A deep learning-based auxiliary screening anticancer drug combination synergistic prediction method

Abstract

The application discloses an auxiliary screening anticancer drug combination synergistic prediction method based on deep learning, adopts a TIAM model, the TIAM model further considers a ternary relationship, the TIAM uses three inputs to calculate a three-dimensional attention tensor Attention 3d to represent the ternary relationship among drugs-drugs-cell lines, so there are two Queries in the TIAM, namely Query1, Query2, Key, Value.TIAM requires three inputs, which are drug a, drug b and cell line c, represented by X a , X b and X c respectively.After processing and outputting by the TIAM, X′ a , X′ b and X′ c are obtained.The application realizes feature extraction of drugs-drugs-cell lines by adopting a ternary attention mechanism, and is more accurate and relevant than existing technologies.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, and particularly relates to a synergistic prediction method for auxiliary screening of anticancer drug combinations based on deep learning. BACKGROUND

[0002] Drug combination therapy, i.e., simultaneous treatment with multiple drugs, generally has the advantages of better efficacy, reduced toxicity and side effects, and reduced or even overcome drug resistance. The discovery of synergistic drug combinations can expand the scope of new drug screening, reduce the development of ineffective drugs, and is of great significance. Due to the explosive growth of drug combination space, in vitro screening of synergistic drug combinations is time-consuming and laborious. With the development of high-throughput technology, large-scale anticancer drug combination datasets such as Oncology and DrugCombDB have emerged, and a large number of synergistic scores of drug combinations and cancer cell lines measured in experiments are provided. In the context of these database resources, using machine learning algorithms and deep learning algorithms to predict synergistic drugs is a very promising direction.

[0003] In recent years, people have proposed a large number of algorithms based on machine learning and deep learning to predict drug synergy by establishing the relationship between the input features (chemical structure, omics features, etc.) of drug combinations and cell lines and the expected output (synergy / non-synergy classification or quantitative synergy score). Celebi et al. constructed the features of drugs and cell lines, and then used five machine learning methods such as XGBoost model to predict the results, but machine learning methods often have the disadvantages of poor prediction performance, not fully utilizing biochemical information as features, etc.; Preuer et al. first used a deep learning method (multi-layer linear perception) to predict drug synergy; Wang et al. used GCN and GAT to learn drug molecular graph features; Yang et al. considered PPI networks to generate drug and cell line features and considered drug toxicity to predict drug synergy; Liu et al. introduced hypergraphs to learn the three-way relationship between drug pairs and cell lines; Rafiei et al. used multi-modal input and used Transformer to learn drug features to predict drug synergy, and Zhang et al. constructed a knowledge graph and used an attention mechanism to perform a drug synergy prediction task. SynergyX uses a multi-omics cross-attention mechanism to predict drug synergy. Among them, the machine learning algorithm model has the problems of limited prediction accuracy and failure to deeply mine the feature information of drugs and cell lines; the deep learning algorithm model has the problems. SUMMARY

[0004] To solve the technical problems in the background art, the present application provides a synergistic prediction method for auxiliary screening of anticancer drug combinations based on deep learning.

[0005] The application provides an auxiliary screening anticancer drug combination synergistic prediction method based on deep learning.

[0006] S1, drug substructure feature extraction, to obtain the vector of the drug;

[0007] S2, cell line multi-omics feature extraction, to obtain the vector of the cell line;

[0008] S3, the input feature of drug a is X a , the input feature of drug b is X b , and the input feature of cell line c is X c ; X a , X b and X c are input into a TIAM model, X a , X b and X c are obtained after processing and output of the TIAM model, the TIAM model calculates a three-dimensional attention tensor Attention 3d to represent the three-way relationship among the drug-drug-cell line based on the three inputs;

[0009] S4: X a , X b and X c are taken as the input of a prediction module to obtain the output.

