A drug-target interaction prediction method based on large language model and image representation

Abstract

The present invention discloses a drug-target interaction prediction method based on a large language model and image representation. The large language model ESM-2 is used to encode the target sequence, and the large language model X-MOL is used to encode the drug SMILES to generate a feature matrix. A two-dimensional image representation of the drug and target is generated through image representation technology. A drug-target interaction prediction model MapCPI is constructed, and the two-dimensional image representation of the drug and target is input into MapCPI to obtain the drug-target interaction prediction probability and determine whether the drug-target binds. The model uses this encoded information to effectively extract drug-target interaction features without relying on target structure information, thereby realizing the prediction of drug-target interactions and showing good generalization performance. The present invention improves the accuracy and efficiency of drug-target interaction prediction by combining a large language model and image representation technology.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug-target affinity prediction, and in particular relates to a drug-target interaction prediction method based on a large language model and image representation. Background Art

[0002] Understanding and predicting drug-target protein interactions is a crucial area of ​​research in drug discovery and development. While traditional experimental methods can provide reliable data, their high cost, time consumption, and resource requirements limit the implementation of large-scale screening. Leveraging computational methods to predict drug-target protein interactions early in drug development holds promise for improving the success rate of drug discovery.

[0003] In recent years, with the rapid development of big data and computer technology, artificial intelligence methods have been widely used in the development of compound-protein interaction (CPI) prediction models. However, existing CPI models generally rely on known ligand information of proteins, have weak generalization capabilities and have several limitations. First, some existing methods rely on the three-dimensional structure of known proteins, and for proteins with unresolved structures, these methods cannot effectively predict their interactions. Second, the protein and compound representations in existing sequence-based CPI prediction methods are not accurate enough, and the correlation information between features is ignored, and the potential relationship between coding features is not paid attention to. These problems lead to the lack of prediction accuracy and generalization ability of existing CPI models, limiting their application in real drug discovery. Summary of the Invention

[0004] This paper proposes a drug-target interaction prediction method based on a large language model and image representation, aiming to overcome the limitations of existing methods and provide a more accurate prediction tool for drug discovery.

[0005] A drug-target interaction prediction method based on a large language model and image representation comprises the following steps:

[0006] S1: Collect drug-target interaction benchmark datasets and obtain drug SMILES and target sequences;

[0007] S2: Use the large language model ESM-2 to encode the target sequence, and use the large language model X-MOL to encode the drug SMILES to generate a feature matrix. Then, use image representation technology to generate a two-dimensional image representation of the drug and target.

[0008] S3: Construct a drug-target interaction prediction model MapCPI, input the two-dimensional image representation of the drug and target into MapCPI, and obtain the predicted probability of drug-target interaction;

[0009] S4: Determine whether drug-target binding occurs based on the predicted probability of drug-target interaction.

[0010] In step S1, the drug-target interaction benchmark dataset includes a BindingDB dataset and a Human dataset.

[0011] In step S2, the target sequence is encoded using the large language model ESM-2, and the drug SMILES is encoded using the large language model X-MOL to generate a feature matrix. The two-dimensional image representation of the drug and target is generated using image representation technology, specifically including:

[0012] 2.1) Use the large language model ESM-2 to encode protein sequences in the protein database to generate a protein encoding matrix. Use the large language model X-MOL to encode all bioactive compounds in the compound database to generate a compound encoding matrix. Use the average pooling operation to convert the original feature matrices of targets and drugs into vectors.

[0013] 2.2) The pairwise distances between protein encoding matrices are calculated based on cosine similarity to generate a protein feature distance matrix. Similarly, the pairwise distances between compound encoding matrices are calculated based on cosine similarity to generate a compound feature distance matrix. The calculation formula is:

[0014]

[0015] Where a and b represent different features, f a and f b Represents the eigenvalue vector composed of all samples on this dimension feature;

[0016] 2.3) Reducing the dimensionality of the protein feature distance matrix or the compound feature distance matrix using a uniform manifold approximation and projection algorithm, and projecting the matrix into a two-dimensional space to obtain a scatter distribution of the protein features or the compound features, respectively;

[0017] 2.4) Using the Jonker-Volgenant algorithm, the scattered distribution of protein features or compound features is linearly distributed to obtain a regularized protein template image or a regularized compound template image. The regularized protein template image and the regularized compound template image constitute a two-dimensional image representation of the drug and target.

