Attention-based neural network system predicting the stability of compounds in human plasma and controlling method for the same

Abstract

The present invention relates to an attention-based neural network system for predicting the stability of a compound in human plasma and a control method thereof. By connecting the molecular fingerprint notation of a test compound to a molecular feature vector and simulating it with a fully connected layer to predict the stability of the compound in human plasma through a binary classification of stability or instability, the stability of the compound in human plasma can be easily predicted in the early stages of drug discovery and development. Consequently, the usefulness of the drug development platform is maximized. Furthermore, by providing a plasma stability prediction device that can repeatedly simulate whether a compound exhibits stability in human plasma without requiring a complex and expensive system to predict stability in human plasma, the reaction of an inherently unstable compound or prodrug in human plasma can be rapidly confirmed, thereby significantly improving the designability of novel drugs.

Description

Technology Field

[0001] The present invention relates to an attention-based neural network system for predicting the stability of compounds in human plasma and a control method thereof. In particular, the invention relates to an attention-based neural network system for predicting the stability of compounds in human plasma and a control method thereof, which can predict the stability of compounds in human plasma through binary classification of stability or instability by connecting the molecular fingerprint notation of a test compound to a molecular feature vector and simulating it with a fully connected layer. Background Technology

[0002] In general, the stability of compounds in human plasma plays a very important role in drug discovery and development. Compounds that degrade rapidly in the plasma tend to have low bioavailability and low efficacy in vivo. Additionally, if plasma stability is poor during sample storage or analysis, in vivo drug concentrations may be inaccurately measured.

[0003] Furthermore, plasma stability as described above is considered an important factor in the new drug development process because accurate measurement of drug concentrations in biological samples is crucial for pharmacodynamic-pharmacokinetic studies in preclinical and clinical trials. In addition, the plasma stability profile of the compound can help medicinal chemistry researchers modify the molecular structure to improve physicochemical properties and prioritize the molecule for subsequent development.

[0004] Therefore, plasma stability analysis must be performed early in the drug development process to reach optimal therapeutic concentrations in the clinical phase.

[0005] Then, looking further into the conventional plasma stability analysis methods described above, molecular characteristics such as absorption, distribution, metabolism, excretion, and toxicity (ADMET), including plasma stability, are evaluated for drug discovery and development (Daina et al., 2017; Xiong et al., 2021; Yang et al., 2019).

[0006] For example, ML and DL models predicting molecular properties such as blood-brain barrier permeability (Shaker et al., 2021; Tang et al., 2022), cardiotoxicity (Lee et al., 2019; Ryu et al., 2020), metabolic stability (Ryu et al., 2021; Sakiyama et al., 2008), and solubility (Lusci et al., 2013) have been developed to accelerate drug discovery. Recently, there have been numerous advancements in message delivery methods trained to predict molecular properties (Coley et al., 2019; Kireev, 1995; Torng and Altman, 2019). Graph networks such as those mentioned above use dense layers of neural networks as nonlinear functions for message delivery convolution and are generally also referred to as GCNNs (graph-convolutional neural networks).

[0007] Generally, plasma stability analysis methods comprehensively evaluate plasma stability across vast chemical libraries using in vitro or in vivo analysis. However, experimental plasma stability analysis methods have been difficult to develop due to their high cost, time-consuming nature, and labor-intensive nature. Furthermore, because they cannot perform a complete investigation of vast chemical libraries, there is a need to develop in-silico-based plasma stability prediction methods. Particularly in the early stages of new drug development, there is a need to rapidly review the plasma stability of numerous compounds using in-silico plasma stability prediction tools. Despite the importance of this in drug development, the lack of a methodology to predict human plasma stability has posed a significant challenge to plasma stability analysis.

[0008] In addition, regarding the conventional plasma stability analysis methods described above, since most drugs in the human body are chemically converted through hepatic metabolism, computer-based technology has been developed to predict the metabolic stability of human liver microsomes to evaluate this chemical stability. In addition to hepatic metabolism, the degradation of compounds in plasma can be catalyzed by various enzymes such as hydrolases and esterases; furthermore, the stability of liver microsomes may differ from that of plasma because the microsome enzymes differ from those in plasma, and blood contains various hydrolytic enzymes such as cholinesterase, aldolase, lipase, dehydropeptidase, and alkaline and acidic phosphatases (Altman and Dittmer, 1961; Hess, 2016). Therefore, it is necessary to develop a separate prediction system capable of evaluating the plasma stability of drugs.

[0009] In addition, if such compounds have an affinity for a hydrolyzable group at an appropriate position with one of these plasma enzymes, plasma degradation is possible, and certain classes of drug molecules, including esters, amides, lactones, lactams, carbamides, sulfonamides, and peptide mimics, tend to be hydrolyzed by plasma esterases, amidases, or proteases (Di et al., 2005).

[0010] However, the plasma stability of most compounds has not yet been elucidated, and furthermore, interspecies differences in the distribution and activity of these enzymes in plasma can lead to differences in the plasma stability profiles of animal species (Liederer and Borchardt, 2006; Yang et al., 2011), and such plasma stability is difficult to predict because it is significantly influenced by surrounding atoms such as steric hindrance and electron-withdrawing groups (Borthwick et al., 2003; Sawa et al., 2002).

[0011] Therefore, there was a problem that in order to predict the stability of plasma in human plasma, it was necessary to develop a deep learning-based analysis system trained based on local and global information of compounds. The problem to be solved

[0012] Accordingly, the present invention was developed to solve the various problems of the prior art as described above. Its purpose is to provide an attention-based neural network system for predicting the stability of a compound in human plasma and a control method thereof, which can easily predict the stability of a compound in human plasma during the early stages of drug discovery and development by connecting the molecular fingerprint notation of a test compound to a molecular feature vector and simulating it with a fully connected layer to predict the stability of the compound in human plasma through a binary classification of stability or instability.

[0013] Another objective of the present invention as described above is to provide a plasma stability prediction device capable of repeatedly simulating whether a compound exhibits stability in human plasma without requiring a complex and expensive system to predict stability in human plasma, thereby providing an attention-based neural network system for predicting the stability of a compound in human plasma and a method for controlling the same, which can rapidly confirm the reaction exhibited by an inherently unstable prodrug in human plasma. means of solving the problem

[0014] The present invention for achieving the above-mentioned purpose comprises: a SMILES module that, when compound data for which a plasma stability test is desired is input through a set input window, converts the input compound data into a molecular fingerprint representation and a molecular graph representation, respectively, and outputs them;

[0015] A CMPNN encoder (Communicative message passing neural network encoder) that updates all node and edge functions for a test compound indicated on a molecular graph output from the above SMILE module;

[0016] A self-attention mechanism module that generates and outputs a molecular feature vector for a test compound updated by the above CMPNN encoder;

[0017] A simulation module that outputs a result value by simulating the molecular fingerprint notation of a test compound converted and output by the above-mentioned SMILE module with the molecular feature vector of a test compound output by the self-attention mechanism module, and connecting the function to the fully connected layer;

[0018] The present invention provides an attention-based neural network system for predicting the stability of compounds in human plasma, comprising a PredPS neural network control module that predicts human plasma stability by analyzing result values ​​obtained from the simulation module above and classifying stability or instability into binary classifications.

[0019] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the PredPS neural network control module uses a dense layer of the neural network as a non-linear function for message transmission convolution.

[0020] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the SMILE module generates in-house data on the plasma stability of a set of 932 compounds using in vitro analysis of human plasma (785 stable compounds and 147 unstable compounds) under the functional control of the PredPS neural network control module, and further includes the function of collecting and processing open source data on 2,166 compounds (647 stable compounds and 1,519 unstable compounds) for human plasma stability, and then evaluating them all using 5-fold cross-validation through the RS (Random Split) module.

[0021] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the PredPS neural network control module further includes a function to test four ML and DL algorithms—RF (Random Forest), SVM (Support Vector Machine), DMPNN (Directed Message Passing Neural Network) (Yang et al., 2019), and CMPNN (Communicative message passing neural network) (Song et al., 2020)—to predict human plasma stability and evaluate the performance of the attention-based graph neural network.

