the pathway dimension, resulting in preliminary pathway features.

[0033] The preliminary pathway characteristics are compared with the biological pathway activation weight S. gate By coupling, the first path feature P is obtained. A Second pathway characteristics P B and cell pathway characteristics P cell .

[0034] In the aforementioned method for predicting drug synergy based on cell semantic-guided dynamic biological pathway weights, the neural network weights are hard-constrained in the following way to obtain the hard-constrained weights:

[0035] W masked =W path Mask

[0036] Where Mask is the binary mask matrix, W path These are the learnable weights of a neural network;

[0037] The first pathway feature P is obtained in the following manner. A Second pathway characteristics P B and cell pathway characteristics P cell :

[0038] P=Linear(X, W masked )·S gate

[0039] S gate This indicates the activation weight of the biological pathway;

[0040] Linear(X, W masked ), indicating a linear transformation of X;

[0041] X-ray the genetic relationship characteristics of the first drug A Gene relationship characteristics of the second drug X B Gene expression characteristics of the target cell line X cell Substituting X into each, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell .

[0042] In the above-mentioned drug synergy prediction method based on cell semantic-guided dynamic biological pathway weights, the fusion features are obtained in the following manner:

[0043] The first pathway feature P A Second pathway characteristics P B and cell pathway characteristics P cell Concatenate them into an input sequence of length 3;

[0044] The input sequence is fed into a multi-head self-attention layer, which captures the interaction relationships among the three components at the pathway level through a self-attention mechanism, and outputs the fused features:

[0045] Attn_Out=MultiHead(Q, K, V)

[0046] Where Q, K, V=[P A P B P cell W Q , [P A P B P cell W K , [P A P B P cell W V .

[0047] In the aforementioned drug synergy prediction method based on cell semantic-guided dynamic biological pathway weights, predicting the synergy score between the first drug and the second drug on the target cell line includes:

[0048] The fused features are flattened and input into the multilayer perceptron prediction head;

[0049] The multilayer perceptron prediction head outputs the cooperative score:

[0050] Score = MLP(Flatten(Attn_Out))

[0051] The multilayer perceptron prediction head (MLP) comprises three linear mapping layers.

[0052] A drug synergy prediction system based on cell semantic-guided dynamic biological pathway weights includes:

[0053] The semantic extraction module is used to obtain the semantic features E of the target cell line. cell ;

[0054] The semantic path gating module is used to control the semantic path based on the semantic feature E. cell Generate biological pathway activation weights S corresponding to the target cell line. gate ;

[0055] The data acquisition module is used to acquire the gene relationship characteristics X of the first drug. A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line. cell ;

[0056] The biological pathway mask gating module is used to construct a binary mask matrix based on the hierarchical relationship between genes and biological pathways. This binary mask matrix is ​​then used to apply hard constraints to the weights of the neural network, and the hard-constrained weights are combined with the activation weights S of the biological pathways. gate X-ray the genetic relationship characteristics of the first drug and the second drug. A X Band the gene expression characteristics X of the target cell line cell Mapping to the pathway level respectively, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell ;

[0057] The interactive fusion module is used to integrate the first path feature P A Second pathway characteristics P B and cell pathway characteristics P cell Interactive fusion is performed to obtain fusion features;

[0058] The prediction output module is used to predict the synergistic effect score of the first drug and the second drug on the target cell line based on the fusion characteristics.

[0059] The advantages of this invention are:

[0060] (1) Using cell semantic embedding as a control signal to dynamically generate biological pathway activation weights, the biological network topology is adapted to the background features of different cell lines, which solves the problem that traditional static structural modeling cannot express the changes in cell line background weights, and greatly improves the adaptability of the model to different cell microenvironments.

[0061] (2) Combining the biological hard constraints of gene-pathway attribution with the soft gating weights assigned by cell semantics allows the feature extraction process to follow the underlying facts of molecular biology and capture the heterogeneity information of the cell microenvironment. This avoids the problems of traditional methods being detached from biological priors or unable to adapt to environmental differences, and improves the scientificity and rationality of model prediction.

