Cancer cell line multi-omics based drug response prediction method and system

Abstract

This invention discloses a drug response prediction method based on multi-omics of cancer cell lines, including acquiring and preprocessing known cancer cell line data, drug data, and corresponding response information; calculating the pathway enrichment score of the cancer cell line and the molecular fingerprint of the drug and constructing a training dataset; constructing an initial drug response prediction model and training it to obtain the drug response prediction model; and using the drug response prediction model to perform actual drug response prediction. This invention also discloses a system for implementing the aforementioned drug response prediction method based on multi-omics of cancer cell lines. This invention fully utilizes multiple omics information, pathway enrichment information, and molecular fingerprints of the cancer cell line to obtain representation vectors for the cancer cell line and the drug, thereby using a bilinear attention network to obtain a fused representation of the cancer cell line and the drug and ultimately complete the prediction; therefore, this invention has higher reliability and better accuracy.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a method and system for predicting drug response based on multi-omics of cancer cell lines. Background Technology

[0002] Accurate prediction of drug response outcomes is of great significance for improving the efficiency of drug design and development.

[0003] The focus of drug response prediction research is the relationship between the drug entity and the cancer cell line entity. Typically, drug response prediction algorithms take the characteristics of various drugs and the omics data of cancer cell lines as input, learn and characterize the relationship between cancer cell lines and drugs, and output drug sensitivity values.

[0004] In recent years, deep learning-based approaches have proven better at handling high-dimensional, large-scale, and complex data, leading to a surge in researchers using deep learning for drug response prediction. Existing deep learning-based drug response prediction schemes demonstrate that utilizing diverse omics data from cancer cell lines and drug information can improve predictive performance. However, existing deep learning-based drug response prediction schemes still have some shortcomings. First, current schemes primarily treat gene expression as an independent variable, neglecting the specific biological pathway information it contains. Second, some studies indicate that current drug response prediction schemes suffer from insufficient ability to characterize different cancer cell lines. These shortcomings result in persistent issues of poor reliability and accuracy in existing deep learning-based drug response prediction schemes. Summary of the Invention

[0005] One of the objectives of this invention is to provide a highly reliable and accurate method for predicting drug response based on multi-omics in cancer cell lines.

[0006] The second objective of this invention is to provide a system for implementing the drug response prediction method based on cancer cell line multi-omics.

[0007] The drug response prediction method based on multi-omics of cancer cell lines provided by this invention includes the following steps:

[0008] S1. Obtain known cancer cell line data, drug data, and corresponding response information;

[0009] S2. Preprocess the data information obtained in step S1;

[0010] S3. Based on the preprocessed data obtained in step S2, calculate the pathway enrichment score of the cancer cell line and the molecular fingerprint of the drug, and construct a training dataset;

[0011] S4. Construct an initial model for drug response prediction based on a multilayer perceptron and a bilinear attention network;

[0012] S5. Using the training dataset obtained in step S3, train the initial drug response prediction model constructed in step S4 to obtain the drug response prediction model.

[0013] S6. Using the drug response prediction model obtained in step S5, perform actual drug response prediction.

[0014] Step S1, which involves acquiring known cancer cell line data, drug data, and corresponding response information, specifically includes the following steps:

[0015] Based on the GDSCv2 database, obtain known cancer cell line-drug response information;

[0016] Based on the CCLE database, omics data of known cancer cell lines are obtained; the omics data of the cancer cell lines include gene expression, genomic mutation and DNA methylation data of the cancer cell lines.

[0017] SMILES data for drugs were obtained using the Pubchem database.

[0018] Step S2, which involves preprocessing the data obtained in step S1, specifically includes the following steps:

[0019] For the acquired cancer cell line data, cancer cell lines with missing data in any omics domain were removed;

[0020] For the acquired cancer cell line multi-omics data, genes from the cancer-related human gene set in the COSMIC database were retained;

[0021] For the acquired drug data, remove drugs that are duplicates in the reaction information or whose labels do not match;

[0022] For the acquired cancer cell line-drug response information, the response information of the pretreated cancer cell line and the corresponding drug is retained.

[0023] Step S3, which involves calculating the pathway enrichment fraction of cancer cell lines and the molecular fingerprint of drugs based on the preprocessed data obtained in step S2, specifically includes the following steps:

[0024] The pathway enrichment fraction of cancer cell lines was calculated using gene set variation analysis.