[0022] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, wherein the SMILE module collects human plasma stability data from public databases PubChem and ChEMBL for model training under the functional control of the PredPS neural network control module, and the collected information provides two types of human plasma stability data, namely the concentration (%) of the compound remaining after a certain period of time and the half-life (t1 / 2).

[0023] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the PredPS neural network control module considers compounds remaining in human plasma at 85% or more after 3 hours to be stable and classifies compounds remaining at less than 85% to be unstable.

[0024] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the PredPS neural network control module classifies compounds with a half-life of 10 hours or more in human plasma as stable and compounds with a half-life of less than 10 hours as unstable.

[0025] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, wherein the PredPS neural network control module maps input data in a high-dimensional feature space and allows linear regression to be performed in said feature space.

[0026] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that both RF and SVM are evaluated using 5-fold cross-validation.

[0027] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of a compound in human plasma, wherein the CMPNN encoder is characterized by encoding structural information of the compound by an MPNN to predict molecular characteristics.

[0028] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the PredPS neural network control module further includes a self-attention layer module and a fully connected layer module optimized for molecular fingerprint notation linked to graph-based notation.

[0029] In addition, another feature of the present invention is to provide an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the SMILE module executes two types of molecular representations, namely molecular fingerprint representation and molecular graph representation, under the functional control of the PredPS neural network control module.

[0030] In addition, another feature of the present invention is a first step in which, under the functional control of the PredPS neural network control module, when compound data to be tested for plasma stability is input through an input window configured by the SMILE module, the input compound data is converted into molecular fingerprint representation and molecular graph representation, respectively, and outputs them;

[0031] A second step in which, after the first step above, the CMPNN encoder updates all node and edge functions for the test compound labeled on the molecular graph output from the SMILE module under the functional control of the PredPS neural network control module;

[0032] A third step in which, after the above second step, the self-attention mechanism module generates and outputs a molecular feature vector for a test compound updated by the CMPNN encoder under the functional control of the PredPS neural network control module;

[0033] A fourth step in which, after the above third step, under the functional control of the PredPS neural network control module, the simulation module connects the molecular fingerprint notation of the test compound converted and output by the SMILE module to the molecular feature vector of the test compound output by the self-attention mechanism module, simulates it with a fully connected layer, and outputs a result value;

[0034] The present invention provides a control method for an attention-based neural network system for predicting the stability of compounds in human plasma, comprising a fifth step in which, after the above-mentioned fourth step, the PredPS neural network control module analyzes the result obtained from the simulation module to predict human plasma stability through a binary classification of stability or instability.

[0035] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that in the fifth step, the PredPS neural network control module uses a dense layer of the neural network as a non-linear function for message transmission convolution.

[0036] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the PredPS neural network control module uses a dense layer of the neural network as a non-linear function for message transmission convolution.

[0037] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the first to fifth steps further include a test step in which the PredPS neural network control module tests four ML and DL algorithms—RF (Random Forest), SVM (Support Vector Machine), DMPNN (Directed Message Passing Neural Network) (Yang et al., 2019), and CMPNN (Communicative message passing neural network) (Song et al., 2020)—to predict human plasma stability and evaluate performance.

[0038] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that the first step further includes a step in which, when the PredPS neural network control module collects human plasma stability data through the SMILE module, the collected information provides two types of human plasma stability data, namely the concentration (%) of the compound remaining after a certain period of time and the half-life (t1 / 2).

[0039] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the fifth step further includes a step in which a PredPS neural network control module considers compounds with a half-life of 10 hours or more in human plasma as stable, while classifying compounds with a half-life of less than 10 hours as unstable.

[0040] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the fifth step further includes a step in which a PredPS neural network control module sets a number of tree structures proposed in metabolic stability for an RF model to predict plasma stability, and an SVM model with a radial basis function (RBF) and a kernel is used for plasma stability.

[0041] In addition, another feature of the present invention is to provide a method for controlling an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that the fifth step further includes a step in which the PredPS neural network control module evaluates both RF and SVM using 5-fold cross-validation.

[0042] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized by further including a step in which structural information of the compound is encoded by an MPNN (Gilmer et al., 2017) to predict molecular characteristics during the evaluation step.

[0043] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that the fifth step further includes an update step in which the PredPS neural network control module updates the edge message vector by subtracting the previous hidden state of the backconnection from the hidden state of the node, and then updates the hidden state of the edge by supplying it to a fully connected layer using the initial hidden state as a bias and a ReLU activation function.

[0044] In addition, another feature of the present invention is to provide a control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that the fifth step further includes a step of predicting the stability of human plasma by binary classification of stability or instability by connecting a representation of a molecular graph to a molecular fingerprint under the functional control of a PredPS neural network control module.

[0046] Specific details of other embodiments are included in "Specific details for implementing the invention" and the attached "drawings".

[0047] The advantages and / or features of the present invention and the methods for achieving them will become clear by referring to the various embodiments described below in detail together with the accompanying drawings.

[0048] However, it should be understood that the present invention is not limited to the configurations of each embodiment disclosed below, but may be implemented in various different forms, and that each embodiment disclosed in this specification is provided merely to make the disclosure of the present invention complete and to fully inform those skilled in the art of the scope of the present invention, and that the present invention is defined only by the scope of each claim of the claims. Effects of the invention

[0049] According to the present invention, by connecting the molecular fingerprint notation of a test compound to a molecular feature vector and simulating it as a fully connected layer to predict the stability of the compound in human plasma through a binary classification of stability or instability, the stability of the compound in human plasma can be easily predicted in the early stages of drug discovery and development, thereby maximizing the usefulness of the drug development platform.

[0050] In addition, according to the present invention as described above, by providing a plasma stability prediction device capable of repeatedly simulating whether a compound exhibits stability in human plasma without requiring a complex and expensive system to predict stability in human plasma, the reaction of an inherently unstable prodrug in human plasma can be rapidly confirmed, thereby having the effect of significantly improving the designability of the prodrug. Brief explanation of the drawing

[0051] FIG. 1 is a diagram schematically illustrating an attention-based neural network system for predicting the stability of a compound in human plasma according to the present invention. FIG. 2 is a flowchart diagram that clearly explains a control method of an attention-based neural network system for predicting the stability of a compound in human plasma according to the present invention. FIG. 3 is a diagram of a relational workflow applied to an attention-based neural network system according to the present invention. FIG. 4 is a diagram showing the overall architecture of a neural network for predicting the stability of a compound in human plasma applied to an attention-based neural network system according to the present invention. FIG. 5 is a diagram illustrating the description of atomic and bonding features applied to an attention-based neural network system according to the present invention. FIG. 6 shows an attention-based neural network system applied to the present invention. A diagram illustrating hyperparameter optimization. FIG. 7 is a diagram showing a graph illustrating the AUC value according to the depth of a message transmission neural network applied to an attention-based neural network system according to the present invention. FIG. 8 is a diagram showing the visualization results of chemical diversity using t-variance stochastic neighbor embeddings (t-SNE) applied to an attention-based neural network system according to the present invention. FIG. 9 is a diagram showing the performance results for an internal data set applied to an attention-based neural network system according to the present invention and the performance results of an existing molecular representation method. FIG. 10 is a diagram showing the stability of plasma in color applied to an attention-based neural network system according to the present invention. FIG. 11 is a drawing showing a user interface of a web server applied to an attention-based neural network system according to the present invention. Specific details for implementing the invention

[0052] Before describing the present invention in detail, it should be understood that the terms and words used in this specification should not be interpreted as being limited to their ordinary or dictionary meanings, and that the inventor of the present invention may appropriately define and use the concepts of various terms to best describe their invention, and furthermore, that these terms and words should be interpreted in a meaning and concept consistent with the technical spirit of the present invention.

