KR102576033B1 - Protein-ligand binding affinity prediction using ensemble of 3d convolutional neural network and system therefor - Google Patents

Abstract

Accurately predicting the binding affinity of protein-ligand complexes is essential for the efficiency and success of rational drug design. Examples provide a new neural network model for predicting the binding affinity of protein-ligand complexes. The new model provided by the embodiment can predict the binding affinity of the complex by ensembleing multiple independently learned networks composed of multiple channels of 3D convolutional neural network layers.

Description

Protein-ligand binding affinity prediction method using an ensemble of 3D-convolutional neural networks and a system for the same {PROTEIN-LIGAND BINDING AFFINITY PREDICTION USING ENSEMBLE OF 3D CONVOLUTIONAL NEURAL NETWORK AND SYSTEM THEREFOR}

The present invention provides a method for accurately predicting the binding affinity of a protein-ligand complex.

Predicting the binding affinity of protein-ligand complexes plays a very important role in drug design and drug discovery. In general, strong binding to the target protein is required to become a lead molecule for drug discovery. However, experimental measurement of protein-ligand binding affinity is a difficult and time-consuming task, and is one of the major obstacles in the drug discovery process. If the affinity between a target protein and a specific ligand can be quickly and accurately predicted, the efficiency of virtual screening can be significantly improved. Therefore, various computational methods for predicting binding affinity are being studied to optimize the drug discovery process.

Generally, computational methods for predicting binding affinity can be divided into three categories. These methods are 1) physics based, 2) experimental data based, and 3) structural data based (knowledge based) methods. The first method is a coupled degree-of-freedom calculation that is faithful to theory, mainly based on the forcefield model. The main advantage of physics-based methods is that, using state-of-the-art force field models, the binding free energy of arbitrary small molecule ligands to proteins can be predicted with an average error of approximately 1 kcal/mol. However, rigorous free energy calculations require a significant amount of computational resources. In state-of-the-art molecular dynamics (MD) code implementations via graphics processing units, the binding free energies of one or two ligands can be calculated depending on the size of the ligands and the protein. This computational burden is an obstacle to the use of molecular dynamics (MD) and also to free energy calculations of molecules such as drugs, which require high-throughput screening.

Empirical scoring functions have been widely used in many protein-ligand docking programs and virtual screening processes. This approximates the protein-ligand interaction using an equation consisting of various physics-based conditions such as van der Waals interaction, solvation free energy, and electrostatic interaction. The parameters of the physics-based conditions are usually tailored to experimental data to reproduce the measured binding affinity values. Because of their computational simplicity and close relationship to physics-based interactions, evaluation functions based on experimental data are still being actively developed.

nez-Rosell, G.; De Fabritiis, G. KDEEP: Protein-Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks. Journal of Chemical Information and Modeling. 2018, pp 287-296] 에서는 본래 이미지 분류를 위해 고안된 SqueezeNet 아키텍처에 기초하여 결합 친화도 예측모델, K_deep을 개발하였다. 상기 K_deep 모델은 약 1.3 백만 개의 파라미터들이 다중 3D-CNN으로 구성된다. 문헌 [Zhang, H.; Liao, L.; Saravanan, K. M.; Yin, P.; Wei, Y. DeepBindRG: A Deep Learning Based Method for Estimating Effective Protein-Ligand Affinity. PeerJ 2019, 7, e7362] 에서는 단백질-리간드 인터페이스의 2D 표현과 ResNet 아키텍처를 이용한 DeepBindRG 모델을 개발하였다. 유사하게, 문헌 [Zheng, L.; Fan, J.; Mu, Y. OnionNet: A Multiple-Layer Intermolecular-Contact-Based Convolutional Neural Network for Protein-Ligand Binding Affinity Prediction. ACS Omega 2019, 4 (14), 15956-15965] 에서는 단백질-리간드 결합 구조를 하나의 채널을 가진 2D 텐서(tensor)로 변환하고 이를 세 개의 2D-CNN 레이어들과 네 개의 밀집 레이어들로 프로세싱하였다. Recently, deep learning methods have made more accurate data-driven predictions possible in various scientific fields. Many deep learning-based methods have been proposed to predict protein-ligand binding affinity. See, for example, Ragoza, M.; Hochuli, J.; Idrobo, E.; Sunseri, J.; Koes, D. R. Protein-Ligand Scoring with Convolutional Neural Networks. J. Chem. Inf. Model. 2017, 57 (4), pp 942-957], a small network consisting of three sequential layers of a 3D convolutional neural network (3D-CNN) with pooling layers was proposed. Similarly, Stepniewska-Dziubinska, MM; Zielenkiewicz, P.; Siedlecki, P. Development and Evaluation of a Deep Learning Model for Protein-ligand Binding Affinity Prediction.Bioinformatics. 2018, pp 3666-3674] developed a binding affinity prediction model consisting of three consecutive 3D-CNN layers and three dense layers following them. Jimenez, J.; skalic, M.; Mart

