Characterization optimization and soft label construction method of drug target interaction prediction model

Abstract

The invention discloses a characterization optimization and soft label construction method for a drug-target interaction prediction model, and the method comprises the following steps: (1) constructing a molecular characterization space based on a Molformer pre-training model, carrying out the feature space similarity analysis of a negative sample through employing a K-Means clustering algorithm, and eliminating redundant samples through setting a dynamic threshold value, thereby achieving the equalization of positive and negative samples; (2) designing a soft label optimization mechanism based on a logarithmic function, converting a pKd value in a DAVIS data set into soft labels continuously distributed in a (0, 1) interval, and reserving binding strength gradient information to enhance the sensitivity of the model to a weak binding compound; and (3) integrating the above data enhancement strategies, and training a drug target interaction prediction model through a five-fold cross validation method to improve the model prediction performance. The method verifies the significant advantages of the method in relieving the problems of data offset and information loss, and also provides a high-robustness data processing framework for virtual drug screening of deep learning.

Description

Technical Field

[0001] The present invention relates to the intersection of biomedicine and artificial intelligence, and specifically provides a method for characterization optimization and soft label construction of a drug-target interaction prediction model to improve the performance of the drug-target interaction prediction model. Background Art

[0002] In the field of drug development, studying the interaction mechanism between target proteins and ligand molecules is a key step in discovering lead compounds. Although computational methods such as molecular docking and molecular dynamics simulations have significantly improved the efficiency of compound screening, traditional methods are still limited by high computing resource consumption and long screening cycles (usually taking months to years) due to the large search space of compounds.

[0003] In recent years, deep learning technology has provided efficient solutions for virtual drug screening by modeling drug-target interaction (DTI) prediction as a binary classification task or converting drug-target affinity (DTA) prediction into a regression problem. Some researchers have proposed a visual question-answering framework that maps target protein distance graphs into image features and combines them with ligand SMILES (Simplified Molecular Input Line Entry System) strings to predict interactions. Others, drawing on the machine translation paradigm, use a one-dimensional convolutional network to encode the target amino acid sequence and a graph convolutional network (GCN) to extract the topological features of the ligand molecules, ultimately outputting the interaction probability using a Transformer decoding layer.

[0004] However, the generalization ability of existing deep learning models is still limited by two core challenges. First, the ratio of positive and negative samples in DTI public datasets (such as DAVIS and BindingDB) is seriously unbalanced, which makes the model prone to "shortcut learning", that is, it performs well on the validation set and test set, but is difficult to generalize to new targets or new compounds. Although studies have attempted to optimize the quality of negative samples, such as screening high-confidence negative samples based on compound structural similarity, mining potential non-interacting pairs with the "Guilt-by-Association" algorithm, and enhancing the representativeness of negative samples using network pharmacology knowledge, random downsampling may lose key chemical space information, and similarity screening is prone to introduce redundant features. Secondly, existing methods often use thresholding to binarize continuous affinity values ​​into binding / non-binding labels, resulting in the loss of binding strength information of weakly bound or potentially optimizable compounds, significantly weakening the model's ability to analyze complex interaction patterns. Summary of the Invention

[0005] This paper addresses the two core bottlenecks of data distribution imbalance and label information discretization in current DTI prediction models, and proposes a joint data processing framework that integrates molecular representation optimization and soft label learning, aiming to enhance the generalization prediction ability of deep learning models for unknown drug-target pairs.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a method for characterization optimization and soft label construction of a drug-target interaction prediction model, comprising the following steps:

[0007] S1. Data acquisition and preprocessing:

[0008] Obtain the DAVIS dataset of the interaction between the target protein and the ligand compound, and calculate the original dissociation constant K d The labels are logarithmically transformed to obtain pK d Value, and divide the samples into positive samples and negative samples through the threshold;

[0009] S2, negative sample optimization sampling:

[0010] To address the problem of redundant negative samples in the DAVIS dataset, the following two strategies are used to generate optimized datasets D1 and D2 respectively;

[0011] S3, label conversion and training dataset construction:

[0012] For the optimized dataset D1 and optimized dataset D2, different types of labels are generated by using threshold division and logarithmic function mapping, and three training datasets are constructed.