[0018] In step 2.1), the protein database is the TTD database and the UniProt database, and the compound database is the ChEMBL database.

[0019] In step S3, a dual-channel convolutional neural network is used to construct a drug-target interaction prediction model MapCPI.

[0020] In step S4, whether the drug binds to the target is evaluated using a benchmark dataset, which is the BindingDB and Human databases.

[0021] Furthermore, a drug-target interaction prediction method based on a large language model and image representation includes the following steps:

[0022] S1: Collect drug-target interaction benchmark datasets and obtain drug SMILES and target sequences;

[0023] S2: Use the large language model ESM-2 and X-MOL to encode the target sequence and drug SMILES respectively, generate feature matrices, and generate two-dimensional image representations through image representation technology;

[0024] S3: Construct a drug-target interaction prediction model MapCPI, input the two-dimensional image representation of drugs and targets into MapCPI, and obtain the drug-target interaction prediction probability.

[0025] S4: Determine whether drug-target binding occurs based on the predicted probability of drug-target interaction.

[0026] In step S1, the dataset includes BindingDB and Human datasets. First, the two datasets are divided into training set, validation set and test set respectively. The BindingDB dataset sets random partitioning of interaction pairs and unseen target partitioning, and the Human dataset sets random partitioning of interaction pairs and cold start partitioning.

[0027] Furthermore, in step S2, the large language models ESM-2 and X-MOL are used to encode the target sequence and drug SMILES respectively to generate a feature matrix, and the two-dimensional image representation of the target and drug is achieved through the two steps of template image generation and image conversion.

[0028] The template image generation technology includes:

[0029] Large language model encoding: Use ESM-2 to encode large-scale protein sequences in the TTD (Therapeutic Target Database) and UniProt databases to generate encoding matrices. Use X-MOL to encode all bioactive compounds in the ChEMBL database to generate encoding matrices. Use average pooling operations to convert the original feature matrices of targets and drugs into vectors.

[0030] Feature distance calculation: The pairwise distances between protein coding features were calculated based on cosine similarity to generate a feature distance matrix. Similarly, the pairwise distances between compound coding features were calculated based on cosine similarity to generate a feature distance matrix. The calculation formula is:

[0031]

[0032] Where a and b represent different features, f a and f b Represents the eigenvalue vector composed of all samples on this dimension feature.

[0033] Data dimensionality reduction: High-dimensional features are projected into two-dimensional space through the Uniform Manifold Approximation and Projection Algorithm (UMAP). The equivalent fuzzy topological structure with the closest distance relationship to the original is searched for and the coordinates of the scattered points are calculated to generate the corresponding scattered point distribution.

[0034] Optimal grid assignment: Using the Jonker-Volgenant (JV) algorithm, the scattered points of all features are linearly assigned to obtain a regularized template image. The template image records the spatial position of each feature. This process minimizes the square distance between the scattered points and the grid coordinates. The calculation formula is as follows:

[0035]

[0036] where x scatter ,y grid Represent the scattered coordinate matrix and grid position matrix of all features respectively;

[0037] The image conversion technology first performs large language model encoding on the drug-target interaction pair under study to obtain encoding vectors, and then maps the eigenvalues ​​of the encoding vectors to corresponding positions based on the template image obtained above, finally obtaining the image representation ESM2image (size 51×51) of the target and the image representation XMOLimage (size 28×28) of the drug.

