Large-scale virtual screening method and system based on drug effect sorting

Through the pharmacodynamic sequencing method based on Transformer architecture and cross attention mechanism, integrating omnis and molecular characteristics, the problem that existing virtual screening methods cannot reflect the actual effect and individual differences of drugs in vivo is solved, achieving a more accurate and explainable Drug screening effect.

CN119092000APending Publication Date: 2024-12-06SOUTHEAST UNIV NANJING INST OF BIOMATERIALS & MEDICAL DEVICES
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411012382.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-26
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing virtual screening methods cannot effectively consider the actual effect of drugs in the body and the individual differences in drug use, resulting in a low success rate of drug activity experiments and clinical trials, and lack accuracy and interpretability.

Method used

The molecular coding model based on the Transformer architecture is used to extract molecular structural features, combine cross attention mechanisms to integrate omnis and molecular features, and use Pairwise sorting to train the drug efficacy prediction model to avoid predicting the absolute value of drug efficacy, and improve the accuracy and reliability of drug screening. sex.

Benefits of technology

Capture small differences between individuals through omics information encoding, improve the accuracy and reliability of drug screening, provide more convincing drug screening results, reduce calculation costs, and improve the scalability of the method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119092000A_ABST
    Figure CN119092000A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of drug molecule screening, in particular to a large-scale virtual screening method and system based on drug effect sorting. The method comprises the steps that S1, a data set of drug effects is constructed, and data in the data set is preprocessed; s2, performing omics feature extraction and molecular feature extraction on the data in the data set; s3, fusing molecular features and omics features by using cross attention; s4, a drug effect prediction model is obtained based on Pairwise sorting training; s5, inputting the molecular queue into a drug effect prediction model; and S6, outputting the drug effect sequence of the molecular queue. According to the method, prediction of the absolute value of the drug effect is avoided through pairwise sorting, and the accuracy of drug screening is improved; using cross attention to fuse omics features and molecular features, and extracting richer and more useful features as input of a drug effect prediction module; omics information codes are used as individual features, the defect that phenotypes are neglected in a traditional virtual screening method is overcome, tiny differences among individuals can be captured, and the reliability and accuracy of drug screening are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of drug molecule screening, and in particular to a large-scale virtual screening method and system based on drug efficacy ranking. Background Art

[0002] Drug development is a complex and lengthy process that involves multiple stages, including the discovery, design, synthesis, and clinical trials of new drugs. This process requires not only a large amount of capital investment, but also a lot of time. Usually, it may take 10 years or even longer for a new drug to go from discovery to market.

[0003] Existing drug development methods cover a variety of means from traditional experiments to modern computing technologies, including laboratory synthesis and testing of compounds, high-throughput screening technology, and virtual screening using computer-aided drug design (CADD) technology. Virtual screening predicts the activity of drugs by simulating the interaction between drug molecules and biological targets, thereby quickly screening out potential drug candidate molecules in the early stages and reducing the number and cost of laboratory experiments. This method not only improves the efficiency of drug discovery, but also makes the drug development process more accurate and personalized. At the same time, with the application of big data and artificial intelligence technologies, drug development is gradually shifting from experience-driven to data-driven, further accelerating the discovery and optimization of new drugs. Despite this, drug development still faces challenges such as strict regulatory requirements, high R&D costs, and long R&D cycles.

[0004] Virtual screening is a method that uses computer technology to assist in drug discovery. It predicts the activity of drugs by simulating the interaction between drug molecules and biological targets, and can screen millions of compounds at a time. However, traditional virtual screening is based on the properties of the compounds and targets themselves, without considering the actual effects of drugs in the body and the differences in individual medication, resulting in a low success rate in actual activity experiments and clinical trials.