[0053] In other words, it should be understood that the terms used in this specification are used merely to describe preferred embodiments of the present invention and are not intended to specifically limit the content of the present invention, and that these terms are defined in consideration of various possibilities of the present invention.

[0054] In addition, it should be noted that in this specification, singular expressions may include plural expressions unless the context clearly indicates a different meaning, and that even if they are expressed in a similarly plural form, they may include a singular meaning.

[0055] Throughout this specification, where it is stated that a component "includes" another component, unless specifically stated otherwise, this may mean that it does not exclude any other component but may include any other component.

[0056] Furthermore, it should be noted that in cases where it is stated that a component "exists inside or is installed in connection with" another component, this component may be installed in direct connection or contact with the other component, or it may be installed at a certain distance apart, and in the case where it is installed at a certain distance apart, there may be a third component or means for fixing or connecting the component to the other component, and a description of this third component or means may be omitted.

[0057] On the other hand, if it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there is no third component or means.

[0058] Likewise, other expressions describing the relationship between each component, such as “between” and “right between”, or “adjacent to” and “directly adjacent to”, should be interpreted as having the same intent.

[0059] In addition, it should be understood that in this specification, terms such as “one side,” “other side,” “one side,” “other side,” “first,” “second,” etc., are used to clearly distinguish one component from another component, and that the meaning of the component is not restricted by such terms.

[0060] In addition, position-related terms such as "up," "down," "left," and "right" used in this specification should be understood as indicating the relative position of the corresponding component in the drawing, and unless an absolute position is specified, these position-related terms should not be understood as referring to an absolute position.

[0061] Furthermore, in specifying the reference numerals for each component of each drawing in this specification, the same component has the same reference numeral even if it is shown in different drawings; that is, the same reference numeral throughout the specification indicates the same component.

[0062] In the drawings attached to this specification, the size, position, connection relationships, etc., of each component constituting the present invention may be described in a partially exaggerated, reduced, or omitted manner for the convenience of explanation or to sufficiently clearly convey the concept of the present invention, and therefore, the proportions or scale may not be strictly accurate.

[0063] In addition, in describing the present invention below, detailed descriptions of components, such as prior art and known technology, that are deemed to unnecessarily obscure the essence of the invention may be omitted.

[0065] Hereinafter, embodiments of the present invention will be described in detail with reference to the relevant drawings.

[0066] FIG. 1 is a diagram schematically illustrating an attention-based neural network system for predicting the stability of a compound in human plasma according to the present invention.

[0067] Referring to FIGS. 2 to 11, the system of the present invention comprises a SMILES module (2) that, when compound data for which a plasma stability test is desired is input through a set input window (1), converts the input compound data into a molecular fingerprint representation and a molecular graph representation, respectively, and outputs them;

[0068] A CMPNN encoder (Communicative message passing neural network encoder: 3) that updates all node and edge functions for a test compound represented by a molecular graph output from the above-mentioned smile module (2);

[0069] A self-attention mechanism module (4) that generates and outputs a molecular feature vector for a test compound updated by the above CMPNN encoder (3);

[0070] A simulation module (5) that outputs a result value by simulating the molecular fingerprint notation of the test compound converted and output by the above-mentioned smile module (2) with the molecular feature vector of the test compound output by the self-attention mechanism module (4) and simulating it into a completely connected layer;

[0071] It is configured to include a PredPS neural network control module (6) that predicts human plasma (or plasma) stability by analyzing the result value obtained from the simulation module (5) and classifying the stability or instability in binary.

[0072] Here, the PredPS neural network control module (6) further includes using a dense layer (7) of the neural network as a non-linear function for message transmission convolution. Additionally, the PredPS neural network control module (6) controls the overall function of an attention-based neural network system for predicting the stability of compounds in human plasma according to the present invention. Furthermore, the PredPS neural network control module (6) may be configured to further include a memory (8) that stores instructions to perform at least one step under the functional control of the PredPS neural network control module (6) and a program (including artificial intelligence action) set to predict the stability of compounds in human plasma, a key panel (9) for inputting user command signals, and a display (10) that displays process signals processed within the system externally.

[0073] Here, SMILES is an abbreviation for Simplified Molecular Input Line Entry System, a notation that represents molecular structures in the form of strings, enabling the application of various machine learning / deep learning algorithms used in NLP to molecular structures (refer to Google search).

[0074] And the above-mentioned Smile Module (2) further includes the function of generating in-house data on the plasma stability of various sets of 932 compounds using in vitro analysis of human plasma (785 stable compounds and 147 unstable compounds) as shown in FIG. 3 under the functional control of the PredPS neural network control module (6), and also collecting and processing open source data on 2166 compounds (647 stable compounds and 1519 unstable compounds) for human plasma stability, and then evaluating them all using 5-fold cross-validation through the RS (Random Split) module.

[0075] Furthermore, the attention-based neural network system according to the present invention comprises an attention-based graph neural network (PredPS) to predict human plasma stability, and additionally includes a function to test four ML and DL algorithms, such as RF (Random Forest), SVM (Support Vector Machine), DMPNN (Directed Message Passing Neural Network) (Yang et al., 2019) and CMPNN (Communicative message passing neural network) (Song et al., 2020), to evaluate the performance of the attention-based graph neural network.

[0076] Here, the attention-based neural network system according to the present invention, which includes a neural network (CMPNN) (Song et al., 2020), exhibits the highest region under the receiver operating characteristic curve (AUC) of 0.901±0.006 when evaluated, for example, using 5-fold cross-validation.

[0077] And the attention-based neural network system according to the present invention is utilized, for example, to predict human plasma stability through a deep learning-based model. At this time, the model used in the attention-based neural network system according to the present invention can be used for binary class prediction of compounds and high-throughput screening of compounds in the early stages of drug discovery.

[0078] In addition, the above-mentioned Smile Module (2) collects human plasma stability data from the public databases PubChem (Kim et al., 2021) and ChEMBL (Mendez et al., 2019) for model training under the functional control of the PredPS neural network control module (6). The information collected at this time provides two types of human plasma stability data: the concentration (%) of the compound remaining after a certain period of time and the half-life (t1 / 2).

[0079] In addition, the above SMILES module (2) first standardizes the format of the Simplified Molecular Input Line Input System (SMILES) of all collected compounds using RDKit (www.rdkit.org) and MolVS (github.com / mcs07 / MolVS) after data collection under the functional control of the PredPS neural network control module (6).

[0080] And the above-mentioned in-house and open-source datasets. The standardization process involves selecting the largest fragment, clear hydrogen removal, ionization, and stereochemical calculations. In the case of the above-mentioned in-house dataset, compounds remaining in human plasma at a rate of 85% or more after 3 hours were considered to have a stable structure, while compounds remaining at a rate of less than 85% were considered to have an unstable structure (Di et al., 2005; Kang et al., 2019). In the case of the above-mentioned open-source dataset, a compound was considered stable if 85% or more of it remained in human plasma after 3 or 4 hours.

[0081] Meanwhile, the above PredPS neural network control module (6) considers compounds with a half-life of 10 hours or more in human plasma to be stable and classifies compounds with a half-life of less than 10 hours to be unstable.

[0082] In addition, the attention-based neural network system according to the present invention uses RF, SVM, DMPNN, and CMPNN (Song et al., 2020) as four basic methods and compares them with conventional ML methods and conventional graph convolution networks.

[0083] Here, the RF is set as a supervised learning algorithm that includes an ensemble of decision trees generated from bootstrap sampling of features.

[0084] In addition, the above RF is considered the "gold standard" in structure-property relationship studies due to its robustness, ease of application, and high predictive accuracy (Ma et al., 2015; Polishchuk, 2017; Ryu et al., 2021).

[0085] In addition, the above SVM method was proposed by Vapnik and is based on the structural risk minimization principle (Vapnik, 1999). The estimation function is a linear extension of a function defined for a specific data set (support vector).

[0086] The input data of the attention-based neural network system according to the present invention is mapped to a high-dimensional feature space, and linear regression further includes a function performed in the feature space.