[0013] S4. Model training and performance evaluation:

[0014] For the three training datasets mentioned above, different operations were performed to train the drug-target interaction prediction model and evaluate the performance, and finally the dataset with the best performance was selected as the result.

[0015] In step S1, the logarithmic transformation formula is:

[0016]

[0017] where pK d Samples with >5 are defined as positive samples, and the rest are negative samples.

[0018] Step S2, negative sample optimization sampling: To address the redundant negative sample problem in the DAVIS dataset, the following two strategies are used to generate optimized datasets D1 and D2, respectively, including the following steps:

[0019] To address the problem of redundant negative samples in the DAVIS dataset, the following two strategies are used to generate optimized datasets D1 and D2 respectively;

[0020] (a) For the positive and negative samples of the DAVIS dataset obtained in step S1, redundant negative samples are eliminated by random sampling, maintaining a positive to negative sample ratio of 1:1, and forming dataset D1;

[0021] (b) Perform cluster screening based on molecular characterization information on the positive and negative samples of the DAVIS dataset obtained in step S1, specifically:

[0022] ①Convert all ligand compounds in the DAVIS dataset into SMILES string format;

[0023] ②Use the Molformer model to extract the molecular representation vector from the SMILES string;

[0024] ③ Perform K-Means clustering on the molecular representation vector of the positive sample and extract the core cluster center;

[0025] ④ Calculate the minimum Euclidean distance between negative samples and the core cluster center, and eliminate redundant samples that exceed the set value;

[0026] ⑤Merge the positive samples with the negative samples filtered in step ④ and adjust the category ratio to 1:1 to form dataset D2.

[0027] In step S2(b), the molecular representation vector is generated by a Molformer model based on the Transformer architecture to capture the compound structure information; the criterion for eliminating negative samples is: the minimum Euclidean distance between it and the cluster center of the positive sample exceeds a set threshold.

[0028] Step S3, label conversion and training data set construction: For the optimized data set D1 and the optimized data set D2, different types of labels are generated by using threshold division and logarithmic function mapping, and three training data sets are constructed, including the following steps:

[0029] (i) Perform hard label conversion on the D1 dataset: If pK d >5, marked as 1, otherwise 0;

[0030] (ii) Perform hard label conversion on the D2 dataset: If pK d >5, marked as 1, otherwise 0;

[0031] (iii) Perform soft label transformation on the D2 dataset by logarithmic function projection, and convert pK into d Mapping to the (0,1) interval:

[0032]

[0033] in, represents the soft label, pK d is the sample label, and the hyperparameters a and b are optimized by the grid search method.

[0034] Step S4, model training and performance evaluation: For the three training datasets mentioned above, different operations are performed to train the drug-target interaction prediction model and evaluate the performance, and the dataset with the best performance is finally selected as the result, including the following steps:

[0035] (a) Using five-fold cross validation, each dataset is divided into training set, validation set, and test set according to different proportions;

[0036] (b) Using mean square error (MSE) as the loss function, the Adam optimizer is used for model training.

[0037] (c) Record the model performance indicators on the test set, including the area under the ROC curve (AUROC) and the area under the precision-recall curve (AUPRC), and finally select the dataset with the best performance as the result.

[0038] In step S4, the drug-target interaction prediction model is a HyperAttentionDTI model, which adopts a structure combining a convolutional neural network and an attention mechanism, and the input features include the amino acid sequence of the target protein and the SMILES string of the ligand compound.