nez-Rosell, G.; De Fabritiis, G. KDEEP: Protein-Ligand Absolute Binding Affinity Prediction via 3D-Convolutional Neural Networks. Journal of Chemical Information and Modeling. 2018, pp 287-296] developed a binding affinity prediction model, K _deep, based on the SqueezeNet architecture originally designed for image classification. The K _deep model consists of multiple 3D-CNNs with approximately 1.3 million parameters. References [Zhang, H.; Liao, L.; Saravanan, K.M.; Yin, P.; Wei, Y. DeepBindRG: A Deep Learning Based Method for Estimating Effective Protein-Ligand Affinity. PeerJ 2019, 7, e7362] developed the DeepBindRG model using 2D representation of the protein-ligand interface and ResNet architecture. Similarly, Zheng, L.; Fan, J.; Mu, Y. OnionNet: A Multiple-Layer Intermolecular-Contact-Based Convolutional Neural Network for Protein-Ligand Binding Affinity Prediction. [ACS Omega 2019, 4 (14), 15956-15965] converted the protein-ligand binding structure into a 2D tensor with one channel and processed it with three 2D-CNN layers and four dense layers. .

In predicting protein-ligand binding affinity, a model with high accuracy and low computational throughput is required. To this end, the application of deep learning-based technologies in various fields can be proposed, and in particular, an optimized network architecture is required to predict the binding affinity of protein-ligand complexes. Additionally, the introduction of a new evaluation function for evaluating new network architectures is required. Additionally, a method is required to identify the most important physical properties in determining binding affinity.

One embodiment of the present invention is a method of predicting protein-ligand binding affinity through a prediction system including a control unit and a memory unit, where the structure of the protein-ligand complex is determined with the center of mass of the ligand in the protein-ligand complex as the origin. and storing the surrounding atomic environment as 3D information embedded in a 3D grid in the memory unit; storing a protein-ligand binding affinity value corresponding to the three-dimensional information in the memory unit; A step of learning a prediction model through a control unit, wherein the prediction model uses the stored three-dimensional information as an input value and the corresponding protein-ligand binding affinity value as a calculation value, and the prediction model is processed independently. training the prediction model, comprising one or more 3D convolutional neural networks; And generating a protein-ligand binding affinity predicted value using the prediction model, wherein each 3D convolutional neural network generates a protein-ligand binding affinity prediction value for a new protein-ligand complex, A method for predicting protein-ligand binding affinity is provided, comprising generating the predicted protein-ligand binding affinity, generating a final predicted value based on each protein-ligand binding affinity predicted value.

Additionally, the final prediction value generated by the prediction model may be the average of each protein-ligand binding affinity prediction value generated through each 3D convolutional neural network.

In addition, the three-dimensional information provides a method for predicting protein-ligand binding affinity, including the pattern of protein-ligand interaction calculated through the density function below.

Here, n(r) is the atomic number density, r _VDW is the van der Waals radius of the atom, and r is the distance from the atom to the center of the grid.

Additionally, an embodiment of the present invention may further include classifying the 3D information into a plurality of classes and expressing them through different channels of the input value.

Additionally, an embodiment of the present invention may further include the step of rotating the 3D information using 24 rotation operations and adding an input value.

In addition, in one embodiment of the present invention, the 3D convolutional neural network includes one or more stacked ensemble-based residual block layers, and each residual block layer is one or more stacked convolutional layers combining batch normalization and ReLU activation layers. may also include

Additionally, one or more parallel processed 3D convolutional neural network layers may be included in the middle of the one or more stacked convolutional layers.

In addition, a protein-ligand binding affinity prediction method is provided in which the results of the one or more parallel processed 3D convolutional neural network layers are concatenated and input to the remaining block layer.

Additionally, in one embodiment of the present invention, the number of the one or more 3D convolutional neural networks may be 5 to 25.

Another embodiment of the present invention includes a communication unit capable of communicating with an external server; An input/output interface unit that controls the input unit and display unit; a memory unit for storing data; and a control unit for performing a protein-ligand binding affinity prediction model, wherein the memory unit determines the structure of the protein-ligand complex with the center of mass of the ligand in the protein-ligand complex as the origin. and includes 3D information embedding the surrounding atomic environment in a 3D grid and a protein-ligand binding affinity value corresponding to the 3D information, and the control unit takes the 3D information as an input value and Using the protein-ligand binding affinity value as a calculated value, one or more 3D convolutional neural networks are independently processed to learn the prediction model, and the control unit generates a new protein-ligand through each 3D convolutional neural network. A protein-ligand binding affinity prediction system is provided, which generates predicted protein-ligand binding affinity values for each complex and generates a final predicted value based on each protein-ligand binding affinity predicted value.

Additionally, the final prediction value generated by the prediction model may be the average of each protein-ligand binding affinity prediction value generated through each 3D convolutional neural network.

Additionally, the three-dimensional information provides a protein-ligand binding affinity prediction system that includes the pattern of protein-ligand interaction calculated through the density function below.

Here, n(r) is the atomic number density, r _VDW is the van der Waals radius of the atom, and r is the distance from the atom to the center of the grid.

Additionally, the control unit may classify the 3D information into a plurality of classes and express them through different channels of the input value.

Additionally, the control unit can rotate the 3D information using 24 rotation operations and add input values.

Additionally, the 3D convolutional neural network includes one or more stacked ensemble-based residual block layers, and each residual block layer may include one or more stacked convolutional layers combining batch normalization and ReLU activation layers.