[0038] Furthermore, in step S3, a two-channel convolutional neural network is used to extract features from both ESM2image and XMOLimage simultaneously. The first layer is a convolutional layer, using multiple convolution kernels with a stride of 1 (proteins use a 13×13 convolution kernel and compounds use a 9×9 convolution kernel). The second layer is a maximum pooling layer with a stride of 2 (proteins use a 5×5 pooling kernel and compounds use a 3×3 pooling kernel). The third layer is a convolutional base, which contains three parallel convolution operations, each using a different convolution kernel (proteins use 1×1, 5×5, and 9×9 convolution kernels, and compounds use 1×1, 3×3, and 5×5 convolution kernels).

[0039] The outputs of the convolutional bases are concatenated and fed into a max pooling layer, where a second convolutional base is used to further enhance feature extraction, followed by a global max pooling layer to extract the final embedding. The output of each convolutional layer is then passed through a ReLU activation function to introduce nonlinearity.

[0040] The protein and compound embeddings generated by the two-channel convolutional neural network are concatenated to obtain the embedding vector z of the interaction pair. The classifier consists of four fully connected layers. The outputs of the first three layers are nonlinearly transformed using the ReLU activation function, and the output of the last layer is normalized using the softmax function to obtain the probability of drug-target interaction.

[0041] Furthermore, in step S4, the present invention and five advanced drug-target interaction prediction models are compared and evaluated on the BindingDB and Human benchmark datasets, and the performance of the present invention is comprehensively evaluated by comprehensively considering evaluation indicators such as the Matthews Correlation Coefficient (MCC), the Area Under the Receiver Operating Characteristic Curve (AUROC), and the Area Under the Precision-Recall Curve (AUPRC).

[0042] MCC: A comprehensive model performance evaluation indicator used to evaluate the overall prediction performance of a binary classification model. The value ranges from -1 to 1, where 1 indicates a completely correct prediction, 0 indicates a random prediction, and -1 indicates a completely wrong prediction. The calculation formula is:

[0043]

[0044] AUROC: It is used to measure the ability of a classification model to distinguish between positive and negative classes. It represents the area under the ROC curve. The closer its value is to 1, the better the model performance. The horizontal axis of the ROC curve is the false positive rate FPR, and the vertical axis is the true positive rate TPR. The calculation formula is:

[0045] AUROC = ∫0 1 TPR(FPR)d(FPR)

[0046]

[0047] AUPRC: It is used to measure the balance between precision and recall of the classification model at different thresholds. The closer the value is to 1, the better the performance. The horizontal axis of the curve is recall, and the vertical axis is precision. The calculation formula is:

[0048] AUPRC=∫0 1 Precision(Recall)d(Recall)

[0049]

[0050] In the above formula, TP represents the number of true positive samples (the number of positive samples correctly predicted), FP represents the number of false positive samples (the number of negative samples incorrectly predicted as positive), TN represents the number of true negative samples (the number of negative samples correctly predicted), and FN represents the number of false negative samples (the number of positive samples incorrectly predicted as negative).

[0051] The present invention is used to evaluate drug-target interactions derived from the Human database, with a predicted probability value ranging from 0 to 1, with 0.5 as the threshold. A predicted probability higher than 0.5 indicates that the drug-target interaction can occur, and a predicted probability lower than 0.5 indicates that the drug-target interaction cannot occur.

[0052] Compared with the prior art, the present invention has the following advantages:

[0053] The present invention abandons protein structure information and directly extracts interaction features from protein sequences and compound SMILES for the prediction of drug-target interactions; secondly, the present invention combines biological large language model encoding with image-like dimension expansion technology to innovatively perform image representation of proteins and compounds based on feature correlation. This image representation strategy aggregates the features of proteins and compounds encoded by the large language model according to the potential relationship between the encoding features, and can accurately characterize interacting proteins and compounds. By combining the large language model and image representation technology, the present invention has developed a drug-target interaction prediction model, which improves the prediction accuracy and efficiency and exhibits good generalization performance. The present invention belongs to the field of drug-target affinity prediction technology and provides a new tool for drug discovery. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 Schematic diagram of the process of the drug-target interaction prediction method based on large language model and image representation in the present invention;

[0055] Figure 2 Schematic diagram of compound / protein data embedding based on large language model and image representation method in the present invention;