[0025] Input biological pathway background gene set and cancer cell line gene g i ={g i1 ,g i2 ,...,gin The expression values ​​of g in n different cell lines in For the nth specific gene;

[0026] The gene g of each cancer cell line was obtained by estimating the nuclear density. i The probability density function value;

[0027] The standardized rank is used to replace the gene distribution function value. By calculating the Kolmokolov-Smilov random walk statistic, the statistical value of gene expression is transformed into the enrichment score of different pathways in the background gene set. Finally, the enrichment score of each cancer cell line in different biological pathways is output.

[0028] The molecular fingerprint of the drug was calculated using Python's RDKit and PyBioMed libraries; the molecular fingerprint included extended connectivity fingerprint, interpretable substructure molecular fingerprint, and PubChem substructure fingerprint.

[0029] Step S4, which involves constructing an initial model for drug response prediction based on a multilayer perceptron and a bilinear attention network, includes the following steps:

[0030] A cancer cell line feature representation module is constructed based on a multilayer perceptron; the cancer cell line feature representation module is used to learn the latent feature representation of cancer cell line omics data;

[0031] A drug feature representation module is constructed based on a multilayer perceptron; the drug feature representation module is used to learn the latent feature representation of drug data.

[0032] Based on the latent feature representations of the obtained cancer cell line omics data and drug data, a bilinear attention mapping is used to construct the interaction matrix of the cell line and drug pair. Then, a bilinear pooling layer is used to process the interaction matrix to obtain a joint representation. Finally, the final characterization is obtained through pooling.

[0033] A prediction module is constructed based on a multilayer perceptron. The final characterization is input into the prediction module to obtain the final prediction result of the drug response.

[0034] Step S4 specifically includes the following steps:

[0035] A cancer cell line feature representation module based on multilayer perceptron:

[0036] The cancer cell lineage set C is represented as {c1, c2, ..., c...} n}, where the i-th cancer cell line c i It includes four types of feature data, namely gene expression features. Gene mutation characteristics DNA methylation characteristics and pathway characteristics

[0037] For each feature of all cancer cell lines included in the cancer cell line set C, a corresponding latent representation is obtained by passing it through a multilayer perceptron: cancer cell line c i Potential characterization of gene expression features Cancer cell line c i Potential characterization of gene mutation features Cancer cell line c i Potential characterization of DNA methylation features and cancer cell line c i Potential representation of pathway characteristics And the cancer cell line c was obtained by splicing. i Latent feature representation

[0038] The computational formula for a multilayer perceptron is expressed as follows:

[0039]

[0040] In the formula Potential characterization of cancer cell line omics features; σ() is the activation function; w mlp b is the first parameter to be learned. mlp The second parameter to be learned; For cancer cell line c i The omics feature vector; the values ​​of mlp are exp, mut, meth, and path;

[0041] A drug feature representation module based on a multilayer perceptron:

[0042] The drug set D is represented as D = {d1, d2, ..., d...} m}; where the j-th drug d j It includes three types of molecular fingerprints, namely extended connectivity fingerprints. Molecular fingerprints that can explain substructure and PubChem substructure fingerprint

[0043] By passing the molecular fingerprint of each drug in the drug set D through a multilayer perceptron, the corresponding latent representation of drug d is obtained: j Potential representation of extended connectivity fingerprints Drug d j Potential representation of the molecular fingerprint of interpretable substructure and drug d j Potential representation of PubChem substructure fingerprints And by splicing together, drug d was obtained. jLatent feature representation for

[0044] The computational formula for a multilayer perceptron is expressed as follows:

[0045]

[0046] In the formula For potential representation of drug data; σ() is the activation function; w fp b is the third parameter to be learned. fp The fourth parameter to be learned; For drug d j Molecular fingerprint; fp can take the values ​​of ECFP, ESFP, and PSFP;

[0047] The interaction matrix between cell lines and drug pairs was constructed using bilinear attention mapping:

[0048] Representation of latent characteristics of cancer cell lineages and potential feature representation of drug sets Interaction matrix II of cancer cell line and drug pair was constructed using bilinear attention mapping:

[0049]

[0050] In the formula, I is a unit vector; q is the weight vector to be learned; The product is the Hadamard product; U is the first weight matrix to be learned; V is the second weight matrix to be learned.