[0039] A system for characterization optimization and soft label construction of drug-target interaction prediction models, including:

[0040] Data acquisition and preprocessing module is used to obtain DAVIS data sets of target protein and ligand compound interactions, and to calculate the original dissociation constant K d The labels are logarithmically transformed to obtain pK d Value, and divide the samples into positive samples and negative samples through the threshold;

[0041] The negative sample optimization sampling module is used to address the redundant negative sample problem in the DAVIS dataset and uses the following two strategies to generate optimized datasets D1 and D2;

[0042] The label conversion and training dataset construction module is used to generate different types of labels for the optimized dataset D1 and the optimized dataset D2 using threshold division and logarithmic function mapping, respectively, and to construct three training datasets;

[0043] The model training and performance evaluation module is used to perform different operations on the three training data sets mentioned above to train the drug-target interaction prediction model and evaluate the performance, and finally select the data set with the best performance as the result.

[0044] The present invention has the following beneficial effects and advantages:

[0045] 1. Addressing data bias: Through a K-Means clustering screening strategy based on molecular representation, redundant negative samples are dynamically eliminated. While ensuring a balanced ratio of positive and negative samples (1:1), this strategy avoids the drawbacks of traditional random sampling that loses chemical space information or similarity screening that introduces redundant features, significantly improving the model's ability to generalize to new targets or compounds.

[0046] 2. Preserve continuous binding strength information: Design a soft label mapping mechanism based on logarithmic function to convert pK d The values ​​are converted into continuous values ​​distributed in the (0,1) interval, overcoming the problem of information loss of weakly bound compounds caused by traditional binary labels and enhancing the model's ability to analyze complex interaction patterns.

[0047] 3. Optimize the robustness of molecular representation: Use the Molformer pre-trained model to extract standardized molecular feature vectors, and combine the K-Means clustering center to dynamically calibrate the negative sample feature space, eliminating feature shifts caused by structural ambiguity or incomplete representation, and improving the reliability of model input data.

[0048] 4. Full-process performance improvement: By integrating a sample balancing strategy with a soft labeling mechanism and optimizing hyperparameters through five-fold cross-validation and grid search, the drug-target interaction prediction model HyperAttentionDTI achieved AUROC and AUPRC scores of 0.928 and 0.931, respectively, representing performance improvements over baseline methods. This validates the high precision and robustness of the scheme in drug virtual screening scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 is a flow chart of the present invention;

[0050] Figure 2 Schematic diagram of the logarithmic function curve. DETAILED DESCRIPTION

[0051] To make the above-mentioned objects, features, and advantages of the present invention more readily apparent, the specific implementation methods of the present invention are described in detail below with reference to the accompanying drawings. The following description sets forth many specific details to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways than those described herein, and those skilled in the art may make similar modifications without departing from the scope of the invention. Therefore, the present invention is not limited to the specific implementation methods disclosed below.

[0052] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art of the art to which the present invention pertains. The terms used in the specification of the invention herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention.

[0053] The present invention discloses a data processing method for drug-target interaction (DTI) prediction, which is used to solve the problem of insufficient model generalization ability caused by data distribution imbalance and label discretization in virtual screening. The method includes the following innovative steps: (1) constructing a molecular representation space based on the Molformer pre-training model, using the K-Means clustering algorithm to perform feature space similarity analysis on negative samples, and eliminating redundant samples by setting a dynamic threshold to achieve positive and negative sample balance; (2) designing a soft label optimization mechanism based on the logarithmic function to convert the pK d The values ​​are converted into soft labels with a continuous distribution in the (0,1) interval, retaining the binding strength gradient information to enhance the model's sensitivity to weakly binding compounds; (3) The above data augmentation strategy is integrated to train the drug-target interaction prediction model HyperAttentionDTI through a five-fold cross-validation method. Experimental results show that compared with traditional random sampling and binary labeling methods, this scheme achieves model area under the receiver operating characteristic (AUROC) and area under the precision-recall curve (AUPRC) of 0.928 (±0.002) and 0.931 (±0.006), respectively, verifying its significant advantages in alleviating data bias and information loss problems. This method also provides a highly robust data processing framework for deep learning-based drug virtual screening.

