KR20200017653A - Method for prediction of drug-target interactions - Google Patents

Abstract

According to the present invention, a method for predicting a protein-target interaction may comprise the steps of: learning a protein sequence using a convolutional neural network in a processor to extract a local residual pattern; learning a drug fingerprint through a first fully connected layer in the processor to extract a drug fingerprint pattern; concatenating the local residual pattern and the drug fingerprint pattern in the processor; and learning the concatenated pattern through a second fully connected layer in the processor.

Description

Method for predicting drug-target interactions {METHOD FOR PREDICTION OF DRUG-TARGET INTERACTIONS}

The present invention relates to a method for predicting interaction of drug-target proteins.

In the early stages of drug discovery, the identification of drug-target interactions (DTIs) plays an important role because drugs inhibit or activate the biological activity of the target through physical binding. Thus, drug developers screen compounds that interact with specific targets with the biological activity of interest. However, the identification of DTI in large-scale chemical or biological experiments typically requires two to three years of testing time and is expensive. Therefore, various in silico methods are being developed to predict possible DTIs to aid drug discovery due to the accumulation of drug, target and interaction data.

Many similarity-based methods were first studied. The drug is assumed to bind to proteins similar to known targets and vice versa. One of the best established methods is to combine chemical and genomic spaces with pharmacological spaces. Yamanashi et al. Use a kernel regression method that uses information about drug interactions known as inputs to identify DTI. To overcome the enormous computational power requirements, Beakley et al. Developed a partial model that trains the interaction model locally, but not as a whole. In addition to significantly reducing computational complexity, this model outperformed the previous model. Similarity-based methods, however, work well for DTI within certain protein families, but not in other classes. Similarity-based methods are currently not commonly used to predict DTI. Thus, there is a need for a method of predicting DTI regardless of the similarity between the protein class and the subject or drug. This method requires considerable computational power.

Japanese Patent Application Laid-Open No. 2017-520868, published on July 27, 2017, title: Binding affinity prediction system and method. Korean Patent Laid-Open Publication No. 10-2018-0017827, published on February 21, 2018, title: A method and system for predicting RNA sequence region binding to a protein using a base profile and a composition. Korean Patent Laid-Open Publication No. 10-2018-0052959, published on May 21, 2018, title: Method and apparatus for predicting binding affinity between MHC and peptide.

It is an object of the present invention to provide a method for predicting novel protein-target interactions.

According to an embodiment of the present invention, a method for predicting protein-target interaction includes: learning a protein sequence using a convolution neural network in a processor to extract a local residual pattern; Learning a drug fingerprint through a first fully connected layer in the processor to extract a drug fingerprint pattern; Concatenation of the local residual pattern and the drug fingerprint pattern in the processor; And learning, by the processor, the concatenated pattern through a second fully connected layer.

In an embodiment, the learning of the protein sequence may include: learning a local residue pattern for a plurality of protein sequences through a convolution layer; And extracting a maximum value as the local residual pattern from a result of the learned local sequence pattern through a max pooling layer.

In an embodiment, the max pooling layer is a global max pooling layer.

In an embodiment, the plurality of protein sequences are trained from at least one database of DrugBank, International Union of Basic and Clinical Pharmacology (IUPHAR), and Kyoto Encyclopedia of Genes and Genomes (KEGG).

In an embodiment, the size of the convolutional layer may be a value obtained by subtracting the window size (WS) from the predetermined protein sequence size (MPL) plus 1.

In an embodiment, the learning of the protein sequence may further include setting the size of the insertion vector for the plurality of protein sequences to the predetermined size of the protein sequence (MPL) through an embedding layer. have.

In an embodiment, the processor may further include generating a score corresponding to the protein-target interaction through the output layer.

The generating of the score may include: setting a weight on each of the local residual pattern and the drug fingerprint pattern; And generating the score by activating an activation function that takes the set local residual pattern and the set drug fingerprint pattern as input.

In an embodiment, the activation function is characterized in that the sigmoid function.

