A multi-modal information fusion model and method for DTA prediction

Abstract

The application provides a multimodal information fusion model and method for DTA prediction, the model comprising a drug molecule structure information encoder, a target structure information encoder, a multimodal balance module and a drug target fusion module; the drug molecule structure information encoder uses a Transformer model to encode drug string modal information and uses a GIN model to extract drug graph modal information features; the target structure information encoder uses a Transformer model to encode target string modal information and uses a GCN model to extract drug graph modal information features; the multimodal balance module uses a contrast learning method to balance and integrate drug string and graph modal information and balance and integrate target string and graph modal information; and the drug target fusion module connects the two modal features of the drug and target obtained by the multimodal balance module and is used for DTA prediction.

Description

Technical Field

[0001] This invention relates to the field of drug target binding affinity prediction technology, and in particular to a multimodal information fusion model and method for DTA prediction. Background Technology

[0002] Drug discovery is the process of finding potential new drugs, involving multiple fields such as pharmacology, chemistry, and biology, and typically requires enormous economic and time costs. It is estimated that developing a new drug costs approximately $2.6 billion, and obtaining FDA approval takes 17 years. Over the years, with the development of computer technology, computer-aided drug discovery has become a trend, making it imperative to develop computational models to advance the drug discovery process. Successfully identifying drug-target interactions is a crucial step in drug discovery, and accurately identifying the affinity of drug-target interactions is even more important for drug development. DTA represents the strength of the binding relationship between a drug molecule and its target. Generally, the stronger the binding between a compound molecule and its target, the more likely the compound is to affect the biological function of the target and the more likely it is to be a suitable drug candidate. Therefore, establishing computational models to accurately predict DTA can accelerate the drug molecule screening process, minimize unnecessary in vitro screening experiments, and is of great significance to drug development.

[0003] Many computational methods and models for DTA prediction have been proposed, such as traditional molecular docking techniques. These techniques are based on the 3D structures of the target and compound molecules, using computer simulations to predict the binding modes and binding affinity between the drug and the target. Many mature molecular docking algorithms are developed as software, such as Gold and Dock, but these techniques are very time-consuming. With the development of computer technology, molecular dynamics simulation techniques have emerged. For example, Elanie et al. combined a fast geometric docking algorithm with molecular mechanical interaction energy assessment to calculate and score the potential of each ligand atom. This approach is more flexible and provides more accurate predictions, but at the cost of high computational and time costs.

[0004] Most early machine learning methods were based on matrix computations for prediction using structural similarity calculations, which significantly reduced costs. For example, He et al. proposed a method called SimBoost, which predicts continuous values ​​of the binding affinity between compounds and proteins. Li et al. proposed a molecular docking method based on random forests, which predicts by applying the Kronecker similarity matrix product. However, these methods rely excessively on the structural data features of molecules, and obtaining this data is both difficult and time-consuming. With the rapid development of deep learning and the era of big data, convolutional neural networks (CNNs), graph neural networks (GNNs), and their variants have been applied to the field of drug discovery. Since the structural information of drugs and targets plays a crucial role in DTA prediction, most existing DTA prediction methods are based on the structural information of drugs and targets, and they can be divided into string-based modalities and graph-based modalities.

[0005] String-based modal methods learn features from sequence data. For example, DeepDTA uses CNNs to extract features from one-dimensional representations of target sequences and drug SMILES. WideDTA supplements this with information on protein domains, motifs, and the largest common substructure word, and introduces a word-based sequence representation for DTA prediction. In contrast, AttentionDTA focuses more on important key subsequences in drug and target sequences and introduces a bilateral multi-head attention mechanism to predict DTA. These methods only focus on the string modality of drug SMILES and target information, ignoring spatial structure and hydrogen atom information. Furthermore, the fixed string length is considered during embedding, leading to the loss of some useful information. To address this drawback, graph-based modal methods have emerged. GraphDTA proposes representing drug molecular structure information as a graph and using GNNs to extract features from the drug molecule graph and CNNs to extract features from the target sequence. DGraphDTA uses the drug molecule graph and target structure graph for DTA prediction and a graph convolutional neural network (GCN) model for feature extraction. However, drug molecule maps lack contextual semantic information and atomic positional arrangement. Furthermore, the target structure map in this method only considers the spatial structure of the target, neglecting the arrangement order of target residues and ignoring the positional information of peptide chain residues. Therefore, it is necessary to systematically consider the multimodal information of drug and target structures to obtain more complete information for better DTA prediction.