[0010] Preferably, in step S3, the Query Matrix is generated based on the features X a and X b of drug a and drug b, and the KeyMatrix and Value Matrix are generated based on the cell line feature X c , as shown in the following formula:

[0011] Q a =X a W Q1 (1)

[0012] Q b =X b W Q2 (2)

[0013] K c =X c W K (3)

[0014] V c =X c W V (4)

[0015] wherein W Q1 , WQ2 , W K , W V are parameters that can be automatically learned in the deep learning model; Q a , Q b , K c and V c are Query a , Query b , Key and Value, respectively; and where I, J, M are the feature dimensions of drug a, drug b, and cell line c extracted by the feature extraction module; K is the hidden layer dimension; then the einstein summation convention einsum is used for calculation:

[0016] Attention 3d = einsum('ik,jk,mk→ijm', Q a , Q b , K c ) (5)

[0017] where: the tensor and 'ik,jk,mk→ijm' is the calculation rule of formula (5), where i, j, m are free indices, and k is the summation index, representing each element of the result Attention 3d is the sum of these characteristic values, which preliminarily represents the attention degree of the i-th feature of drug a and the j-th feature of drug b to the m-th feature of the cell line;

[0018]

[0019] where: a ik represents the i-th row and k-th column characteristic value of the drug a feature matrix, b jk represents the j-th row and k-th column characteristic value of the drug b feature matrix, and c mk represents the m-th row and k-th column characteristic value of the cell line c matrix, is the sum of these characteristic values, which preliminarily represents the attention degree of the i-th feature of drug a and the j-th feature of drug b to the m-th feature of the cell line;

[0020] In order to obtain the attention score of drug a and drug b to different features of cell line c, the global average pooling operation is used on the first two dimensions of the tensor Attention 3d , and the vector Attention ab2c is obtained:

[0021]

[0022] where I and J represent the number of features of drug a and drug b, respectively; ∑ represents the summation operation, and Softmax is the softmax activation function;​

[0023] Finally using Attention ab2c with V c weighted operation, get X' after adjusting the characteristics c ,

[0024] X' c

[0025] = Attention ab2c

[0026] ⊙ V c

[0027] where ⊙ represents Hadamard product; TIAM model processing output to get X" c :

[0028] X" c = X' c + X c (27).

[0029] Preferably, Query Matrix is generated based on the characteristics X b and X c of drug b and cell line c, Key Matrix and Value Matrix are generated based on the characteristics X a of drug a, as shown in the following formula:

[0030] Q b = X b W Q1 (9)

[0031] Q c = X c W Q2 (10)

[0032] K a = X a W K (11)

[0033] V a = X a W V (12)

[0034] wherein W Q1 , W Q2 , W K , W V are parameters that can be automatically learned in a deep learning model; Q b , Q c , K a and V a are Query1, Query2, Key and Value, respectively;

[0035] Attention 3d = einsum('jk,mk,ik→jmi', Q b ,Q c ,K a ) (13)

[0036] where: three-dimensional attention tensor J,M,I are the feature dimensions of drug b, cell line c, drug a; and 'ijk,mk,ik→jmi' is the calculation rule of formula (13), which can be formalized as:

[0037]

[0038] The attention scores of drug b and cell line c to different features of drug a are calculated:

[0039]

[0040] Finally, using Attention bc2a and V a weighted operation, get the adjusted feature X' a :

[0041] X' a

[0042] = Attention bc2a

[0043] ⊙V a

[0044] where: ⊙ represents Hadamard product; TIAM model processing output to get X" a :

[0045] X" a = X' a +X a (25)。

[0046] Preferably, Query Matrix is generated based on the features X a and X c of drug a and cell line c, Key Matrix and Value Matrix are generated based on drug b, X b , as shown in the following formula:

[0047] Q a = X a W Q1 (17)

[0048] Q c = Xc W Q2 (18)

[0049] K b =X b W K (19)

[0050] V b =X b W V (20)

[0051] where W Q1 , W Q2 , W K , W V are the parameters that can be automatically learned in the deep learning model; Q a , Q c , K b and V b are Query1, Query2, Key and Value respectively; the three-dimensional attention tensor Attention 3d is:

[0052] Attention 3d = einsum('ik,mk,jk→imj', Q a , Q c , K b ) (21)

[0053] where: the three-dimensional attention tensor I, M, J are the feature dimensions of drug a, cell line c, and drug b. And 'ik,mk,jk→imj' is the calculation rule of formula (21), and the specific calculation method can be formalized as follows:

[0054]

[0055] The attention scores of drug b and cell line c on different features of drug a are calculated:

[0056]

[0057] Finally, Attention ac2b is used to perform weighted operation with V b to obtain X' b after adjusting the features:

[0058] X' b

[0059] = Attention ac2b

[0060] ⊙V b

[0061] Where: ⊙ represents the Hadmar product, which is obtained after the TIAM model processes the output to get X″. b :

[0062] X″ b =X′ b +X b (26).

[0063] Preferably, the drug feature in step S1 is represented as a shape of The vector.

[0064] Preferably, in step S2, the cell line characteristics are represented by a shape of The vector.

[0065] Preferably, the prediction module includes a fully connected layer, a ReLU activation function, and a Softmax activation function.

[0066] Preferably, a dropout layer is provided between the ReLU activation function and the Softmax activation function to prevent overfitting.

[0067] The proposed deep learning-based method for assisted screening of anticancer drug combinations has higher accuracy in drug synergy prediction. It employs a ternary attention mechanism to achieve feature extraction of drug-drug-cell line, which is more accurate and relevant than existing technologies.

[0068] Additional aspects and advantages 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

[0069] Figure 1 This is a diagram of the TSYnergy model of the present invention;

[0070] Figure 2 This is a diagram of the TIAM model of the present invention;

[0071] Figure 3 This is a schematic diagram of the prediction module. Detailed Implementation

[0072] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0073] like Figures 1-3An auxiliary screening anticancer drug combination synergy prediction method based on deep learning is shown, which comprises the following steps:

[0074] Drug substructure feature extraction:

[0075] For the feature extraction process of the drug, the method of drug substructure feature extraction of SynergyX is adopted: the ESPF algorithm is used to divide the string into discrete and medium-sized substructures. About 2700 valuable substructures are confirmed and selected to form a substructure dictionary. Finally, each drug is fixed as a 165-dimensional vector. If the length of the drug substructure is less than 165, the method of zero padding is adopted. For details, please refer to the paper SynergyX;

[0076] Cell line multi-omics feature extraction:

[0077] For the feature processing of the cell line, considering the richness of the cell line data, the multi-omics data of the cell line sorted by SynergyX is followed: 4079 genes related to cancer cell lines, drug target relationships, and protein-protein interaction correlations are sorted; At the same time, the multi-omics data related to these genes is downloaded from DepMap, including gene expression, gene mutation, gene copy number, gene methylation, gene effect and gene dependence. If there is any missing genomics data, the average value of the gene in the remaining cell lines is used to fill in, then the tanh normalization is applied, and finally for each cell line, a fixed order is integrated to obtain an input feature

[0078] Fusion of drug-drug-cell line ternary information based on ternary interaction attention:

[0079] The traditional self-attention mechanism and cross-attention mechanism use Query, Key to calculate the attention score matrix, and then use the calculated attention score to weight the final feature, so this can only calculate the attention weighting of the input to itself or another input, only considering one-dimensional relationship or two-dimensional relationship. TIAM further considers the three-dimensional relationship, TIAM uses three inputs to calculate a three-dimensional attention tensor Attention 3d to represent the ternary relationship between drug-drug-cell line, so there are two Query in TIAM, Query1, Query2, Key, Value. TIAM requires three inputs, drug a, drug b, and cell line c, represented by X a , X b and X c , respectively, after processing by TIAM, X′ a , X′ b and X′ c. Obtain X' a , X' b and X' c The process of obtaining X' a , X' b and X' c is three similar processes, which are described in detail as follows.

[0080] First, Query Matrix is generated based on the features X a and X b of drug a and drug b, and Key Matrix and Value Matrix are generated based on the features X c of cell lines, as shown in the following formulas.