In an embodiment, the method may further include adjusting a learning speed or a window size during cross validation.

In the method of predicting protein-target interaction according to an embodiment of the present invention, a local residual pattern is extracted using CNN, and a corresponding drug fingerprint pattern is learned and concatenated to efficiently predict the result of protein-target interaction. Can be.

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings are provided to assist in understanding the present embodiment, and provide embodiments with a detailed description. However, the technical features of the present embodiment are not limited to the specific drawings, and the features disclosed in the drawings may be combined with each other to constitute a new embodiment.
1 is a diagram conceptually illustrating a deep learning apparatus for predicting drug-target interaction according to an embodiment of the present invention.
2 is a diagram illustrating a deep learning model for predicting drug-target interaction according to an embodiment of the present invention.
3 exemplarily shows the entire schema for extracting local residual patterns from the entire protein sequence.
4 illustratively shows a precision-recall curve for an optimization model of a protein descriptor.
FIG. 5 illustratively shows a performance comparison of an independent data set of PubChem data sets.
FIG. 6 illustratively shows a performance comparison of an independent data set of PubChem data sets.
7 is an exemplary comparison of performance between the convolution model of the present invention and the model of Wen.
FIG. 8A shows the binding site of the mast / stem cell growth factor receptor kit protein, and FIG. 8B is an illustration showing the binding site of the 5-hydroxytryptamine receptor 1B protein.
9 is a diagram illustrating a method for predicting drug-target interaction according to an embodiment of the present invention.

DETAILED DESCRIPTION Hereinafter, the contents of the present invention will be described clearly and in detail so that those skilled in the art can easily carry out the drawings.

As the inventive concept allows for various changes and numerous embodiments, particular embodiments will be illustrated in the drawings and described in detail in the text. However, this is not intended to limit the present invention to a specific disclosed form, it should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention. Terms such as first and second may be used to describe various components, but the components should not be limited by the terms.

1 is a diagram conceptually illustrating a deep learning apparatus 100 for predicting drug-target interaction (DTI) according to an embodiment of the present invention. Referring to FIG. 1, the deep learning apparatus 100 includes a CNN unit 110, a first FC unit 120, a concatenation unit 130, a second FC unit 140, and a DTI prediction unit 150. can do.

The CNN unit 110 may be implemented to extract a local residue pattern of a protein by receiving the protein sequence and learning according to a convolution layer and a pooling layer. The convolutional layer may generate a plurality of feature maps by performing a convolution operation using a plurality of filters (or kernels) on the protein sequence. The pooling layer may extract a specific value in a predetermined area in each of the feature maps to simplify the output of the convolutional layer. For example, the pooling layer may extract a specific value in various ways such as a maximum value, a minimum value, and an average value in the unit region. In the following description, it is assumed that the pooling layer of the present invention is a max pooling layer for extracting a maximum value. In summary, the CNN unit 110 may extract a local residual pattern using a filter.

The first FC unit 120 may be implemented to receive the drug fingerprint and learn through a fully connected layer. The first FC unit 120 may organize and abstract existing feature information by using a weight vector and an activation function. It can be expressed as the formula O (result layer) = activation_function (W (weight)) x H (input layer) + b (bias), so that the value of each dimension of the result layer is multiplied by the weight The result is a nonlinearity derived by using the activation function.The result of this result layer is that each dimension of the result layer is organized by weighting each dimension of the input layer. In the case of drug fully-connected layers, each dimension indicates whether or not a particular substructure exists in the compound. The combination of specific infrastructures derived from the use of weights, ie learning through a fully connected layer, is a specific infrastructure that is important for drug-target prediction. Learning means that the action in the model.

The concatenation unit 130 may be implemented to concatenate the local residual pattern output from the CNN unit 110 and the drug fingerprint pattern output from the first FC unit 120.

The second FC unit 140 may be implemented to learn through a fully connected layer about the pattern output from the connection unit 130. Here, significant combinations can be learned from the purified drug and protein features in order to predict the interaction of the drug with the target protein. The second FC unit 140 may be driven in a similar manner as the first FC unit 120.

