Method for predicting compound target based on compound perturbation transcriptome and protein sequence and application

Abstract

The invention provides a method for predicting a compound target based on a compound perturbation transcriptome and a protein sequence and application, and the method creatively combines the characteristics of the compound perturbation transcriptome and the characteristics of the protein target sequence as input characteristics and combines a deep learning model to realize prediction of the compound target. Specifically, perturbation transcriptome and active and inactive target spot information data of a compound are collected, transcriptome features are extracted through principal component analysis and linear discriminant analysis, meanwhile, a protein language model is used for representing target spot protein features, the transcriptome and a target spot of the compound are converted into learnable embedding, and the target spot information of the compound is obtained. And the classification model is used for training and learning potential association of the two, and is finally used for compound target prediction. According to the method, an external test set is used for proving high accuracy and practical applicability of the model. The method has the advantages of being low in cost, high in efficiency and wide in range, the compound target spot discovery efficiency can be remarkably improved, and the research and development process of drugs can be accelerated.

Description

Technical Field

[0001] The present invention relates to the technical field of compound target prediction in the drug discovery process, and in particular to a method for predicting compound targets based on a combination of perturbation transcriptome data, protein sequences and deep learning. Background Art

[0002] Drug development is a vital task in the field of modern medicine, and it is of great significance for protecting and improving human health. However, drug development also faces many challenges and difficulties. Drug development is a long and expensive process. According to statistics, the development of a drug usually spans 10-15 years and costs billions of dollars. The process of drug discovery is an intricate process that involves the identification of new drug targets, the design and synthesis of compounds, and the evaluation of their efficacy and safety. Target discovery is a crucial first step in the modern drug development process. Drug targets usually refer to the binding sites of drugs in the body. However, due to the complex biological systems in the human body and the complexity of compound-protein interactions, there are still many challenges in determining drug targets.

[0003] Traditional experimental methods usually identify targets based on changes in relevant characteristics after protein binding to drugs. For example, the cellular thermal shift assay (CETSA) uses the principle that target proteins usually become more stable after binding to drug molecules; the drug affinity reaction target stability analysis (DARTS) is based on the phenomenon that target proteins are resistant to proteases after binding to drugs, and is used to identify the binding relationship between drugs and target proteins. In addition, there are methods such as stability analysis based on protein oxidation rate. However, the above methods are time-consuming and labor-intensive. Computational target prediction methods have become an important supplementary tool for experiments due to their low cost and high throughput advantages. Structure-based target prediction methods, such as reverse virtual screening, are based on the principle of molecular docking. For a given ligand molecule, a molecular docking program is used to dock the small molecule with the protein receptors in the database one by one, and finally identify the drug target protein. However, they are limited by protein crystal structure data and computing power requirements.

[0004] The rapid accumulation of perturbation transcriptional profile data provides new perspectives for computational target prediction. One strategy is to link the transcriptional profile of the query compound with that of known drugs and infer targets based on similarity of gene features. One example is the connectivity map (cMAP) study, which assigns new molecules to drug targets with the most similar transcriptional profiles by comparing the transcriptional profiles of known drugs and query compounds. Another strategy is protein network-based analysis, such as ProTINA, which uses cell type-specific protein-gene transcriptional regulation dynamic models to infer drug targets from gene transcription profiles.

[0005] Although the above strategies successfully use perturbed transcriptional profiles to predict compound targets, their accuracy still has room for improvement. Perturbed transcriptional profile data have been subjected to multiple experiments under different perturbation conditions, which inevitably produces data noise due to environmental factors, technical operations, etc. At the same time, how to characterize protein targets has always been a challenge, making it difficult for the model to learn the potential relationship between the two. Summary of the invention

[0006] The purpose of the present invention is to address the deficiencies of the prior art. Principal component analysis (PCA) and linear discriminant analysis (LDA) can be used to extract features from transcriptome data with noisy backgrounds, while the protein language model, which has been popular in recent years, can characterize the target protein sequence. Therefore, the two are used together in combination with a deep learning algorithm to provide a method for predicting compound targets based on compound-perturbed transcriptomes and protein sequences, which assists in drug target discovery and solves the deficiencies of current prediction technologies in accuracy and efficiency.