[0062] (3) By reading the pathway activation weights generated by the dynamic pathway gating module, the key biological pathways that contribute the most to the synergistic effect of drugs can be traced, and the core biological mechanism of the synergistic effect of drug combinations can be clarified. This solves the problem of poor interpretability and inability to intuitively reflect the synergistic logic at the pathway level in traditional methods, and provides a feasible biological basis for clinical combination drug use and new drug development.

[0063] (4) It does not require a large amount of additional auxiliary annotation data such as drug sensitivity and target prediction, thus avoiding the problem of poor quality auxiliary task data interfering with the main task. Attached Figure Description

[0064] Figure 1 This is a flowchart illustrating the construction of the prediction model in the drug synergy prediction method based on cell semantics-guided dynamic biological pathway weights, as described in this embodiment of the invention.

[0065] Figure 2This is a diagram of the architecture of the prediction model in the drug synergy prediction method based on cell semantic-guided dynamic biological pathway weights provided in the embodiments of the present invention.

[0066] Figure 3 The diagram shown is a biological pathway mask gating architecture that integrates hard constraint mapping and semantically gated biological relationship mapping provided in an embodiment of the present invention.

[0067] Figure 4 This is an architecture diagram of the serialized triple interaction attention mechanism layer provided in an embodiment of the present invention;

[0068] Figure 5 The flowchart of the drug synergy prediction method based on cell semantic-guided dynamic biological pathway weights provided in the embodiments of the present invention is shown.

[0069] Figure 6 This is a system module diagram of a drug synergy prediction system based on cell semantic-guided dynamic biological pathway weights.

[0070] Figure 7 The performance test results of the prediction model implemented in this scheme on three indicators based on five-fold cross-validation are shown in the figure.

[0071] Figure 8 This is a scatter plot showing the predicted values ​​and actual values ​​from the prediction model of this scheme.

[0072] Figure 9 This is a graph showing the distribution of prediction errors in the prediction model of this scheme;

[0073] Figure 10 The graph shows the results of the comparative experiment on the Spearman index;

[0074] Figure 11 The graph shows the results of the comparative experiment on the Pearson index;

[0075] Figure 12 The figure shows the results of the comparative experiment on the MSE index. Detailed Implementation

[0076] 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.

[0077] like Figure 1 and Figure 2 As shown, the drug synergy prediction method based on cell semantic-guided dynamic biological pathway weights in this embodiment includes:

[0078] Predictive model building

[0079] S1. Collect the following three types of data

[0080] (1) Drug synergy dataset: This embodiment was obtained from the public database DrugComb, which contains 1,693 drugs and 184,320 drug interaction data. Each data sample records the synergy score of two drugs on a specific cell line.

[0081] (2) Genetic relationship data used to align drug and cellular characteristics from a biological perspective:

[0082] Drug-gene relationship data, obtained from the STITCH database, is used to characterize the interaction between drugs and genes;

[0083] Cell gene expression data were obtained from the GDSC database to characterize the basic gene expression levels of each cell line. Then, based on the 978 landmark genes defined in the L1000 database, the corresponding gene expression values ​​were selected from the above data. The landmark gene expression values ​​of individual cell line samples were then sorted from high to low to form the expression sequence, which is used as the cell gene expression data required in this embodiment.

[0084] (3) Gene pathway relationship data: obtained from the KEGG database, which records the subordinate relationships between genes and biological pathways, and is used to construct the subsequent hard constraint mask matrix.

[0085] S2. Feature preprocessing: The collected data is processed into three types of input features required by the prediction model.

[0086] (1) Cell semantic features: Cell gene expression data of cell lines are input into the pre-trained Cell2Sentence large language model. The attention mechanism of this model is used to capture deep semantic associations between genes and output cell semantic features E with a dimension of 1024. cell ;

[0087] In another embodiment, the cellular large language model can also use scGPT, etc.

