Deep learning-based collaborative prediction method for auxiliary screening of anticancer drug combination

Abstract

The invention discloses an auxiliary screening anti-cancer drug combination collaborative prediction method based on deep learning, a TIAM model is adopted, the TIAM model further considers a ternary relationship, the TIAM uses three inputs to calculate a three-dimensional attention tensor Attention3d to represent the ternary relationship among a drug-drug-cell line, and therefore two Quries, namely Query 1, Query 2, Key and Value, exist in the TIAM model. TIAM requires three inputs, namely a drug a, a drug b and a cell line c which are represented by Xa, Xb and Xc respectively, and X'a, X'b and X'c are obtained after TIAM treatment and output. According to the method, the feature extraction of the drug-drug-cell line is realized by adopting a ternary attention mechanism, and compared with the prior art, the method is more accurate and has relevance.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method for assisted screening of anticancer drug combination synergy prediction based on deep learning. Background Art

[0002] Drug combination therapy, i.e., simultaneous treatment with multiple drugs, usually has the advantages of better efficacy, reduced toxicity and side effects, and reduced or even overcoming drug resistance. The discovery of synergistic drug combinations can expand the scope of new drug screening and reduce the development of ineffective drugs. It is very meaningful to find synergistic drug combinations. Due to the explosive growth of drug combination space, in vitro screening of synergistic drug combinations is time-consuming and laborious. At the same time, with the development of high-throughput technology, large anticancer drug combination data sets such as Oncology and DrugCombDB have emerged, providing a large number of drug combinations and synergistic scores measured by cancer cell line experiments. In the context of these database resources, the use of machine learning algorithms and deep learning algorithms to predict synergistic drugs has become a very promising direction.

[0003] In recent years, a large number of algorithms based on machine learning and deep learning have been proposed 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 (synergistic / non-synergistic classification or quantitative synergistic score). Celebi et al. constructed the features of drugs and cell lines, and then used 5 machine learning methods such as XGBoost model to predict the results, but machine learning methods often have shortcomings such as poor prediction performance and failure to fully utilize biochemical information as features; Preuer et al. used deep learning methods (multi-layer linear perceptron) to predict drug synergy for the first time; 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 ternary relationship between drug pairs and cell lines; Rafiei et al. used multimodal input and used Transformer to learn drug features to predict drug synergy; Zhang et al. constructed knowledge graphs and used attention mechanisms for drug synergy prediction tasks. SynergyX uses multi-omics cross-attention mechanisms to predict drug synergy. Among them, there are problems with the machine learning algorithm model: the prediction accuracy is relatively limited, and it fails to deeply explore the characteristic information of drugs and cell lines; there are problems with the deep learning algorithm model. Summary of the invention

[0004] In order to solve the technical problems existing in the background technology, the present invention proposes a synergistic prediction method for auxiliary screening of anti-cancer drug combinations based on deep learning.

[0005] The present invention proposes a deep learning-based assisted screening method for synergistic prediction of anticancer drug combinations, comprising the following steps:

[0006] S1, extract drug substructure features to obtain drug vectors;

[0007] S2, extract multi-omics features of cell lines to obtain vectors of cell lines;

[0008] S3. The input feature of drug a is X a , the input feature of drug b is X b , the input features of cell line c are X c , X a , X b , X c Input to the TIAM model, the TIAM model processes the output to get X″ a , X″ b and X″ c , the TIAM model uses three inputs to calculate a three-dimensional attention tensor Attention 3d to represent the ternary relationship between drug-drug-cell line;

[0009] S4: Output X″ a , X″ b and X″ c The output is obtained as input to the prediction module.