[0087] The above ECFP with a fixed length of 1024 is used with an RF model and SVM implemented in Python 3.6.13 with Scikit-learn package version 0.24.2 (Pedregosa et al., 2011).

[0088] For example, in the case of the above RF model, 500 trees proposed for metabolic stability were established (Ryu et al., 2021). The radial basis function (RBF) and the SVM model with the kernel further include functions used for plasma stability.

[0089] The above RF and SVM both include additional features evaluated using 5-fold cross-validation.

[0090] It also includes a feature in which structural information of compounds is encoded by MPNN (Gilmer et al., 2017) to predict molecular properties.

[0091] In addition, the above CMPNN was developed to improve the insufficient representation of molecular graph properties in DMPNN (Song et al., 2020) and has the ability to enhance node-edge interactions using a 'communication' kernel.

[0092] Meanwhile, the attention-based neural network system according to the present invention is largely configured to include a CMPNN encoder (3), a self-attention layer, and a fully connected layer module optimized for molecular fingerprint notation connected to graph-based notation (see FIG. 4).

[0093] In addition, the above-mentioned smile module (2) further includes the function of executing two types of molecular representations, molecular fingerprint representation and molecular graph representation, under the functional control of the PredPS neural network control module (6).

[0094] In this case, for the molecular fingerprint, the SMILES string is converted into a 2048-bit binary feature vector using the Morgan fingerprint (Morgan, 1965) provided by the Python package RDKit (http: / / www.rdkit.org). On the other hand, for the molecular graph, all node and edge features are initialized using atomic and bonding properties, respectively (see Fig. 5).

[0095] For all nodes in the graph above, the node message vector is updated by an aggregation function based on a message booster (Song et al., 2020) using the previous hidden state of all neighboring nodes. The hidden state of each node is updated by a communication function using the message vector and the previous hidden state.

[0096] The attention-based neural network system according to the present invention adopts a multilayer perceptron as the communication function that demonstrates the best performance benchmarked by Song et al. (Song et al, 2020). Subsequently, the edge message vector is updated by subtracting the previous hidden state of the backconnection from the hidden state of the node. Then, the hidden state of the edge is updated by supplying it to a fully connected layer using the initial hidden state as a bias and a ReLU activation function.

[0097] The above procedure updates the hidden state of the nodes and edges 5 times, and the final message vector and hidden state vector are calculated using aggregation and communication functions, respectively.

[0098] In addition, the reading procedure for all the above nodes generates molecular feature vectors by applying a self-attention mechanism method (Velickovic et al., 2017).

[0099] Meanwhile, the attention-based neural network system according to the present invention further includes a function to predict human plasma stability by binary classification of stability or instability by connecting a representation of a molecular graph to a molecular fingerprint and training a fully connected layer under the functional control of the PredPS neural network control module (6).

[0100] Accordingly, the attention-based neural network system according to the present invention integrated in-house and open-source datasets to prepare training datasets with various complex structures, taking into account the problems of the prior art described above. At the same time, the ratio of stable to unstable compounds was also prepared similarly.

[0101] Therefore, the attention-based neural network system according to the present invention achieves, for example, an overall accuracy of 0.835±0.007, an AUC of 0.901±0.006, a sensitivity of 0.823±0.054, and a specificity of 0.846±0.049 in 5-fold cross-validation (see FIG. 9).

[0102] In addition, the attention-based neural network system according to the present invention produced results that surpassed existing ML models (RF and SVM) and existing graph-based neural networks (MPNN and CMPNN). Furthermore, the attention-based neural network system according to the present invention as described above is configured by connecting a self-attention layer to the CMPNN model architecture.

[0103] In particular, the attention-based neural network system according to the present invention further includes a function applied by an attention layer to capture the importance of substructures when determining plasma stability, instead of simply combining all learned representations by sum pooling after a message-transmitting neural encoder.

[0104] And it was confirmed that the attention-based neural network system according to the present invention using attention pooling has a higher AUC value than CMPNN alone.

[0105] In addition, all performance metrics were improved when fingerprint features obtained using the Morgan algorithm (Morgan, 1965) were connected to the final graph representation (see Fig. 9).

[0106] The molecular fingerprint of the attention-based neural network system according to the present invention can provide clear structural information by capturing characteristics related to the molecular substructure, including aromatic rings and functional groups.

[0107] Here, the attention-based neural network system according to the present invention obtains attention weight scores by a mechanism performed by a self-attention module to identify learned functions. It further includes a function to examine attention patterns to evaluate whether the model has focused on specific molecular substructures to predict plasma stability.

[0108] At this point, six unstable chemicals were randomly selected from the training set to analyze the attention patterns. The most unstable compounds in plasma contain ester bonds (Di et al., 2005). Visualizing the attention weight scores confirms that there is a local concentration on atoms constituting ester bonds, consistent with the model's predictions (see Fig. 10).

[0109] In addition, the attention-based neural network system according to the present invention further includes the function of creating a publicly accessible web server to predict the human plasma stability of a requested compound.

[0110] The web server according to the present invention as described above accepts the SMILES format of a query compound as input and returns a binary classification result as stable or unstable, and the user interface of the web server according to the present invention can be configured as shown in FIG. 11.

[0111] Next, a control method of an embodiment of the present invention configured as described above will be explained.

[0112] FIG. 2 is a flowchart diagram that clearly explains the control method of an attention-based neural network system for predicting the stability of a compound in human plasma according to the present invention.

[0113] Referring to FIG. 2, the method of the present invention comprises a first step (S101) in which, when compound data to be tested for plasma stability is input through an input window (1) in which a smile module (2) is configured under the functional control of a PredPS neural network control module (6), the input compound data is converted into a molecular fingerprint representation and a molecular graph representation, respectively, and outputs them;

[0114] A second step (S102) in which, after the first step (S101) above, the CMPNN encoder updates all node and edge functions for the test compound labeled on the molecular graph output from the SMILE module under the functional control of the PredPS neural network control module;

[0115] A third step (S103) in which, after the second step (S102) above, the self-attention mechanism module generates and outputs a molecular feature vector for a test compound updated by the CMPNN encoder under the functional control of the PredPS neural network control module;

[0116] A fourth step (S104) in which, after the above third step (S103), under the functional control of the PredPS neural network control module, the simulation module connects the molecular fingerprint notation of the test compound converted and output by the SMILE module to the molecular feature vector of the test compound output by the self-attention mechanism module, simulates it with a fully connected layer, and outputs a result value;

[0117] After the above fourth step (S104), the PredPS neural network control module analyzes the result obtained from the simulation module and predicts human plasma stability through binary classification of stability or instability, comprising a fifth step (S105).

[0118] Here, step 5 (S105) further includes the PredPS neural network control module using a dense layer of the neural network as a non-linear function for message passing convolution.

[0119] Here, the PredPS neural network control module further includes using a dense layer of the neural network as a non-linear function for message passing convolution.

[0120] And the above first step (S101) to fifth step (S105) further include a test step in which the PredPS neural network control module tests four ML and DL algorithms, RF (Random Forest), SVM (Support Vector Machine), DMPNN (Directed Message Passing Neural Network) (Yang et al., 2019) and CMPNN (Communicative message passing neural network) (Song et al., 2020), to predict human plasma stability and evaluate performance.

[0121] In addition, the first step (S101) further includes an evaluation step in which the PredPS neural network control module generates in-house data on the plasma stability of various sets of 932 compounds using in vitro analysis of human plasma (785 stable compounds and 147 unstable compounds) through the SMILE module, and also collects and processes open source data on 2166 compounds (647 stable compounds and 1519 unstable compounds) for human plasma stability prediction, and then evaluates all of them using 5-fold cross-validation through the RS (Random Split) module.

[0122] In addition, the first step (S101) further includes a step in which, when the PredPS neural network control module collects human plasma stability data through the Smile module, the collected information provides two types of human plasma stability data, namely the concentration (%) of the compound remaining after a certain period of time and the half-life (t1 / 2).