Additionally, one or more parallel processed 3D convolutional neural network layers may be included in the middle of the one or more stacked convolutional layers.

In addition, a protein-ligand binding affinity prediction system is provided that concatenates the results of the one or more parallel processed 3D convolutional neural network layers and inputs them to the remaining block layer.

Additionally, in another embodiment of the present invention, the plurality of 3D convolutional neural networks may be 5 or more and 25 or less.

One of the various embodiments of the present specification discloses a new protein-ligand binding affinity prediction model from the ResNext architecture, which was previously developed for accurate image classification. In addition, compared to existing models, prediction quality can be significantly improved through the use of a new network architecture and an ensemble-based approach using multiple predictors instead of a single predictor. For example, prediction quality can be improved with an ensemble-based approach that uses the average value of multiple predictors. The ensemble approach has the advantage of not requiring additional modifications to the network architecture and being easily applicable to most existing models. Benchmark results using the CASF-2016 data set show that the performance of the present invention is comparable to the best existing evaluation function. Additionally, the relative importance of factors can be analyzed to identify the most essential physical properties in determining binding affinity.

1 is a diagram schematically showing a prediction system according to an embodiment of the present invention.
Figure 2 is a schematic diagram of the overall network structure of an embodiment of the present invention.
Figure 3 is a schematic diagram showing the structure of each remaining block.
Figure 4 shows prediction quality versus number of networks.
Figure 5 shows the benchmark results of existing protein-ligand binding affinity evaluation functions and AK-score.
Figure 6 is a diagram showing a comparison between experimental binding affinity and predicted values obtained through AK-score-ensemble, X-score, and Autodock vina.
Figure 7 shows the calculation results of factor importance.

The present invention will become clear by referring to the embodiments described in detail below along with the drawings attached to this specification. However, the present invention is not limited to the embodiments disclosed below and will be implemented in various different forms. The present embodiments only serve to ensure that the disclosure of the present invention is complete and that common knowledge in the technical field to which the present invention pertains is not limited. It is provided to fully inform those who have the scope of the invention, and the present invention is only defined by the scope of the claims. Meanwhile, the terms used in this specification are for describing embodiments and are not intended to limit the present invention. As used herein, singular forms also include plural forms, unless specifically stated otherwise in the context. As used herein, “comprises” or “comprising” does not exclude the presence or addition of one or more other components or steps in addition to the mentioned components or steps. Terms such as first, second, etc. may be used to describe various components, but the components should not be limited by the terms. Terms are used only to distinguish one component from another.

Hereinafter, various embodiments of the present invention will be described by way of example with reference to the drawings.

1 is a block diagram showing an example of the configuration of a prediction system according to an embodiment, and conceptually shows parts related to the present embodiment. Each component may be provided in one device and processed independently, but the present invention is not limited to this, and may also include that each component is performed on a separate device connected through a network.

The external server 20 can be connected to the prediction system 10 through a network, and provides protein structure information, ligand structure information, protein-ligand complex structure information, binding affinity information according to the protein-ligand complex binding site, and gene Information such as information, intermolecular interaction information, and/or protein structure similarity information may be provided. For example, the external server 20 may be a database for prediction processing of the prediction system 10 or a server that provides the same.

The prediction system 10 may include a control unit 11, a communication unit 12, an input/output interface unit 13, and a memory unit 14.

The control unit 11 is a component that controls the entire prediction system 10 and may include, for example, a processing unit such as a CPU or GPU. The control unit 11 can train models to be described later using information stored in the memory unit 14, and can also calculate a predicted value for a new input through the learned model. Specifically, the control unit 11 may control a model that calculates a prediction of protein-ligand binding affinity. To this end, the control unit 11 may include an internal memory for storing control programs such as an operating system (OS), programs defining various processing sequences, and data. And, the control unit 11 can perform information processing to execute various processes using these programs, etc.

Additionally, the communication unit 12 may include an interface that can be connected to a communication device such as a router connected to a communication line, and can control communication between the prediction system 10 and the external server 20. .

The input/output interface unit 13 may be an interface connected to the input unit 15 and/or the display unit 16. The prediction system 10 and the user can communicate through the input/output interface unit 13. For example, the display unit 16 may be a display means (for example, a display made of liquid crystal or organic EL, a monitor, a touch panel, etc.) that displays a display screen such as an application. Additionally, the input unit 15 may be, for example, a key input unit, a touch panel, a control pad (eg, a touch pad, a game pad, etc.), a mouse, a keyboard, a microphone, etc.

Additionally, the memory unit 14 may be a device that stores various databases, tables, etc. For example, the memory unit contains protein structure information, ligand structure information, protein-ligand complex structure information, binding affinity information according to the protein-ligand complex binding site, genetic information, intermolecular interaction information, and/or protein structure similarity information, etc. may include information.

data set

The PDBBind database was used for protein-ligand binding affinity data for network training and testing. As of August 2018, the refined set of databases has experimental binding affinities for 3767 protein-ligand complexes. Among them, a manually selected core set containing 290 high-resolution data was used for testing. The remaining 4055 complexes grouped by purified binding affinity were used as the training set.