[0006] Multimodal technologies can systematically consider information from multiple different modalities. Over the past decade, information fusion technology has successfully achieved the fusion of multimodal information, and its utilization has attracted researchers' attention. For example, Tuan et al. proposed a new method for detecting fake news by fusing multimodal features from text and visual data. Mou et al. proposed a deep learning model to fuse data from multiple modalities, including eye data, vehicle data, and environmental data. This demonstrates that multimodal information fusion technology has been widely applied in various fields. Similarly, the fusion and utilization of multimodal information can also be applied in drug discovery. For example, Deng et al. developed Graph2MDA, a method based on multimodal variational graph embedding, which fuses multiple attributes and features of microorganisms and drugs to predict microbe-drug associations. Lyu et al. considered the potential correlations between drugs and multimodal data such as targets and enzymes, and designed an MDNN dual-channel framework to obtain multimodal representations of drugs. It can be seen that these drug discovery methods use embedded representations of different drug attributes expressed in different modalities, without simultaneously focusing on multiple modalities of a single attribute. Moreover, existing DTA methods only consider the single structural properties of drugs and targets, without taking into account the multiple attribute information of their different modes. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a multimodal information fusion model and method for DTA prediction. The model can embed string and graph modal information from drugs and targets, and balance the feature representations of different modalities through contrastive learning methods to output richer information for DTA prediction.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical method: a multimodal information fusion model for DTA prediction, comprising: a drug molecule structure information encoder, a target structure information encoder, a multimodal balancing module, and a drug-target fusion module;

[0009] The drug molecular structure information encoder uses the Transformer model to encode the modal information of the drug string and uses the GIN model to extract the modal information features of the drug graph.

[0010] The target structure information encoder uses the Transformer model to encode the modal information of the target string and uses the GCN model to extract the modal information features of the drug map.

[0011] The multimodal balancing module uses contrastive learning to balance and integrate drug strings and graph modal information, as well as target strings and graph modal information.

[0012] The drug-target fusion module connects the two modal features of the drug and the target obtained from the multimodal balancing module for DTA prediction.

[0013] As another aspect of the present invention, a multimodal information fusion method for DTA prediction includes:

[0014] Step S1, embedding of string modalities;

[0015] The drug SMILES code is treated as a string, encoded with integers, and the positional encoding incorporated into this encoding is used to obtain a vector representation. The Transformer model is then used to extract features from this vector to obtain the final vector representation of the SMILES string.

[0016] The target sequence is treated as a string, encoded with integers, and the positional encoding incorporated into this encoding is used to obtain a vector representation. The Transformer model is then used to extract features from this vector to obtain the final vector representation of the target string.

[0017] Step S2, embedding of graph modes;

[0018] Each atom is treated as a node in the drug molecule graph, the relationships between atoms are treated as the adjacency matrix of the drug molecule graph, and the properties of atoms are treated as the attribute features of the nodes in the drug molecule graph. The feature vectors of the drug molecule graph and its nodes are used as inputs, and node embedding is performed through the GIN model to obtain the representation vector of the drug molecule graph.

[0019] Each residue is treated as a node in the target structure graph, and the probability of contact between residue pairs is used as the adjacency matrix of the target structure graph. The position of each residue is scored based on the sequence alignment results and used as the attribute feature of the node in the target structure graph. The feature vectors of the target structure graph and its nodes are used as inputs and node embedding is performed through the GCN model to obtain the representation vector of the target structure graph.

[0020] Step S3: Contrastive learning and fusion of multimodal representations;

[0021] Feature representations are learned by maximizing the consistency between string modalities and graph modalities. After obtaining the final representations of the two modalities for drug and target respectively, they are concatenated to obtain drug and target modal information for DTA prediction.

[0022] Furthermore, in step S1, after integer encoding of the drug and target strings, the positional information of the string modality is captured by using the arrangement information of drug atoms and target residues, and different levels of abstract features are learned from the input through the Transformer model. Finally, a max pooling layer is applied to obtain the final vector representation of the drug and target strings.

[0023] Furthermore, in step S1, the position information of the string modality is represented by the following formula:

[0024] PE (pos,2i) =sin(pos / 10000) 2i / dmodel (1)

[0025] PE (pos,2i+1) =cos(pos / 10000) 2i / dmodel (2)

[0026] Where pos represents a character in the string, i is the dimension of the character encoding, and dmodel is the encoding of the character.