[0005] Drug effect prediction is an important part of the drug development process. It aims to evaluate the potential effects of drug molecules on biological targets and predict their therapeutic effects. With the development of bioinformatics and computational chemistry, drug effect prediction technology has changed from traditional experimental methods to modern technologies that combine computational models and machine learning. The mainstream drug effect prediction direction is to model and predict drug activity based on the chemical structure of the drug and the three-dimensional structure of the protein target and their interactions, which can process a large amount of biomedical data. The limitation of this method is that it cannot reflect the real drug effect of the human body. Therefore, more and more models have introduced omics data, namely genomics, transcriptomics, proteomics, etc., to obtain more comprehensive drug effect information, which helps to reveal the molecular mechanism of drug action and improve the accuracy of prediction. When this type of model is used, omics information and drug structure are usually input, and the model will return the absolute value of drug effect. The result is very dependent on the accuracy of the training data and is not very interpretable. However, in actual situations, even multiple repetitions cannot guarantee that the drug effect data obtained from the experiment is accurate and correct, so more accurate and more interpretable methods need to be discovered and used for drug effect prediction. Summary of the invention

[0006] The purpose of the present invention is to address the problems existing in the background technology and to propose a large-scale virtual screening method and system based on drug efficacy ranking.

[0007] The technical solution of the present invention is a large-scale virtual screening method based on drug efficacy ranking, comprising the following specific steps:

[0008] S1. Construct a drug effect dataset and preprocess the data in the dataset;

[0009] The drug effect dataset collected cell line experimental data including genomic mutation information and drug sensitivity test from the GDSC and CTRP databases. The medication information, genomic mutation information and drug effect of each cell line corresponded one by one, and those from different databases were recorded as different batches. In the end, a total of about 820,000 data were collected.

[0010] The preprocessing step includes: combining the drug sensitivity data of each sample in the batch in pairs; using the obtained molecular pairs as the input of the model, and the drug sensitivity results corresponding to the molecular pairs as the output of the model for training;

[0011] S2, extract omics features and molecular features from the data in the dataset;

[0012] The steps of omics feature extraction include: processing genomic mutation data, using 0 to indicate that the gene has not mutated and 1 to indicate that the gene has mutated; referring to DrugCell, the mutation information of 01 encoding of 3008 genes was obtained; using the DrugCell model weight to extract mutation features, and finally obtaining the mutation features of each cell line, which were used to characterize individuals;

[0013] The steps of molecular feature extraction include: extracting the structural features of each molecule in the molecular pair using a pre-trained molecular model, wherein the structural features of each molecule are represented by a vector;

[0014] The molecular model is based on the Transformer architecture, which includes 12 layers of encoders and 12 layers of decoders. Each layer of encoder and decoder consists of components such as multi-head self-attention mechanism, feed-forward neural network, residual connection and layer normalization. The encoder encodes the input sequence into hidden representation, and the decoder decodes these hidden representations into output sequence;

[0015] S3. Use the cross-attention mechanism to fuse the omics features with each molecular feature in the corresponding molecular pair to obtain a fused feature vector, which is used as input to train the drug efficacy prediction model.

[0016] S4, a drug efficacy prediction model is obtained based on Pairwise sorting training;

[0017] Pairwise sorting is a pairwise comparison method that decomposes the overall sorting problem into multiple pairwise comparison problems. In the context of drug effect prediction, this method involves comparing the effects of each pair of drugs (or drug combinations) to determine their relative effects. During training, drug sensitivity results are used in the forward propagation of the model, and sorting results are used when calculating loss. Adam is used as the optimization function, and the learning rate is set to 0.0001. Finally, a drug effect prediction model is trained.

[0018] S5, inputting the molecular cohort into the drug efficacy prediction model;

[0019] S6. Output the efficacy ranking of the molecular queue.