[0088] (2) Drug characteristics: The gene relationship data of drugs are organized into a gene relationship matrix, and the gene relationship matrix is ​​standardized to obtain the gene relationship characteristics. The standardization process refers to the normalization operation of the drug / cell-gene expression value matrix. Specifically, the mean and standard deviation are calculated according to the gene dimension, (x-mean) / standard deviation is calculated, and then Min-Max Scaling normalization is performed.

[0089] This scheme is used to predict the synergistic effect scores of two drugs in a specific cell line. Therefore, drug characteristics are processed in pairs, such as drug A and drug B as a pair, and the gene relationship characteristics X of drug A are obtained separately. A Genetic relationship characteristics X with drug B B The dimension is the number of genes, for example, 574 genes;

[0090] (3) Cell characteristics: The gene expression matrix of the cell line was standardized to obtain the gene expression characteristics X of the cell line. cell The dimension is also the number of genes.

[0091] Genetic relationship characteristics of each pair of drugs X A X B Gene expression characteristics of specific cell lines X cell Cellular semantic features E cell The scores of the synergistic effects of the two drugs with the specific cell line in the drug synergy dataset constitute a sample data bar, which is used to train the prediction model.

[0092] S3. Constructing a predictive model based on cellular semantic guidance and dynamic semantic pathway gating.

[0093] Based on semantic feature E cell Dynamically generate pathway activation weights S gate The specific implementation method is as follows:

[0094] First, let's consider the 1024-dimensional E. cel Input a two-layer neural network,

[0095] The first layer maps the dimension to 256 and uses GELU as the activation function.

[0096] The second layer maps the dimension to 300, corresponding to 300 KEGG pathways. The activation function is Sigmoid, and the output values ​​are normalized to the [0,1] interval, generating a biological pathway activation weight vector representing the activation level of the 300 pathways. This guides the model to focus on key biological functions in the current cell line during the feature extraction stage. The calculation formula is as follows:

[0097] S gate =Sigmoid(W2*GELU(W1*E cell +b1)+b2)

[0098] Where W1 and W2 are learnable weight matrices, b1 and b2 are bias terms, and S gate It is a 300-dimensional vector, where each dimension represents the physiological activity or functional priority of a specific pathway in the current cellular environment, serving as a dynamic control signal for subsequent pathway feature extraction.

[0099] The above calculation realizes the cellular semantic feature E cell The transformation to dynamic weights of biological pathways allows pathway activation weights to adapt to different cell lines.

[0100] S4. Construct a biological pathway mask gating system that integrates hard constraint mapping and semantic-gated biological relationship mapping for the prediction model. Its structure is as follows: Figure 3 As shown, the specific steps include the following:

[0101] (1) Construct a hard constraint mask matrix: Based on the gene pathway relationship data obtained in step S1, construct a binary mask matrix Mask∈R300×574, where 574 is the number of genes and 300 is the number of pathways. If gene j belongs to biological pathway i, then Maskj=1, otherwise it is 0. This matrix remains fixed during the training process and serves as the topological hard constraint of the neural network.

[0102] (2) Generating the weight matrix after hard constraints: the neural network weight matrix W for the biological pathway mask gating layer path Initialize using the Kaiming normal distribution, and set W path Multiplying element-wise with the mask matrix yields the hard-constrained weight matrix W. masked :

[0103] W masked =W path Mask

[0104] The above processing, by using a binary mask matrix to impose hard constraints on the neural network weight matrix, ensures that only neuron connections that conform to biological hierarchical relationships are preserved, while connections that do not conform are forcibly set to zero.

[0105] (3) Pathway feature extraction: using W masked Gene relationship characteristics X A Drug-gene relationship characteristics X B Gene expression characteristics X cell A linear transformation is performed to map the features from the gene dimension to the pathway dimension, resulting in preliminary pathway features.

[0106] The result is then compared with the dynamic weight S generated in step S3. gate Coupling, i.e., soft gating, is performed to obtain the path feature P, specifically implemented as follows:

[0107] P=Linear(X, W masked )·S gate

[0108] Linear(X, W masked ), indicating a linear transformation of X;

[0109] X-ray the genetic relationship characteristics of drugs A X B Gene expression characteristics of the target cell line X cell Substituting X into each, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell .

[0110] This pathway masking gating layer acts on the expression profiles of two drugs and cell lines, outputting a feature vector with three pathway dimensions, namely P. A P B P cell This allows the same drug to produce heterogeneous pathway response characteristics in different cellular semantic contexts.

[0111] S5. Construct a sequential triplet interaction attention mechanism for the prediction model to capture the dynamic interactions between drug A, drug B, and cell lines at the pathway level. Its structure is as follows: Figure 4 As shown, the specific implementation is as follows:

[0112] Construct a multi-head self-attention layer, and extract P from S4. A P B P cell Concatenate them into an interactive sequence of shape [Batch, 3, 300];

[0113] The above sequence is input into a multi-head self-attention layer, which captures the interaction between the three at the pathway level through the self-attention mechanism, and outputs fused features:

[0114] Attn_Out=MultiHead(Q, K, V)

[0115] Where Q, K, and V are obtained by linear transformation of the input sequence:

[0116] Q=[PA,PB,Pcell]WQ,K=[PA,PB,Pcell]WK,V=[PA,PB,Pcell]WV,

[0117] WQ, WK, and WV are the learnable projection matrices.

[0118] The above process captures the dynamic compensatory effects and feedback regulation relationships of drug-drug-cell at the pathway level by introducing a triplet sequenced self-attention mechanism. The model can automatically learn the interference patterns at the pathway level under the combined action of drugs. Compared with the traditional simple feature splicing method, this method realizes the deep interaction and organic integration of multimodal features, enhances the model's ability to model the complex nonlinear relationship of drug synergy, and significantly improves the prediction accuracy.

[0119] S6. Construct the output layer of the prediction model

[0120] Constructing a multilayer perceptron prediction head consisting of three linear layers:

[0121] First layer: Map 900 dimensions to 256 dimensions, use ReLU as the activation function, and introduce Dropout (probability 0.2) to prevent overfitting;

[0122] The second layer maps 256 dimensions to 128 dimensions, uses ReLU as the activation function, and also introduces Dropout (probability 0.2).

[0123] The third layer maps 128 dimensions to 1 dimension and outputs a continuous synergy score.

[0124] The feature Attn_Out output by the attention mechanism is flattened to obtain a global feature vector of shape [Batch, 900]. This vector is then input into the aforementioned collaborative prediction head to provide the predicted regression score. The entire process can be represented as follows:

[0125] Score = MLP(Flatten(Attn_Out))

[0126] Specifically, this prediction model adopts an end-to-end training method, with the loss function being the mean squared error (MSE) and the optimizer being Adam. During training, backpropagation is performed based on the MSE of the predicted score and the actual label to update all learnable parameters.

[0127] like Figure 5 As shown, after the prediction model is trained, the synergistic effect of any two drugs on any target cell line can be predicted in the following way. Here, the two drugs to be predicted are referred to as the first drug and the second drug, and the cell line to be predicted is referred to as the target cell line:

[0128] Obtain pharmacogenomic relationship data, cell line gene expression data, and gene pathway relationship data of the first and second drugs from various publicly available bioinformatics databases;

[0129] The cell semantic features E are obtained by processing the data in a manner similar to the training process described above. cell The genetic relationship characteristics of the first drug X A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line cell ,

[0130] Based on semantic feature E cell Generate biological pathway activation weights S corresponding to the target cell line. gate ;

[0131] S gate =Sigmoid(W2*GELU(W1*E cell +b1)+b2)

[0132] A binary mask matrix is ​​constructed based on the hierarchical relationship between genes and biological pathways, and the binary mask matrix is ​​used to apply hard constraints to the weights of the neural network.