[0081] Q a = X a W Q1 (1)

[0082] Q b = X b W Q2 (2)

[0083] K c = X c W K (3)

[0084] V c = X c W V (4)

[0085] wherein W Q1 , W Q2 , W K , W V are parameters that can be automatically learned in a deep learning model; Q a , Q b , K c and V c are Query a , Query b , Key and Value respectively.

[0086] For the sake of explanation, we ignore Batch_size here, and in order to obtain a three-dimensional attention tensor Attention 3d we are based on the above obtained Q a , Q b and K c , wherein and where I, J, M are the feature dimensions of drug a, drug b, cell line c extracted by the feature extraction module; K is the hidden layer dimension, which can be adjusted freely according to the model. Then the einsum is used for calculation, as shown in formula (5).

[0087] Attention 3d = einsum('ik,jk,mk→ijm', Q a , Q b , K c ) (5)

[0088] where the tensor and 'ik,jk,mk→ijm' is the calculation rule of formula 5, where i, j, m are free indices, and k is the summation index, representing each element of the result Attention 3d . is the sum of these characteristic values, which preliminarily represents the attention degree of the i-th feature of drug a and the j-th feature of drug b to the m-th feature of the cell line.

[0089]

[0090] where a ik represents the i-th row and k-th column characteristic value of the drug a characteristic matrix, b jk represents the j-th row and k-th column characteristic value of the drug b characteristic matrix, and c mk represents the m-th row and k-th column characteristic value of the cell line c matrix, is the sum of these characteristic values, which preliminarily represents the attention degree of the i-th feature of drug a and the j-th feature of drug b to the m-th feature of the cell line.

[0091] In order to obtain the attention score of drug a and drug b to different features of cell line c, we use the global average pooling operation on the first two dimensions of the tensor Attention 3d to obtain the vector Attention ab2c , as shown in formula (7):

[0092]

[0093] where I and J represent the number of features of drug a and drug b respectively. ∑ represents the summation operation, and Softmax is the softmax activation function.

[0094] Finally, Attention ab2c is used for weighted operation with V c to obtain X′ c after adjusting the features.

[0095] X′ c

[0096] = Attention ab2c

[0097] ⊙V c

[0098] where ⊙ represents Hadamard product.

[0099] X′ is obtained a The process is also similar, first, Query Matrix is generated based on the characteristics X b and X c of drug b and cell line c, Key Matrix and Value Matrix are generated based on drug a X a , as shown in the following formula.

[0100] Q b = X b W Q1 (9)

[0101] Q c = X c W Q2 (10)

[0102] K a = X a W K (11)

[0103] V a = X a W V (12)

[0104] where W Q1 , W Q2 , W K , W V are automatically learned parameters in deep learning model; Q b , Q c , K a and V a are Query1, Query2, Key and Value respectively.

[0105] Similar to formula (5) and formula (6), we use the following calculation method. As shown in formula (13) and (14)

[0106] Attention 3d = einsum('jk,mk,ik→jmi', Q b , Q c , K a ) (13)

[0107] where three-dimensional attention tensor J, M, I are the feature dimensions of drug b, cell line c, drug a. And 'ijk, mk, ik→jmi' is the calculation rule of formula 13, which can be formalized as formula (14) as follows.

[0108]

[0109] As shown in formula (15), the attention scores of drug b and cell line c on different features of drug a can be calculated:

[0110] Finally, use Attention bc2a to perform a weighting operation. Get X' a after adjusting the features. a As shown in formula (16).

[0111] X' a

[0112] = Attention bc2a

[0113] ⊙V a

[0114] Where ⊙ represents Hadamard product.

[0115] Get X' b , the process is the same. First, the Query Matrix is generated based on the features X a and X c of drug a and cell line c, and the Key Matrix and Value Matrix are generated based on drug b X b , as shown in the following formula.