Convolutional Neural Network

To take advantage of the convolutional neural network, the structure of the protein-ligand complex was represented as a 3-dimensional (3D) grid. For each protein-ligand complex, the center of mass of the ligand in the complex was set as the origin, and the surrounding atomic environment was embedded into a 3D-grid with an edge length of 30 Å. Along the X, Y and Z axes, 30 grid boxes were created at 1.0 Å spacing. To collect the pattern of protein-ligand interactions for each grid box, the atomic number density was calculated using the density function below (Equation 1).

Formula 1

here, is the van der Waals radius of the atom, is the distance from the atom to the center of the grid.

Atoms are classified into eight classes, which represent different channels of input data. The example treats the atoms of the protein and ligand separately, resulting in 16 real-valued channels representing the aggregated number density of each protein-ligand complex. Table 1 below shows a description of the atom types used in the examples.

Atom type used to classify the atoms that form the protein-ligand binding site Atom type Definition Hydrophobic Aliphatic or aromatic C Aromatic Aromatic C Hydrogen bond donor Donor 1 H-bond or Donor S
Spherical H with NA, NS, OA, OS, SA Hydrogen bond acceptor Acceptor 1 H-bond or S Spherical N;
Acceptor 2 H-bonds or S Spherical O;
Acceptor 2 H-bonds S Positive ionizable Gasteiger positive charge Negative ionizable Gasteiger negative charge Metallic MG, Zn, Mn, Ca, or Fe Excluded Volume All atom types

The examples added to the number of data sets by rotating the grid through all 24 possible rotation operations to reduce the orientation dependence of the composite structure.

Network architecture

The main configuration of the network provided in the embodiment is an ensemble-based residual network used in the ResNext model for image recognition. Hereinafter, this will be described in detail in FIGS. 1 and 2.

Figure 2 is a schematic diagram of the overall network structure of the embodiment.

The entire network mainly consists of 14 stacked layers of ensemble-based residual blocks.

Figure 3 shows the structure of each remaining block. The remaining block consists of three stacked convolutional layers combining batch normalization and ReLU (Rectified Linear Units) activation layers. In the middle of the block, each channel is distributed into 16 3D convolutional layers and processed in parallel.

In Figures 2 and 3, Conv3D refers to a 3D convolutional neural network layer, BN (batch normalization) refers to batch normalization, and RL (residual layer) refers to a residual layer. The number of remaining networks deployed in parallel is also called cardinality.

In the example, 16 Conv3D layers were used for each residual block. In this specification, the network architecture of the embodiment is referred to as AK-score (Arontier-Kangwon docking scoring function: AK-score).

In the embodiment, the ReLU activation function is used. All weight parameters are initialized using the He_normal initialization method. Model loss is calculated as the mean absolute error (MAE) between experimental and predicted binding affinity values in kcal/mol.

For parameter optimization, the Adam optimizer is used with parameters beta-1 = 0.99 and beta-2 = 0.999. The model was trained at multiple learning rates to evaluate the impact of learning rate on the final prediction quality. Running rates of 0.0001, 0.0005, 0.0007 and 0.0010 were evaluated. While training the network, the entire data set consisted of random permutations.

Ensemble prediction

To enhance prediction accuracy, the embodiment adopted an ensemble prediction method, and the final prediction value was derived from the average of multiple independently learned models. For many machine learning tasks, the parameters for each prediction model are optimized from initial random values. As the number of parameters becomes large, the final parameter set generally does not converge. To reduce this tendency, the embodiment trained multiple networks independently and checked whether averaging multiple predictions yielded a better prediction. In the embodiment, the final predicted value was derived from the average of multiple independently learned models, but this is for illustrative purposes, and the present invention is not limited to this, and other values such as the median may be used.

performance evaluation

To evaluate the performance of the example, the model of the example was compared with a previously proposed 3D-CNN-based binding affinity prediction model. The K_deep model based on the SqueezeNet architecture used for image classification was used as a comparison model. The models were trained using the same parameters. All models disclosed in this specification used Keras-2.2.4 with Tensorflow-1.13.1 as the backend.

The examples compared the performance of models from various perspectives. The CASF-2016 benchmark set has been used as a reference point for comparison of various docking models. The accuracy of protein-ligand docking prediction is generally measured from three perspectives. 1) Scoring, how well do the predicted binding affinity values correlate with experimental values? 2) Ranking, how well are relevant combined features predicted? 3) Docking, has the correct docking pose been accurately identified to have lower energy than the decoy?

To evaluate the scoring power of the model, the Pearson correlation coefficient was calculated. Additionally, Spearman, Kendall tau, and PI (Predictive index) values were used to evaluate ranking power. To evaluate docking power, we checked whether the ligand with the actual bound ligand was included in the top1, top2, and top3 percent of possible decoys.

Results and Analysis

Accuracy of binding affinity predictions

The examples performed three different types of experiments based on the AK-Score architecture. AK-score-single, AK-score-small, and AK-score-ensemble. AK-score-single used one prediction network shown in Figures 1 and 2. AK-score-small uses only 10 features, excluding the relatively rare ones (e.g., positive, negative, and metallic in Table 1) compared to others. AK-score-ensemble used the average of 20 independently trained networks as the final prediction value. Table 2 below provides the mean absolute error (MAE) and root mean squared error (RMSE) between the predicted and experimental values of various models trained at different running rates. For comparison, the results of another 3D-CNN-based deep learning model, K _deep model, to which the same group was applied were also disclosed.