[0123] Meanwhile, the above 5th step (S105) further includes a step in which the PredPS neural network control module considers compounds with a half-life of 10 hours or more in human plasma as stable, while classifying compounds with a half-life of less than 10 hours as unstable.

[0124] And the above 5th step (S105) further includes a step of comparing the PredPS neural network control module with conventional ML methods and conventional graph convolution networks using four basic methods: RF, SVM, DMPNN, and CMPNN (Song et al., 2020).

[0125] In addition, the above 5th step (S105) further includes a step in which the PredPS neural network control module sets up a number of trees (e.g., 500 trees) proposed in the case of the RF model for metabolic stability to predict plasma stability, and an SVM model with a radial basis function (RBF) and a kernel is used for plasma stability.

[0126] And the above 5th step (S105) further includes a step of ensuring that the PredPS neural network control module evaluates both RF and SVM using 5-fold cross-validation.

[0127] In addition, the above evaluation step further includes a step in which structural information of the compound is encoded by MPNN (Gilmer et al., 2017) to predict molecular characteristics.

[0128] In addition, the above 5th step (S105) further includes an update step in which the PredPS neural network control module updates the edge message vector by subtracting the previous hidden state of the backconnection from the hidden state of the node, and then updates the hidden state of the edge by supplying it to the fully connected layer using the initial hidden state as a bias and a ReLU activation function.

[0129] And in the updater step, the hidden state of the node and edge is updated 5 times, and the final message vector and the hidden state vector are calculated using aggregation and communication functions, respectively, and the reading procedure for all the nodes further includes a step of generating a molecular feature vector by applying a self-attention mechanism method.

[0130] Meanwhile, the above 5th step (S105) further includes a step of predicting human plasma stability by binary classification of stability or instability by connecting the representation of the molecular graph to the molecular fingerprint under the functional control of the PredPS neural network control module and training the fully connected layer.

[0131] Furthermore, the above 5th step (S105) further includes a step of applying an attention layer to capture the importance of substructures when determining plasma stability, instead of simply combining all learned representations by sum pooling after message transmission neural encoding under the functional control of the PredPS neural network control module.

[0132] And the above 5th step (S105) further includes a step of providing clear structural information by capturing characteristics related to the molecular substructure, including aromatic rings and functional groups, of the molecular fingerprint under the functional control of the PredPS neural network control module.

[0133] Here, the above 5th step (S105) further includes the step of the PredPS neural network control module obtaining an attention weight score by means of a mechanism performed by the self-attention module to identify learned features and examining the attention pattern to evaluate whether the model has focused on a specific molecular substructure to predict plasma stability.

[0134] In other words, the method of an attention-based neural network system for predicting the stability of a compound in human plasma according to the present invention first, in the first step (S101), when compound data to be tested for plasma stability is input through an input window set by a SMILES module under the functional control of a PredPS neural network control module (6), the input compound data is converted into a molecular fingerprint notation and a molecular graph notation, respectively, and output.

[0135] And in the second step (S102), the CMPNN encoder (3) updates all node and edge functions for the test compound labeled in the molecular graph output from the smile module (2) under the functional control of the PredPS neural network control module (6).

[0136] In addition, in the third step (S103), the self-attention mechanism module (4) generates and outputs a molecular feature vector for the test compound updated by the CMPNN encoder under the functional control of the PredPS neural network control module (6).

[0137] Furthermore, in the fourth step (S103), the simulation module (5), under the functional control of the PredPS neural network control module (6), connects the molecular fingerprint notation of the test compound converted and output by the smile module (2) to the molecular feature vector of the test compound output by the self-attention mechanism module (4), simulates it as a fully connected layer, and outputs a result value.

[0138] In addition, in step 5 (S105), the PredPS neural network control module (6) analyzes the result obtained from the simulation module (5) and predicts human plasma stability through a binary classification of stability or instability.

[0139] Meanwhile, the development, structure, and implementation process of the attention-based neural network system for predicting the stability of compounds in human plasma according to the present invention are described in more detail as follows.

[0140] The stability of compounds in human plasma is important for maintaining sufficient systemic drug exposure and is considered an essential factor in the early stages of drug discovery and development. Rapid degradation of compounds in plasma can lead to reduced in vivo efficacy.

[0141] However, since there are currently no open source software programs available to predict human plasma stability, the attention-based neural network system according to the present invention proposes a human plasma stability prediction model that can be accessed through a web service.

[0142] That is, the attention-based neural network system according to the present invention as described above proposes an attention-based graph neural network to predict the plasma stability of a compound in human plasma using in-house and open source datasets.

[0143] The attention-based neural network system according to the present invention exhibited superior performance compared to two machine learning and two deep learning algorithms used in a comparison representing stability prediction efficiency.

[0144] In addition, the attention-based neural network system according to the present invention achieved, for example, a region under the receiver operating characteristic curve of 0.901±0.006, an accuracy of 0.835±0.007, a sensitivity of 0.823±0.054, and a specificity of 0.846±0.049 when evaluated using 5-fold cross-validation.

[0145] Meanwhile, to summarize the process of developing the attention-based neural network system according to the present invention,

[0146] The stability of compounds in human plasma plays a crucial role in drug discovery and development; compounds that degrade rapidly in plasma tend to have low bioavailability and low in vivo efficacy. Furthermore, poor plasma stability during sample storage or analysis can lead to inaccurate measurements of in vivo drug concentrations. This plasma stability is considered a critical factor in new drug development because the accurate determination of drug concentrations in biological samples is essential for pharmacodynamic-pharmacokinetic studies in preclinical and clinical trials. Additionally, the plasma stability profile of a compound can assist medicinal chemistry researchers in modifying its molecular structure to improve physicochemical properties and prioritizing molecules for subsequent development (Di et al., 2005).

[0147] Therefore, plasma stability analysis as described above must be performed early in the drug development process to reach the optimal therapeutic concentration in the clinical stage.

[0148] However, comprehensively evaluating plasma stability for vast chemical libraries using in vitro or in vivo assays is difficult because it is costly, time-consuming, and laborious. In particular, the development of in-silico plasma stability prediction methods is necessary because a complete investigation of vast chemical libraries cannot be performed.

[0149] Therefore, it is necessary to rapidly review the plasma stability of numerous compounds using in silico plasma stability prediction tools during the early stages of new drug development. However, despite its importance in drug development, open-source software capable of predicting human plasma stability has not been developed to date.

[0150] Most drugs in the human body are chemically converted through hepatic metabolism, and computer-based technology has recently been developed to predict the metabolic stability of human liver microsomes to evaluate chemical stability (Li et al., 2022; Ryu et al., 2021). In addition to the aforementioned hepatic metabolism, the degradation of compounds in plasma can be catalyzed by various enzymes such as hydrolases and esterases. The stability of liver microsomes may differ from that of plasma because the enzymes in the microsomes are different from those in the plasma.

[0151] Human blood contains various hydrolytic enzymes such as cholinesterase, aldolase, lipase, dehydrogenpeptidase, alkaline and acidic phosphatase (Altman and Dittmer, 1961; Hess, 2016).

[0152] And plasma degradation is possible if the above compound has an affinity for a hydrolyzable group at an appropriate position with one of these plasma enzymes. Certain classes of drug molecules, including esters, amides, lactones, lactams, carbamides, sulfonamides, and peptide mimics, tend to be hydrolyzed by plasma esterases, amidases, or proteases (Di et al., 2005).

[0153] However, the plasma stability of many other structures has not yet been elucidated. Furthermore, interspecies differences in the distribution and activity of these enzymes in plasma, as described above, can lead to differences in the plasma stability profiles of animal species (Liederer and Borchardt, 2006; Yang et al., 2011).

[0154] In addition, plasma stability as described above is difficult to predict because it is significantly affected by surrounding atoms such as steric hindrance and electron-drawout groups (Borthwick et al., 2003; Sawa et al., 2002).

[0155] Therefore, to predict plasma stability in human plasma, it is necessary to develop deep learning-based computational tools trained based on local and global information of compounds.