[0116] Q a = X a W Q1 (17)

[0117] Q c = X c W Q2 (18)

[0118] K b = X b W K (19)

[0119] V b = X b W V (20)

[0120] Where, W Q1 , W Q2 , W K , WV are the parameters that can be automatically learned in the deep learning model. Q a , Q c , K b and V b are Query1, Query2, Key and Value respectively.

[0121] Similar to formula (5) and formula (6), we use the following calculation method. As shown in formula (21) and (22)

[0122] Attention 3d = einsum('ik,mk,jk→imj', Q a , Q c , K b ) (21)

[0123] where the three-dimensional attention tensor I, M, J are the feature dimensions of drug a, cell line c, and drug b. And 'ik,mk,jk→imj' is the calculation rule of formula (21), and the specific calculation method can be formalized as shown in formula (22).

[0124]

[0125] As shown in formula (23), the attention score of drug b and cell line c to different features of drug a can be calculated.

[0126]

[0127] Finally, use Attention ac2b and V b to perform weighted operation. Get X' a after adjusting the features, as shown in formula (15).

[0128] X' a

[0129] = Attention ac2b

[0130] ⊙ V a

[0131] where ⊙ represents Hadamard product.

[0132] Considering that factors such as gradient explosion may occur, consider adding a residual structure to optimize the model, as shown in the following formula.

[0133] X" a

[0134] = X' a + Xa

[0135] X" b

[0136] = X' b + X b

[0137] X" c

[0138] = X' c + X c

[0139] Thus, we get the output X" of TIAM a , X" b , X" c .

[0140] After passing through TIAM, we get three final hidden layer feature representations of drug a, drug b and cell line c: X" a , X" b and X" c , flatten these features and concatenate them from head to tail, and then put them into the prediction module to predict the final synergistic prediction result y'. The prediction module is composed of a fully connected layer, a Relu activation function and a Softmax activation function. Finally, we add a dropout layer to prevent overfitting. Details are shown in Figure 3 . Specifically, as shown in Figure 3 , a dropout layer is provided between the Relu activation function and the Softmax activation function to prevent overfitting.

[0141] Interpretable model CIMI paper citation:

[0142] Wu C, Wang X, Lian D, et al. Acausality inspired framework for model interpretation [C] / / Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2023: 2731-2741.

[0143] Previous work often used attention scores for interpretability analysis. However, attention scores based on correlation are often less logical than those based on causality. Therefore, we utilize the general encoder-decoder interpretable model CIMI to provide a causal explanation for our black-box model TSYnergy. The model is divided into three modules. Only the decoder part of the interpreter is trainable; the parameter values ​​of the other parts are fixed. The specific process is as follows: Figure 1 As shown.

[0144] CIMI's input is a matrix x, which is a concatenation of drug a, drug b, and cell line c. The interpreter g(·) generates the explanation mask g(x). Specifically, if g(x) i If one element corresponds to the i-th gene in cell line c, then g(x) i This can be understood as the causal probability of drug a and drug b synergistically acting on this gene to synergistically treat cancer. CIMI designed three modules to obtain the causal explanation mask g(x): the causal sufficiency module, the causal intervention module, and the causal priority module.

[0145] Causal Sufficiency Module.

[0146] Causal features x are obtained using the interpretation mask g(x). e Non-causal feature x u , where x e =g(x)⊙x,x u =(1-g(x))⊙x, in order to ensure the causal characteristic x e Predicted labels A sufficient cause, which requires ensuring the causal characteristic x u It is used to predict labels. The most suitable feature, not others. In other words, x e The label can always be predicted using an optimal function f′. Non-causal features for predicting labels There is no meaningful information. Therefore, the loss of causal adequacy can be modeled as shown in Equation (9):

[0147]

[0148] in It is the mean squared error loss, x e =g(x)⊙x,x u= (1-g(x))⊙x. CIMI points out that due to the huge and continuous input space, it is difficult to find the optimal function f′. A good approach is to set the optimal function f′ as f, which is the black-box model TSYnergy. Therefore, the loss can be rewritten as shown in formula (10).