[0007] The present invention is achieved through the following technical solutions:

[0008] A method for predicting compound targets based on transcriptome and protein sequence perturbations by compounds, comprising:

[0009] Obtain perturbed transcriptome data of the compound to be predicted and extract perturbed transcriptome features through principal component analysis and linear discriminant analysis;

[0010] Collect a number of protein sequence data and use protein language models to convert protein sequences into embeddings that can be used for deep learning to represent protein features;

[0011] Connect the perturbed transcriptome features one by one with the protein features to form a new feature vector;

[0012] The new feature vectors are input one by one into a trained classification model, and the classification model outputs a prediction result of whether the protein corresponding to the input new feature vector is a target of the compound.

[0013] Furthermore, the perturbation transcriptome features are extracted by principal component analysis and linear discriminant analysis, specifically as follows: first, principal component analysis is performed on the perturbation transcriptome data, the principal components of the data are calculated and the dimension is reduced, and the high-dimensional data is mapped to a low-dimensional space to remove noise and redundant information while retaining the main variation characteristics; then, based on the low-dimensional data after dimension reduction, a linear discriminant analysis is used to learn a linear transformation matrix that maximizes intra-class similarity and inter-class difference; finally, the data after PCA dimension reduction is projected based on the linear transformation matrix and mapped to a new feature space, thereby obtaining perturbation transcriptome features that characterize the perturbation pattern. In the new feature space, perturbation transcriptome data points with the same label are clustered together, while data points with different labels are separated.

[0014] Furthermore, the perturbed transcriptome data of the compound to be predicted comes from the CMAP LINCS2020Level 5 dataset, and the protein sequence data comes from UniProt.

[0015] Furthermore, the classification model adopts multi-layer perceptron, random forest, support vector machine, convolutional neural network, etc.

[0016] Furthermore, the classification model is obtained by training based on the data set with the goal of minimizing the error between the classification model output and the label.

[0017] Furthermore, each sample of the data set includes a new feature vector and a label formed by connecting the transcriptome feature and protein feature of the corresponding compound perturbation, and the label is set in the following manner:

[0018] If the IC50, EC50 and Ki values ​​of the corresponding compound to the protein are less than or equal to 10 μM, the protein is the active target of the corresponding compound and the label is 1; if the IC50, EC50 and Ki values ​​of the corresponding compound to the protein are greater than 100 μM, the protein is the inactive target of the corresponding compound and the label is 0.

[0019] Furthermore, in the data set, among samples containing the same compound, the ratio of the number of samples with label 0 to the number of samples with label 1 is greater than 3:1.

[0020] Furthermore, the classification model also includes performing hyperparameter combination search optimization on the validation set.

[0021] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a method for predicting compound targets based on transcriptome and protein sequence perturbed by compounds is implemented.

[0022] A storage medium containing computer executable instructions, which, when executed by a computer processor, implements a method for predicting compound targets based on transcriptome and protein sequence perturbations by compounds.

[0023] The beneficial effects of the present invention are as follows: the drug target prediction method of the present invention can theoretically be applied to any compound, with compound perturbation transcriptome data and target protein sequence as input features of the model, the model has high accuracy, and the model can be used to perform batch screening on a large number of compound databases, with the advantages of low cost and high efficiency. This method has broad application prospects in the direction of auxiliary drug targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 : Flow chart of the construction of compound target prediction model in the present invention.

[0025] Figure 2 : In an optional embodiment proposed by the present invention, a graph showing the changes in F1 value and AUPRC value of different parameter combinations on the validation set for each of 20 hyperparameter searches using Optuna.

[0026] Figure 3 : In an optional embodiment proposed by the present invention, the training process of the hyperparameter training model of Trial9 is used to show the changes in AUPRC on the training set (Training) and the test set (Test) as the number of training rounds increases, as well as the evaluation results of the fixed model on the test set.

[0027] Figure 4 : In an optional embodiment proposed by the present invention, a Top100 accuracy comparison result diagram of two data forms (log2FC and MODZ) in actual application of predicting natural product targets.

[0028] Figure 5 : In an optional embodiment proposed by the present invention, the model predicts 4 natural products with unknown target information and displays their Top 30 predicted target results. DETAILED DESCRIPTION

[0029] In order to better illustrate the purpose, technical solutions and advantages of the present invention, the present invention is further described below in conjunction with specific embodiments.

[0030] The present invention provides a method for predicting compound targets based on transcriptome and protein sequence perturbations by compounds, the method comprising:

[0031] Obtain perturbed transcriptome data of the compound to be predicted and extract perturbed transcriptome features through principal component analysis and linear discriminant analysis;

[0032] Collect a number of protein sequence data and use protein language models to convert protein sequences into embeddings that can be used for deep learning to represent protein features;

[0033] Connect the perturbed transcriptome features one by one with the protein features to form a new feature vector;

[0034] The new feature vectors are input one by one into a trained classification model, and the classification model outputs a prediction result of whether the protein corresponding to the input new feature vector is a target of the compound.