[0051] The elements in Interaction Matrix II represent the interaction relationships between different characteristics of the drug and the cancer cell line;

[0052] The interaction matrix is ​​processed using a bilinear pooling layer to obtain a joint representation:

[0053] A bilinear pooling layer is used to process the interaction matrix II to obtain the joint representation f'; where the k-th element in f' is represented as U k Let V be the k-th column of U. k The k-th column of V;

[0054] The final characterization is obtained through pooling:

[0055] Through pooling, the final representation f is obtained as f = SumPool(f',s), where SumPool() is a one-dimensional non-overlapping pooling function with a span of s;

[0056] A prediction module is constructed based on a multilayer perceptron. The final characterization is input into the prediction module to obtain the final prediction result of the drug response.

[0057] The constructed prediction module is a multilayer perceptron with an output layer dimension of 1. The final representation f is input into the prediction module to obtain the final prediction result of the drug response.

[0058] The training described in step S5 specifically includes the following steps:

[0059] The following function is used as the loss function for model training:

[0060]

[0061] In the formula, nn represents the number of cancer cell line-drug response values; Y ii This represents the true value of the drug response value for the ii-th cancer cell line. This is the predicted value of the drug response value for the ii-th cancer cell line.

[0062] The Adam optimizer is used to iteratively update the model parameters.

[0063] Step S6, which uses the drug response prediction model obtained in step S5 to predict the actual drug response, specifically includes the following steps:

[0064] For the new cancer cell line or new drug to be predicted, the pathway enrichment fraction of the cancer cell line and the molecular fingerprint of the drug are calculated using the content described in step S3, and then input into the drug response prediction model obtained in step S5.

[0065] The prediction results of the drug response prediction model are sorted in descending order to generate a list of response values ​​for cancer cell line-drug pairs, thus completing the drug response prediction for the actual drug.

[0066] This invention also provides a system for implementing the drug response prediction method based on multi-omics of cancer cell lines, comprising a data acquisition module, a data processing module, a dataset construction module, a model construction module, a model training module, and a response prediction module; the data acquisition module, data processing module, dataset construction module, model construction module, model training module, and response prediction module are connected in series; the data acquisition module is used to acquire known cancer cell line data information, drug data information, and corresponding response information, and upload the data information to the data processing module; the data processing module is used to preprocess the acquired data information according to the received data information, and upload the data information to the dataset construction module; the dataset construction module is used to preprocess the acquired data information according to the received data information. According to the information, based on the preprocessed data, the pathway enrichment score of cancer cell lines and the molecular fingerprint of drugs are calculated, and a training dataset is constructed. The data information is then uploaded to the model building module. The model building module is used to construct an initial drug response prediction model based on a multilayer perceptron and a bilinear attention network according to the received data information, and the data information is uploaded to the model training module. The model training module is used to train the constructed initial drug response prediction model using the obtained training dataset according to the received data information, to obtain a drug response prediction model, and the data information is uploaded to the response prediction module. The response prediction module is used to perform actual drug response prediction based on the received data information and the obtained drug response prediction model.

[0067] The drug response prediction method and system based on multi-omics of cancer cell lines provided by this invention fully utilizes the multi-omics information, pathway enrichment information and molecular fingerprint of the cancer cell lines to obtain the characterization vectors of the cancer cell lines and drugs, and then uses a bilinear attention network to obtain the fused characterization of the cancer cell lines and drugs and finally complete the prediction; therefore, this invention has higher reliability and better accuracy. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0069] Figure 2 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0070] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The drug response prediction method based on multi-omics of cancer cell lines disclosed in this invention includes the following steps:

[0071] S1. Obtain known cancer cell line data, drug data, and corresponding response information; specifically including the following steps:

[0072] Based on the GDSCv2 database, obtain known cancer cell line-drug response information;

[0073] Based on the CCLE database, omics data of known cancer cell lines are obtained; the omics data of the cancer cell lines include gene expression, genomic mutation and DNA methylation data of the cancer cell lines.

[0074] SMILES data for drugs are obtained based on the Pubchem database;

[0075] S2. Preprocess the data obtained in step S1; specifically, this includes the following steps:

[0076] For the acquired cancer cell line data, cancer cell lines with missing data in any omics domain were removed;

[0077] For the acquired cancer cell line multi-omics data, the genetic data (approximately 733) from the cancer-related human gene set in the COSMIC database were retained;

[0078] For the acquired drug data, remove drugs that are duplicates in the reaction information or whose labels do not match;

[0079] For the acquired cancer cell line-drug response information, the response information of the pretreated cancer cell lines and the corresponding drugs is retained;

[0080] S3. Based on the preprocessed data obtained in step S2, calculate the pathway enrichment score of the cancer cell lines and the molecular fingerprint of the drugs, and construct a training dataset; specifically including the following steps:

[0081] Gene set variation analysis (GSVA) ​​was used; GSVA is a non-parametric, unsupervised analysis method to calculate the pathway enrichment scores of cancer cell lines.