The benchmark results show that the AK-score-ensemble model produces the most accurate prediction results. Among the evaluated models, AK-score-ensemble has the lowest accuracy measure, with a MAE of 1.01 kcal/mol and an RMSE of 1.29 kcal/mol. Compared to the single model, the average error of the ensemble model is lower by approximately 0.1 kcal/mol. Additionally, when compared to the K _deep model, AK-score-ensemble has a lower average error of 0.2 kcal/mol. The results of the example also show that by selecting the optimal running rate, the MAE can be improved by about 0.05 kcal/mol and the RMSE by 0.10 kcal/mol.

Results of evaluating the prediction accuracy of ResNext-ensemble, ResNext, and K _deep using the PDBbind-2016 data set and mean absolute error and root mean square error measures. Model Learning Rate M.A.E. RMSE Kdeep 0.0001 1.131 1.462 0.0005 1.200 1.519 0.0006 1.164 1.534 0.0010 1.219 1.536 AK-score single 0.0001 1.159 1.511 0.0005 1.101 1.415 0.0007 1.130 1.425 0.0010 1.110 1.406 AK-score small 0.0006 1.128 1.462 AK-score ensemble 0.0007 1.014 1.293

The example also evaluated the performance of the model based on three criteria applied to the CASF-2016 data set: scoring, ranking, and docking power. Table 3 below shows this in detail. In all three criteria, the AK-score-ensemble model shows the best performance. Scoring power is indicated by the Pearson correlation coefficient between predicted and experimental values. Overall, AK-score models yielded higher correlation values than K _deep models. Among all evaluated models, only the AK-score-ensemble model yielded correlation coefficient values higher than 0.8. In terms of ranking power, AK-score models outperformed K _deep models on average. Among these, the AK-score-ensemble model derived the highest rank correlation coefficient. In terms of docking power, the difference between the prediction performance of the K _deep model and the AK-score-single model was not noticeable compared to other criteria. However, the prediction results of the AK-score-ensemble model were significantly better than the K _deep model.

Comparison of prediction accuracy for CASF-2016 dataset Model Scoring Ranking Docking Learning rate Pearson
(R) Spearman
(SP) Kendall
(tau) Predictive
(PI) Top1
(%) Top2
(%) Top3
(%) K _deep 0.0001 0.738 0.539 0.435 0.559 24.8 38.5 52.2 0.0005 0.709 0.486 0.389 0.535 29.1 39.9 49.6 0.0006 0.701 0.528 0.439 0.558 29.1 39.9 49.6 0.0010 0.715 0.479 0.400 0.492 24.8 36.3 44.6 AK-score-single 0.0001 0.719 0.572 0.456 0.600 34.9 48.6 56.1 0.0005 0.755 0.596 0.512 0.616 29.9 43.2 54.0 0.0007 0.759 0.616 0.526 0.640 31.3 47.1 57.9 0.0010 0.760 0.598 0.505 0.627 26.4 43.9 54.0 AK-score-small 0.0006 0.741 0.567 0.495 0.586 28.4 46.8 56.5 AK-score-ensemble 0.0007 0.812 0.670 0.589 0.698 36.0 51.4 59.7

Ensembles of networks improve prediction quality

Figure 4 shows that prediction quality changes depending on the number of networks.

Figure 4(a) shows the scoring power measured by the Pearson correlation coefficient between experimental and predicted values for binding affinity. Figure 4 (b) shows the ranking power measured by three rank correlation coefficients: Spearman (SP), Kendall tau (tau), and Predictive Index (PI).

Results according to the embodiment show that the overall prediction accuracy improves as more networks are used. In all three quality measures: scoring, ranking, and docking power, accuracy increased dramatically as we moved from one network to five networks. In terms of scoring power, the Pearson correlation coefficient (R) between experimental and predicted values increased until the network reached 25. Figure 4(a) is a diagram showing this. When one network is used, the correlation coefficient is 0.74. However, when the average of five networks is used, values greater than 0.80 are obtained. After 10 networks, the improvement is gradual but continues to increase until 25 networks are used. Similarly, ranking power increases until 25 networks are used. All three ranking measurement factors, SP, tau, and PI, consistently increase when the ensemble average of the network is averaged. Figure 4(b) shows these results. These results clearly show that using an ensemble of prediction networks significantly improves prediction quality. This simply and clearly improves prediction accuracy without having to reinvent various network architectures.

Comparison with existing scoring functions

Figure 5 shows the benchmark results of existing protein-ligand binding affinity evaluation functions and AK-score. Figure 5(a) shows that AK-Score shows the highest scoring power compared to existing evaluation functions. Figure 5(b) shows that AK-Score ranks right after vina-RF ₂₀ , which has the highest performance currently known.