The DTI prediction unit 150 may be implemented to generate a score corresponding to the drug-target interaction (DTI) for the output of the second FC unit 140. The final drug-target interaction can be predicted through the weights for the organized combinations of the second FC unit 140. For example, an organized combination may use a Sigmoid function as an activation function so that values from 0 to 1 can be scored. Classification may be made according to this score value. Such scoring may output a value that predicts the final DTI through weights to the organized combinations in the second fully connected layer.

In general, the identification of drug-target interactions (DTIs) plays an important role in drug discovery. Drugs perform specific functions by interacting with target proteins in different ways. Due to the high cost and labor intensive nature of in vitro and in vivo experiments, the importance of silico-based DTI prediction approaches is emphasized. Therefore, the deep learning apparatus 100 according to an embodiment of the present invention may use a convolutional neural network (CNN) for protein raw sequence learning to capture local residual patterns participating in DTI. In addition, the deep learning apparatus 100 according to an embodiment of the present invention may detect the binding region of the protein to the DTI by examining the pooled convolution results. In conclusion, the deep learning apparatus 100 of the present invention is a predictive model for detecting a local residual pattern of a target protein. The deep learning apparatus 100 may well display protein features of an unprocessed protein sequence and may yield better prediction results than the previous approach. .

In the case of a general feature-based DTI prediction model, the drug fingerprint is the most commonly used descriptor of the drug substructure. The drug fingerprint can be transformed into a binary vector with an indicator value indicating the presence of the substructure of the drug. In the case of proteins, CTD (composition, transition and distribution) descriptors can be used as a general computational expression. Various protein and chemical descriptors are introduced, but feature-based models do not show good predictive performance. In a typical machine learning model, features must be read from the original primitive forms such as SMILES and amino acid sequences by modeling, where local residue pattern information is lost during conversion, and local residue pattern information is lost. It is also difficult to recover.

The deep learning apparatus 100 according to an embodiment of the present invention may extract the local residual pattern for the DTI without biasing the maximum length without losing the local residual pattern information. The deep learning apparatus 100 according to the embodiment of the present invention may predict large-capacity DTI by using raw protein sequences for various protein lengths as well as various target protein classes.

2 is a diagram illustrating a deep learning model for predicting drug-target interaction according to an embodiment of the present invention. Referring to FIG. 2, the deep learning model for predicting drug-target interactions employs a convolution filter over the entire sequence of the protein to extract key protein local residue patterns that participate in DTI. By pooling the maximum results of convolution, it can be determined how a given protein sequence matches the local residual pattern of participating in DTI. Deep learning models using this data as input variables in higher layers can construct abstracted and organized features for proteins.

Finally, the deep learning model of the present invention may concatenate drug features (or drug fingerprint patterns) to protein features (or protein sequence patterns). The drug features may be calculated by learning from drug fingerprints through a fully connected layer (first fully connected layer). And the concatenated features can predict the likelihood of DTIs through a higher fully connected layer (second fully connected layer).

The deep learning model according to an embodiment of the present invention may be trained with integrated large DTI information from various DTI databases. Here, the DTI databases may include DrugBank, International Union of Basic and Clinical Pharmacology (IUPHAR), Kyoto Encyclopedia of Genes and Genomes (KEGG), and the like. The deep learning model of the present invention can be optimized from the negative interactions predicted from MATADOR and Liu et al. By using optimized models, DTI can be predicted from biological assays such as PubChem BioAssays and KinaseSARfari for model performance evaluation.

On the other hand, a known DTI can be obtained from three databases, DrugBank, KEGG, and IUPHAR, to build a training data set. And duplicate DTI of three databases can be eliminated. As a result, a total of 12,859 compounds, 5163 proteins, and 35,145 DTIs can be obtained. All collected DTIs can be considered positive samples for training. The negative sample may consist of a positive sample and an exclusive random DTI selected by twice the positive sample. Also, since voice samples are randomly selected, ten voice data sets can be generated.