[0020] A large-scale virtual screening system based on drug efficacy ranking, using the above virtual screening method, including an input end, an output end, an encoding module, an interaction module, and a prediction module;

[0021] The input end is used to input drug molecule pairs into the encoding module;

[0022] The encoding module is used to extract omics features and molecular features of drug molecule pairs;

[0023] The interaction module is used to fuse the omics features and molecular features of the drug molecule pair and generate a fusion feature vector; the generated fusion feature vector is input into the prediction module;

[0024] The prediction module uses the trained drug efficacy prediction model to predict the fused feature vector, and the generated prediction result is output through the output end.

[0025] Preferably, the input end uses a new sample and a molecular queue of unknown effect as input, predicts the activity rate ranking of the queue for the sample, and outputs the preferred compound through the output end.

[0026] An application is to use the above method to screen drug molecules; or to use the above system to screen drug molecules.

[0027] Compared with the prior art, the present invention has the following beneficial technical effects:

[0028] The present invention uses a molecular encoding model based on the Transformer architecture to extract molecular structural features; uses cross-attention to fuse omics features and molecular features to extract richer and more useful features as the input of the drug efficacy prediction module; avoids predicting the absolute value of drug efficacy through pairwise sorting, thereby improving the accuracy of drug screening.

[0029] The present invention uses omics information encoding as individual characteristics, overcomes the neglect of phenotypes by traditional virtual screening methods, can capture subtle differences between individuals, and improve the reliability and accuracy of drug screening;

[0030] The training data set used in the present invention comes from the drug use results of different cell lines collected by GDSC, which is a true reflection of the drug effect, and its results are more convincing than those of dry experiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the screening method according to an embodiment of the present invention;

[0032] Figure 2 Schematic diagram of the structure of the screening system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0033] Example 1

[0034] This embodiment provides a large-scale virtual screening method based on drug efficacy ranking. Figure 1 As shown, the specific steps include:

[0035] S1. Construct a drug effect dataset and preprocess the data in the dataset;

[0036] The drug effect dataset collected cell line experimental data including genomic mutation information and drug sensitivity test from the GDSC and CTRP databases. The medication information, genomic mutation information and drug effect of each cell line corresponded one by one, and those from different databases were recorded as different batches. In the end, a total of about 820,000 data were collected.

[0037] The preprocessing step includes: combining the drug sensitivity data of each sample in the batch in pairs; using the obtained molecular pairs as the input of the model, and the drug sensitivity results corresponding to the molecular pairs as the output of the model for training;

[0038] S2, extract omics features and molecular features from the data in the dataset;

[0039] The steps of omics feature extraction include: processing genomic mutation data, using 0 to indicate that the gene has not mutated and 1 to indicate that the gene has mutated; referring to DrugCell, the 01-encoded mutation information of 3008 genes was obtained; using the DrugCell model weights to extract mutation features, and finally obtaining the mutation features of each cell line, which are used to characterize individuals; the training method and training process of the DrugCell model have been described in detail in the published paper "Predicting Drug Response and Synergy Using a Deep Learning Model of Human Cancer Cells", and will not be repeated here.

[0040] The steps of molecular feature extraction include: using a pre-trained molecular model to extract the structural features of each molecule in the molecular pair, and the structural features of each molecule are represented by a vector; the training method and training process of the pre-trained molecular model have been disclosed in Chinese patents CN117238436A and CN117334271B, and will not be repeated in this application.

[0041] The molecular model is based on the Transformer architecture, which includes 12 layers of encoders and 12 layers of decoders. Each layer of encoder and decoder consists of components such as multi-head self-attention mechanism, feed-forward neural network, residual connection and layer normalization. The encoder encodes the input sequence into hidden representation, and the decoder decodes these hidden representations into output sequence;

[0042] S3. Use the cross-attention mechanism to fuse the omics features with each molecular feature in the corresponding molecular pair to obtain a fused feature vector, which is used as input to train the drug efficacy prediction model;

[0043] S4, a drug efficacy prediction model is obtained based on Pairwise sorting training;

[0044] Pairwise sorting is a pairwise comparison method that decomposes the overall sorting problem into multiple pairwise comparison problems. In the context of drug effect prediction, this method involves comparing the effects of each pair of drugs (or drug combinations) to determine their relative effects. During training, drug sensitivity results are used in the forward propagation of the model, and sorting results are used when calculating loss. Adam is used as the optimization function, and the learning rate is set to 0.0001. Finally, a drug effect prediction model is trained.