[0010] Preferably, in step S3, the Query Matrix is ​​based on the features X of drug a and drug b. a and X b Generate, KeyMatrix and Value Matrix based on cell line feature X c Generate, 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] Among them, W Q1 , WQ2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q a , Q b , K c and V c Query a , Query b , Key and Value; and Where I, J, and 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; and 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] Among them: tensor And ′ik,jk,mk→ijm′ is the calculation rule of formula (5), where i,j,m are free indices and k is the sum index, indicating the result Attention 3d Each element of It needs to be summed over dimension k:

[0018]

[0019] Among them: a ik represents the eigenvalue of the i-th row and k-th column of the feature matrix of drug a, b jk represents the eigenvalue of the jth row and kth column of the feature matrix of drug b, c mk represents the eigenvalue of the mth row and kth column of the cell line c matrix, It is the sum of the products of these feature values, which preliminarily indicates the degree of attention paid by 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 scores of drug a and drug b on different features of cell line c, in the tensor Attention 3d The first two dimensions use global average pooling to obtain vector Attention ab2c :

[0021]

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

[0023] Finally, use Attention ab2c With V c Perform weighted operation to obtain X′ after adjusting the features c ,

[0024] X′ c

[0025] =Attention ab2c

[0026] ⊙V c

[0027] Where ⊙ represents the Hadamard product; the output of the TIAM model is X″ c :

[0028] X″ c =X′ c +X c (27).

[0029] Preferably, the Query Matrix is ​​based on the feature X of drug b and cell line c. b and X c Generate, Key Matrix and Value Matrix based on drug a feature X a Generate, 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] Among them, W Q1 , W Q2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q b , Q c , K a and V a They are Query1, Query2, Key and Value;

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

[0036] Where: 3D attention tensor J, M, I are the characteristic dimensions of drug b, cell line c, and drug a; and ′ijk,mk,ik→jmi′ is the calculation rule of formula (13), which can be formalized as follows:

[0037]

[0038] Calculate the attention scores of drug b and cell line c for different features of drug a:

[0039]

[0040] Finally, use Attention bc2a With V a Perform weighted operation to obtain X′ after adjusting the features a :

[0041] X′ a

[0042] =Attention bc2a

[0043] ⊙V a

[0044] Where: ⊙ represents the Hadamard product; X″ is obtained after the TIAM model is processed and outputted a :

[0045] X″ a =X′ a +X a (25).

[0046] Preferably, the Query Matrix is ​​based on the feature X of drug a and cell line c. a and X c Generate, Key Matrix and Value Matrix based on drug b, X b Generate, 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] Among them, W Q1 , W Q2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q a , Q c , K b and V b They are Query1, Query2, Key and Value; three-dimensional attention tensor Attention 3d for:

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

[0053] Where: 3D attention tensor I, M, J are the characteristic 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] Calculate the attention scores of drug b and cell line c for different features of drug a:

[0056]

[0057] Finally, use Attention ac2b With V b Perform weighted operation to obtain X′ after adjusting the features b :

[0058] X′ b

[0059] =Attention ac2b

[0060] ⊙V b

[0061] Where: ⊙ represents the Hadamard product, and the output of the TIAM model is X″ b :

[0062] X″ b =X′ b +X b (26).

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

[0064] Preferably, in step S2, the cell line feature is represented by a shape of 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 for preventing overfitting is provided between the Relu activation function and the Softmax activation function.

[0067] In the present invention, the proposed deep learning-based assisted screening anti-cancer drug combination synergy prediction method has higher drug synergy prediction accuracy; the ternary attention mechanism is used to realize the feature extraction of drug-drug-cell line, which is more accurate and relevant than the existing technology.

[0068] Additional aspects and advantages of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

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

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

[0071] Figure 3 Schematic diagram of the prediction module. DETAILED DESCRIPTION

[0072] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar symbols throughout represent the same or similar elements or elements with the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0073] like Figure 1-Figure 3A method for assisted screening of anticancer drug combination synergy prediction based on deep learning is shown, comprising the following steps:

[0074] Drug substructure feature extraction:

[0075] For drug feature extraction, SynergyX's drug substructure feature extraction method is used: the ESPF algorithm is used to segment the string into discrete, medium-sized substructures. About 2,700 valuable substructures are identified and selected to form a substructure dictionary. Finally, each drug is fixed to a 165-dimensional vector. If the length of the drug substructure is less than 165, the zero padding method is adopted. For detailed process, please refer to the paper SynergyX;

[0076] Cell line multi-omics feature extraction:

[0077] Considering the richness of cell line data, we used the multi-omics data of cell lines organized by SynergyX for feature processing of cell lines: 4079 genes with high correlation with cancer cell lines, drug target relationships, and protein-protein interactions were organized; at the same time, multi-omics data related to these genes were downloaded from DepMap, including gene expression, gene mutation, gene copy number, gene methylation, gene effect, and gene dependency. If any genomic data is missing, the average value of the gene in the remaining cell lines is used to fill it, and then tanh normalization is applied. Finally, for each cell line, an input feature is obtained by integration in a fixed order.