[0035] The classification model is obtained by training based on the data set with the goal of minimizing the error between the classification model output and the label. The effect of the present invention is further described below in conjunction with a specific embodiment.

[0036] First, build and train a classification model such as Figure 1 As shown, the following steps are included:

[0037] Step 1: construct a data set, wherein each sample of the data set includes a new feature vector and a label formed by connecting the transcriptome feature and protein feature of the corresponding compound perturbation, specifically including the following sub-steps:

[0038] Step 1.1: Data collection and processing:

[0039] The CMAPLINCS2020 level 5 dataset was downloaded from the clue database (https: / / clue.io / data / CMap2020#LINCS2020), which contains 720,216 perturbed transcriptomes obtained from 34,419 small molecule compounds in hundreds of cell lines and under different perturbation conditions. The compound and protein target affinity data were downloaded from the ChEMBL database (https: / / www.ebi.ac.uk / chembl / ) and screened. The screening conditions were as follows: only human target data were used, and the protein with IC50, EC50 and Ki values ​​of the compound to the protein was defined as the active target of the corresponding compound when it was less than or equal to 10 μM, and the protein with a value greater than 100 μM was defined as the inactive target of the corresponding compound. After screening, 553,347 compounds and their target information were finally obtained.

[0040] As a preferred embodiment, in order to match the distribution of most proteins that are not compound targets in reality, for each compound, all inactive targets are randomly selected to construct a ratio of active targets to inactive targets of 1:3. For a few compounds with a ratio of inactive targets to active targets greater than 3:1, all their target information is retained.

[0041] Furthermore, InChIKey is a standardized hash value based on the International Chemical Identifier (InChI) developed by the International Union of Pure and Applied Chemistry (IUPAC). When the InChIKey of two compounds are exactly the same, it means that they are the same compounds. Using InChIKey to match the above two database compounds, we finally obtained 3141 compounds that contain both perturbed transcriptome and target information.

[0042] Step 1.2, transcriptome feature extraction:

[0043] The reported literature 1 (DOI: 10.1038 / s43587-021-00138-z) proposed to use metric learning to extract features from transcriptome data with noisy background. Specifically, the principal component analysis (PCA) is used to reduce the dimension first, and then the linear discriminant analysis (LDA) is used to learn a transformation matrix that maximizes the similarity within the class and the difference between the classes. This matrix can project the data into a new space so that the perturbed transcriptome data points with the same label are clustered together as much as possible, while the data points with different labels are separated as much as possible. Based on the transformation matrix, the original perturbed transcriptome data after PCA dimensionality reduction is projected to obtain the perturbed transcriptome features.

[0044] The literature also evaluated the effects of factors such as cell type, compound, time point and dose on transcriptional markers, focusing on how to minimize the interference of these factors on similarity calculations. The results show that the effect of dose on marker distribution is small, so different doses of compounds are usually regarded as the same treatment conditions. In actual experiments, when the experiment is repeated for the same compound under the same perturbation conditions, data noise is inevitable. Therefore, the present invention first uses PCA for dimensionality reduction (n_components = 0.98), and then learns the transformation matrix through LDA metric learning. Before the learning process, a perturbation treatment label is assigned to each transcriptome data, including compound, cell line and treatment time information, and the same perturbation treatment label is regarded as the same class. PCA dimensionality reduction and LDA metric learning can extract the most significant features that distinguish different categories from the original data and project the original data into a lower-dimensional feature space. This method makes the perturbed transcriptome data points of the same label as aggregated as much as possible, while the data points of different labels are as dispersed as possible. Finally, the present invention reduces the 12328-dimensional original features of the CMAP LINCS2020level5 dataset to a 1240-dimensional embedding vector.

[0045] Step 1.3: Protein target feature extraction:

[0046] The target protein sequence was queried and downloaded through UniProt, and the protein language model pre-trained in ConPLex was used to convert the protein sequence into a 1024-dimensional embedding for subsequent training.

[0047] Step 1.4: Build the dataset:

[0048] The new feature vector formed by connecting the transcriptome features of the compound disturbance corresponding to each of the 3141 compounds collected with the protein features one by one, that is, obtaining 2264-dimensional training data, and giving a label as a sample according to whether the protein feature is the active target of the compound, 1 represents the active target of the compound, and 0 represents the inactive target, as shown in Table 1. Finally, the data set was constructed.