The deep learning model according to an embodiment of the present invention may build two independent test data sets from the PubChem BioAssay database and ChEMBL KinaseSARfari for evaluation. This data set may consist of the results of experimental analysis. To obtain positive DTI from PubChem, "active" DTIs can be collected from assays with dissociation constants (Kd <10 μm). Since predicting whether a drug binds to a protein, evaluation of dissociation constants (Kd) among the various types of assays (IC50, EC50, Kd, Ki, AC50) is the most appropriate to obtain a positive sample.

For negative samples, samples annotated as "inactive" in other assay types may be used. However, too many negative samples can be collected in the PubChem bioassay. First, only negative samples with drug and targets included in the positive samples of PubChem bioassays are collected. Second, by adding random negative samples containing drug or protein in the positive sample, the number of negative samples will be equal to the positive sample. As a result, 12,906 positive and negative samples can be composed of 14,343 drugs and 714 proteins.

Samples can also be collected from KinaseSARfari. KinaseSARfari may consist of an assay comprising a compound that binds to a kinase domain. Each assay result with a dissociation constant (Kd <10 μm) that is small enough to be considered positive to obtain a positive sample from KinaseSARfari can be considered positive. In contrast to PubChem bioassays, the number of negative samples may be similar to the number of positive samples of KinaseSARfari. Therefore, voice samples are not sampled. The 3835 positive samples and 5520 negative samples may consist of 3379 compounds and 389 proteins.

On the other hand, the deep learning model according to the embodiment of the present invention can use the raw protein sequence as the input of the protein. For drugs, ECFP drug fingerprints can be used that graphically analyze molecules and search for substructures of the molecular structure in subgraphs of the overall molecular graph. In particular, using RDKit, ECFP fingerprints with a radius of 2 can be generated from a raw SMILES string. Finally, each drug can be reproduced as a binary vector of length 2048. Such indicators may indicate the presence of a particular infrastructure.

Meanwhile, the deep learning model according to the embodiment of the present invention extracts local residue patterns from the entire protein sequence through CNN, and calculates potential expression of drug fingerprints through fully connected layers. can do. After processing both the drug and protein layers, the result can be produced in the fully connected layer by concatenating the processed layers. All layers except the output layer can be activated with an exponential linear unit (ELU) function as in the following equation.

The output layer can be activated with sigmoid functions for classification. The entire neural network model can be implemented in Keras.

On the other hand, one of the difficulties in explaining the protein characteristics of the machine learning model and the deep learning model is that the protein lengths are all different. Another difficulty is that only certain parts of the protein, such as specific domains or motifs, are involved in the DTI, not the entire protein structure. As a result, the physicochemical properties of the entire protein sequence are not suitable for proactively describing the DTI due to noise information from portions of the sequence that are not involved in the DTI. Therefore, extraction of local residual patterns related to DTI requires accurate prediction. Convolutional neural networks can capture important local residual patterns throughout the space.

3 exemplarily shows the entire schema for extracting local residual patterns from the entire protein sequence. Referring to FIG. 3, to enable convolution of the protein sequence, the amino acid residue pattern of the protein sequence can be modified from the amino acid label to an insertion vector having an insertion size of 20 (ES). During this modification, the length of the insertion vector for all proteins can be set to the maximum protein length (MPL), or 2500. The margin is filled with the null label ($) and its include vector. This is a meaningless convolution result. As a result, an embedding layer of 20 × 2500 can be configured for protein function. Convolution may be performed on protein sequences in a 1D manner using striding 1 with convolution from J th to j + ws th amino acids. This may be defined by the following equation.

Convolutional operations on the entire protein sequence can produce a convolutional layer of (MPL-WS + 1) size for each convolution filter. Where WS is the window size. Finally, global max-pooling can be performed for each filter to extract the most significant local residual pattern. This may be defined by the following equation.