[0045] S5, inputting the molecular cohort into the drug efficacy prediction model;

[0046] S6. Output the efficacy ranking of the molecular queue.

[0047] In this embodiment, Transformer is a deep learning model based on the attention mechanism, which was originally proposed by Vaswani et al. in 2017 to process sequence-to-sequence tasks such as machine translation and natural language generation. Compared with traditional recurrent neural networks (RNN) and long short-term memory networks (LSTM), Transformer is more suitable for parallel computing, can handle long-distance dependency problems, and is easy to train. The core of the Transformer model is the Self-Attention Mechanism, which allows the model to consider all positions of the input sequence at the same time and automatically adjust the attention weights. Transformer consists of two parts: an encoder and a decoder.

[0048] Cross-Attention Mechanism is a method used in deep learning to process sequence data. The main idea is to allow one sequence (usually called the "query" sequence) to interact with another sequence (usually called the "key-value" sequence) to generate a richer and more contextual representation, which is especially effective when processing multimodal data or combining multiple sequences together. It is widely used in the Transformer architecture, such as in tasks such as machine translation, text generation, and image description, which can help the model better capture the correlation and dependency between sequences. The cross-attention mechanism can play an important role in drug efficacy prediction, especially when processing complex multimodal data, such as combining information such as chemical molecular structure, gene expression data, and drug-target interaction. Through the cross-attention mechanism, the characteristics of drug molecules and target proteins can interact to capture the complex relationship between the two, thereby improving the accuracy of drug-target interaction prediction.

[0049] Pairwise ranking is a method that compares two items and ranks them according to specific criteria. In the fields of recommender systems, information retrieval, and machine learning, pairwise ranking has been proven to be an effective technique that can improve the accuracy and efficiency of ranking. The core idea of ​​pairwise ranking is to decompose the overall ranking problem into multiple pairwise comparison problems, which can reduce the computational complexity while maintaining high prediction accuracy. Applying pairwise ranking to drug effect prediction can effectively address some limitations of existing methods. By comparing the effects of different drugs in pairs, this method can capture the relative differences between drugs in more detail, thereby providing more accurate prediction results. In addition, the decomposition characteristics of pairwise ranking also help to reduce computational costs and improve the scalability of the method. It can be applied not only to drug effects, but also to predictions such as drug toxicity.

[0050] This embodiment provides a drug screening method based on drug efficacy ranking, by collecting drug sensitivity test data and genomic data from public data sets to establish a training data set; using the omics encoding model based on graph neural network to extract omics features, using the molecular encoding model based on Transformer architecture to extract molecular structure features; using cross-attention to fuse omics features and molecular features, extracting richer and more useful features as input for the drug efficacy prediction module; training in the form of two-drug combinations to obtain an AI model, which can be used for large-scale virtual drug screening. This method is based on omics information, overcoming the neglect of phenotypes by traditional virtual screening methods; the training data comes from wet experiments, providing stronger persuasiveness; the sorting-based training strategy improves the accuracy and interpretability of predictions.