[0082] Input biological pathway background gene set and cancer cell line gene g i ={g i1 ,g i2 ,...,g in The expression values ​​of g in n different cell lines in For the nth specific gene;

[0083] The gene g of each cancer cell line was obtained by estimating the nuclear density. i The probability density function value;

[0084] To reduce errors, the standardized rank is used instead of the gene distribution function value. By calculating the Kolmokolov-Smilov random walk statistic, the statistical value of gene expression is transformed into the enrichment score of different pathways in the background gene set. Finally, the enrichment score of each cancer cell line in different biological pathways is output.

[0085] The molecular fingerprints of drugs were calculated using Python's RDKit and PyBioMed libraries. These molecular fingerprints included Extended Connectivity Fingerprints (ECFP), Explainable Substructure Partition Fingerprints (ESPF), and PubChem Substructure Fingerprints (PSFP).

[0086] S4. Construct an initial model for drug response prediction based on a multilayer perceptron and a bilinear attention network; including the following steps:

[0087] A cancer cell line feature representation module is constructed based on a multilayer perceptron; the cancer cell line feature representation module is used to learn the latent feature representation of cancer cell line omics data;

[0088] A drug feature representation module is constructed based on a multilayer perceptron; the drug feature representation module is used to learn the latent feature representation of drug data.

[0089] Based on the latent feature representations of the obtained cancer cell line omics data and drug data, a bilinear attention mapping is used to construct the interaction matrix of the cell line and drug pair. Then, a bilinear pooling layer is used to process the interaction matrix to obtain a joint representation. Finally, the final characterization is obtained through pooling.

[0090] A prediction module is constructed based on a multilayer perceptron. The final characterization is input into the prediction module to obtain the final prediction result of the drug response.

[0091] In practice, this step includes the following:

[0092] A cancer cell line feature representation module based on multilayer perceptron:

[0093] The cancer cell lineage set C is represented as {c1, c2, ..., c...} n}, where the i-th cancer cell line c i It includes four types of feature data, namely gene expression features. Gene mutation characteristics DNA methylation characteristics and pathway characteristics

[0094] For each feature of all cancer cell lines included in the cancer cell line set C, a corresponding latent representation (with the same dimension) is obtained by passing it through a multilayer perceptron (there are four independent multilayer perceptrons for the four features, each feature passing through its own corresponding multilayer perceptron): cancer cell line c i Potential characterization of gene expression features Cancer cell line c i Potential characterization of gene mutation features Cancer cell line c i Potential characterization of DNA methylation features and cancer cell line c i Potential representation of pathway characteristics And the cancer cell line c was obtained by splicing. i Latent feature representation

[0095] The computational formula for a multilayer perceptron is expressed as follows:

[0096]

[0097] In the formula Potential characterization of cancer cell line omics features; σ() is the activation function; w mlp b is the first parameter to be learned. mlp The second parameter to be learned; For cancer cell line c i The omics feature vector; the values ​​of mlp are exp, mut, meth, and path;

[0098] A drug feature representation module based on a multilayer perceptron:

[0099] The drug set D is represented as D = {d1, d2, ..., d...} m}; where the j-th drug d j It includes three types of molecular fingerprints, namely extended connectivity fingerprints. Molecular fingerprints that can explain substructure and PubChem substructure fingerprint

[0100] For each molecular fingerprint of all drugs included in drug set D, pass it through a multilayer perceptron (3 molecular fingerprints correspond to 3 independent multilayer perceptrons, each molecular fingerprint is passed through its own corresponding multilayer perceptron) to obtain the corresponding latent representation: drug d j Potential representation of extended connectivity fingerprints Drug d j Potential representation of the molecular fingerprint of interpretable substructure and drug d jPotential representation of PubChem substructure fingerprints And by splicing together, drug d was obtained. j Latent feature representation for

[0101]

[0102] The computational formula for a multilayer perceptron is expressed as follows:

[0103]

[0104] In the formula For potential representation of drug data; σ() is the activation function; w fp b is the third parameter to be learned. fp The fourth parameter to be learned; For drug d j Molecular fingerprint; fp can take the values ​​of ECFP, ESFP, and PSFP;

[0105] The interaction matrix between cell lines and drug pairs was constructed using bilinear attention mapping:

[0106] Representation of latent characteristics of cancer cell lineages and potential feature representation of drug sets Interaction matrix II of cancer cell line and drug pair was constructed using bilinear attention mapping:

[0107]

[0108] In the formula, I is a unit vector; q is the weight vector to be learned; The product is the Hadamard product; U is the first weight matrix to be learned; V is the second weight matrix to be learned.