AK-Score's benchmark results show that AK-Score's prediction accuracy is comparable to the best existing scoring functions for the CASF-2016 dataset. The CASF-2016 dataset provides prior prediction results of known evaluation functions. This allows a fair comparison between the model of the embodiment and the existing evaluation function using exactly the same test set. In terms of scoring power, the highest correlation coefficient produced by AK-Score is 0.828. This is a higher value than the highest value calculated with vina-RF ₂₀ , which is 0.816. Additionally, the ranking power of AK-Score is 0.736, which is slightly lower than the highest value of 0.761 for vina-RF ₂₀ , and ranks second among the evaluation functions tested.

Figure 6 is a diagram showing a comparison between experimental binding affinity and predicted values obtained through AK-score-ensemble, X-score, and Autodock vina. Specifically, Figure 6 (a) shows a scatter plot of the predicted value obtained through AK-score-ensemble and the experimental binding affinity. Figures 6 (b) and (c) show scatter plots of the predicted values obtained through Autodock vina and the predicted values obtained through X-score, respectively.

Overall, the results of the AK-score-ensemble model show a high correlation with the experimental values, showing a correlation coefficient of 0.827 for all complexes tested. On the other hand, the results of X-score and Autodock are clearly biased. On average, the X-score estimates the absolute binding affinity significantly lower, which is indicated by the small slope coefficient of the regression line. Autodock-vina shows better correlation than X-score. However, this is also evaluated somewhat lower than the experimental value. In summary, we clearly show that AK-score-ensemble outperforms commonly used empirical data-based evaluation functions in terms of absolute binding affinity.

Evaluation of factor importance

In order to gain chemical and biological insights through learned networks, it is necessary to identify which atomic factors play an important role in determining the binding affinity of protein-ligand complexes. To this end, the embodiment conducted additional experiments to 1) create a specific channel zero value, and 2) perform prediction by randomly shuffling the channel values. The logic of the second experiment is based on the conjecture that making all values in the channel zero would result in too much information loss, and that preserving the mean and variance of the channel values may be important in making reasonable predictions.

Overall, from the two experiments, it was confirmed that the excluded volume and binding site of the ligand are the most important factors in determining the binding affinity of the protein-ligand complex.

Figure 7 is a diagram showing such results. Figure 7 shows the calculation results of factor importance measured by loss of prediction accuracy. The mean absolute error (MAE) of the prediction is shown in units of kcal/mol on the y-axis. In Figure 7(a), the channel corresponding to the factor on the X-axis is filled with zero, and in Figure 7(b), the values of the channel corresponding to the factor are randomly shuffled.

In other words, shape complementarity between the binding site and the ligand is most important in determining the binding affinity of the complex. As shown in (a) of Figure 7, when there is no exclusion volume information of the ligand, the average binding affinity prediction accuracy is lowered by 1.4 kcal/mol. When following the exclusion volume information, the hydrophobic atom information of the ligand and the binding site are identified as the second important factors. Regarding the binding site, the hydrogen-soluble atoms of the binding site play a third important role. Interestingly, for ligands, aromatic atoms play a third important role. As can be seen in (b) of Figure 7, in the shuffling experiment, the overall directionality is similar to that of the zero experiment, but the average decrease in prediction accuracy is smaller. The most striking difference for the binding site is that the relative importance of the hydrogen bond acceptor atoms becomes greater than that of the hydrophobic atoms.

conclusion

The embodiment provides AK-score, a new binding affinity prediction model that combines a multi-branch deep learning network architecture and ensemble prediction approach. The examples predict the binding affinity of protein-ligand complexes with significantly high accuracy, providing results comparable to the best existing scoring functions in terms of scoring and ranking power. Using the ensemble-based approach provided by the embodiment and the average of multiple independently learned models is a very simple yet powerful approach. The embodiment is not limited to a specific network model, and various existing machine learning-based models may be applied. Additionally, the examples provide insight into the relative importance of atoms based on their chemical properties. As experiments on factor importance show, for a ligand, the exclusion volume of atoms, the spatial distribution of hydrophobic and aromatic atoms play a very important role in determining the binding affinity. For proteins, the exclusion volume of atoms and the distribution of hydrophobic atoms and hydrogen bond acceptors have been identified as important factors.

Embodiments according to the present invention described above may be implemented in the form of program instructions that can be executed through various computer components and recorded on a computer-readable recording medium. A computer-readable recording medium may include program instructions, data files, data structures, etc., singly or in combination. Program instructions recorded on a computer-readable recording medium may be specially designed and configured for the present invention, or may be known and usable by those skilled in the computer software field. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical recording media such as CD-ROMs and DVDs, and magneto-optical media such as floptical disks. media), and hardware devices specifically configured to store and perform program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include not only machine language code such as that created by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. The hardware device may be configured to operate as one or more software modules to perform processing according to the invention and vice versa.

Additionally, the embodiments according to the present invention described above may be a set of program instructions that can be executed through various computer components and a user application for executing them. Specifically, it may be a program itself that can be downloaded from a server or via a storage medium and installed on a client computer.

In the above, the present invention has been described with specific details such as specific components and limited embodiments and drawings, but this is only provided to facilitate a more general understanding of the present invention, and the present invention is not limited to the above embodiments. , a person skilled in the art to which the present invention pertains can make various modifications and variations from this description.

Accordingly, the spirit of the present invention should not be limited to the above-described embodiments, and the scope of the patent claims described below as well as all modifications equivalent to or equivalent to the scope of the patent claims fall within the scope of the spirit of the present invention. They will say they do it.