[0149]

[0150] Where x e =g(x)⊙x,x u =(1-g(x))⊙x

[0151] Causal Intervention Module:

[0152] Hope to find x e and x u They are independent, which allows us to find invariant interpretations x for similar input instances. e This improves the generalization ability of the interpreter. Based on prior knowledge: for the non-interpreted part x... u The intervention has no effect on the explanatory part. Therefore, an intervention loss was designed to separate the explanatory part.

[0153] First, following the common practice in causal reasoning, we examine the non-explanatory part x. u Intervention: By calculating the non-explanatory portion x u And linear interpolation of the non-explanatory part of another variable randomly drawn from X to perturb the explanatory part x u Specifically, as shown in formula (11).

[0154]

[0155] Where λ ~ U(0, ∈), and ∈ restricts the magnitude of the perturbation. Furthermore, we can ensure x by optimizing the intervention loss in formula (12). e and x u It is independent.

[0156]

[0157] This loss guarantees the generated explanation x e In the unexplained part of the perturbation x u The invariance of the input and output ensures local consistency in the interpretation. This means the interpreter generates consistent interpretations for similar inputs. This helps capture more generalizable features and improves the generalization ability of the interpretable.

[0158] Causal Prior Module:

[0159] We help the interpreter learn in two ways. 1) We add latent causal hints to the neural network of the interpreter. 2) We design a weakly supervised loss for the causal mask g(x).

[0160] Design of the interpreter neural network: One difficulty in the design of the interpreter is the lack of prior knowledge on how to choose the network structure. That is, when we use a neural network as the interpreter, we do not know what network structure to choose. If the interpreter is not as complex as the black box model f, how can we guarantee that the interpreter can understand the black box model? If the interpreter is more complex than the black box model f, this may lead to overfitting.

[0161] CIMI directly uses the parameters of the black box model f to generate the explanation mask g(x). Specifically, the encoding part of the black box model f is used as the encoder of the interpretable model, and the decoding part of the interpretable model uses a simple neural network, denoted as φ.

[0162] The output of the encoder is concatenated with the original input instance on the y-axis to obtain as the input of the decoder φ, as shown in equation (13)

[0163] g(x)

[0164] = φ([f e (x); v x ]) (13)

[0165] is the explanation mask. Using the encoder of the black box model f as the encoder of the interpreter makes the interpreter as complex as the black box model, so that it has enough ability to understand the decision mechanism of the black box model. Because the parameters of f e are fixed, we only need to train the decoder of the interpreter to train a good explanation, avoiding a large training cost.

[0166] Weakly supervised loss: If there is no regularization constraint on the causal mask, there is a result that the value of the explanation mask is all 1, which makes the interpreter meaningless. To solve this problem, we use the true value label with noise as the prior of learning the causal mask to guide the learning process. Based on this method, a weakly supervised loss is designed as shown in equation (14).

[0167]

[0168] where x' represents another instance other than x. g x′(x) represents mapping x' through g(x) to the explanation mask part, g x′ (x) = φ([f e (x'; x') ). On the contrary, g(x) = φ([f e (x); x]).

[0169] This weakly supervised loss can prevent the explainer from being overly optimistic in predicting all features as explanation features, which helps to reduce meaningless results.

[0170] Overall Framework and Optimization of the Interpretable Explainer.

[0171] Combining the above three modules, the overall optimization objective of the explainer CIMI is summarized as shown in the following formula (15).

[0172]

[0173] wherein α is a balance parameter. The weakly supervised loss is introduced to make it easier to adjust the regularization parameter, so this item does not require a balance parameter.

[0174] Experimental results:

[0175] Based on the Oncology dataset, experiments were conducted.

[0176] Comparative experiments were conducted on the model, and the proposed PITSynergy model was compared with four baseline models in terms of performance, including DeepSynergy, GraphSynergy, KGANSynergy, and SynergyX. The specific performance of each model on the four evaluation indicators: AUC, AUPR, ACC, and Recall, is listed in the table. First, in terms of the key indicator AUC (Area Under Curve), the PITSynergy model achieved the highest value of 0.8763, which was about 3.19% higher than the best baseline model SynergyX (0.8492) and about 5.57% higher than the worst baseline model GraphSynergy (0.8301), showing a significant advantage.