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

[0079] The traditional self-attention mechanism and cross-attention mechanism use Query and Key to calculate the attention score matrix, and then use the calculated attention score and Value to weight the final feature. Therefore, this can only calculate the attention weight of the input to itself or another input, and only considers unary or binary relationships. TIAM further considers ternary relationships. TIAM uses three inputs to calculate a three-dimensional attention tensor Attention 3d To represent the ternary relationship between drug-drug-cell line, there are two queries in TIAM, namely Query1, Query2, Key, Value. TIAM requires three inputs, namely drug a, drug b, and cell line c, respectively represented by X a , X b and X c It means that after TIAM processing, the output is X′ a , X′ b and X′ c. Get X′ a , X′ b and X′ c The process of obtaining X′ is three similar processes, which are described in detail below. a , X′ b and X′ c process.

[0080] First, the Query Matrix is ​​based on the features X of drug a and drug b. a and X b Generate, Key Matrix and ValueMatrix based on cell line feature X c Generate as shown in the following formula.

[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] Among them, W Q1 , W Q2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q a , Q b , K c and V c Query a , Query b , Key and Value.

[0086] For the sake of explanation, we ignore Batch_size here, in order to get the three-dimensional attention tensor Attention 3d Based on the Q obtained above, a , Q b and K c ,in and Where I, J, and 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, which can be freely adjusted according to the model. Then the Einstein summation convention 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] The tensor And 'ik,jk,mk→ijm' is the calculation rule of formula 5, where i,j,m are free indexes and k is the sum index, indicating the result Attention 3d Each element of It needs to be summed over dimension k, which can be formalized as shown in formula (6).

[0089]

[0090] where a ik represents the eigenvalue of the i-th row and k-th column of the feature matrix of drug a, b jk represents the eigenvalue of the jth row and kth column of the feature matrix of drug b, c mk represents the eigenvalue of the mth row and kth column of the cell line c matrix, It is the sum of the products of these feature values, which preliminarily represents the degree of attention paid by 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 scores of drug a and drug b on different features of cell line c, we use the tensor Attention 3d The first two dimensions use global average pooling to obtain 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, use Attention ab2c With V c Perform weighted operation. Get X′ after adjusting the features c .

[0095] X′ c

[0096] =Attention ab2c

[0097] ⊙V c

[0098] Here ⊙ represents the Hademann product.

[0099] Get X′ a The process is similar. First, the Query Matrix is ​​based on the feature X of drug b and cell line c. b and X c Generate, Key Matrix and Value Matrix based on drug a X a Generate 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] Among them, W Q1 , W Q2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q b , Q c , K a and V a They 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] The three-dimensional attention tensor J, M, and I are the characteristic dimensions of drug b, cell line c, and drug a. ′ijk, mk, ik→jmi′ is the calculation rule of formula 13, which can be formalized as shown in formula (14).

[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 With V a Perform weighted operation. Get X′ after adjusting the features a , as shown in formula (16).

[0111] X′ a

[0112] =Attention bc2a

[0113] ⊙V a

[0114] Here ⊙ represents the Hademann product.

[0115] Get X′ b The process is similar. First, the Query Matrix is ​​based on the feature X of drug a and cell line c. a and X c Generate, Key Matrix and Value Matrix based on drug b X b Generate 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] Among them, W Q1 , W Q2 , W K , WV It is a parameter that can be automatically learned in the deep learning model. a , Q c , K b and V b They 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 ) (twenty one)

[0123] The three-dimensional attention tensor I, M, J are the characteristic dimensions of drug a, cell line c, and drug b. ′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 scores of drug b and cell line c on different features of drug a can be calculated.