[0156] The above plasma stability can be used to profile prodrugs that require rapid conversion from plasma. And medicinal chemists can utilize plasma reactions as part of a prodrug approach.

[0157] In addition, the prodrug improves permeability and oral bioavailability to allow high concentrations of the prodrug to reach the bloodstream. The hydrolytic enzymes mentioned above cleave the prodrug to release the active drug into the blood.

[0158] For example, remdesivir, recently approved as an antiviral for COVID-19, utilizes a novel approach of delivering a phosphate prodrug via intravenous administration to minimize payload release during transport and maximize tissue payload distribution (Wiemer, 2020). Therefore, the plasma stability prediction tool mentioned above can help in rapidly designing appropriate prodrugs that are inherently unstable in plasma.

[0159] The above plasma stability test is used to evaluate molecular properties such as absorption, distribution, metabolism, excretion, and toxicity (ADMET) for drug discovery and development (Daina et al., 2017; Xiong et al., 2021; Yang et al., 2019).

[0160] For example, ML and DL models predicting molecular properties such as blood-brain barrier permeability (Shaker et al., 2021; Tang et al., 2022) and cardiotoxicity (Lee et al., 2019; Ryu et al., 2020), metabolic stability (Ryu et al., 2021; Sakiyama et al., 2008), and solubility (Lusci et al., 2013) have been developed to accelerate drug discovery. Recently, there have been numerous advancements in message delivery methods trained to predict molecular properties (Coley et al., 2019; Kireev, 1995; Torng and Altman, 2019).

[0161] These graph networks use dense layers of neural networks as non-linear functions for message passing convolution and are generally also called GCNN (graph-convolutional neural network).

[0162] Furthermore, graph-based models are naturally suitable for molecular modeling because they can model atoms as nodes and bonds as edges in a mathematical graph to represent molecules.

[0163] The main advantage of the aforementioned GCNN is that it considers distant information through iterative message passing while avoiding local dependencies of descriptor-based models, such as molecular fingerprints. The GCNN outperforms previous descriptor-based ML approaches in various molecular feature prediction tasks (Song et al., 2020; Wang et al., 2022; Wu et al., 2018; Yang et al., 2019). Additionally, there have been various attempts to combine attention mechanism modules with GCNN to enhance prediction performance by capturing global dependencies between features of the substructure (Du et al., 2022; Kim et al., 2022; Li et al., 2019; Tang et al., 2020).

[0164] Therefore, the attention-based neural network system according to the present invention has developed an attention-based graph neural network that predicts the plasma stability of a given compound in human plasma and classifies the compound as stable or unstable, taking into account the above needs (see FIG. 3).

[0165] And, as shown in FIG. 3, the attention-based neural network system according to the present invention first generated in-house data on the plasma stability of various sets of 932 compounds using in vitro analysis of human plasma (785 stable compounds and 147 unstable compounds), and also collected open source data on 2,166 compounds (647 stable compounds and 1,519 unstable compounds) for human plasma stability.

[0166] Afterwards, the attention-based neural network system according to the present invention configured an attention-based graph neural network (PredPS) to predict human plasma stability, and to evaluate the performance of the attention-based graph neural network, four ML and DL algorithms were also tested: RF (Random Forest), SVM (Support Vector Machine), DMPNN (Directed Message Passing Neural Network) (Yang et al., 2019), and CMPNN (Communicative message passing neural network) (Song et al., 2020).

[0167] Here, it was confirmed that the attention-based neural network system according to the present invention, which includes the neural network (CMPNN) as described above (Song et al., 2020), exhibited the highest region under the receiver operating characteristic curve (AUC) of 0.901±0.006 when evaluated, for example, using 5-fold cross-validation.

[0168] The attention-based neural network system according to the present invention is utilized, for example, to predict human plasma stability through a deep learning-based model. In this case, the model used in the attention-based neural network system according to the present invention can be used for binary class prediction of compounds and large-scale screening of compounds in the early stages of drug discovery.

[0169] [In Vitro Human Plasma Stability Analysis]

[0170] All compounds used for the human plasma safety analysis in the attention-based neural network system according to the present invention were obtained from the Korea Research Institute of Chemical Technology (Daejeon, Korea). Pooled plasma was purchased from Innovative Research, Inc. (Novie, Michigan, USA). The compounds for testing were added to pre-cultured 100% plasma (pH 7.4) to produce a final compound concentration of 2 μM with a final DMSO concentration of 2%. Additionally, the spiked plasma samples were cultured at 37°C, and the reaction was terminated by adding a sufficient volume of acetonitrile containing disopyramide as an internal standard. Furthermore, the compound concentration of the supernatant was analyzed by LC-MS / MS after centrifugation at 4000 rpm for 20 minutes at 4°C.

[0171] [Open Source Human Plasma Safety]

[0172] The attention-based neural network system according to the present invention collected human plasma stability data from the public databases PubChem (Kim et al., 2021) and ChEMBL (Mendez et al., 2019) for model training. The information collected provides two types of human plasma stability data: the concentration (%) of the compound remaining after a certain period of time and the half-life (t1 / 2).

[0173] [Data Preparation]

[0174] The attention-based neural network system according to the present invention first standardized the Simplified Molecular Input Line System (SMILES) format of all collected compounds using RDKit (www.rdkit.org) and MolVS (https: / / github.com / mcs07 / MolVS) after data collection, and in-house and open source datasets. The standardization process included the selection of the largest fragment, clear hydrogen removal, ionization, and stereochemical calculations.

[0175] In this case, for the aforementioned in-house dataset, compounds remaining in human plasma at a rate of 85% or more after 3 hours were considered to have a stable structure, whereas compounds remaining at a rate of less than 85% were considered to have an unstable structure (Di et al., 2005; Kang et al., 2019). For the aforementioned open source dataset, a compound was considered stable if 85% or more of it remained in human plasma after 3 or 4 hours.

[0176] For example, assuming that the above compound decreases linearly to 85% in plasma within 3 to 4 hours, the half-life is about 10 hours.

[0177] Accordingly, in the attention-based neural network system according to the present invention, compounds with a half-life of 10 hours or more in human plasma are considered stable, and compounds with a half-life of less than 10 hours are classified as unstable.

[0178] In addition, the attention-based neural network system according to the present invention was compared with the ML method of the prior art and the graph convolution network of the prior art by using RF, SVM, DMPNN, and CMPNN (Song et al., 2020) as four basic methods as baselines.

[0179] Here, the RF is set as a supervised learning algorithm comprising an ensemble of decision trees generated from bootstrap sampling of features. Furthermore, the RF is considered the "gold standard" in structure-feature relationship research due to its robustness, ease of application, and high prediction accuracy (Ma et al., 2015; Polishchuk, 2017; Ryu et al., 2021).

[0180] Meanwhile, the above SVM method was proposed by Vapnik and is based on the structural risk minimization principle (Vapnik, 1999). The estimation function is a linear extension of a function defined for a specific data set (support vector).

[0181] The input data of the attention-based neural network system according to the present invention was mapped to a high-dimensional feature space, and linear regression was performed in the feature space. The ECFP with a fixed length of 1024 was used with an RF model and an SVM implemented in Python 3.6.13 with the Scikit-learn package version 0.24.2 (Pedregosa et al., 2011).

[0182] For example, in the case of the above RF model, 500 trees proposed for metabolic stability were established (Ryu et al., 2021). A radial basis function (RBF) and an SVM model with the above kernel were used for plasma stability. Both the RF and SVM were evaluated using 5-fold cross-validation.

[0183] In addition, structural information of the compound was encoded by MPNN (Gilmer et al., 2017) to predict molecular properties.

[0184] Here, the aforementioned MPNN refers to a method in graph convolution that continuously updates node information corresponding to atoms when a molecular structure is represented as a graph.

[0185] The attention-based neural network system according to the present invention compares the performance of the MPNN variants DMPNN and CMPNN by setting them as graph-based baseline methods. The MPNN focuses on achieving node (atom) embeddings while ignoring information transmitted by edges (connections).