[0056] Figure 3 Schematic diagram of the structure of the drug-target interaction prediction method based on a large language model and image representation in the present invention;

[0057] Figure 4 This is a visualization result image of the drug-target interaction prediction performed in the present invention. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments derived by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0059] In this implementation case, Figure 1 As shown, the present invention provides a drug-target interaction prediction method MapCPI based on a large language model and image representation, comprising the following steps:

[0060] Step S1: Collect drug-target interaction benchmark datasets and obtain drug SMILES and target sequences

[0061] First, two drug-target interaction benchmark datasets, BindingDB and Human, were collected from public literature, and the SMILES and target sequence data of the drugs were obtained. The BindingDB dataset contains 39,747 positive interaction pairs and 31,218 negative interaction pairs. It was randomly divided into training, validation, and test sets, and a subset of interaction pairs whose targets were not seen in the training set was extracted from the test set. For the Human dataset, it was first randomly divided into training, validation, and test sets according to the ratio of 8:1:1. Secondly, a cold start partition was set for the Human dataset. First, 5% and 10% of the interaction pairs were randomly placed in the validation and test sets, respectively. From the remaining 85% of interaction pairs, all interaction pairs that repeated the drug and target in the validation or test sets were removed to form the training set.

[0062] Step S2: Use the large language model ESM-2 and X-MOL to encode the target sequence and drug SMILES respectively, generate a feature matrix, and generate a two-dimensional image representation through image representation technology;

[0063] For all drug-target interaction pairs, the drug and target are encoded using a large language model to obtain feature vectors, and then converted into two-dimensional image representations using image representation technology, that is, the drug and target are represented as XMOLimage and ESM2image respectively. Figure 2 As shown in Figure 2, the specific image representation process mainly includes two stages: template image generation and image conversion. The detailed steps are as follows:

[0064] 1) Template image generation phase, including large language model encoding, feature distance calculation, data dimensionality reduction, and optimal grid allocation, as described below:

[0065] 1.1 Large Language Model Encoding: Full proteome sequences of all drug target species were collected from the TTD (Therapeutic Target Database) and UniProt databases, and SMILES of all drug-like molecules with activity data were collected from the ChEMBL database. The protein sequences and compound SMILES were encoded using the Transformer-based large language model (ESM-2) with a parameter size of 3 billion, and the compound SMILES were encoded using X-MOL, respectively, to generate raw feature matrices. Using average pooling, the raw feature matrices of proteins and compounds were converted into encoding vectors of length 2,560 and 768, respectively. The 1,300,449 protein sequences collected were encoded using ESM-2, resulting in a 1,300,449 × 2,560 encoding matrix. The 2,231,654 bioactive compound SMILES were encoded using X-MOL, resulting in a 2,231,654 × 768 encoding matrix.

[0066] 1.2 Feature distance calculation: Based on the encoding results of 1.3 million protein sequences, the pairwise distances between protein coding features were calculated based on cosine similarity to generate a 2,560×2,560 feature distance matrix. Based on the encoding results of 2.2 million compound SMILES, the pairwise distances between compound coding features were calculated based on cosine similarity to generate a 768×768 feature distance matrix. The calculation formula for this process is:

[0067]

[0068] Where a and b represent different features, f a and f b Represents the eigenvalue vector composed of all samples on this dimension feature.

[0069] 1.3 Data dimensionality reduction: Based on the above feature distance matrix, the uniform manifold approximation and projection algorithm (UMAP) is used to project the high-dimensional features into two-dimensional space. The equivalent fuzzy topological structure with the closest relationship to the original distance is searched to calculate the coordinates of the scattered points and generate the corresponding scattered point distribution.

[0070] 1.4 Optimal Grid Allocation: Using the Jonker-Volgenant (JV) algorithm, the scattered points of all features are linearly distributed to obtain a regularized template image. The template image records the spatial position of each feature. This process minimizes the square distance between the scattered points and the grid coordinates. The calculation formula is as follows:

[0071]

[0072] where x scatter ,y gridRepresent the scatter coordinate matrix and grid position matrix of all features respectively.