[0133] Based on the weights after hard constraints and the biological pathway activation weight S gate X-ray the genetic relationship characteristics of the first and second drugs. A X B and the gene expression characteristics X of the target cell line cell Mapping to the pathway level respectively, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell ;

[0134] P=Linear(X, W masked )·S gate

[0135] The multi-head self-attention layer will use the first path feature P A Second pathway characteristics P B and cell pathway characteristics P cell Interactive fusion is performed to obtain fusion features;

[0136] Attn_Out=MultiHead(Q, K, V)

[0137] The multilayer perceptron prediction head predicts the synergistic effect score of the first and second drugs on the target cell line based on this fusion feature:

[0138] Score=MLP(Flatten(Attn_Out)).

[0139] Furthermore, the above methods can be used to obtain a drug synergy prediction system based on cellular semantic-guided dynamic biological pathway weights, such as... Figure 6 As shown, the system includes:

[0140] Semantic extraction module 1 integrates large-scale cellular language models, such as scGPT and Cell2Sentence, to obtain the semantic features E of the target cell line. cell ;

[0141] Semantic path gating module 2 is used to determine the semantic feature E. cell Generate biological pathway activation weights S corresponding to the target cell line. gate ;

[0142] Data acquisition module 3 is used to acquire the gene relationship characteristics X of the first drug. A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line cell ;

[0143] The biological pathway mask gating module 4 is used to construct a binary mask matrix based on the hierarchical relationship between genes and biological pathways. This binary mask matrix is ​​then used to apply hard constraints to the weights of the neural network, and the hard-constrained weights are combined with the biological pathway activation weights S. gate X-ray the genetic relationship characteristics of the first and second drugs. A X B and the gene expression characteristics X of the target cell line cell Mapping to the pathway level respectively, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell ;

[0144] Interactive fusion module 5 is used to integrate the first path feature P A Second pathway characteristics P B and cell pathway characteristics P cell Interactive fusion is performed to obtain fusion features;

[0145] Prediction output module 6 is used to predict the synergistic score of the first drug and the second drug on the target cell line based on the fusion characteristics.

[0146] After predicting synergistic effects, the pathway activation weights S generated by the semantic pathway module can be read. gate Perform interpretability analysis. For example, for a given drug combination and a specific cell line, S gate The pathways corresponding to dimensions with higher median values ​​(such as the PI3K-Akt signaling pathway, apoptosis pathway, cell cycle pathway, etc.) are the key pathways that contribute the most to the synergistic effect in this cellular environment, which can provide interpretable evidence at the biological mechanism level for clinical combination drug use.

[0147] To verify the effectiveness and practicality of this scheme, this embodiment studies the performance of the constructed prediction model. The dataset is randomly divided into 5 parts, with each part used as the test set in turn, and the remaining 4 parts used as the training set. The training and testing are repeated five times, and the average result is taken as the final evaluation metric. MSE (mean squared error), Spearman correlation coefficient, and Pearson correlation coefficient are used as evaluation metrics. The dataset partitioning strategy is the same for all comparative experiments.

[0148] The performance evaluation results of the prediction model in this scheme are as follows: Mean: MSE=57.5830, Spearman=0.5056, Pearson=0.8949; Standard Deviation: MSE=1.8283, Spearman=0.0051, Pearson=0.0034. The experimental results of 5-fold cross-validation are as follows. Figure 7 As shown, the extremely small standard deviations of the Spearman and Pearson coefficients indicate that the model's predictive performance is highly stable across different data subsets. To visually verify the model's predictive effectiveness, a scatter plot was created to compare the predicted and actual results for each sample. The results are shown below. Figure 8 As shown, all scatter points are distributed along the diagonal, indicating a high degree of consistency between the model's predicted values ​​and the actual values. Furthermore, a visual analysis of the error distribution between all sample predicted values ​​and actual values ​​is performed, as shown below. Figure 9 The results showed that the errors were mainly concentrated in the small value range, further verifying the accuracy of the model's predictions.