[0186] The above D-MPNN uses messages associated with directional edges (connections) (Yang et al., 2019) to compensate for the problems of the MPNN mentioned above.

[0187] At this time, the central node of the graph updates its information by mixing edge information of neighboring nodes connected by molecular bonds.

[0188] Furthermore, the above information is transmitted according to the surrounding environment of each node, and by repeating this process multiple times, the structural features of the molecule can be effectively encoded.

[0189] In addition, the above CMPNN was developed to improve the insufficient representation of molecular graph properties in DMPNN (Song et al., 2020). The method of the above CMPNN is to enhance node-edge interactions using a 'communication' kernel.

[0190] The above DMPNN was implemented using source code obtained from ChemProp (https: / / github.com / chemprop / chemprop), and the above CMPNN was implemented using source code obtained from https: / / github.com / SY575 / CMPNN.

[0191] Meanwhile, the attention-based neural network system according to the present invention is largely composed of a CMPNN encoder (3), a self-attention layer, and a fully connected layer module optimized for molecular fingerprint representation connected to a graph-based representation (see FIG. 4).

[0192] The attention-based neural network system according to the present invention considered two types of molecular representations, molecular fingerprints and graphs. For the molecular fingerprint, the SMILES string was converted into a 2048-bit binary feature vector using the Morgan fingerprint (Morgan, 1965) provided by the Python package RDKit (http: / / www.rdkit.org).

[0193] In addition, for the molecular graph, all node and edge functions were initialized using atomic and bonding properties, respectively (see Fig. 5).

[0194] For all nodes in the graph above, the node message vector is updated by an aggregation function based on a message booster (Song et al., 2020) using the previous hidden state of all neighboring nodes. The hidden state of each node is updated by a communication function using the message vector and the previous hidden state.

[0195] The attention-based neural network system according to the present invention adopts a multilayer perceptron as the communication function that demonstrates the best performance benchmarked by Song et al. (Song et al, 2020). Subsequently, the edge message vector is updated by subtracting the previous hidden state of the backconnection from the hidden state of the node. Then, the hidden state of the edge is updated by supplying it to the fully connected layer using the initial hidden state as the bias and the ReLU activation function.

[0196] The above procedure updates the hidden state of the nodes and edges 5 times, and the final message vector and hidden state vector are calculated using aggregation and communication functions, respectively.

[0197] Next, a molecular feature vector was generated by applying a self-attention method to the readout procedure for all nodes (Velickovic et al., 2017).

[0198] Finally, the attention-based neural network system according to the present invention connected a representation of a molecular graph to a molecular fingerprint and trained a fully connected layer to predict human plasma stability as a binary classification of stability or instability.

[0199] Here, more detailed information regarding the embedding information and hyperparameters as described above is provided in FIGS. 6 and FIGS. 7.

[0200] Meanwhile, to summarize the attention-based neural network system according to the present invention more simply, in-house and open source datasets are prepared for training a human plasma stability prediction model. In order to build a generalizable and powerful deep learning model, a sufficient amount of unbiased input data with various characteristics is required. At this time, high-quality input data was prepared by integrating an in-house dataset obtained from in vitro human plasma stability measurements with a human plasma stability dataset from a public database.

[0201] In particular, regarding the in-house dataset, the quality of the dataset was determined to be excellent as a result of measuring the human plasma stability of compounds with various scaffolds under consistent analytical conditions at the Korea Chemical Bank (www.chembank.org).

[0202] To investigate the diversity of molecular characteristics as described above, high-dimensional data were projected into a low-dimensional space using t-distributed stochastic neighbor embedding (t-SNE), a dimensionality reduction method. Molecular representations based on Morgan fingerprints were used as inputs for t-SNE for 3,098 compounds (see Fig. 8). Furthermore, as can be seen in Fig. 8, compounds from the open-source dataset appeared as clusters, indicating that several compounds possess similar molecular characteristics. In contrast, compounds from the in-house dataset were relatively more evenly distributed in the t-SNE.

[0203] On the other hand, plasma stability data from open-source datasets showed a high proportion of unstable compounds, primarily prodrugs. In contrast, the proportion of stable compounds is high in the in-house dataset.

[0204] Therefore, training data with low chemical diversity or class imbalance can lead to overfitting and model generalization problems (Ali et al., 2013).

[0205] Accordingly, the attention-based neural network system according to the present invention integrated in-house and open-source datasets to prepare training datasets with various complex structures, taking into account the problems of the prior art described above. At the same time, the ratio of stable to unstable compounds was also prepared similarly.

[0206] Therefore, the attention-based neural network system according to the present invention achieved, for example, an overall accuracy of 0.835±0.007, an AUC of 0.901±0.006, a sensitivity of 0.823±0.054, and a specificity of 0.846±0.049 in 5-fold cross-validation (Table 1-Fig. 9).

[0207] In addition, the attention-based neural network system according to the present invention produced results that surpassed existing ML models (RF and SVM) and existing graph-based neural networks (MPNN and CMPNN). Furthermore, the attention-based neural network system according to the present invention as described above is configured by connecting a self-attention layer module to the CMPNN model architecture.

[0208] In particular, the attention-based neural network system according to the present invention applies an attention layer to capture the importance of substructures when determining plasma stability, instead of simply combining all learned representations with sum pooling after the message transmission neural encoder.

[0209] And it was confirmed that the attention-based neural network system according to the present invention using attention pooling has a higher AUC value than CMPNN alone.

[0210] In addition, all performance metrics were improved when fingerprint features obtained using the Morgan algorithm (Morgan, 1965) were connected to the final graph representation (see Fig. 9).

[0211] The molecular fingerprint of the attention-based neural network system according to the present invention can provide clear structural information by capturing characteristics related to the molecular substructure, including aromatic rings and functional groups.

[0212] In addition, the attention-based neural network system according to the present invention calculated sensitivity (recall) to estimate the risk of false negatives, as predicting actual unstable compounds as stable is more serious. False negatives can mislead medicinal chemists, causing them to continue futile work on unstable compounds and waste time and resources.

[0213] The attention-based neural network system according to the present invention had a sensitivity of 82.3% on the test set. The higher the sensitivity score as described above, the lower the risk of false negatives. High AUC and sensitivity values ​​indicate that the attention-based neural network system according to the present invention demonstrates high accuracy in predicting human plasma safety, which is sufficient for ADMET screening in the early stages of drug discovery.

[0214] Here, the attention-based neural network system according to the present invention obtained attention weight scores by a mechanism performed by a self-attention module to identify learned features. Attention patterns were investigated to evaluate whether the model focused on specific molecular substructures to predict plasma stability.

[0215] At this time, six unstable chemicals were randomly selected from the training set to analyze the attention patterns. The most unstable compounds in plasma contain ester bonds (Di et al., 2005). Visualization of the attention weight scores confirmed that there is a local concentration on atoms constituting ester bonds, consistent with the model's predictions (see Fig. 10).

[0216] In addition, the attention-based neural network system according to the present invention created a publicly accessible web server to predict the human plasma stability of a requested compound. The web server according to the present invention, as described above, accepts the SMILES format of the query compound as input and returns a binary classification result as stable or unstable, and the user interface of the web server according to the present invention can be configured as shown in FIG. 11.

[0217] Meanwhile, to summarize the attention-based neural network system according to the present invention more conclusively

[0218] We propose a plasma stability prediction tool that classifies input compounds in human plasma into stable and unstable types. Furthermore, the attention-based neural network system according to the present invention

[0219] A module comprising a CMPNN encoder and a self-attention layer is included. In this case, the attention-based neural network system according to the present invention exhibits the highest accuracy, sensitivity, and AUC as a result of comparative analysis using existing ML methods and existing graph-based neural networks. Additionally, a publicly accessible web server is provided to predict human plasma safety. Furthermore, the attention-based neural network system according to the present invention can serve as a useful tool for predicting the human plasma safety of compounds in the early stages of drug discovery and development.