[0177] Secondly, in terms of the AUPR (Area Under the Precision-Recall Curve) indicator, PITSynergy also performed well, reaching 0.8921, which was about 2.97% higher than SynergyX's 0.8651. This result indicates that PITSynergy has stronger ability in handling imbalanced data.

[0178] In terms of ACC (Accuracy), PITSynergy achieved an accuracy of 0.7846, which is 1.68% higher than SynergyX (0.7716), indicating an improvement in overall prediction accuracy.

[0179] Finally, in terms of Recall, PITSynergy achieved 0.7893, which is 6.47% higher than GraphSynergy (0.7414), indicating that PITSynergy significantly outperforms other models in the ability to detect positive samples. In summary, PITSynergy outperforms existing baseline models in all evaluation metrics, with significant improvements in AUC and AUPR, demonstrating the effectiveness and superiority of the proposed method in drug combination prediction tasks.

[0180] To verify the contribution of the proposed modules to model performance, we conducted ablation experiments by removing the ternary attention module (TIA) and residual connection (Residule), respectively. The performance results under different configurations are shown in the table. The experimental conclusions are as follows:

[0181] Ternary Attention Module (TIA): When TIA is removed, the model's AUC decreases from 0.8763 to 0.8636, and AUPR decreases from 0.8921 to 0.8764, indicating a slight decrease in performance. This shows that the TIA module can effectively improve the model's discriminative ability, especially in AUC and AUPR.

[0182] Residual Connection (Residule): After removing the residual connection, AUC decreases to 0.866, and AUPR decreases to 0.8782, which is lower than the complete model, especially in the Recall metric, which decreases from 0.7893 to 0.7733. This indicates that the residual connection plays an important role in optimizing the recall ability of the model.

[0183] Complete Model (ours): The complete model containing TIA and residual connection performs best in all metrics, especially in AUC and AUPR, reaching 0.8763 and 0.8921 respectively, demonstrating the synergistic effect of these two modules on improving model performance.

[0184] Conclusion: Ablation experiment results show that both the ternary attention module and the residual connection significantly improve model performance. The complete model achieves the best performance in AUC, AUPR, ACC, and Recall, verifying the effectiveness of these two modules in drug combination prediction tasks.