[0126]

[0127] Finally, use Attention ac2b With V b Perform weighted operation. Get X′ after adjusting the features a , as shown in formula (15).

[0128] X′ a

[0129] =Attention ac2b

[0130] ⊙V a

[0131] Here ⊙ represents the Hademann product.

[0132] Taking into account factors such as possible gradient explosion, 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] The output X″ of TIAM is obtained a , X″ b , X″ c .

[0140] After TIAM, we obtain the three final hidden layer feature representations of drug a, drug b, and cell line c: a , X″ b and X″ c , flatten these features and concatenate them head to tail, and then put them into the prediction module to predict the final collaborative prediction result y′. The prediction module consists of a fully connected layer, a Relu activation function, and a Softmax activation function. Finally, we add a dropout layer to prevent overfitting. The details are as follows Figure 3 Specifically, Figure 3 As shown in the figure, a dropout layer is set between the Relu activation function and the Softmax activation function to prevent overfitting.

[0141] Interpretable Model CIMI Paper Citations:

[0142] Wu C, Wang

[0143] Previous work often uses attention scores for interpretability analysis. However, attention scores based on correlation are often not as reasonable as causal relationships. Therefore, we use the universal encoder-decoder interpretable model CIMI to perform causal interpretation on our black box model TSynergy. The model is generally divided into three modules. The only trainable parameters of the model are the decoder part of the interpreter, and the parameter values ​​of other parts are fixed. The specific process is as follows: Figure 1 shown.

[0144] The input of CIMI is the matrix x which is the concatenation of drug a and drug b cell line c, where After the interpreter g(·) is used, the explanation mask g(x) is obtained. Specifically, if g(x) i An element of corresponds to the i-th gene of cell line c, then g(x) i It can be understood as the causal probability that drug a and drug b act synergistically on this gene to synergistically treat cancer. CIMI designed three modules to obtain the causal explanation mask g(x). They are the causal sufficiency module, the causal intervention module, and the causal priority module.

[0145] Causal Sufficiency Module.

[0146] Use the explanation mask g(x) to get the causal feature x e and the non-causal feature x u , where x e =g(x)⊙x,x u =(1-g(x))⊙x, in order to ensure the causal characteristics x e is the predicted label This requires that the causal feature x u It is used to predict labels The most suitable feature, rather than other features. In other words, x e The label can always be predicted by an optimal function f′ Non-causal features are important for predicting labels There is no meaningful information. Therefore, the causal sufficiency loss can be modeled as shown in formula (9):

[0147]

[0148] in is the mean square error loss, x e =g(x)⊙x,x u=(1-g(x))⊙x. CIMI points out that since the input space is huge and continuous, it is difficult to find the best function f′. A good method is to set the best function f′ to f, that is, the black box model TSynergy, so 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 are independent, so that we can find invariant explanations x for similar input instances e , improve the generalization of the interpreter. Based on prior knowledge: for the non-interpreted part x u The intervention has no effect on the explanation part. Therefore, an intervention loss is designed to separate the explanation part.

[0153] First, according to the common practice in causal reasoning, the non-explanatory part x u Intervention: By calculating the non-explained part x u and linear interpolation of the non-explained portion of another variable randomly sampled from X to perturb the explained portion x u Specific examples are shown in formula (11):

[0154]

[0155] where λ~U(0,∈), and ∈ limits the size of the perturbation. Further, we can ensure x by optimizing the intervention loss of formula (12) e and x u Is independent.

[0156]

[0157] This loss ensures that the generated explanation x e In the perturbation non-interpreted part x u This invariance ensures the local consistency of the explanation. That is, the explanation generated by the interpreter for similar inputs is consistent. This helps capture more generalizable features and improves the generalization of the interpretable.

[0158] Causal prior Module:

[0159] There are two ways to help the interpreter learn: 1) Add potential causal hints to the interpreter's neural network. 2) Design a weak supervision loss for the causal mask g(x).

[0160] Design of interpreter neural network: One difficulty in interpreter design is the lack of prior knowledge of what network structure to choose. In other words, when we use neural network as interpreter, we don’t know what network structure to choose. If the interpreter is not as complex as the black box model f, how can we ensure that the interpreter is capable of understanding 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, and the decoding part of the interpreter uses a simple neural network, denoted as φ.