Where j includes all convolution results for protein Pk. For each window a vector of filter magnitudes with the maximum convolution results can be calculated. No bias is derived from the local residual pattern and the location of the maximum protein length. Finally, by concatenating all the max-pooling results, a latent expression of the protein can be constructed. This indicates how important the local residual pattern for the interaction is to the sequence. For more organization and abstraction of protein features, linked maximum pooling results can be provided to the full connectivity layer.

As mentioned above, latent representations of drug fingerprint descriptors are more useful for predicting DTI. After the protein and drug features have been purified with neural networks, a complete connectivity layer can be implemented to concatenate these drug features and obtain final output to determine if the drug and target interact.

Using the configured deep neural model, the input can be delivered to the output layer in a feed-forward manner. The deep neural model can calculate losses with binary crossover entropy as follows.

The loss function can be applied to L2-norm to prevent overfitting.

Finally, by using the Adam optimizer to update weights, generalized predictions for deep learning models can be provided.

In addition, a batch normalization technique may be used to prevent overfitting.

The deep learning model according to an embodiment of the present invention may adjust hyperparameters such as learning speed and window size affecting performance during cross-validation. An external validation data set can be constructed to select the most appropriate hyperparameters. A positive DTI is collected from the MATADOR database, which is exclusive to the training data set. A negative DTI with a high negative score (> 0.93) of Liu et al's data can be obtained to make a reliable speech data set. As a result, 400 positive DTIs and 404 negative DTIs can be generated as the external validation set. After the external verification data set is built, a grid retrieval method can be used to identify optimal hyperparameters that present the area under precision-recall (AUPR). When AUPR is measured, the optimal threshold value can be given as EER (Equal Error Rate).

Where θ is the classification threshold and α is a constant that determines the ratio of costs for accuracy and false classification in recall.

On the other hand, a deep neural model based on independent test data sets after defining classification thresholds by determining sensitivity (Sen), specificity (Spe), precision (Pre), accuracy (Acc), and F1 measurement (F1) The predictive performance of can be measured. As a general step of hyperparameter setting, the learning rate of weight update may first be adjusted to 0.0001. After the learning rate is fixed, the window size, number, and hidden layer of drug features can be benchmarked using AUPR.

4 illustratively shows a precision-recall curve for an optimization model of a protein descriptor. As shown in FIG. 4, finally, an optimized hyperparameter variable of the model can be selected to obtain AUPR 0.817 for the validation data set. As shown in FIG. 2, the fully optimized model is visualized graphically. In the same way, models using different protein descriptors can be built and optimized.

On the other hand, after the hyperparameters are adjusted, the performance for different protein descriptors, CTD descriptors (typically commonly used in chemical-genomic models), normalized SW scores, and the convolution methods of the present invention can be compared.

FIG. 5 illustratively shows a performance comparison of an independent data set of PubChem data sets. Referring to FIG. 5, the results show that the convolutional model of the present invention has better performance than other protein descriptors for all data sets. When the threshold value is selected by EER, the convolution model of the present invention can be generally applied by performing the same on both the PubChem and KinaseSARfari data sets. In contrast, the selected threshold does not work in the evaluation of an independent test data set, since the similarity descriptors show an imbalance between the metrics, thus losing generalization. Similarity descriptors also provide an AUPR of about 0.65 in the validation phase, but an accuracy of about 0.52 in the PubChem data set. This is close to random when evaluating independent data sets.

The SW score function only works if the protein class is limited and specified. Thus, SW scores are not suitable for large DTI predictions. CTD descriptors perform better than similar preferences but are poorer than the convolution models of the present invention. However, as shown in FIG. 6 in the KinaseSARfari data set, the performance of the CTD descriptor is significantly reduced, especially in the F1 score. Because the KinaseSARfari data set provides compound domain biological assays, CTD models trained with the entire protein sequence cannot predict compound-domain interactions. Thus, the F1 score and sensitivity change are greatly reduced. In contrast to CTD, the convolution model of the present invention allows the model to reliably predict compound-domain interactions by learning local residual patterns during training.