[0027] Furthermore, the Transformer model includes an MSA layer and an MLP block, and the function representation of the MSA layer is as follows:

[0028] z m =MSA(LN(z) l-1 ))+z l-1 ,l=1…L (3)

[0029] Among them, z l-1 z represents the input to the MSA layer. m LN represents the output of MSA, L represents the normalization layer, and L represents the batch length of the input.

[0030] The MLP block contains two CNN layers and one normalization layer, and its function is expressed as:

[0031] z l =MLP(LN(z) m ))+z m ,l=1…L (4)

[0032] Among them, z l This represents the output of the MLP.

[0033] Furthermore, the GIN model comprises five GIN layers, each followed by a batch normalization layer, and the last batch normalization layer is connected to a global max pooling layer. Each GIN layer uses a multilayer perceptron model to process node features x. i Updated to:

[0034]

[0035] Where k represents the GIN layer, ∈ is a learnable parameter or a fixed scalar, and N(i) is the set of neighbors of node i.

[0036] Furthermore, the GCN model comprises three GCN layers, each activated by the ReLU function, with a global max pooling layer following the last GCN layer. Each GCN layer performs one convolution operation, as shown in the following equation:

[0037]

[0038] in, It is the adjacency matrix of the target structure graph. σ is the degree matrix of the target structure graph, σ is the activation function, and W is the learnable weight matrix.

[0039] Preferably, in step S3, when learning feature representations by maximizing the consistency between string modalities and graph modalities, any drug or target itself is fixed as anchor point A, resulting in a set of positive samples P composed of multiple modalities of that drug or target itself. The representations of all modalities of other drugs or targets are regarded as negative samples N, generating each positive pair (A,P) and negative pair (A,N). Contrastive learning is used to force all modal representations of anchor point A to be consistent and distinguishable from all modal representations of negative samples. The calculation of Loss is as follows:

[0040]

[0041] Wherein, for each sample i, P(i) is its set of positive samples, |P(i)| is the number of positive samples, p is one of the positive samples, N(i) is the set of negative samples, n is one of the negative samples, and T is the temperature coefficient.

[0042] Compared to traditional models, the multimodal information fusion model and method provided in this invention have superior feature capture capabilities and can provide richer and more complete modal information for DTA prediction. Specifically, this invention fully considers string and graph modal information in drugs and targets. String modalities include drug smiles and target sequences, while graph modalities include drug molecule graphs and target structure graphs. Different deep learning models are selected to extract data features from different modalities, and contrastive learning methods are used to balance the feature representations of different modalities. Furthermore, the arrangement information of drug atoms and target residues is utilized to capture the positional information of string modalities. All of these methods can extract more useful feature information from smiles and target sequences, effectively improving DTA prediction performance. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the application of the multimodal information fusion model involved in this invention to DTA prediction;

[0044] Figure 2 This is a diagram of the contrastive learning framework in the multimodal information fusion method of this invention;

[0045] Figure 3 This is a comparison chart of the MSE and CI metrics of each model on the Davis dataset in the evaluation experiment of this invention;

[0046] Figure 4 This is a comparison chart of the MSE and CI metrics of each model on the KIBA dataset in the evaluation experiment of this invention. Detailed Implementation

[0047] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0048] Traditional drug-target affinity prediction methods based on biological experiments cannot meet the needs of drug R&D in the era of big data. Although deep learning-based drug-target affinity prediction models have achieved significant success, these models only consider single-modal features of drug and target information, resulting in incomplete and insufficient information. In fact, information from different modalities of drugs and targets can be complementary, and fusing information from different modalities can yield more valuable insights. Based on this, this invention designs a multimodal information fusion model called FMDTA for DTA prediction, as detailed below.

[0049] I. FMDTA

[0050] Since both string-based and graph-based data have their own advantages and disadvantages, this invention merges and fuses information from both modalities to obtain more complete information. FMDTA consists of four parts: a drug molecule structure information encoder, a target structure information encoder, a multimodal balancing module, and a drug-target fusion module. Among them:

[0051] The drug molecular structure information encoder uses the Transformer model to encode the modal information of the drug string and uses the GIN model to extract the modal information features of the drug graph.

[0052] The target structure information encoder uses the Transformer model to encode the modal information of the target string and uses the GCN model to extract the modal information features of the drug map.

[0053] The multimodal balancing module uses contrastive learning to balance and integrate drug string and graph modal information, as well as target string and graph modal information.

[0054] The drug-target fusion module connects the two modal features of the drug and the target obtained from the multimodal balancing module for DTA prediction.