[0109] The elements in Interaction Matrix II represent the interaction relationships between different characteristics of the drug and the cancer cell line;

[0110] The interaction matrix is ​​processed using a bilinear pooling layer to obtain a joint representation:

[0111] A bilinear pooling layer is used to process the interaction matrix II to obtain the joint representation f'; where the k-th element in f' is represented as U k Let V be the k-th column of U. k The k-th column of V;

[0112] The final characterization is obtained through pooling:

[0113] Through pooling, the final representation f is obtained as f = SumPool(f',s), where SumPool() is a one-dimensional non-overlapping pooling function with a span of s;

[0114] A prediction module is constructed based on a multilayer perceptron. The final characterization is input into the prediction module to obtain the final prediction result of the drug response.

[0115] The constructed prediction module is a multilayer perceptron with an output layer dimension of 1. The final representation f is input into the prediction module to obtain the final prediction result of the drug response.

[0116] S5. Using the training dataset obtained in step S3, train the initial drug response prediction model constructed in step S4 to obtain the drug response prediction model; wherein, the training specifically includes the following steps:

[0117] The following function is used as the loss function for model training:

[0118]

[0119] In the formula, nn represents the number of cancer cell line-drug response values; Y ii This represents the true value of the drug response value for the ii-th cancer cell line. The predicted value of the drug response value for the ii-th cancer cell line.

[0120] The Adam optimizer is used to iteratively update the model parameters;

[0121] S6. Using the drug response prediction model obtained in step S5, perform actual drug response prediction; specifically including the following steps:

[0122] For the new cancer cell line or new drug to be predicted, the pathway enrichment fraction of the cancer cell line and the molecular fingerprint of the drug are calculated using the content described in step S3, and then input into the drug response prediction model obtained in step S5.

[0123] The prediction results of the drug response prediction model are sorted in descending order to generate a list of response values ​​for cancer cell line-drug pairs, thus completing the drug response prediction for the actual drug.

[0124] The bilinear attention network model is a model proposed to solve visual question answering problems. It extends a single attention network through a bilinear attention mechanism to adapt to multimodal features of paired inputs. Compared with single attention, bilinear attention can provide richer fusion information at the same computational cost.

[0125] To address the shortcomings of current deep learning-based drug response prediction schemes, this invention utilizes multi-omics data from cancer cell lines, the enrichment scores of cancer cell lines in different biological pathways, and multiple molecular fingerprints of drugs. By fusing multi-dimensional information from cancer cell lines and drugs using a bilinear attention module, the above-mentioned drug response prediction method based on multi-omics of cancer cell lines was designed.

[0126] The following simulation examples compare the effectiveness of the method of this invention with existing methods:

[0127] To verify the effectiveness of this invention, three verification methods were adopted, referring to the verification standards of computational models in other related fields: (1) comparative experiments; (2) independent testing; (3) Denovo experiments and five evaluation indicators: root mean square error (RMSE), mean absolute error (MAE), coefficient of determination (R²), and five evaluation indicators. 2 The scheme was evaluated using Pearson correlation coefficient (PCC) and Spearman correlation coefficient (SCC).

[0128] A benchmark dataset containing 81,467 IC50 values ​​of 169 drugs and 536 cancer cell lines was collected from the GDSCv2 database. Additionally, an independent test set containing 8,909 IC50 values ​​of 551 cell lines and 184 drugs was collected from the CCLE database. The cancer cell lines and drugs in the independent test set were not included in the benchmark dataset. In the comparative experiments, the benchmark dataset was split into training, validation, and test sets in an 8:1:1 ratio for both training and testing. The model obtained from the comparative experiments was then tested on the independent test dataset. In the Denovo experiments, two strategies were used to test the model: the first strategy split the cancer cell lines in the benchmark dataset into training, validation, and test sets in an 8:1:1 ratio, excluding drug responses related to cell lines in the validation and test sets from the training set; the second strategy split the drugs in the benchmark dataset into training, validation, and test sets in an 8:1:1 ratio, excluding drug-related response values ​​from the validation and test sets from the training set.

[0129] To verify the effectiveness of the proposed method in drug response prediction, the method of this invention was compared with five existing deep learning-based drug response prediction methods: tCNNs, DeepCDR, Precily, DeepTTA, and GADRP. Table 1 shows the comparison results between the proposed method and the baseline methods.