[0073] 2) Image conversion stage

[0074] First, the drug-target interaction pair under study is encoded by the ESM-2 and X-MOL large language models to obtain the encoding vectors respectively. Based on the template images obtained above, the eigenvalues ​​of the encoding vectors are mapped to the corresponding positions respectively, and finally the image representation ESM2image (size 51×51) of the target and the image representation XMOLimage (size 28×28) of the drug are obtained.

[0075] Step S3: Construct a drug-target interaction prediction model MapCPI, input the two-dimensional image representation of the drug and target into MapCPI, and obtain the drug-target interaction prediction probability.

[0076] Use a two-channel convolutional neural network to extract features from ESM2image and XMOLimage simultaneously, such as Figure 3 As shown. The first layer is a convolutional layer, which uses a variety of convolutional kernels with a stride of 1 (72 convolutional kernels of size 13×13 are used for proteins, and 48 convolutional kernels of size 9×9 are used for compounds), followed by a maximum pooling layer with a stride of 2 to reduce computational cost (5×5 pooling kernels are used for proteins, and 3×3 pooling kernels are used for compounds). The third layer is a convolutional base, which contains three parallel convolution operations, each using a different convolutional kernel (1×1, 5×5, and 9×9 convolutional kernels are used for proteins, and 1×1, 3×3, and 5×5 convolutional kernels are used for compounds). The outputs of the convolutional bases are concatenated and input into a maximum pooling layer. Another convolutional base is then used to further enhance the representation capability. After the outputs of the second convolutional base are concatenated, global maximum pooling is applied to extract the final embedded representation. The output of each convolutional layer introduces nonlinear characteristics through the ReLU activation function.

[0077] The protein and compound embeddings generated by the two-channel convolutional neural network are concatenated into a joint embedding vector z for downstream drug-target interaction prediction, such as Figure 3 The classifier consists of four fully connected layers, with 512, 256, 64, and 2 neurons, respectively. The outputs of the first three layers undergo a nonlinear transformation using the ReLU activation function. The output of the last layer is normalized using the softmax function to obtain the probability of a drug-target interaction. A binary cross-entropy loss function is then used to simultaneously optimize the parameters of the dual-channel convolutional neural network and the classifier.

[0078] S4: Determine whether drug-target binding occurs based on the predicted probability of drug-target interaction.

[0079] The present invention comprehensively considers three evaluation indicators: Matthews Correlation Coefficient (MCC), Area Under the Receiver Operating Characteristic Curve (AUROC), and Area Under the Precision-Recall Curve (AUPRC). The larger the value of these indicators, the better the model prediction performance.

[0080] MCC: A comprehensive model performance evaluation indicator used to evaluate the overall prediction performance of a binary classification model. The value ranges from -1 to 1, where 1 indicates a completely correct prediction, 0 indicates a random prediction, and -1 indicates a completely wrong prediction. The calculation formula is:

[0081]

[0082] AUROC: It is used to measure the ability of a classification model to distinguish between positive and negative classes. It represents the area under the ROC curve. The closer its value is to 1, the better the model performance. The horizontal axis of the ROC curve is the false positive rate FPR, and the vertical axis is the true positive rate TPR. The calculation formula is:

[0083] AUROC = ∫0 1 TPR(FPR)d(FPR)

[0084]

[0085] AUPRC: Used to measure the balance between precision and recall of a classification model at different thresholds. The larger the value, the better the performance. The horizontal axis of the curve is Recall, and the vertical axis is Precision. The calculation formula is:

[0086] AUPRC=∫0 1 Preision(Recall)d(Recall)

[0087]

[0088] In the above formula, TP represents the number of true positive samples (the number of positive samples correctly predicted), FP represents the number of false positive samples (the number of negative samples incorrectly predicted as positive), TN represents the number of true negative samples (the number of negative samples correctly predicted), and FN represents the number of false negative samples (the number of positive samples incorrectly predicted as negative).