[0055] II. The workflow of FMDTA is as follows Figure 1As shown, the multimodal information fusion method provided by this invention mainly includes three steps:

[0056] Step S1, Embedding of string modalities

[0057] S11, embedding of drug SMILES

[0058] Each drug SMILES code is treated as a string and encoded using integers, with integers used as characters to encode the input. It's worth noting that the model needs to be trained before practical application. To facilitate training with the FMDAT provided in this invention, each SMILES string should be cut or padded to a fixed length of 100 characters. These integer sequences are used as input to the embedding layer, which returns a 128-dimensional vector representation. A Transformer model is then used to learn different levels of abstract features from the input.

[0059] Considering that the order of the SMILES strings is crucial for the characteristic representation of drug molecules, this invention incorporates the positional information of SMILES, and uses the following formula to represent the positional encoding:

[0060] PE (pos,2i) =sin(pos / 10000) 2i / dmodel (1)

[0061] PE (pos,2i+1) =cos(pos / 10000) 2i / dmodel (2)

[0062] Where pos represents a character in the SMILE string, i is the dimension of the character encoding, and dmodel is the encoding of that character.

[0063] After obtaining the location encoding, the drug SMLIES encoding information is added to the location encoding to obtain the complete encoding information. Then, the Transformer model learns different levels of abstract features from the input.

[0064] For the Transformer model, this invention follows the original Transformer design, which consists of a Multi-Head Attention (MSA) layer and an MLP block. The function representation of the MSA layer is as follows:

[0065] z m =MSA(LN(z) l-1 ))+z l-1 ,l=1…L (3)

[0066] Among them, z l-1 z represents the input to the MSA layer. mThis represents the output of MSA, LN represents the normalization layer, and L represents the batch length of the input.

[0067] Following the MSA layer, the MLP block contains two CNN layers and one linear normalization layer, whose function is expressed as:

[0068] z l =MLP(LN(z) m ))+z m ,l=1…L (4)

[0069] Among them, z l This represents the output of the MLP.

[0070] Finally, a max pooling layer is applied to obtain the final vector representation of the drug string.

[0071] S12, Embedding of target sequence

[0072] Each target sequence is treated as a string, and the characters are encoded as integers and used as input. Similarly, for ease of training, each target sequence is pruned or padded to a fixed length of 1000 residues. These integer sequences are used as input to the embedding layer, which returns a 128-dimensional vector representation. Likewise, considering that the order of target residues is crucial for representing the structural features of the target, positional information is encoded, and then a Transformer model learns different levels of abstract features from the input. Finally, a max-pooling layer is applied to obtain the final vector representation of the target string.

[0073] Step S2, Embedding of graph modes

[0074] S21, Embedding of drug molecule map

[0075] In this invention, the drug molecule graph data comes from GraphDTA [from reference

[15] Jiang, M.;Li, Z.;Zhang, S.;Wang, S.;Wang, X.;Yuan, Q.;Wei, Z. Drug–target affinity prediction using graphneural network and contact maps. RSC Advances 2020, 10, 20701–20712.]. In this invention, each atom is used as a node in the drug molecule graph, and the connections between atoms are used as the adjacency matrix of the drug molecule graph. At the same time, the relevant attributes of the atoms are used as the attribute features of the nodes in the drug molecule graph. The feature vectors of the drug molecule graph and its nodes are used as inputs, and node embedding is performed through the GIN (Graph Isomorphic Networks) model.

[0076] Specifically, the GIN model consists of five GIN layers, each of which uses a multilayer perceptron (MLP) model to process node features x i Updated to:

[0077]

[0078] Where k represents the GIN layer, ∈ is a learnable parameter or a fixed scalar, and N(i) is the set of neighbors of node i.

[0079] In this GIN model, each GIN layer is followed by a batch normalization layer, and finally, a global max pooling layer is added to obtain the representation vector of the drug molecule graph.

[0080] S22, Embedding of target structure map

[0081] In this invention, the target structure map data comes from DGraphDTA [derived from the literature Jiang, M.; Li, Z.; Zhang, S.; Wang, S.; Wang, X.; Yuan, Q.; Wei, Z. Drug–targetaffinity prediction using graphneural network and contact maps. RSCA Advances 2020, 10, 20701–20712.]. This invention uses each residue as a node in the target structure map, and the probability of contact between residue pairs is used as the adjacency matrix of the target map, thus effectively preserving the spatial information of the protein. Simultaneously, the position of each residue is scored based on sequence alignment results and used as the feature vector of the residue node. Then, the target structure map and its node feature vectors are used as input, and node embedding is performed using a GCN model.