[0162] The encoder output With the original input example Splicing on the y axis As the input of decoderφ, as shown in formula (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. e The parameters of are fixed, so we only need to train the decoder of the interpreter to train a good explanation, avoiding a lot of training costs.

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

[0167]

[0168] Where x′ represents another instance other than x. x′(x) means mapping x′ to the interpretation mask part through g(x), g x′ (x)=φ([f e (x′); x′]). Relative g(x)=φ([f e (x); x]).

[0169] This weakly supervised loss can prevent the explainer from over-optimistically predicting all features as explanatory features, which helps reduce meaningless results.

[0170] Overall Framework and Optimization.

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

[0172]

[0173] Among them, α is a balance parameter. The weak supervision 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] Experiments were conducted based on the Oncology dataset.

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

[0177] Secondly, in terms of AUPR (Area Under the Precision-Recall Curve), PITSynergy also performed well, reaching 0.8921, surpassing SynergyX's 0.8651, an increase of about 2.97%. This result shows that PITSynergy has stronger capabilities in processing imbalanced data.

[0178] In terms of ACC (Accuracy), the accuracy of PIT Synergy is 0.7846, which is 1.68% higher than SynergyX (0.7716), showing the improvement of the model in overall prediction accuracy.

[0179] Finally, in terms of the Recall index, PITSynergy reached 0.7893, which is 6.47% higher than GraphSynergy (0.7414), indicating that PITSynergy is significantly better than other models in detecting positive samples. In summary, PITSynergy surpasses the existing baseline models in all evaluation indicators, especially in AUC and AUPR, which verifies the effectiveness and superiority of the proposed method in drug combination prediction tasks.

[0180] In order to verify the contribution of the proposed module to the model performance, we conducted an ablation experiment and removed the ternary attention module (TIA) and residual connection (Residule) respectively. The performance results under different configurations are given in the table. The experimental conclusions are as follows:

[0181] Ternary Attention Module (TIA): After removing TIA, the AUC of the model dropped from 0.8763 to 0.8636, and the AUPR dropped from 0.8921 to 0.8764. The performance decreased, indicating that the TIA module can effectively improve the discrimination ability of the model, especially in terms of AUC and AUPR.

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

[0183] Complete model (ours): The complete model including TIA and residual connection performs best in all indicators, especially in AUC and AUPR indicators, reaching 0.8763 and 0.8921 respectively, proving that the synergy of these two modules is crucial to the improvement of model performance.

[0184] Conclusion: The ablation experiment results show that the ternary attention module and residual connection have a significant effect on improving the model performance. The complete model achieves the best performance in terms of AUC, AUPR, ACC and Recall, which verifies the effectiveness of these two modules in the drug combination prediction task.

[0185] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A deep learning-based assisted screening method for synergistic prediction of anticancer drug combinations, characterized in that: The steps include: S1, extract drug substructure features to obtain drug vectors; S2, extract multi-omics features of cell lines to obtain vectors of cell lines; S3. The input feature of drug a is X a , the input feature of drug b is X b , the input features of cell line c are X c , X a , X b , X c Input to the TIAM model, the TIAM model processes the output to get X" a , X” b and X” c , the TIAM model uses three inputs to calculate a three-dimensional attention tensor Attention 3d to represent the ternary relationship between drug-drug-cell line; S4: Output X" a , X” b and X” c The output is obtained as input to the prediction module.