7 is an exemplary comparison of performance between the convolution model of the present invention and the model of Wen. In addition to comparing the convolution and other protein descriptors of the convolution model of the present invention, the convolution model is compared with the previous model. The previous model chosen for comparison is by Wen et al. Other studies find it difficult to compare performance because the purpose and data set are different from those of our model. DTI is presented using optimized models and descriptors. After the pre-training and fine tuning steps, the PubChem data set is evaluated and the results are compared with the model. In comparison, the model of the convolution of the present invention performs better than the previous model, as shown in FIG. The previous model is written in a DBN, which is an RBM stack. However, RBM is now considered outdated and is only used as a weight initialization technique in the deep learning method.

On the other hand, the detection of the binding region by the CNN can be made sequentially. Because the maximum results are collected for each filter for each window, the pooled convolution results can highlight areas that match the local residual pattern. The pooled value is not directly related to the DTI prediction result because it passes through the higher fully connected layer, but a larger value may affect the prediction result by including the score of the matching local residual pattern. Therefore, if the convolution model of the present invention can capture the local residual pattern, it can give high value to the actual coupling area.

As a result of examining the middle layer of the complex result of the prediction, it shows that the convolution model of the present invention can be used as a feature to predict the DTI by collecting the local residual pattern. The sc-PDB database provides atom descriptions of proteins, ligands and binding sites of complex structure. By parsing binding site annotations, the binding site between the protein domain and the physiological ligand can be queried.

DTI predictions are investigated with high scores in the PubChem independent data set. Interestingly, convolutional results provide high value when binding regions are included. For example, mast / stem cell growth factor receptor kits (P10721, KIT_HUMAN) have many sc-PDB annotations such as physiological ligands 3GOF, 4UOI, 4HVS, 3GOE and 1PKG as shown in FIG. 8A. Have In FIG. 8A, the tinned bond residues are colored red for each sc-PDB tin. The various sc-PDB comments are slightly different but have common areas. As expected, some pooled results of the convolutional layer accurately cover the binding region with a high ranking between the filters affecting the high prediction score.

In addition, these convolution results are not limited to protein classes. Two sc-PDB data (4IAQ, 4IAR) can be found for the protein 5-hydroxytryptamine (serotonin) receptor 1B (P28222, 5HT1B_HUMAN), a G-protein receptor (GPCR). From the above data, the four binding regions forming the alpha-helical structure can be covered by the pooled convolution results as shown in FIG. 8B.

The method for predicting protein-target interaction according to an embodiment of the present invention proposes a new DTI prediction model for extracting local residual patterns of entire target protein sequences with CNN. The model of the present invention can train the model with DTI of various drug databases and optimize the model with DTI of MATADOR. As a result, the detected local characteristics of the protein sequence are superior to other protein descriptors such as CTD and SW scores. In addition, the model of the present invention has better performance than the previous model built on the DBN. The model also has the ability to detect binding regions in protein sequences by reviewing the pooled convolution results and comparing them with the annotations of sc-PDB. Finally, our approach to collecting local residual patterns with CNNs provides better character than the previous approach by enlarging the protein features of the native sequence.

9 is a diagram illustrating a method for predicting drug-target interaction according to an embodiment of the present invention. 1 to 9, the method for predicting drug-target interaction may proceed as follows.

A protein sequence pattern is learned through a convolution neural network (CNN), and as a result, a local residual pattern can be captured (S110). Local residual patterns from protein patterns can be learned via CNN. The CNN may consist of a convolution layer and a global max polling layer. The convolutional layer can be formed by performing a convolution operation on the protein pattern. The global max pooling layer can be formed by integrating the output values of the convolutional layer using the maximum value. The drug fingerprint internal pattern may be learned (S120). For example, a drug fingerprint pattern may be learned for a drug fingerprint through a fully connected layer (FC); or an inner product operation. In an embodiment, the drug fingerprint is represented by a binary vector of 2048 length (ECFP4), which can be learned through a fully connected layer (FC).