[0082] Specifically, the GCN model consists of three GCN layers, each of which performs a convolution operation, as shown below:

[0083]

[0084] in, It is the adjacency matrix of the target structure graph. σ is the degree matrix of the target structure graph, σ is the activation function, and W is the learnable weight matrix.

[0085] Each GCN layer in this GCN model is activated by the ReLU function, and a global max pooling layer is added after the last GCN layer to obtain the representation vector of the target structure graph.

[0086] S3, Contrastive Learning and Representation Fusion of Multimodal Representations

[0087] In the process of extracting features from data of different modalities, on the one hand, because the internal structures of different drugs or targets may contain information about multiple similar functional groups or residues, the features between different drugs / targets become blurred and similar after encoding. On the other hand, due to the significant differences in the embedding methods of different modalities, the feature representations of the same drug / target in the two modalities show significant differences, while the feature representations of different drugs / targets in the same modality show slight differences. Therefore, it is necessary to select an appropriate fusion method to balance the information embedding between different modalities, so that the information of multiple modalities can be complementary.

[0088] To balance the information embedding between different modalities, this invention employs a contrastive learning method to capture the interaction relationships between modal information. For example... Figure 2 As shown, in FMDTA, we learn feature representations by maximizing the consistency between string modalities and graph modalities. Specifically, for example, for drug i, we fix itself as anchor A, obtaining a set of positive samples P composed of its own multiple modalities. The representations of all modalities of other drugs are considered as negative samples N, generating each positive pair (A, P) and negative pair (A, N). Then, contrastive learning is used to force the consistency of all modal representations of anchor A and distinguish it from all modal representations of negative samples. In this contrastive learning process, the embedding representation of drugs (targets) is iteratively updated by calculating the loss, making the representations of the two modalities of the same drug (target) closer, and the representations of different drugs (targets) of the same modality and different drugs (targets) of different modalities more differentiated, in order to balance the embedding representation of drugs (targets) and obtain the final embedding representation. The calculation of the loss is as follows:

[0089]

[0090] Wherein, for each sample i, P(i) is its set of positive samples, |P(i)| is the number of positive samples, p is one of the positive samples, N(i) is the set of negative samples, n is one of the negative samples, and T is the temperature coefficient.

[0091] The same principle applies to comparative learning of embedded information from different modalities of the target.

[0092] After obtaining the final representations of the drug and target modalities, they are spliced ​​together to obtain drug and target modal information for DTA prediction. DTA can then be predicted through two fully connected layers.

[0093] Third, in order to verify the feasibility and superiority of the model and method involved in this invention, a verification experiment will be conducted next.

[0094] 3.1 Experimental Data

[0095] In this experimental evaluation, data from DeepDTA [derived from the literature] were used. H.; A.; Ozkirimli, E. DeepDTA: deep drug–target binding affinity prediction. Bioinformatics 2018, 34, i821–i829.] The benchmark datasets include Davis [from Davis, MI; Hunt, JP; Herrgard, S.; Ciceri, P.; Wodicka, LM; Pallares, G.; Hocker, M.; Treiber, DK; Zarrinkar, P. Comprehensive analysis of kinase inhibitor selectivity. Nature biotechnology 2011, 29, 1046–1051.] and KIBA [from Tang, J.; Szwajda, A.; Shakyawar, S.; Xu, T.; Hintsanen, P.; Wennerberg, K.; Aittokallio, T. Making sense of large-scale kinase inhibitor bioactivity data sets: a comparative and integrative analysis. Journal of Chemical Information] and Modeling 2014, 54, 735–743.] dataset.

[0096] The Davis dataset contains 72 compounds and 442 proteins, along with their corresponding affinity values, measured by the Kd value (kinase dissociation constant), ranging from 5.0 to 10.8. The average length of the SMILES strings for compounds is 64, and the average length of the target sequences is 788.

[0097] The KIBA dataset contains bioactivity of combination kinase inhibitors from various sources (e.g., Ki, Kd, ​​and IC50), including binding affinity for 2,116 drugs and 229 targets, measured by KIBA scores ranging from 0.0 to 17.2. The average length of the compound SMILES strings is 58, and the average length of the target sequences is 728. Data summary is shown in Table 1.