2. The method for assisted screening of anticancer drug combinations based on deep learning according to claim 1, characterized in that: In step S3, the Query Matrix is ​​based on the features X of drug a and drug b. a and X b Generate, Key Matrix and ValueMatrix based on cell line feature X c Generate, as shown in the following formula: 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) Among them, W Q1 , W Q2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q a , Q b , K c and V c Query a , Query b , Key and Value; and Where I, J, and 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; and then the Einstein summation convention einsum is used for calculation: Attention 3d =einsum('ik,jk,mk→ijm',Q a ,Q b ,K c ) (5) Among them: tensor And 'ik,jk,mk→ijm' is the calculation rule of formula (5), where i,j,m are free indexes and k is the sum index, indicating the result Attention 3d Each element of It needs to be summed over dimension k: Among them: a ik represents the eigenvalue of the i-th row and k-th column of the feature matrix of drug a, b jk represents the eigenvalue of the jth row and kth column of the feature matrix of drug b, c mk represents the eigenvalue of the mth row and kth column of the cell line c matrix, It is the sum of the products of these feature values, which preliminarily indicates the degree of attention paid by the i-th feature of drug a and the j-th feature of drug b to the m-th feature of the cell line; In order to obtain the attention scores of drug a and drug b on different features of cell line c, in the tensor Attention 3d The first two dimensions use global average pooling to obtain vector Attention ab2c : Where I and J represent the number of features of drug a and drug b, respectively; ∑ represents the sum operation, and Softmax is the softmax activation function; Finally, use Attention ab2c With V c Perform weighted operation to obtain X' after adjusting the features c , X' c =Attention ab2c ⊙V c (8) Where ⊙ represents the Hadamard product; the output of the TIAM model is processed to obtain X” c : X” c =X' c +X c (27)。 3. The method for assisted screening of anticancer drug combination synergy prediction based on deep learning according to claim 1, characterized in that: Query Matrix based on feature X of drug b and cell line c b and X c Generate, Key Matrix and Value Matrix based on drug a feature X a Generate, as shown in the following formula: 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) Among them, W Q1 , W Q2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q b , Q c , K a and V a They are Query1, Query2, Key and Value; Attention 3d =einsum('jk,mk,ik→jmi',Q b ,Q c ,K a ) (13) Where: 3D attention tensor J, M, I are the characteristic dimensions of drug b, cell line c, and drug a; and 'ijk, mk, ik→jmi' is the calculation rule of formula (13), which can be formalized as follows: Calculate the attention scores of drug b and cell line c for different features of drug a: Finally, use Attention bc2a With V a Perform weighted operation to obtain X' after adjusting the features a : X' a =Attention bc2a ⊙V a (16) Where: ⊙ represents the Hadamard product; the output of the TIAM model is obtained by processing X” a : X” a =X' a +X a (25)。 4. The method for assisted screening of anticancer drug combination synergy prediction based on deep learning according to claim 1, characterized in that: Query Matrix based on feature X of drug a and cell line c a and X c Generate, Key Matrix and Value Matrix based on drug b, X b Generate, as shown in the following formula: 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) Among them, W Q1 , W Q2 , W K , W V It is a parameter that can be automatically learned in the deep learning model; Q a , Q c , K b and V b They are Query1, Query2, Key and Value; three-dimensional attention tensor Attention 3d for: Attention 3d =einsum('ik,mk,jk→imj',Q a ,Q c ,K b ) (21) Where: 3D attention tensor I, M, J are the characteristic dimensions of drug a, cell line c, and drug b. And 'ij, mk, jk→imj' is the calculation rule of formula (21), and the specific calculation method can be formalized as follows: Calculate the attention scores of drug b and cell line c for different features of drug a: Finally, use Attention ac2b With V b Perform weighted operation to obtain X' after adjusting the features b : X’ b =Attention ac2b ⊙V b (24) Where: ⊙ represents the Hadamard product, and the output of the TIAM model is obtained by processing X” b : X” b =X’ b +X b (26)。 5. The method for assisted screening of anticancer drug combination synergy prediction based on deep learning according to claim 1, characterized in that: The drug feature in step S1 is represented as a shape of Vector.

6. The method for assisted screening of anticancer drug combination synergy prediction based on deep learning according to claim 1, characterized in that: In step S2, the cell line feature is represented by a shape of Vector.

7. The method for assisted screening of anticancer drug combination synergy prediction based on deep learning according to claim 1, characterized in that: The prediction module includes a fully connected layer, a Relu activation function and a Softmax activation function.

8. The method for synergistic prediction of anticancer drug combination based on deep learning for auxiliary screening according to claim 7, characterized in that: A dropout layer is provided between the Relu activation function and the Softmax activation function to prevent overfitting.

Images