[0049] Table 1: Example of model training input

[0050] Compound Label Compound perturbation transcriptome signature Protein Target Active Tags CHEMBL10 PBIOA016_PC3_24H:C03 CHEMBL215 1 CHEMBL10 PBIOA016_PC3_24H:C03 CHEMBL221 1 CHEMBL10 PBIOA016_PC3_24H:C03 CHEMBL335 0 CHEMBL10 LJP009_PC3_24H:G08 CHEMBL2223 0 CHEMBL423 CPD001_PC3_24H:BRD-A02759312-003-17-9:10 CHEMBL1075145 0 CHEMBL423 CPC011_PC3_24H:BRD-A02759312-003-19-5:10 CHEMBL1075145 0 CHEMBL17860 REP.A012_PC3_24H:M09 CHEMBL1075113 0 CHEMBL17860 REP.A012_PC3_24H:M09 CHEMBL3708265 0 CHEMBL17860 REP.A012_PC3_24H:M11 CHEMBL2121 0 CHEMBL17860 REP.A012_PC3_24H:M11 CHEMBL214 1

[0051] In this embodiment, the data set is divided into training set: validation set: test set = 80%: 10%: 10% using the train_test_split function of the Python machine learning framework software package scikit-learn. The validation set is used for model optimization, and the test set is used to evaluate the final performance of the model.

[0052] Step 2: Train the classification model as follows:

[0053] The classification model can use a multi-layer perceptron, random forest, support vector machine, convolutional neural network, etc. In this embodiment, a multi-layer perceptron (MLP) model is used as a classification model. Specifically, a multi-layer perceptron is an artificial neural network that learns complex patterns and functions in data through multiple neuron layers. Due to its hierarchical structure, it can learn nonlinear representations of input data. It is a powerful model that is commonly used for classification, regression and other tasks. Among them, the input of the classification model is a new feature vector, and the output is a prediction result of whether the corresponding protein is the target of the corresponding compound. Each sample of the training set is used as input, and training is performed with the goal of minimizing the error between the classification model output and the label to obtain a trained classification model.

[0054] After training is completed, this embodiment uses six indicators, namely, accuracy, precision, recall, F1 score, AUROC, and AUPRC, to evaluate the performance of the model. Their calculation formulas are as follows, where TP is true positive, TN is true negative, FP is false positive, and FN is false negative. AUROC is the area under the ROC curve, and AUPRC refers to the area under the precision-recall curve, which is expressed as follows:

[0055]

[0056] The MLP model is sensitive to the combination of hyperparameters. As a preferred solution, this embodiment uses optuna to perform hyperparameter combination search and optimization on the validation set. Since the number of inactive targets of the compound is more than the active targets (negative samples> positive samples), the present invention selects F1 and AUPRC to evaluate the classification performance of the model. Specifically, Optuna is a hyperparameter optimization framework based on Bayesian optimization, which can find the best hyperparameter combination in the fewest number of experiments. Since the dimension of the training data is 2264, this embodiment comprehensively considers data characteristics, computing resources and other issues, and sets the search range of the model hidden layer to 1 to 4 layers, dropout to 0.1 to 0.4, the number of neurons in each layer to 100 to 200, the learning rate to 1e-4, 1e-2, the maximum number of rounds to 150, and the optimizer to be Adam or SGD, which is commonly used to solve binary classification problems. F1 is used as the optimization indicator, the number of searches is 20 times, and some results are as follows Figure 2 And as shown in Table 2.

[0057] Table 2: Scores of six indicators for 20 parameter searches

[0058]

[0059] As shown in Table 2, in the 20 searches, the F1 and AUPRC of the model in Trial9 on the validation set were the highest. Therefore, this embodiment uses this parameter for training and fixes it as the final trained classification model. Its specific hyperparameters are hidden layer: [2264,1809], [1809,1977], [1977,990], [990,835], [835,1], Dropout: 0.26051198024596567, 0.16616269153317942, 0.1351244003622237, 0.34033060495716216, learning rate: 0.0012212088107909464, optimizer: Adam, loss function: BCELoss, and training for 149 rounds. The model is fixed and named PTPCM, and the prediction ability of the model is evaluated using the test set, such as Figure 3 As shown, the improvement of AUPRC during the training process of the model is demonstrated, and the final model performs well on the test set, with F1 of 0.971, AUROC of 0.998, and AUPRC of 0.995, indicating that the model can accurately distinguish between active and inactive targets and can be used for target prediction.