[0098] Table 1 Data Statistics

[0099]

[0100] 3.2 Evaluation Indicators

[0101] This experiment uses the mean squared error (MSE), the concordance index (CI), and... The index serves as an evaluation metric for model performance.

[0102] MSE is a commonly used evaluation metric for regression tasks. It measures the performance of a sample by calculating the average of the sum of squared differences between the predicted and actual values. i Predicted value and actual sample value y i The difference lies in the MSE. A smaller MSE indicates a better effect. The formula for calculating MSE is as follows:

[0103]

[0104] CI (Corrective Conformity) is a metric that evaluates the consistency between the ranking of predicted values ​​and the ranking of actual values. A higher CI value indicates better predictive performance. The formula for calculating CI is as follows:

[0105]

[0106] Among them, b x It has a relatively high affinity. x The predicted value, b y It has a lower affinity d y The predicted value is given by Z, which is the normalization constant; h(x) is the step function, as shown in the following formula:

[0107]

[0108] The index is used to evaluate the external predictive performance of the QSAR (Quantitative Structure-Activity Relationship) model. The formula is shown below:

[0109]

[0110] Where r 2 and These are the squared correlation coefficients with and without intercepts, respectively.

[0111] For training purposes, this experiment used a server with two Intel(R) Xeon(R) Gold 5215 2.5GHz CPUs, 256GB of RAM, and three NVIDIA Corporation GV100GL GPUs. The hyperparameters used in the experiment are shown in Table 2.

[0112] Table 2 Hyperparameters

[0113] Hyper-parameters Setting Learning rate 0.0005 Batch size 512 Optimizer Adam GIN layers 5 GCN layers 3 Transformer_embedding_dim 128 GNNS_embedding_dim 128

[0114] 3.3 Evaluation Strategy

[0115] This experiment uses 5-fold cross-validation to verify the model's performance. In each fold, the dataset is randomly divided into training, validation, and test sets in a 3:1:1 ratio. Finally, the model selected through validation is evaluated on the test set.

[0116] 3.4 Model Comparison

[0117] In this experiment, classic DTA prediction models were divided into two categories and their performance was compared. The methods that only considered the string modality were DeepDTA, WideDTA, and AttentionDTA. These models all used CNN models to encode the strings.

[0118] DeepDTA consists of two independent CNN modules, which are used to learn feature representations of drug SMILES and target sequence information, respectively, and input the representations into three fully connected layers for DTA prediction.

[0119] WideDTA: Based on DeepDTA, it introduces information on protein domains and motifs and maximum common substructure words to better predict DTA.

[0120] AttentionDTA: Introduces a bilateral multi-head attention mechanism to focus on key subsequences that are important for drug and protein sequences when predicting their affinity.

[0121] Methods that take graph modalities into account include GraphDTA, DGraphDTA, DeepGLSTM, and SAG-DTA, all of which use GNNs to encode graphs.

[0122] GraphDTA introduces a drug molecule graph, uses GNNs to learn the molecular graph features of the drug, uses CNN to learn the feature representation of the target sequence, and concatenates these two parts to predict the DTA through a fully connected layer.

[0123] DGraphDTA: Based on GraphDTA, a target structure graph is introduced. GNNs are used to learn the features of drugs and targets, and these two parts are concatenated and predicted through a fully connected layer.

[0124] SAG-DTA: Based on GraphDTA, it improves the representation of drug molecule graph features for DTA prediction through a self-attention mechanism.

[0125] DeepGLSTM: A method based on graph convolutional networks and LSTM to encode drugs and targets respectively to predict DTA.

[0126] Table 3 shows the results of the DTA prediction model on the DAVIS dataset.

[0127]

[0128]

[0129] Table 3 presents the performance of FMDTA and the baseline method on the Davis dataset. To maintain fairness in the comparison, the same hyperparameters and evaluation metrics were used for all models. It can be seen that FMDTA significantly outperforms methods that rely solely on string or graph modalities across all three metrics, with an MSE of 0.195, a CI of 0.909, and an R-value of [missing value]. 2 m is 0.748. Furthermore, graph-based methods generally outperform string-based methods because the crude truncation of string-based data to a fixed length results in significant information loss. Although graph-based data is more complete, features are often smoothed during graph embedding. Therefore, it is essential to fuse multimodal information from drugs and targets using appropriate methods to ensure their information complements each other.