[0054] like Figure 1 、 Figure 2 As shown, a method for characterization optimization and soft label construction of a drug-target interaction prediction model includes the following steps:

[0055] Step S1, data acquisition and preprocessing: obtain the DAVIS data set of the interaction between the target protein and the ligand compound, and calculate the original dissociation constant K d The labels are logarithmically transformed to obtain pK d Value, and divide the samples into positive samples and negative samples through the threshold;

[0056] Step S2, negative sample optimization sampling: To address the redundant negative sample problem in the DAVIS dataset, two strategies, random sampling and cluster screening based on molecular characterization information, are used to generate optimized datasets D1 and D2 respectively;

[0057] Step S3, label conversion and training data set construction: For the optimized data set D1 and the optimized data set D2, different types of labels are generated by using threshold division and logarithmic function mapping respectively, and three training data sets are constructed;

[0058] Step S4: Model training and performance evaluation: For the three training data sets mentioned above, the drug-target interaction prediction model is trained and the performance is evaluated.

[0059] In step S1, the logarithmic transformation formula is:

[0060]

[0061] where pK d Samples with >5 are defined as positive samples, and the rest are negative samples.

[0062] The step S2 comprises the following steps:

[0063] (a) For the positive and negative samples of the DAVIS dataset obtained in step S1, redundant negative samples are eliminated by random sampling, maintaining a positive to negative sample ratio of 1:1, and forming dataset D1;

[0064] (b) Perform cluster screening based on molecular characterization information on the positive and negative samples of the DAVIS dataset obtained in step S1, specifically:

[0065] ①Convert all ligand compounds in the DAVIS dataset into SMILES string format;

[0066] ②Use the Molformer model to extract the molecular representation vector from the SMILES string;

[0067] ③ Perform K-Means clustering on the molecular representation vector of the positive sample and extract the core cluster center;

[0068] ④ Calculate the minimum Euclidean distance between negative samples and the core cluster center, and eliminate redundant samples that exceed the set value;

[0069] ⑤Merge the positive samples with the negative samples filtered in step ④ and adjust the category ratio to 1:1 to form dataset D2.

[0070] In step S3, the three training data sets are constructed in the following manner:

[0071] (i) Perform hard label conversion on the D1 dataset: If pK d >5, marked as 1, otherwise 0;

[0072] (ii) Perform hard label conversion on the D2 dataset: If pK d >5, marked as 1, otherwise 0;

[0073] (iii) Perform soft label transformation on the D2 dataset by logarithmic function projection, and convert pK into d Mapping to the (0,1) interval:

[0074]

[0075] in, represents the soft label, pK d is the sample label, and the hyperparameters a and b are optimized by the grid search method.

[0076] In step S4, the drug-target interaction model may adopt the HyperAttentionDTI model, and its training and performance evaluation methods are as follows:

[0077] (a) Using five-fold cross-validation, each dataset was divided into 64% training set, 16% validation set, and 20% test set;

[0078] (b) The model is trained on an NVIDIA A800 GPU using the Adam optimizer (learning rate 0.0001) with mean squared error (MSE) as the loss function.

[0079] (c) Record the model performance indicators on the test set, including the area under the ROC curve (AUROC) and the area under the precision-recall curve (AUPRC), and finally select the dataset with the best performance as the result.

[0080] The present invention uses drug-target interaction data as the training platform configuration as follows: an Intel(R) Xeon(R) Platinum 8358P CPU @ 2.60GHz, 512GB of memory, an NVIDIA A800 GPU with 80GB of video memory, and a Linux operating environment using Python version 3.11. The method was implemented on the deep learning framework Pytorch version 2.4.0, CUDA version 11.8, and cuDNN version 9.0.1.

[0081] A method for characterization optimization and soft label construction of a drug-target interaction prediction model includes the following parts:

[0082] (1) Download the DAVIS dataset file and write a Python script to convert all K d The label is converted to pK as in formula (1) d , and all pK d Samples with label values ​​greater than 5 are marked as positive samples, and the others are marked as negative samples.