10: Prediction system 20: External server
11: Control unit 12: Communication unit
13: input/output interface unit 14: memory unit
15: input unit 16: display unit

Claims (18)

A method of predicting protein-ligand binding affinity through a prediction system including a control unit and a memory unit,
Storing the structure of the protein-ligand complex in a memory unit as three-dimensional information with the center of mass of the ligand in the protein-ligand complex as the origin and embedding the surrounding atomic environment in a 3D grid;
storing a protein-ligand binding affinity value corresponding to the three-dimensional information in the memory unit;
A step of learning a prediction model through a control unit, wherein the prediction model uses the stored three-dimensional information as an input value and the corresponding protein-ligand binding affinity value as a calculation value, and the prediction model is processed independently. training the prediction model, comprising one or more 3D convolutional neural networks; and
A step of generating a protein-ligand binding affinity predicted value using the prediction model, generating protein-ligand binding affinity prediction values for a new protein-ligand complex through the one or more 3D convolutional neural networks, respectively, Generating the predicted protein-ligand binding affinity, generating a final predicted value based on the predicted protein-ligand binding affinity of
The three-dimensional information includes the pattern of protein-ligand interaction calculated through the density function below,

Here, n(r) is the atomic number density, r _VDW is the van der Waals radius of the atom, and r is the distance from the atom to the center of the grid. Protein-ligand binding affinity prediction method. According to claim 1,
The final prediction value generated by the prediction model is an average of each protein-ligand binding affinity prediction value generated through the one or more 3D convolutional neural networks. A method of learning a prediction model that predicts protein-ligand binding affinity through a prediction system including a control unit and a memory unit,
Storing the structure of the protein-ligand complex in a memory unit as three-dimensional information with the center of mass of the ligand in the protein-ligand complex as the origin and embedding the surrounding atomic environment in a 3D grid;
storing a protein-ligand binding affinity value corresponding to the three-dimensional information in the memory unit; and
A step of learning a prediction model through a control unit, wherein the prediction model uses the stored three-dimensional information as an input value and the corresponding protein-ligand binding affinity value as a calculation value, and the prediction model is processed independently. A step of training the prediction model, including one or more 3D convolutional neural networks,
The three-dimensional information includes the pattern of protein-ligand interaction calculated through the density function below,

Here, n(r) is the atom number density, r _VDW is the van der Waals radius of the atom, and r is the distance from the atom to the center of the grid. Method of learning a prediction model to predict protein-ligand binding affinity. A method of predicting protein-ligand binding affinity through a prediction system including a control unit and a memory unit,
Storing the structure of the protein-ligand complex in a memory unit as three-dimensional information with the center of mass of the ligand in the protein-ligand complex as the origin and embedding the surrounding atomic environment in a 3D grid;
storing a protein-ligand binding affinity value corresponding to the three-dimensional information in the memory unit;
A step of learning a prediction model through a control unit, wherein the prediction model uses the stored three-dimensional information as an input value and the corresponding protein-ligand binding affinity value as a calculation value, and the prediction model is processed independently. training the prediction model, comprising one or more 3D convolutional neural networks; and
A step of generating a protein-ligand binding affinity predicted value using the prediction model, generating protein-ligand binding affinity prediction values for a new protein-ligand complex through the one or more 3D convolutional neural networks, respectively, Generating the predicted protein-ligand binding affinity, generating a final predicted value based on the predicted protein-ligand binding affinity of
A method for predicting protein-ligand binding affinity, further comprising classifying the three-dimensional information into a plurality of classes and expressing them in different channels of the input value. A method of predicting protein-ligand binding affinity through a prediction system including a control unit and a memory unit,
Storing the structure of the protein-ligand complex in a memory unit as three-dimensional information with the center of mass of the ligand in the protein-ligand complex as the origin and embedding the surrounding atomic environment in a 3D grid;
storing a protein-ligand binding affinity value corresponding to the three-dimensional information in the memory unit;
A step of learning a prediction model through a control unit, wherein the prediction model uses the stored three-dimensional information as an input value and the corresponding protein-ligand binding affinity value as a calculation value, and the prediction model is processed independently. training the prediction model, comprising one or more 3D convolutional neural networks; and
A step of generating a protein-ligand binding affinity predicted value using the prediction model, generating protein-ligand binding affinity prediction values for a new protein-ligand complex through the one or more 3D convolutional neural networks, respectively, Generating the predicted protein-ligand binding affinity, generating a final predicted value based on the predicted protein-ligand binding affinity of
A method for predicting protein-ligand binding affinity, further comprising the step of rotating the three-dimensional information using 24 rotation operations and adding input values. According to claim 2,
The 3D convolutional neural network includes one or more stacked ensemble-based residual block layers,
A protein-ligand binding affinity prediction method, wherein each residual block layer includes one or more stacked convolutional layers combined with a batch normalization and ReLU activation layer. According to claim 6,
A protein-ligand binding affinity prediction method comprising one or more parallel processed 3D convolutional neural network layers in the middle of the one or more stacked convolutional layers. According to claim 7,
A protein-ligand binding affinity prediction method in which the results of the one or more parallel processed 3D convolutional neural network layers are concatenated and input to the remaining block layer. According to claim 1,
A method for predicting protein-ligand binding affinity, wherein the one or more 3D convolutional neural networks are 5 or more and 25 or less. A communication unit capable of communicating with an external server;
An input/output interface unit that controls the input unit and display unit;
a memory unit for storing data; and
A protein-ligand binding affinity prediction system comprising a control unit for performing a protein-ligand binding affinity prediction model,
The memory unit includes three-dimensional information that sets the structure of the protein-ligand complex as the origin of the center of mass of the ligand in the protein-ligand complex and embeds the surrounding atomic environment in a 3D grid, and protein-ligand binding affinity corresponding to the three-dimensional information. Contains degree values,
The control unit uses the 3D information as an input value and the corresponding protein-ligand binding affinity value as a calculation value to independently process one or more 3D convolutional neural networks to learn the prediction model,
The control unit generates predicted protein-ligand binding affinity values for each new protein-ligand complex through each of the one or more 3D convolutional neural networks, and generates a final predicted value based on each predicted protein-ligand binding affinity value. Create a ,
The three-dimensional information includes the pattern of protein-ligand interaction calculated through the density function below,