[0130] To further test the generalization ability of the proposed method, this experiment evaluated the model on the KIBA dataset using the same hyperparameters as in the Davis dataset. The experimental results are shown in Table 4. It can be seen that FMDTA still performs excellently; specifically, FMDTA has an MSE of 0.133, a CI of 0.899, and an R-value of [missing information]. 2 m is 0.801. These results demonstrate the effectiveness and good generalization ability of the model of this invention in DTA prediction.

[0131] Table 4 Results of the DTA prediction model on the KIBA dataset

[0132]

[0133]

[0134] 3.5 Ablation Study

[0135] 3.5.1 Effectiveness Analysis of Multimodal Information Fusion

[0136] The model of this invention was compared with DeepDTA, GraphDTA, and DGraphDTA on the Davis dataset. To ensure fairness in the comparison, the same parameters were used for the same modules of the model of this invention and the comparison models. Here, FMDTA(w / oPC) represents a simple concatenation and fusion of the two modalities of the drug and the target, without including string position information encoding and contrastive learning.

[0137] pass Figure 3 It can be seen that fusing information from both modalities is superior to using only a single modal. Combining the structural diagrams of drugs and targets yields more complete structural and spatial information, overcoming the limitations of DeepDTA's reliance solely on the string modality. The string modality information of drug smiles and target sequences provides the permutation of strings and contextual semantic information, compensating for the limitations of DGraphDTA's reliance solely on structural diagram information. GraphDTA, utilizing both drug molecule structural diagrams and target sequence information, also relies on only a single modality, resulting in incomplete information. This demonstrates that the two modalities can complement each other, and fusing them can effectively improve the model's predictive ability on DTA. Furthermore, GraphDTA outperforms DeepDTA and DGraphDTA, indicating that the different modalities of drug and target protein information lead to better performance.

[0138] To further verify the effectiveness of multimodal information fusion for the DTA task, FMDTA (w / o PC) was compared with DeepDTA, GraphDTA and DGraphDTA on the KIBA dataset. Figure 4 The performance of FMDTA (without PC) and the unimodal model on the KIBA dataset is shown. It can be seen that the fusion of multimodal information also performs excellently on the KIBA dataset, indicating that the proposed method is reasonable and effective.

[0139] 3.5.2 Effectiveness Analysis of String Position Information Encoding and Multimodal Contrastive Learning

[0140] This section aims to demonstrate the importance of positional information encoding in string modalities and the significance of multimodal contrastive learning. As mentioned earlier, the arrangement of atoms in SMILES and the resulting order of target sequence residues are crucial for molecular feature representation. Furthermore, maximizing the consistency between string and graph modalities in learning feature representations is also essential. A comprehensive ablation study is deployed here to explore the necessity of each individual module, and comparisons are made on the Davis and KIBA datasets.

[0141] The three variants of FMDTA are:

[0142] (1) FMDTA (w / o PC): FMDTA without positional information of string modality encoding component and multimodal contrastive learning component, directly splices two separate modalities of drug and target protein.

[0143] (2) FMDTA (w / o C): FMDTA without a multimodal contrastive learning component extracts the positional information of string modalities but does not consider the interaction between modal features.

[0144] (3) FMDTA(w / o P): FMDTA without string modality position information encoding, which balances and fuses string modality and graph modality features through contrastive learning components.

[0145] Table 5. Ablation Study Results Regarding Location Information and Contrast Learning

[0146]

[0147] As shown in Table 5, on the Davis dataset, adding the string modality positional encoding module improved the MSE by 0.5% and the CI by 0.7%. On the KIBA dataset, the improvement in MSE was not significant, while the CI improved by 0.3%. Adding the multimodal contrastive representation learning module on the Davis dataset significantly improved MSE and CI by 2% and 1%, respectively. On the KIBA dataset, MSE and CI significantly improved by 0.4% and 0.7%, respectively. The addition of the positional encoding module and the contrastive learning module led to more significant improvements, with MSE improving by 2.8% and 1.5% and CI by 0.7% and 1.2% on the Davis and KIBA datasets, respectively. Therefore, we can draw the following conclusions:

[0148] (1) Adding string-encoded location information can capture more complete information and improve model performance.

[0149] (2) Contrastive learning allows information from different modalities to interact during encoding, resulting in more balanced representations of features from different modalities and more flexible models.

[0150] The above embodiments are preferred implementations of the present invention. In addition, the present invention can be implemented in other ways. Any obvious substitutions without departing from the concept of the present technical solution are within the protection scope of the present invention.

[0151] To facilitate understanding by those skilled in the art of the improvements of this invention over the prior art, some of the accompanying drawings and descriptions have been simplified, and for clarity, some other elements have been omitted from this application. Those skilled in the art should realize that these omitted elements may also constitute the content of this invention.