[0083] (2) Random sampling is used to remove redundant negative samples from the DAVIS dataset, keeping the positive and negative sample ratio approximately 1:1, to form dataset D1.

[0084] (3) Based on the sampling of molecular representation information, RDKit is used to generate standardized SMILES strings for the compounds to eliminate structural ambiguity; the SMILES strings are converted into molecular representation vectors of fixed dimensions through the Molformer model; K-Means clustering is performed on the representation vectors of positive samples, the core cluster centers are extracted, and the minimum Euclidean distance between the negative samples and the cluster centers is calculated, as shown in formula (3).

[0085]

[0086] Among them, x i represents negative samples, C i is the cluster center, and n is the number of negative samples.

[0087] Redundant negative samples exceeding the threshold are eliminated; the filtered positive and negative samples are merged and the category ratio is adjusted to form a structurally balanced dataset D2, as shown in Table 1.

[0088] Table 1 Summary of dataset D2

[0089]

[0090] (4) For the optimized dataset D1 and the optimized dataset D2, different types of labels are generated by using threshold division and logarithmic function mapping, and three training datasets are constructed;

[0091] (i) Perform hard label conversion on the D1 dataset: If pK d >5, marked as 1, otherwise 0;

[0092] (ii) Perform hard label conversion on the D2 dataset: If pK d >5, marked as 1, otherwise 0;

[0093] (iii) Perform soft label transformation on the D2 dataset by logarithmic function projection, and convert pK into d Mapping to the (0,1) interval:

[0094]

[0095] in represents the soft label, pK d is the sample label.

[0096] (5) Use the PyTorch framework to reproduce the HyperAttentionDTI model and set the number of network layers as shown in Table 2.

[0097] Table 2 HyperAttentionDTI model structure parameters

[0098]

[0099] (6) Five-fold cross validation was used to divide each dataset into a training set (64%), a validation set (16%), and a test set (20%). The mean square error (MSE) function was used as the loss function. The Adam optimizer (learning rate 0.0001) was used. The area under the ROC curve (AUROC) and the area under the precision-recall curve (AUPRC) were used to measure the model performance. The test results are shown in Table 3.

[0100] Table 3 HyperAttentionDTI model prediction results

[0101]

[0102] According to the experimental results in Table 3, the clustering and negative sample screening strategy based on molecular characterization information proposed in the present invention significantly improved the predictive performance of the HyperAttentionDTI model. At the same time, the soft labeling mechanism constructed by the logarithmic function retained the continuous distribution information of the binding strength, which reduced the quantitative prediction error of the model for the drug-target interaction strength and achieved refined modeling of biological activity data.

Claims

1. A method for characterization optimization and soft label construction of a drug-target interaction prediction model, characterized in that: The following steps are involved: S1. Data acquisition and preprocessing: Obtain the DAVIS dataset of the interaction between the target protein and the ligand compound, and calculate the original dissociation constant K d The labels are logarithmically transformed to obtain pK d Value, and divide the samples into positive samples and negative samples through the threshold; S2, negative sample optimization sampling: To address the problem of redundant negative samples in the DAVIS dataset, the following two strategies are used to generate optimized datasets D1 and D2 respectively; S3, label conversion and training dataset construction: For the optimized dataset D1 and optimized dataset D2, different types of labels are generated by using threshold division and logarithmic function mapping, and three training datasets are constructed. S4. Model training and performance evaluation: For the three training datasets mentioned above, different operations were performed to train the drug-target interaction prediction model and evaluate the performance, and finally the dataset with the best performance was selected as the result.

2. The method for characterization optimization and soft label construction of a drug-target interaction prediction model according to claim 1, characterized in that: In step S1, the logarithmic transformation formula is: where pK d Samples with >5 are defined as positive samples, and the rest are negative samples.