[0089] In this embodiment, as shown in Table 1, the developed MapCPI model exhibits excellent prediction performance in the random partitioning task of both the BindingDB and Human datasets, achieving the highest AUROC, AUPRC, and MCC indicators in the BindingDB dataset, and the highest AUROC and AUPRC indicators in the Human dataset. This result demonstrates the excellence and feasibility of the model proposed in the present invention. In addition, Figure 4 As shown in Table 1, the developed MapCPI model achieved optimal prediction performance in all three metrics: MCC, AUROC, and AUPRC, in both the BindingDB unseen target segmentation and the Human cold-start segmentation tasks. This further validates its excellent generalization ability in drug-target interaction prediction. These visualization results intuitively support the conclusions in Table 1 and further demonstrate the superiority of the model in drug-target interaction prediction.

[0090] Table 1 Comparison of experimental data of MapCPI model and other methods

[0091]

[0092] Furthermore, in this example, as shown in Table 2, the MapCPI model's prediction results for 10 drug-target interactions from the Human database were completely consistent with the results recorded in the database. This further demonstrates the superior performance and reliability of the model of the present invention in predicting drug-target interactions.

[0093] Table 2 Results of the MapCPI model in predicting 10 drug-target interactions

[0094]

[0095]

[0096]

[0097] Finally, it should be noted that although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations may be made to these embodiments without departing from the principles and spirit of the present invention.

[0098] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

[0099] The application of the present invention is not limited to the above examples. Those skilled in the art can make improvements or changes based on the above description. All these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A drug-target interaction prediction method based on a large language model and image representation, characterized in that: The following steps are involved: S1: Collect drug-target interaction benchmark datasets and obtain drug SMILES and target sequences; S2: Use the large language model ESM-2 to encode the target sequence, and the large language model X-MOL to encode the drug SMILES, generate a feature matrix, and generate a two-dimensional image representation of the drug and target through image representation technology, including: 2.1) Use the large language model ESM-2 to encode protein sequences in the protein database to generate a protein encoding matrix. Use the large language model X-MOL to encode all bioactive compounds in the compound database to generate a compound encoding matrix. Use the average pooling operation to convert the original feature matrices of targets and drugs into vectors. 2.2) Calculate the pairwise distances between protein encoding matrices based on cosine similarity to generate a protein feature distance matrix. Similarly, calculate the pairwise distances between compound encoding matrices based on cosine similarity to generate a compound feature distance matrix. The calculation formula is: ； Where a and b represent different features, and Represents the eigenvalue vector composed of all samples on this dimension feature; 2.3) Using uniform manifold approximation and projection algorithms, the protein feature distance matrix or compound feature distance matrix is ​​reduced in dimensionality and projected into a two-dimensional space to obtain the scatter distribution of protein features or compound features, respectively; 2.4) Using the Jonker-Volgenant algorithm, linearly distribute the scattered distribution of protein features or compound features to obtain a regularized protein template image or a regularized compound template image. The regularized protein template image and the regularized compound template image constitute a two-dimensional image representation of the drug and target; S3: Construct a drug-target interaction prediction model MapCPI, input the two-dimensional image representation of the drug and target into MapCPI, and obtain the predicted probability of drug-target interaction; S4: Determine whether drug-target binding occurs based on the predicted probability of drug-target interaction.

2. The drug-target interaction prediction method based on a large language model and image representation according to claim 1, characterized in that: In step S1, the drug-target interaction benchmark dataset includes a BindingDB dataset and a Human dataset.

3. The drug-target interaction prediction method based on a large language model and image representation according to claim 1, characterized in that: In step 2.1), the protein databases are the TTD database and the UniProt database.

4. The drug-target interaction prediction method based on a large language model and image representation according to claim 1, characterized in that: In step 2.1), the compound database is the ChEMBL database.

5. The drug-target interaction prediction method based on a large language model and image representation according to claim 1, characterized in that: In step S3, a dual-channel convolutional neural network is used to construct a drug-target interaction prediction model MapCPI.

Images