Claims

1. A multimodal information fusion method for DTA prediction, characterized in that, include: Step S1, embedding of string modalities; The drug SMILES code is treated as a string, encoded with integers, and the positional encoding incorporated into this encoding is used to obtain a vector representation. The Transformer model is then used to extract features from this vector to obtain the final vector representation of the SMILES string. The target sequence is treated as a string, encoded with integers, and the positional encoding incorporated into this encoding is used to obtain a vector representation. The Transformer model is then used to extract features from this vector to obtain the final vector representation of the target string. Step S2, embedding of graph modes; Each atom is treated as a node in the drug molecule graph, the relationships between atoms are treated as the adjacency matrix of the drug molecule graph, and the properties of atoms are treated as the attribute features of the nodes in the drug molecule graph. The feature vectors of the drug molecule graph and its nodes are used as inputs, and node embedding is performed through the GIN model to obtain the representation vector of the drug molecule graph. Each residue is treated as a node in the target structure diagram, the probability of whether residue pairs are in contact is used as the adjacency matrix of the target structure diagram, and the position of each residue is scored based on the sequence alignment results, which is used as the attribute feature of the node in the target structure diagram. The target structure graph and its node feature vectors are used as input, and node embedding is performed through the GCN model to obtain the representation vector of the target structure graph. Step S3: Contrastive learning and fusion of multimodal representations; Feature representations are learned by maximizing the consistency between string modalities and graph modalities. After obtaining the final representations of the two modalities for drug and target respectively, they are concatenated to obtain drug and target modal information for DTA prediction. When learning feature representations by maximizing the consistency between string modalities and graph modalities, an arbitrary drug or target is fixed as anchor point A, resulting in a set of positive samples P composed of multiple modalities of that drug or target. The representations of all modalities of other drugs or targets are considered as negative samples N, generating each positive pair (A,P) and negative pair (A,N). Contrastive learning is used to force all modal representations of anchor point A to be consistent and distinguishable from all modal representations of negative samples. The loss is calculated as follows: Wherein, for each sample i, P(i) is its set of positive samples, |P(i)| is the number of positive samples, p is one of the positive samples, N(i) is the set of negative samples, n is one of the negative samples, and T is the temperature coefficient.

2. The multimodal information fusion method for DTA prediction according to claim 1, characterized in that: In step S1, after integer encoding of the drug and target strings, the positional information of the string modality is captured by the arrangement information of drug atoms and target residues. The Transformer model learns different levels of abstract features from the input, and then the max pooling layer is applied to obtain the final vector representation of the drug and target strings.

3. The multimodal information fusion method for DTA prediction according to claim 2, characterized in that: In step S1, the position information of the string modality is represented by the following formula: ON (pos,2i) =sin(pos / 10000 2i / dmodel) (1) ON (pos,2i+1) =cos(pos / 10000 2i / dmodel ) (2) Where pos represents a character in the string, i is the dimension of the character encoding, and dmodel is the encoding of the character.

4. The multimodal information fusion method for DTA prediction according to claim 3, characterized in that: The Transformer model includes an MSA layer and an MLP block. The function representation of the MSA layer is as follows: z m =MSA(LN(z l-1 ))+z l-1 ,l=1…L (3) Among them, z l-1 z represents the input to the MSA layer. m This represents the output of MSA, LN represents the normalization layer, and L represents the batch length of the input. The MLP block contains two CNN layers and one normalization layer, and its function is expressed as: With l =MLP(LN(z m ))+z m ,l=1…L (4) Among them, z l This represents the output of the MLP.

5. The multimodal information fusion method for DTA prediction according to claim 4, characterized in that: The GIN model comprises five GIN layers, each followed by a batch normalization layer. The final batch normalization layer is connected to a global max pooling layer. Each GIN layer uses a multilayer perceptron model to process node features x. i Updated to: Where k represents the GIN layer, ∈ is a learnable parameter or a fixed scalar, and N(i) is the set of neighbors of node i.

6. The multimodal information fusion method for DTA prediction according to claim 5, characterized in that: The GCN model consists of three GCN layers, each activated by the ReLU function. A global max pooling layer is connected after the last GCN layer. Each GCN layer performs one convolution operation, as shown in the following equation: in, It is the adjacency matrix of the target structure graph. σ is the degree matrix of the target structure graph, σ is the activation function, and W is the learnable weight matrix.

Images