[0149] To further evaluate the performance advantages of the prediction model, the prediction model provided in this solution was compared with benchmark models based on machine learning. The selected comparison models included ridge regression, Bayesian regression, and random forest regression. All of these models predict drug synergistic effects by mapping drug and cellular gene features to regression values. The comparison results based on the three evaluation indicators are as follows: Figures 10-12 As shown in the figure. The results demonstrate that the proposed method has significant advantages. In all experiments, the model achieved the lowest MSE value, indicating optimal prediction accuracy. Simultaneously, it achieved the highest Pearson linear correlation coefficient and Spearman rank correlation coefficient, indicating that the model's predictions exhibited optimal linear and rank correlation with the actual values. This is because the proposed model overcomes the limitations of traditional static biological network modeling by fully integrating prior biological knowledge with deep gene semantic features, taking into account both the underlying molecular biological principles and the heterogeneity of the cellular microenvironment, ultimately achieving improved prediction accuracy and stability.

[0150] The drug synergy prediction method provided in this scheme improves the biological interpretability and generalization ability of the prediction through deep coupling of cross-modal information. Firstly, it overcomes the limitations of traditional static biological networks by introducing semantic embedding E based on a pre-trained cellular large language model. cellAs a control signal, activation weights of pathways are dynamically generated through semantic pathway gating, achieving environment-dependent expression of biological network topology. Secondly, by fully integrating prior biological knowledge with deep semantic features, the model's logical rigor is enhanced. In the feature extraction stage, an intelligent gated mask linear layer is employed, combining strict hard constraints on gene-pathway attribution relationships with semantic importance assigned by a large language model, thus balancing the underlying facts of molecular biology with the heterogeneous information of the cellular microenvironment. Finally, a sequence-based triplet interactive attention mechanism is introduced, changing the traditional simple feature splicing fusion method. Through a multi-head self-attention mechanism, the dynamic compensatory effects and relationships between drug A, drug B, and cell lines at the pathway level are captured, integrating drug perturbation and cellular response data in an optimal nonlinear manner, significantly improving prediction accuracy.

[0151] The predictive model implemented in this solution can be applied to drug combination screening and new drug development in the treatment of various diseases. It can not only provide high-confidence collaborative score prediction, but also trace key biological pathways through gating weights, providing interpretable biological evidence for clinical decision-making, reducing drug development costs and reducing the risk of medication for patients.

[0152] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights, characterized in that, The method includes: Obtain the semantic features E of the target cell line cell ; According to the semantic feature E cell Generate biological pathway activation weights S corresponding to the target cell line. gate ; Obtain the gene relationship characteristics X of the first drug A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line. cell ; A binary mask matrix is ​​constructed based on the hierarchical relationship between genes and biological pathways, and the binary mask matrix is ​​used to apply hard constraints to the weights of the neural network. Based on the hard-constrained neural network weights and the biological pathway activation weights S gate X-ray the genetic relationship characteristics of the first drug and the second drug. A X B and the gene expression characteristics X of the target cell line cell Mapping to the pathway level respectively, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell ; The first pathway feature P A Second pathway characteristics P B and cell pathway characteristics P cell Interactive fusion is performed to obtain fusion features; Based on the fusion characteristics, the synergistic effect score of the first drug and the second drug on the target cell line is predicted.

2. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 1, characterized in that, The semantic features E of the target cell line are obtained as follows: cell : Obtain gene expression data of the target cell line; The gene expression data is input into a pre-trained cellular big language model, and the attention mechanism of the cellular big language model is used to capture the semantic associations between genes, outputting the semantic feature E. cell .

3. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 1, characterized in that, Biological pathway activation weights S corresponding to the target cell line are generated through semantic pathway gating. gate : S gate =σ(f(E cell1 )) Where f is a linear mapping function and σ is an activation function.

4. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 3, characterized in that, The semantic path gating is implemented using a two-layer neural network: S gate =Sigmoid(W2*GELU(W1*E cell +b1)+b2) Where W1 and W2 are learnable weights, b1 and b2 are bias terms, and GELU is the activation function.

5. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 1, characterized in that, The gene relationship characteristics X of the first drug were obtained in the following manner. A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line cell : Obtain gene relationship data for the first drug, gene relationship data for the second drug, and gene expression data for the target cell line; Gene relationship data of the first drug and gene relationship data of the second drug are respectively organized into gene relationship matrices; The gene expression data of the target cell line were organized into a gene expression matrix; Each matrix is ​​standardized to obtain the gene relationship feature X of the first drug. A The genetic relationship characteristics of the second drug X B and gene expression characteristics of the target cell line X cell .

6. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 1, characterized in that, Each element of the binary mask matrix is ​​used to indicate whether the corresponding gene belongs to the corresponding biological pathway, and each feature is mapped to the pathway level in the following way: Using the weights after hard constraints, the first drug gene relationship feature X is analyzed. A Second drug gene relationship characteristics X B Gene expression characteristics X cell Linear transformations were performed to obtain preliminary pathway characteristics; The preliminary pathway characteristics are compared with the biological pathway activation weight S. gate By coupling, the first path feature P is obtained. A Second pathway characteristics P B and cell pathway characteristics P cell .

7. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 6, characterized in that, The weights of the neural network are hard-constrained as follows to obtain the hard-constrained weights: W masked =W path ·Mask Where Mask is the binary mask matrix, W path These are the learnable weights of a neural network; The first pathway feature P is obtained in the following manner. A Second pathway characteristics P B and cell pathway characteristics P cell : P=Linear(X, W masked )·S gate S gate This indicates the activation weight of the biological pathway; Linear(X, W masked ), indicating a linear transformation of X; X-ray the genetic relationship characteristics of the first drug A Gene relationship characteristics of the second drug X B Gene expression characteristics of the target cell line X cell Substituting X into each, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell .

8. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 1, characterized in that, The fusion feature is obtained in the following manner: The first pathway feature P A Second pathway characteristics P B and cell pathway characteristics P cell Concatenate them into the input sequence; The input sequence is fed into a multi-head self-attention layer, which captures the interaction relationships among the three components at the pathway level through a self-attention mechanism, and outputs the fused features: Attn_Out=MultiHead(Q, K, V) where Q, K, V = [P A , P B , P cell W Q , [P A , P B , P cell W K , [P A , P B , P cell W V .

9. The method for predicting drug synergy based on cellular semantic-guided dynamic biological pathway weights according to claim 1, characterized in that, Predicting the synergistic effect score of the first drug and the second drug on the target cell line includes: The fused features are flattened and input into the multilayer perceptron prediction head; The multilayer perceptron prediction head outputs the cooperative score: Score = MLP(Flatten(Attn_Out)) The multilayer perceptron prediction head (MLP) comprises three linear mapping layers.

10. A drug synergy prediction system based on cell semantic-guided dynamic biological pathway weights, characterized in that, include: The semantic extraction module is used to obtain the semantic features E of the target cell line. cell ; The semantic path gating module is used to control the semantic path based on the semantic feature E. cell Generate biological pathway activation weights S corresponding to the target cell line. gate ; The data acquisition module is used to acquire the gene relationship characteristics X of the first drug. A The genetic relationship characteristics of the second drug X B and the gene expression characteristics X of the target cell line. cell ; The biological pathway mask gating module is used to construct a binary mask matrix based on the hierarchical relationship between genes and biological pathways. This binary mask matrix is ​​then used to apply hard constraints to the weights of the neural network, and the hard-constrained weights are combined with the activation weights S of the biological pathways. gate X-ray the genetic relationship characteristics of the first drug and the second drug. A X B and the gene expression characteristics X of the target cell line cell Mapping to the pathway level respectively, we obtain the first pathway feature P. A Second pathway characteristics P B and cell pathway characteristics P cell ; The interactive fusion module is used to integrate the first path feature P A Second pathway characteristics P B and cell pathway characteristics P cell Interactive fusion is performed to obtain fusion features; The prediction output module is used to predict the synergistic effect score of the first drug and the second drug on the target cell line based on the fusion characteristics.

Images