3. The method for characterization optimization and soft label construction of a drug-target interaction prediction model according to claim 1, characterized in that: Step S2, negative sample optimization sampling: To address the redundant negative sample problem in the DAVIS dataset, the following two strategies are used to generate optimized datasets D1 and D2, respectively, including the following steps: To address the problem of redundant negative samples in the DAVIS dataset, the following two strategies are used to generate optimized datasets D1 and D2 respectively; (a) For the positive and negative samples of the DAVIS dataset obtained in step S1, redundant negative samples are eliminated by random sampling, maintaining a positive to negative sample ratio of 1:1, and forming dataset D1; (b) Perform cluster screening based on molecular characterization information on the positive and negative samples of the DAVIS dataset obtained in step S1, specifically: ①Convert all ligand compounds in the DAVIS dataset into SMILES string format; ②Use the Molformer model to extract the molecular representation vector from the SMILES string; ③ Perform K-Means clustering on the molecular representation vector of the positive sample and extract the core cluster center; ④ Calculate the minimum Euclidean distance between negative samples and the core cluster center, and eliminate redundant samples that exceed the set value; ⑤Merge the positive samples with the negative samples filtered in step ④ and adjust the category ratio to 1:1 to form dataset D2.

4. The method for characterization optimization and soft label construction of a drug-target interaction prediction model according to claim 3, characterized in that: In step S2(b), the molecular representation vector is generated by a Molformer model based on the Transformer architecture to capture the compound structure information; the criterion for eliminating negative samples is: the minimum Euclidean distance between it and the cluster center of the positive sample exceeds a set threshold.

5. The method for characterization optimization and soft label construction of a drug-target interaction prediction model according to claim 1, characterized in that: Step S3, label conversion and training data set construction: For the optimized data set D1 and the optimized data set D2, different types of labels are generated by using threshold division and logarithmic function mapping, and three training data sets are constructed, including the following steps: (i) Perform hard label conversion on the D1 dataset: If pK d >5, marked as 1, otherwise 0; (ii) Perform hard label conversion on the D2 dataset: If pK d >5, marked as 1, otherwise 0; (iii) Perform soft label transformation on the D2 dataset by logarithmic function projection, and convert pK into d Mapping to the (0,1) interval: in, represents the soft label, pK d is the sample label, and the hyperparameters a and b are optimized by the grid search method.

6. The method for characterization optimization and soft label construction of a drug-target interaction prediction model according to claim 1, characterized in that: Step S4, model training and performance evaluation: For the three training datasets mentioned above, different operations are performed to train the drug-target interaction prediction model and evaluate the performance, and the dataset with the best performance is finally selected as the result, including the following steps: (a) Using five-fold cross validation, each dataset is divided into training set, validation set, and test set according to different proportions; (b) Using mean square error (MSE) as the loss function, the Adam optimizer is used for model training. (c) Record the model performance indicators on the test set, including the area under the ROC curve (AUROC) and the area under the precision-recall curve (AUPRC), and finally select the dataset with the best performance as the result.

7. The method for characterization optimization and soft label construction of a drug-target interaction prediction model according to claim 1, characterized in that: In step S4, the drug-target interaction prediction model is a HyperAttentionDTI model, which adopts a structure combining a convolutional neural network and an attention mechanism, and the input features include the amino acid sequence of the target protein and the SMILES string of the ligand compound.

8. A drug-target interaction prediction model characterization optimization and soft label construction system, characterized by: include: Data acquisition and preprocessing module is used to obtain DAVIS data sets of target protein and ligand compound interactions, and to calculate the original dissociation constant K d The labels are logarithmically transformed to obtain pK d Value, and divide the samples into positive samples and negative samples through the threshold; The negative sample optimization sampling module is used to address the redundant negative sample problem in the DAVIS dataset and uses the following two strategies to generate optimized datasets D1 and D2; The label conversion and training dataset construction module is used to generate different types of labels for the optimized dataset D1 and the optimized dataset D2 using threshold division and logarithmic function mapping, respectively, and to construct three training datasets; The model training and performance evaluation module is used to perform different operations on the three training data sets mentioned above to train the drug-target interaction prediction model and evaluate the performance, and finally select the data set with the best performance as the result.

Images