[0130] Table 1. Comparison Results (Symptom Table)

[0131] tCNNs 1.1574 0.8509 0.8166 0.9042 0.8776 DeepCDR 1.0189 0.7538 0.8579 0.927 0.9003 Precily 1.1136 0.8345 0.8302 0.9125 0.8792 DeepTTA 0.9859 0.7226 0.8394 0.9167 0.8853 GADRP 0.9953 0.7263 0.8402 0.9298 0.9097 Method of the present invention 0.9573 0.6987 0.8745 0.9353 0.9112

[0132] As can be seen from Table 1, the method of the present invention achieves good results in RMSE, MAE, and R...2 It outperformed other baseline models in PCC and SCC, and improved by 3.8%, 3.8%, 4%, 0.6%, and 0.2% respectively compared to the second-best GADRP scheme.

[0133] To further evaluate the model's performance, this invention also conducted independent tests using an independent dataset from the CCLE database. Since hyperparameter tuning and model training are independent of the independent test set, independent testing can better measure the model's generalization ability to unknown data. Experimental results using a model trained on a benchmark dataset are shown in Table 2 below.

[0134] Table 2. Schematic diagram of independent test results

[0135] tCNNs 2.0174 1.5780 0.2395 0.3706 0.3354 DeepCDR 1.6519 1.2948 0.4399 0.5026 0.4739 Precily 1.9919 1.5618 0.2881 0.4091 0.3698 DeepTTA 1.5436 1.2420 0.5118 0.6469 0.5082 GADRP 1.6465 1.2247 0.4514 0.4717 0.4217 Method of the present invention 1.3981 1.0844 0.5167 0.7561 0.5565

[0136] As shown in Table 2, the method of this invention outperforms other baseline models on the independent test set. The independent test results demonstrate that the method of this invention has good generalization ability.

[0137] To verify the ability of the method of the present invention to predict novel cancer cell line-drug responses, we conducted Denovo experiments on the method of the present invention and five other baseline models under two experimental strategies. The experimental results are shown in Tables 3 and 4 below.

[0138] Table 3. Schematic diagram of Denovo experiment results for cancer cell line segmentation.

[0139] tCNNs 1.6429 1.2348 0.6293 0.7952 0.7408 DeepCDR 1.4068 1.0552 0.7296 0.8604 0.8200 Precily 1.4399 1.0750 0.7154 0.8484 0.8030 DeepTTA 1.3512 1.0114 0.7506 0.8664 0.8288 GADRP 1.3528 1.0166 0.6589 0.8670 0.8289 Method of the present invention 1.2998 0.9692 0.7692 0.8783 0.8445

[0140] Table 4. Schematic diagram of Denovo experimental results for drug fractionation.

[0141] tCNNs 2.0739 1.6611 0.2422 0.5064 0.4298 DeepCDR 1.8569 1.3699 0.5171 0.7399 0.6470 Precily 1.7055 1.2975 0.5543 0.7507 0.6646 DeepTTA 1.8641 1.1292 0.4929 0.7151 0.6647 GADRP 1.7965 1.2970 0.4680 0.7501 0.6789 Method of the present invention 1.6348 1.2739 0.5926 0.7738 0.6969

[0142] Table 3 shows the experimental results based on cancer cell line segmentation. The method of this invention achieved the best performance when faced with data from new cancer cell lines, indicating that the model has good generalization performance due to multi-omics data and pathway information. Table 4 shows the experimental results based on drug segmentation. As can be seen from Table 4, the performance of all models decreased, which may be due to the diversity of drug molecular structures and the similar genetic information in cancer cell lines. However, the performance of the method of this invention is still superior to other baseline models. The above studies demonstrate that the method of this invention has good adaptability and generalization on both new drugs and new cancer cell lines.

[0143] like Figure 2The diagram shows the functional modules of the system of the present invention: The system disclosed in this invention for implementing the drug response prediction method based on multi-omics of cancer cell lines includes a data acquisition module, a data processing module, a dataset construction module, a model construction module, a model training module, and a response prediction module; these modules are connected in series. The data acquisition module acquires known cancer cell line data, drug data, and corresponding response information, and uploads the data to the data processing module. The data processing module preprocesses the acquired data based on the received data and uploads the data to the dataset construction module. The dataset construction module... Based on the received data and preprocessed data, the system calculates the pathway enrichment score of cancer cell lines and the molecular fingerprint of drugs, constructs a training dataset, and uploads the data to the model building module. The model building module, based on the received data and using a multilayer perceptron and bilinear attention network, constructs an initial drug response prediction model and uploads the data to the model training module. The model training module, using the received data and the obtained training dataset, trains the initial drug response prediction model to obtain a drug response prediction model and uploads the data to the response prediction module. The response prediction module, using the obtained drug response prediction model, performs actual drug response prediction based on the received data.