Thereafter, the CNN learned protein sequence pattern and the FC learned drug fingerprint pattern may be concatenated (S130). Simply attach the learned protein local sequence pattern results and drug fingerprint pattern results. For example, the concatenation process may be expressed as follows. concatenating ([0,1], [2,3])-> [0,1,2,3]. Then, the concatenated pattern may be learned through the complete connection layer (S140). At this time, interactions with each other may be learned by using the learned drug pattern and the protein sequence pattern. Drug interactions can then be predicted (S150). The final score can be derived by weighting the protein and drug patterns organized through the complete link layer and activating them with an activation function. Since all features are weighted at this time, it is called fully connected. For example, if the layer is H and the weight matrix is W: sigmoid (H * W) = sigmoid ([0.9, 0.1, 0.5, 0.11] * [0.2, 0.5, 0.8, 0.1]) = sigmoid (0.641 ) = Score can be expressed.

The steps and / or actions according to the invention may occur simultaneously in different embodiments in different order, in parallel, or for other epochs, etc., as would be understood by one skilled in the art. Can be.

In some embodiments, some or all of the steps and / or actions may be directed to instructions, programs, interactive data structures, clients, and / or servers stored on one or more non-transitory computer-readable media. At least some may be implemented or performed using one or more processors. One or more non-transitory computer-readable media may be illustratively software, firmware, hardware, and / or any combination thereof. In addition, the functionality of the "module" discussed herein may be implemented in software, firmware, hardware, and / or any combination thereof.

One or more non-transitory computer-readable media and / or means for implementing / performing one or more operations / steps / modules of embodiments of the present invention may be used in application-specific integrated circuits (ASICs), standard integrated circuits, A controller that performs appropriate instructions, including a microcontroller, and / or an embedded controller, field-programmable gate arrays (FPGAs), complex programmable logic devices (CPLDs), and the like. Does not.

On the other hand, the contents of the present invention described above are only specific embodiments for carrying out the invention. The present invention will include not only specific and practically usable means per se, but also technical ideas as abstract and conceptual ideas that can be utilized in future technologies.

CNN: Convolutional Neural Network
FC: fully connected neural network
100: deep learning device
110: CCN unit
120: first FC unit
130: connection unit
140: second FC unit
150: DTI prediction unit

Claims (10)

In a method for predicting protein-target interactions:
Learning a protein sequence using a convolution neural network in a processor to extract a local residual pattern;
Learning a drug fingerprint through a first fully connected layer in the processor to extract a drug fingerprint pattern;
Concatenation of the local residual pattern and the drug fingerprint pattern in the processor; And
Learning at the processor the concatenated pattern through a second fully connected layer. The method of claim 1,
Learning the protein sequence,
Learning a local residue pattern for the plurality of protein sequences through a convolution layer; And
Extracting a maximum value from the result of the learned local sequence pattern as the local residual pattern through a max pooling layer. The method of claim 2,
And the max pooling layer is a global max pooling layer. The method of claim 2,
Wherein said plurality of protein sequences are trained from a database incorporating DrugBank, International Union of Basic and Clinical Pharmacology (IUPHAR), and Kyoto Encyclopedia of Genes and Genomes (KEGG). The method of claim 2,
Wherein the size of the convolutional layer is a value obtained by subtracting the window size (WS) from the predetermined protein sequence size (MPL) plus one. The method of claim 5, wherein
Learning the protein sequence,
And setting the size of the insertion vector for the plurality of protein sequences through an embedding layer to the size of the predetermined protein sequence (MPL). The method of claim 1,
Generating a score corresponding to a protein-target interaction via an output layer in the processor. The method of claim 7, wherein
Generating the score,
Setting weights to each of the local residual pattern and the drug fingerprint pattern; And
Generating the score by activating an activation function that takes the set local residual pattern and the set drug fingerprint pattern as input. The method of claim 8,
The activation function is a sigmoid function. The method of claim 1,
Adjusting the learning rate or window size during cross validation.

Images