Here, n(r) is the atomic number density, r _VDW is the van der Waals radius of the atom, and r is the distance from the atom to the center of the grid. Protein-ligand binding affinity prediction system. According to claim 10,
The final prediction value generated by the prediction model is an average of each protein-ligand binding affinity prediction value generated through each of the one or more 3D convolutional neural networks. Ministry of Communications;
Input/output interface unit;
a memory unit for storing data; and
A protein-ligand binding affinity prediction system comprising a control unit for performing a protein-ligand binding affinity prediction model,
The memory unit includes three-dimensional information that sets the structure of the protein-ligand complex as the origin of the center of mass of the ligand in the protein-ligand complex and embeds the surrounding atomic environment in a 3D grid, and protein-ligand binding affinity corresponding to the three-dimensional information. Contains degree values,
The control unit uses the 3D information as an input value and the corresponding protein-ligand binding affinity value as a calculation value to independently process one or more 3D convolutional neural networks to learn the prediction model,
The control unit generates a predicted protein-ligand binding affinity value using the prediction model,
The three-dimensional information includes a pattern of protein-ligand interaction calculated through the density function below,

Here, n(r) is the atomic number density, r _VDW is the van der Waals radius of the atom, and r is the distance from the atom to the center of the grid. Protein-ligand binding affinity prediction system. Ministry of Communications;
Input/output interface unit;
a memory unit for storing data; and
A protein-ligand binding affinity prediction system comprising a control unit for performing a protein-ligand binding affinity prediction model,
The memory unit includes three-dimensional information that sets the structure of the protein-ligand complex as the origin of the center of mass of the ligand in the protein-ligand complex and embeds the surrounding atomic environment in a 3D grid, and protein-ligand binding affinity corresponding to the three-dimensional information. Contains degree values,
The control unit uses the 3D information as an input value and the corresponding protein-ligand binding affinity value as a calculation value to independently process one or more 3D convolutional neural networks to learn the prediction model,
The control unit generates predicted protein-ligand binding affinity values for each new protein-ligand complex through each of the one or more 3D convolutional neural networks, and generates a final predicted value based on each predicted protein-ligand binding affinity value. Create a ,
The protein-ligand binding affinity prediction system wherein the control unit classifies the three-dimensional information into a plurality of classes and expresses it through different channels of the input value. Ministry of Communications;
Input/output interface unit;
a memory unit for storing data; and
A protein-ligand binding affinity prediction system comprising a control unit for performing a protein-ligand binding affinity prediction model,
The memory unit includes three-dimensional information that sets the structure of the protein-ligand complex as the origin of the center of mass of the ligand in the protein-ligand complex and embeds the surrounding atomic environment in a 3D grid, and protein-ligand binding affinity corresponding to the three-dimensional information. Contains degree values,
The control unit uses the 3D information as an input value and the corresponding protein-ligand binding affinity value as a calculation value to independently process one or more 3D convolutional neural networks to learn the prediction model,
The control unit generates predicted protein-ligand binding affinity values for each new protein-ligand complex through each of the one or more 3D convolutional neural networks, and generates a final predicted value based on each predicted protein-ligand binding affinity value. Create a ,
A protein-ligand binding affinity prediction system in which the control unit rotates the three-dimensional information using 24 rotation operations and adds input values. According to claim 11,
The 3D convolutional neural network includes one or more stacked ensemble-based residual block layers,
A protein-ligand binding affinity prediction system, where each residual block layer includes one or more stacked convolutional layers combined with a batch normalization and ReLU activation layer. According to claim 15,
A protein-ligand binding affinity prediction system comprising one or more parallel 3D convolutional neural network layers in the middle of the one or more stacked convolutional layers. According to claim 16,
A protein-ligand binding affinity prediction system that concatenates the results of the one or more parallel processed 3D convolutional neural network layers and inputs them to the remaining block layer. According to claim 10,
A protein-ligand binding affinity prediction system wherein the one or more 3D convolutional neural networks are 5 or more and 25 or less.