[0060] Step 3: Target prediction based on the trained classification model. This embodiment takes the prediction of natural product targets as an example, and the details are as follows:

[0061] Natural products usually have complex chemical structures and may contain multiple active ingredients, so it is difficult to determine their targets. This embodiment collects perturbation transcriptome data and target information of 496 natural products as an external test set to verify the actual application effect of the method of the present invention. The LINCS level 5 perturbation transcriptome uses the moderated z-score (MODZ) to represent the changes in genes. However, in real scenarios, log2FC is often used to represent the degree of change in genes, so this implementation uses log2FC for target prediction. At the same time, the method of processing RNAseq to obtain the modulated Z value (MODZ) with reference to the reported document 2 (DOI: 10.1007 / s13238-021-00885-0) also performs target prediction and compares the prediction effects of the two treatment methods. After obtaining the above two types of data, arrange them in the training data format, and use the trained metric learning model to obtain the perturbed transcriptome features and connect them with the protein features one by one to form new features, and then input them into PTPCM respectively, and sort them according to the output score. The higher the score, the more likely the target is to be the target of the natural product. In this embodiment, natural products with less than 50 targets are screened and collected, and this is considered more credible, and finally 210 natural products with target records are obtained. The TopK accuracy is used as an evaluation indicator. Specifically, the TopK accuracy is the ratio of the number of compounds that can hit any of its recorded targets before the model predicts the TopK target to the entire predicted compound set (210 in this case). Figure 4 As shown in the figure, both prediction data forms achieved high accuracy. The Top100 accuracy using log2FC was 0.78, while the Top100 accuracy using MODZ was 0.68. Therefore, Log2FC was used for subsequent predictions. Some prediction results are shown in Table 3. For some compounds, such as fritillary alkaloids, the Top30 targets predicted by the model hit two of its recorded targets. For strychnine, the model even predicted its only recorded target, which proves the high accuracy of the model.

[0062] Table 3: Prediction results of some natural products

[0063]

[0064]

[0065] For natural products without target information, this example also performed target prediction and found some reported literature support, such as Figure 5As shown, quercetin is a flavonoid compound found in many plants, which has antioxidant, anti-inflammatory and antibacterial effects. It has been reported in literature 3 (DOI: 10.1007 / s13238-021-00885-0) that quercetin can inhibit MAPK signaling factors ERK1 / 2, JNK and p38MAPK in adipocytes and macrophages. ERK2 is MAPK1, which appears in the top 30 predicted targets of the method of the present invention.

[0066] Salvianolic acid B is a natural compound extracted from Salvia miltiorrhiza. Recent studies have shown that it has a therapeutic effect on the neurodegenerative disease Alzheimer's disease, which has been reported in literature4

[0067] (DOI:10.1016 / j.tice.2023.102260) It was found that salvianolic acid B had a protective effect on Aβ (amyloid β protein)-induced neuroinflammation by inhibiting the activation of NLRP3 inflammasome and promoting the polarization transition of microglia from M1 (pro-inflammatory) to M2 (anti-inflammatory). Although the key factors mentioned in the article were not included in the prediction results of the present invention, CHRM2 was predicted, which is an important target in the treatment of AD. Acetylcholine can bind to it to improve cognitive function.

[0068] Osteoarthritis is a chronic degenerative joint disease that affects more than 500 million people worldwide and may even lead to disability. It has been reported in document 5 (DOI: 10.16333 / j.1001-6880.2021.8.010) that eleutheroside E can significantly improve inflammatory cell infiltration, fibrous tissue proliferation, and cartilage surface destruction in osteoarthritis sites, and reduce the levels of inflammatory mediators (IL-1β and PGE2) and MMP-3 and MMP-9 in joint fluid to treat osteoarthritis. Multiple MMP family proteins, including MMP-9, appeared in the prediction results of the present invention.

[0069] Berberine is a plant alkaloid extracted from Coptis chinensis, which has multiple pharmacological activities, including hypoglycemic, anti-obesity, anti-inflammatory and other effects. In recent years, many studies have confirmed that it has anti-cancer and anti-depressant effects. It has been reported in document 6 (DOI: 10.1186 / s11658-024-00607-4) that berberine can enhance the transcription of the EGFR gene, thereby upregulating the RAS-RAF1-MEK1 / 2-ERK1 / 2 signaling pathway, inducing tumor cell autophagy, and inhibiting tumor cell proliferation and metastasis. It is worth noting that EGFR appears in the Top30 targets predicted by the present invention. These evidences prove that the target prediction method of the present invention has high accuracy and credibility and can be applied in practice.