[0185] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1.A method for predicting synergistic effects of anticancer drug combinations based on deep learning-assisted screening, characterized in that, Comprising the following steps: S1, drug substructure feature extraction, to obtain the vector of the drug; S2, cell line multi-omics feature extraction, to obtain the vector of the cell line; S3, input feature of drug a is X a , input feature of drug b is X b , input feature of cell line c is X c , X a , X b , X c is input to TIAM model, TIAM model processes output to get X" a , X" b and X" c , TIAM model calculates a three-dimensional attention tensor Attention 3d to represent the triadic relationship among drug-drug-cell line using three inputs; S4: Output X″ a ,X″ b and X″ c The output is derived from the input of the prediction module; In step S3, the Query Matrix is generated based on the features X of drug a and drug b a and X b The Key Matrix and the Value Matrix are generated based on the features X of the cell lines, as shown in the following equations: c The Key Matrix and the Value Matrix are generated based on the features X of the cell lines, as shown in the following equations: Q a = X a W Q1 (1) Q b = X b W Q2 (2) K c = X c W K (3) V c = X c W V (4) where W Q1 , W Q2 , W K , W V are the parameters that can be automatically learned in the deep learning model; Q a , Q b , K c and V c are Query a , Query b , Key and Value, respectively; and where I, J, M are the feature dimensions of drug a, drug b, cell line c extracted by the feature extraction module; K is the hidden layer dimension; then the einstein summation convention einsum is used for calculation: Attention 3d = einsum('ik,jk,mk→ijm', Q a , Q b , K c ) (5) where: Tensor And 'ik,jk,mk→ijm' is the computational rule for equation (5), where i, j, m are free indices, and k is a summation index, denoting the result Attention 3d Each element of is to be summed over dimension k: where: a ik represents the eigenvalue of the characteristic matrix of drug a in the i-th row and k-th column, b jk represents the eigenvalue of the characteristic matrix of drug b in the j-th row and k-th column, c mk represents the eigenvalue of the matrix of cell line c in the m-th row and k-th column, is the sum of the products of these eigenvalues, which represents the degree of attention of the i-th feature of drug a and the j-th feature of drug b to the m-th feature of cell line. To get the attention scores of drug a and drug b on different characteristics of cell line c, the tensor Attention 3d The first two dimensions use the global average pooling operation to get the vector Attention ab2c : Wherein, I and J represent the number of features of drug a and drug b respectively; ∑ represents the summation operation, and Softmax is a softmax activation function; Finally using Attention ab2c with V c weighted operation, get X' after adjusting the characteristics c , X' c = Attention ab2c ⊙V c (8) where represents Hadamard product; TIAM model processing output to get X" c : X" c = X' c + X c (27); Query Matrix is generated based on the features X of drug b and cell line c b and X c Key Matrix and Value Matrix are generated based on the features X of drug a a as shown in the following equations: Q b = X b W Q1 (9) Q c = X c W Q2 (10) K a = X a W K (11) V a = X a W V (12) wherein W Q1 , W Q2 , W K , W V are parameters that can be automatically learned in a deep learning model; Q b , Q c , K a and V a are Query1, Query2, Key and Value, respectively; Attention 3d = einsum('jk,mk,ik→jmi', Q b , Q c , K a ) (13) where: three-dimensional attention tensor J, M, I are the feature dimensions of drug b, cell line c, drug a; and 'ijk, mk, ik→jmi' is the calculation rule of formula (13), which can be formalized as follows: The attention score of drug b and cell line c to different features of drug a is calculated as follows: Finally using Attention bc2a with V a weighted operation, get the adjusted features after X a : X' a = Attention bc2a ⊙V a (16) where: represents the Hadamard product; TIAM model processing output X" a : X" a = X' a + X a (25); Query Matrix is generated based on the features X of drug a and cell line c a and X c Key Matrix and Value Matrix are generated based on drug b, X b as shown in the following equations: Q a = X a W Q1 (17) Q c = X c W Q2 (18) K b = X b W K (19) V b = X b W V (20) where W Q1 , W Q2 , W K , W V are parameters that can be automatically learned in the deep learning model; Q a , Q c , K b and V b are Query1, Query2, Key and Value, respectively; and the three-dimensional attention tensor Attention 3d is: Attention 3d = einsum('ik,mk,jk→imj', Q a , Q c , K b ) (21) where: three-dimensional attention tensor I, M, J are the feature dimensions of drug a, cell line c, drug b, respectively, and 'ik,mk,jk→imj' is the computational rule of equation (21), which can be formalized as follows: The attention score of drug b and cell line c to different features of drug a is calculated as follows: Finally using Attention ac2b with V b weighted operation, get X' after adjusting the characteristics b : X' b = Attention ac2b ⊙V b (24) where: represents the Hadamard product, X" is the output of the TIAM model processing b : X" b = X' b + X b (26). 2.The deep learning-based auxiliary screening anticancer drug combination synergistic prediction method according to claim 1, characterized in that, The drug feature in step S1 is represented as a vector of the form . 3.The deep learning-based auxiliary screening anticancer drug combination synergistic prediction method according to claim 1, characterized in that, The cell line characteristics in step S2 are represented as vectors of the form . 4.The deep learning-based auxiliary screening anticancer drug combination synergistic prediction method according to claim 1, characterized in that, The prediction module comprises a full connection layer, a Relu activation function and a Softmax activation function. 5.The deep learning-based auxiliary screening anticancer drug combination synergistic prediction method according to claim 4, characterized in that, A dropout layer for preventing overfitting is arranged between the Relu activation function and the Softmax activation function.

Images