[0051] Example 2

[0052] This embodiment provides a method for establishing a drug efficacy prediction model based on a CCLE drug sensitivity data set, including:

[0053] Step 1: By searching and collecting public databases, we obtained mutation data of about 1,800 cell lines, which corresponded to about 890,000 drug sensitivity results;

[0054] Step 2: Divide the data into training set, validation set and test set according to the cell line, with a ratio of 6:2:2. Group the drug sensitivity test data by compound in each cell line, list all molecular pair combinations and their corresponding drug sensitivity test results, and set batch_size to 30 during training;

[0055] Step 3: Use the trained DrugCell model weights to extract genomic mutation features and obtain a mutation feature vector with a length of 6;

[0056] Step 4: Use the pre-trained molecular model to extract the structural features of each molecule in the molecular pair. The structural features of each molecule are represented by a vector with a length of 768;

[0057] Step 5: Use the cross-attention mechanism to fuse the omics features with each molecular feature in the corresponding molecular pair to obtain a fused feature vector with a length of 768, and input the feature vector into the drug efficacy prediction module;

[0058] Step six, the drug efficacy prediction module maps the fused feature vector of length 768 to a vector of length 384 through a hidden fully connected layer, uses ReLu as the activation function, and sets Dropout to 0.1. Finally, it outputs the prediction result through a fully connected layer. The forward propagation uses drug sensitivity data as the expected output, and uses the mean square error function (MSE) to calculate the loss between the predicted result and the expected output ranking. The Adam optimization function is used in the back propagation, and the initial learning rate is set to 0.02, which decreases continuously with the gradient. After completing 400 Epoch training, the model with the smallest loss in the validation set is selected as the optimal model. The optimal model was tested with the test set, and the results showed that the correlation between the predicted ranking and the actual ranking reached 0.85240;

[0059] Step seven: load the trained model weights, use the new sample and the molecular queue of unknown effect as input, predict the activity rate ranking of the queue for the sample, and give the preferred compound.

[0060] In this embodiment, a deep learning algorithm is used, and the Transformer self-attention mechanism is applied to extract the structural features of the compound; based on the graph neural network (GNN), feature extraction is performed on genomic mutation information, and high-dimensional mutation information is converted into low-dimensional information, thereby reducing the complexity of calculation and improving the calculation efficiency;

[0061] The cross-attention mechanism is a technique used in deep learning to process multimodal data, and plays an important role in the fusion of molecular structure and omics data. Through this mechanism, the model is able to learn the complex associations between molecular structural features and omics data, so as to better understand and predict the biological activity or drug properties of molecules. Specifically, cross-attention allows the model to dynamically focus on relevant information in another modality (such as omics data) while processing one modality (such as molecular structure), enhancing the model's ability to capture interdependencies between data. This dynamic, context-dependent information integration not only improves the accuracy of predictions, but also helps to reveal the deep connection between molecules and biological processes, providing strong support for drug discovery research.

[0062] In this embodiment, the drug effect is affected by many factors, and the drug reactions of different populations are not the same. Traditional drug effect prediction methods usually score single drugs or drug combinations, and determine the priority of medication based on the scores. This method relies on the absolute value of the drug effect, which fluctuates between individuals. Current experimental techniques cannot provide accurate drug efficacy data, so the prediction results are very limited. The Pairwise sorting mechanism introduced in this embodiment can more carefully capture the relative differences between drugs by comparing the effects of different drugs in pairs and evaluating the pros and cons of drugs, thereby providing more accurate prediction results. For each pair of drugs, Pairwise sorting determines their relative effects by defining a comparison function. This method is also scalable, and the comparison function can be based on the efficacy, side effects, bioavailability or other relevant indicators of the drug. The algorithm gradually builds an overall drug effect ranking by iteratively comparing all possible drug pairs. The comparison results of each pair will affect the overall ranking.

[0063] Example 3

[0064] A large-scale virtual screening system based on drug efficacy ranking, using the virtual screening method of Example 1, comprising an input end, an output end, an encoding module, an interaction module, and a prediction module;

[0065] The input end is used to input drug molecule pairs into the encoding module;

[0066] The encoding module is used to extract omics features and molecular features of drug molecule pairs;

[0067] The interaction module is used to fuse the omics features and molecular features of the drug molecule pair and generate a fusion feature vector; the generated fusion feature vector is input into the prediction module;

[0068] The prediction module uses the trained drug efficacy prediction model to predict the fused feature vector, and the generated prediction result is output through the output end.