Claims

1. A method for predicting drug response based on multi-omics in cancer cell lines, comprising the following steps: S1. Obtain known cancer cell line data, drug data, and corresponding response information; S2. Preprocess the data obtained in step S1; S3. Based on the preprocessed data obtained in step S2, calculate the pathway enrichment score of the cancer cell line and the molecular fingerprint of the drug, and construct a training dataset; S4. Construct an initial model for drug response prediction based on a multilayer perceptron and a bilinear attention network; including the following steps: A cancer cell line feature representation module is constructed based on a multilayer perceptron; the cancer cell line feature representation module is used to learn the latent feature representation of cancer cell line omics data; A drug feature representation module is constructed based on a multilayer perceptron; the drug feature representation module is used to learn the latent feature representation of drug data. Based on the latent feature representations of the obtained cancer cell line omics data and drug data, a bilinear attention mapping is used to construct the interaction matrix of the cell line and drug pair. Then, a bilinear pooling layer is used to process the interaction matrix to obtain a joint representation. Finally, the final characterization is obtained through pooling. A prediction module is constructed based on a multilayer perceptron. The final characterization is input into the prediction module to obtain the final prediction result of the drug response. The specific implementation includes the following steps: A cancer cell line feature representation module based on multilayer perceptron: Cancer cell lineage collection Represented as Among them, the i-th cancer cell line It includes four types of feature data, namely gene expression features. Gene mutation characteristics DNA methylation characteristics and pathway characteristics ; Grouping cancer cell lines Each feature of all included cancer cell lines is processed through a multilayer perceptron to obtain a corresponding latent representation: cancer cell line Potential characterization of gene expression features Cancer cell line Potential characterization of gene mutation features Cancer cell line Potential characterization of DNA methylation features and cancer cell line Potential representation of pathway characteristics And spliced ​​together to obtain cancer cell lines Latent feature representation ; The computational formula for a multilayer perceptron is expressed as follows: In the formula Potential characterization of cancer cell line omics features; For activation functions; The first parameter to be learned; The second parameter to be learned; Cancer cell line omics feature vectors; The value is , , and ; A drug feature representation module based on a multilayer perceptron: Drug Collection Represented as ;in, The jth drug It includes three types of molecular fingerprints, namely extended connectivity fingerprints. Molecular fingerprints that can explain substructure And PubChem substructure fingerprint ; drug collection Each molecular fingerprint of all included drugs is processed through a multilayer perceptron to obtain a corresponding latent representation: drug Potential representation of extended connectivity fingerprints ,drug Potential representation of the molecular fingerprint of interpretable substructure and medicine Potential representation of PubChem substructure fingerprints And pieced together the drugs Latent feature representation for ; The computational formula for a multilayer perceptron is expressed as follows: In the formula A potential representation of drug data; For activation functions; The third parameter to be learned; The fourth parameter to be learned; For drugs Molecular fingerprints; The value is , and ; The interaction matrix between cell lines and drug pairs was constructed using bilinear attention mapping: Representation of latent characteristics of cancer cell lineages Potential characteristics of drug sets The interaction matrix between cancer cell lines and drug pairs was constructed using bilinear attention mapping. : In the formula It is a unit vector; Let be the weight vector to be learned; The product of Hadamard; The first weight matrix to be learned; This is the second weight matrix to be learned; Interaction matrix The elements in the text represent the interaction relationships between different characteristics of the drug and the cancer cell line; The interaction matrix is ​​processed using a bilinear pooling layer to obtain a joint representation: A bilinear pooling layer is used to transform the interaction matrix. Processing yields a joint representation ;in, The k-th element in is represented as , for The kth column, for The kth column; The final characterization is obtained through pooling: The final characterization is obtained through pooling. for ,in For the span is One-dimensional non-overlapping pooling functions; A prediction module is constructed based on a multilayer perceptron. The final characterization is input into the prediction module to obtain the final prediction result of the drug response. The constructed prediction module is a multilayer perceptron with an output layer dimension of 1, which will represent the final data. The data is input into the prediction module to obtain the final predicted drug response. S5. Using the training dataset obtained in step S3, train the initial drug response prediction model constructed in step S4 to obtain the drug response prediction model. S6. Using the drug response prediction model obtained in step S5, perform actual drug response prediction.