[0070] Corresponding to the aforementioned embodiment of a method for predicting compound targets based on compound perturbation of transcriptome and protein sequence, the present invention also provides an electronic device, including one or more processors, for implementing the aforementioned embodiment of a method for predicting compound targets based on compound perturbation of transcriptome and protein sequence.

[0071] The electronic device of the present invention may be any device with data processing capability, and the device with data processing capability may be a device or apparatus such as a computer.

[0072] The device embodiment can be implemented by software, hardware or a combination of software and hardware. Taking software implementation as an example, as a device in a logical sense, it is formed by the processor of any device with data processing capability in which it is located to read the corresponding computer program instructions in the non-volatile memory into the memory for execution. From the hardware level, it includes a processor, memory, network interface, and non-volatile memory. In addition, any device with data processing capability in which the device in the embodiment is located can also include other hardware according to the actual function of the device with data processing capability, which will not be described in detail.

[0073] The implementation process of the functions and effects of each unit in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned method, and will not be repeated here.

[0074] For the device embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment. The device embodiment described above is only schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of the present invention. Ordinary technicians in this field can understand and implement it without paying creative work.

[0075] An embodiment of the present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, a method for predicting compound targets based on compound perturbation transcriptome and protein sequence in the above embodiment is implemented.

[0076] The computer-readable storage medium may be an internal storage unit of any device with data processing capability described in any of the aforementioned embodiments, such as a hard disk or a memory. The computer-readable storage medium may also be any device with data processing capability, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. equipped on the device. Furthermore, the computer-readable storage medium may also include both an internal storage unit of any device with data processing capability and an external storage device. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capability, and may also be used to temporarily store data that has been output or is to be output.

[0077] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the present invention.

Claims

1. A method for predicting compound targets based on compound perturbation transcriptome and protein sequence, characterized in that: include: Obtain perturbed transcriptome data of the compound to be predicted and extract perturbed transcriptome features through principal component analysis and linear discriminant analysis; Collect a number of protein sequence data and use protein language models to convert protein sequences into embeddings that can be used for deep learning to represent protein features; Connect the perturbed transcriptome features one by one with the protein features to form a new feature vector; The new feature vectors are input one by one into a trained classification model, and the classification model outputs a prediction result of whether the protein corresponding to the input new feature vector is a target of the compound.

2. The method according to claim 1, characterized in that: The perturbation transcriptome features are extracted by principal component analysis and linear discriminant analysis, as follows: The perturbed transcriptome data are reduced in dimension by principal component analysis and the main characteristic components are extracted. Then, linear discriminant analysis is used to learn a linear transformation matrix that can maximize the intra-class similarity and inter-class difference. Finally, the linear transformation matrix is ​​used to project the reduced-dimensional data and map it to a new feature space to obtain the perturbation transcriptome features that characterize the perturbation pattern.

3. The method according to claim 1, characterized in that The perturbed transcriptome data of the compound to be predicted comes from the CMAP LINCS2020Level 5 dataset, and the protein sequence data comes from UniProt.

4. The method according to claim 1, characterized in that The classification model uses a multi-layer perceptron, a random forest, a support vector machine or a convolutional neural network.

5. The method according to claim 1, characterized in that The classification model is obtained by training based on a data set with the goal of minimizing the error between the classification model output and the label.

6. The method according to claim 5, characterized in that Each sample of the data set includes a new feature vector and a label formed by connecting the transcriptome feature and protein feature of the corresponding compound perturbation, and the label is set in the following manner: If the IC50, EC50 and Ki values ​​of the corresponding compound to the protein are less than or equal to 10 μM, the protein is the active target of the corresponding compound and the label is 1; if the IC50, EC50 and Ki values ​​of the corresponding compound to the protein are greater than 100 μM, the protein is the inactive target of the corresponding compound and the label is 0.

7. The method according to claim 6, characterized in that In the data set, among samples containing the same compound, the ratio of the number of samples with label 0 to the number of samples with label 1 is greater than 3:

1.

8. The method according to claim 5, characterized in that The classification model also includes performing hyperparameter combination search optimization on a validation set.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the method for predicting compound targets based on compound perturbation transcriptome and protein sequence as described in any one of claims 1 to 8 is implemented.

10. A storage medium comprising computer executable instructions, which, when executed by a computer processor, implement a method for predicting compound targets based on transcriptome and protein sequence perturbations by compounds as described in any one of claims 1 to 8.

Images