[0069] Preferably, the input end uses a new sample and a molecular queue of unknown effect as input, predicts the activity rate ranking of the queue for the sample, and outputs the preferred compound through the output end.

[0070] Example 4

[0071] One application is to use the method of Example 1 to screen drug molecules; or to use the system of Example 3 to screen drug molecules.

[0072] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto, and various changes can be made within the knowledge scope of technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A large-scale virtual screening method based on drug efficacy ranking, characterized in that: The specific steps include: S1. Construct a drug effect dataset and preprocess the data in the dataset; S2, extract omics features and molecular features from the data in the dataset; S3, using cross-attention to fuse molecular and omics features; S4, a drug efficacy prediction model is obtained based on Pairwise sorting training; S5, inputting the molecular cohort into the drug efficacy prediction model; S6. Output the efficacy ranking of the molecular queue.

2. A large-scale virtual screening method based on drug efficacy ranking according to claim 1, characterized in that: The drug effect dataset in S1 collects cell line experimental data including genomic mutation information and drug sensitivity test from the GDSC and CTRP databases; the medication information, genomic mutation information and drug effect of each cell line correspond one to one, and those from different databases are recorded as different batches.

3. A large-scale virtual screening method based on drug efficacy ranking according to claim 1, characterized in that: The preprocessing steps in S1 include: combining the drug sensitivity data of each sample in the batch in pairs; using the obtained molecular pairs as the input of the model, and the drug sensitivity results corresponding to the molecular pairs as the output of the model for training.

4. A large-scale virtual screening method based on drug efficacy ranking according to claim 1, characterized in that: The steps of omics feature extraction in S2 include: processing genomic mutation data, using 0 to indicate that the gene has not mutated and 1 to indicate that the gene has mutated; using the DrugCell model weights to extract mutation features, and finally obtaining the mutation features of each cell line, which are used to characterize individuals.

5. A large-scale virtual screening method based on drug efficacy ranking according to claim 1, characterized in that: The step of molecular feature extraction in S2 includes: using a pre-trained molecular model to extract the structural features of each molecule in the molecular pair, and the structural features of each molecule are represented by a vector.

6. A large-scale virtual screening method based on drug efficacy ranking according to claim 1, characterized in that: Using the cross-attention mechanism, the omics features are fused with each molecular feature in the corresponding molecular pair to obtain the fused feature vector as input for training the drug efficacy prediction model.

7. A large-scale virtual screening method based on drug efficacy ranking according to claim 1, characterized in that: During the training of the prediction model in S4, the drug sensitivity results were used in the forward propagation of the model, and the sorting results were used when calculating the loss. Adam was used as the optimization function, and the learning rate was set to 0.0001. Finally, a drug efficacy prediction model was obtained through training.

8. A large-scale virtual screening system based on drug efficacy ranking, using the virtual screening method according to any one of claims 1 to 7, characterized in that: It includes an input end, an output end, an encoding module, an interaction module, and a prediction module; The input end is used to input drug molecule pairs into the encoding module; The encoding module is used to extract omics features and molecular features of drug molecule pairs; The interaction module is used to fuse the omics features and molecular features of the drug molecule pair and generate a fusion feature vector; the generated fusion feature vector is input into the prediction module; The prediction module uses the trained drug efficacy prediction model to predict the fused feature vector, and the generated prediction result is output through the output end.

9. A large-scale virtual screening system based on drug efficacy ranking according to claim 8, characterized in that: The input end uses a new sample and a molecular queue of unknown effect as input, predicts the activity rate ranking of the queue for the sample, and outputs the preferred compound through the output end.

10. An application, characterized in that: Screening drug molecules using the method described in any one of claims 1-7; or, screening drug molecules using the system described in claims 8-9.