2. The drug response prediction method based on multi-omics of cancer cell lines according to claim 1, characterized in that... Step S1, which involves acquiring known cancer cell line data, drug data, and corresponding response information, specifically includes the following steps: Based on the GDSCv2 database, obtain known cancer cell line-drug response information; Based on the CCLE database, omics data of known cancer cell lines are obtained; the omics data of the cancer cell lines include gene expression, genomic mutation and DNA methylation data of the cancer cell lines. SMILES data for drugs were obtained using the Pubchem database.

3. The drug response prediction method based on multi-omics of cancer cell lines according to claim 2, characterized in that... Step S2, which involves preprocessing the data obtained in step S1, specifically includes the following steps: For the acquired cancer cell line data, cancer cell lines with missing data in any omics domain were removed; For the acquired omics data of cancer cell lines, genes from the cancer-related human gene set in the COSMIC database are retained; For the acquired drug data, remove drugs that are duplicates in the reaction information or whose labels do not match; For the acquired cancer cell line-drug response information, the response information of the pretreated cancer cell line and the corresponding drug is retained.

4. The drug response prediction method based on multi-omics of cancer cell lines according to claim 3, characterized in that... Step S3, which involves calculating the pathway enrichment fraction of cancer cell lines and the molecular fingerprint of drugs based on the preprocessed data obtained in step S2, specifically includes the following steps: The pathway enrichment fraction of cancer cell lines was calculated using gene set variation analysis. Background gene set of input biological pathways and genes of cancer cell lines Expression values ​​in n different cell lines, For the nth specific gene; The genes of each cancer cell line were obtained by kernel density estimation. The probability density function value; The standardized rank is used to replace the gene distribution function value. By calculating the Kolmokolov-Smilov random walk statistic, the statistical value of gene expression is transformed into the enrichment score of different pathways in the background gene set. Finally, the enrichment score of each cancer cell line in different biological pathways is output. The molecular fingerprint of the drug was calculated using Python's RDKit and PyBioMed libraries; the molecular fingerprint included extended connectivity fingerprint, interpretable substructure molecular fingerprint, and PubChem substructure fingerprint.

5. The drug response prediction method based on multi-omics of cancer cell lines according to claim 4, characterized in that... The training described in step S5 specifically includes the following steps: The following function is used as the loss function for model training. : In the formula The number of cancer cell line-drug response values; This represents the true value of the drug response value for the ii-th cancer cell line. This is the predicted value of the drug response value for the ii-th cancer cell line; The Adam optimizer is used to iteratively update the model parameters.

6. The drug response prediction method based on multi-omics of cancer cell lines according to claim 5, characterized in that... Step S6, which uses the drug response prediction model obtained in step S5 to predict the actual drug response, specifically includes the following steps: For the new cancer cell line or new drug to be predicted, the pathway enrichment fraction of the cancer cell line and the molecular fingerprint of the drug are calculated using the content described in step S3, and then input into the drug response prediction model obtained in step S5. The prediction results of the drug response prediction model are sorted in descending order to generate a list of response values ​​for cancer cell line-drug pairs, thus completing the drug response prediction for the actual drug.

7. A system for implementing the drug response prediction method based on multi-omics of cancer cell lines as described in any one of claims 1 to 6, characterized in that... It includes a data acquisition module, a data processing module, a dataset construction module, a model construction module, a model training module, and a response prediction module; the data acquisition module, data processing module, dataset construction module, model construction module, model training module, and response prediction module are connected in series; the data acquisition module is used to acquire known cancer cell line data information, drug data information, and corresponding response information, and upload the data information to the data processing module; The data processing module is used to preprocess the acquired data based on the received data information and upload the data information to the dataset construction module; The dataset construction module is used to calculate the pathway enrichment score of cancer cell lines and the molecular fingerprint of drugs based on the received data and the preprocessed data, and to build a training dataset and upload the data information to the model construction module; the model construction module is used to build an initial drug response prediction model based on a multilayer perceptron and a bilinear attention network based on the received data and to upload the data information to the model training module. The model training module is used to train the initial drug response prediction model based on the received data and the obtained training dataset, thereby obtaining the drug response prediction model, and then uploads the data to the response prediction module; the response prediction module is used to perform actual drug response prediction based on the received data and the obtained drug response prediction model.

Images