[0220] Hereinafter, the drawings are described in more detail according to the present invention. FIG. 3 is a workflow flowchart in which the method of the system of the present invention is applied, which predicts human plasma stability for a given compound as a binary classification (stable or unstable) and is executed based on an attention-based message transmission neural network trained with in-house and open source data consisting of a set of chemically diverse compounds.

[0221] Figure 4 illustrates the model architecture of the system of the present invention, which consists of a communication message transmission neural network (CMPNN) (Song et al., 2020), a self-attention layer, and a fully connected layer. It describes a process in which input compounds (SMILES) are converted into molecular fingerprints and graph representations, in the case of graph representations, all node and edge functions are updated by a CMPNN encoder, then a self-attention method is applied to the reading procedure for all nodes to generate molecular feature vectors, finally the molecular fingerprints and graph functions are connected and simulated (or trained) with a fully connected layer, and the prediction result for human plasma stability is returned as a binary classification (stable or unstable).

[0222] FIG. 5 illustrates the features of atoms and bonds in the system of the present invention, and all functions except atomic mass are included in one-hot encoding.

[0223] FIG. 6 is a table showing hyperparameter optimization for the system of the present invention, including ReLU, rectified linear unit; PReLU, parametric ReLU; tanh, hyperbolic tangent; ELU, exponential linear unit; SELU, scaled exponential linear unit.

[0224] Figure 7 is an explanatory diagram illustrating the AUC value according to the depth of the message transmission neural network in the system of the present invention, and Figure 8 is a diagram illustrating the results of visualizing chemical diversity using t-variance stochastic neighbor embeddings (t-SNE) in the system of the present invention, where blue represents in-house data and red represents open source data on human plasma stability.

[0225] Figure 9 is an explanatory diagram illustrating the performance results of the present invention and existing molecular representation methods on an internal dataset in the system of the present invention. Among the contents, RF (Random Forest) and SVM (Support Vector Machine) were implemented using the Scikit-learn package, DMPNN was implemented using source code obtained from ChemProp (https: / / github.com / chemprop / chemprop), and CMPNN was implemented using source code obtained from https: / / github.com / SY575 / CMPNN. Five-fold cross-validation was used with random splitting, and the mean and standard deviation for each performance metric are provided.

[0226] Figure 10 illustrates the results in which the molecular colors of the heatmap in the system of the present invention vary according to plasma stability, where red indicates the predicted unstable function and blue indicates the predicted stable function.

[0227] FIG. 11 is an explanatory diagram illustrating the user interface of a web server applied to the system of the present invention.

[0228] As such, according to the present invention, by connecting the molecular fingerprint notation of a test compound to a molecular feature vector and simulating it as a fully connected layer to predict the stability of the compound in human plasma through a binary classification of stability or instability, the stability of the compound in human plasma can be easily predicted in the early stages of drug discovery and development, thereby maximizing the utility of the drug development platform.

[0229] In addition, according to the present invention as described above, by providing a plasma stability prediction device capable of repeatedly simulating whether a compound exhibits stability in human plasma without requiring a complex and expensive system to predict stability in human plasma, the reaction exhibited by an inherently unstable prodrug in human plasma can be rapidly confirmed, thereby significantly improving the designability of the prodrug.

[0230] Although various preferred embodiments of the present invention have been described above with some examples, the descriptions of various embodiments described in the "Specific details for carrying out the invention" section are merely illustrative, and those skilled in the art to which the present invention pertains will understand that the present invention can be modified in various ways or equivalent embodiments can be carried out based on the above description.

[0231] In addition, since the present invention can be implemented in various other forms, the present invention is not limited by the description above. The above description is provided merely to make the disclosure of the present invention complete and to fully inform those skilled in the art of the scope of the present invention, and it should be understood that the present invention is defined only by each claim of the claims. Explanation of the symbols

[0232] 1 : Input window 2 : Smile Module 3 : CMPNN Encoder 4: Self-Attention Mechanism Module 5 : Simulation Module 6 : PredPS Neural Network Control Module 7 : Dense Layer 8 : Memory 9: Key Panel 10: Display

Claims

Claim 1 A SMILES module that, when compound data for plasma stability testing is input through a configured input window, converts and outputs the input compound data into molecular fingerprint representation and molecular graph representation, respectively; a CMPNN encoder (Communicative message passing neural network encoder) that updates all node and edge functions for the test compound represented by the molecular graph output from the SMILES module; a self-attention mechanism module that generates and outputs a molecular feature vector for the test compound updated by the CMPNN encoder; and a simulation module that connects the functions of the molecular fingerprint representation of the test compound converted and output by the SMILES module to the molecular feature vector of the test compound output by the self-attention mechanism module, simulates it as a fully connected layer, and outputs a result value; An attention-based neural network system for predicting the stability of compounds in human plasma, comprising a PredPS neural network control module that predicts human plasma stability by binary classification of stability or instability by analyzing result values ​​obtained from the simulation module above, wherein the PredPS neural network control module uses a dense layer of a neural network as a non-linear function for message transmission convolution. Claim 2 delete Claim 3 An attention-based neural network system for predicting the stability of compounds in human plasma according to claim 1, characterized in that the SMILE module further includes the function of generating in-house data on the plasma stability of a set of 932 compounds using in vitro analysis of human plasma (785 stable compounds and 147 unstable compounds) under the functional control of the PredPS neural network control module, and also collecting and processing open source data on 2,166 compounds (647 stable compounds and 1,519 unstable compounds) for human plasma stability, and then learning and evaluating them all using 5-fold cross-validation through the RS (Random Split) module. Claim 4 An attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that, in claim 1, the PredPS neural network control module further includes a function to test four ML and DL algorithms—RF (Random Forest), SVM (Support Vector Machine), DMPNN (Directed Message Passing Neural Network) (Yang et al., 2019), and CMPNN (Communicative message passing neural network) (Song et al., 2020)—to predict human plasma stability and evaluate the performance of an attention-based graph neural network. Claim 5 An attention-based neural network system for predicting the stability of a compound in human plasma according to claim 1, wherein the SMILE module collects human plasma stability data from public databases PubChem and ChEMBL for model training under the functional control of the PredPS neural network control module, and wherein the collected information provides two types of human plasma stability data, namely the concentration (%) of the compound remaining after a certain period of time and the half-life (t1 / 2). Claim 6 An attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that, in claim 1, the PredPS neural network control module considers compounds with a half-life of 10 hours or more in human plasma to be stable and classifies compounds with a half-life of less than 10 hours to be unstable. Claim 7 An attention-based neural network system for predicting the stability of compounds in human plasma according to claim 1, wherein the PredPS neural network control module maps input data in a high-dimensional feature space and allows linear regression to be performed in the feature space. Claim 8 An attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that, in claim 4, both the RF and SVM are evaluated using 5-fold cross-validation. Claim 9 An attention-based neural network system for predicting the stability of a compound in human plasma, wherein the CMPNN encoder is characterized by encoding structural information of the compound by MPNN to predict molecular characteristics in claim 1. Claim 10 An attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that, in claim 1, the PredPS neural network control module further comprises a self-attention layer module and a fully connected layer module optimized for molecular fingerprint notation linked to graph-based notation. Claim 11 An attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that, in claim 1, the SMILE module executes two types of molecular representations, namely molecular fingerprint representation and molecular graph representation, under the functional control of the PredPS neural network control module. Claim 12 A first step in which, under the functional control of the PredPS neural network control module, when compound data to be tested for plasma stability is input through an input window configured by the SMILE module, the input compound data is converted into molecular fingerprint representation and molecular graph representation, respectively, and output; a second step in which, after the first step, under the functional control of the PredPS neural network control module, the CMPNN encoder updates all node and edge functions for the test compound represented by the molecular graph output from the SMILE module; a third step in which, after the second step, under the functional control of the PredPS neural network control module, the self-attention mechanism module generates and outputs a molecular feature vector for the test compound updated by the CMPNN encoder; and after the third step, under the functional control of the PredPS neural network control module, the simulation module connects the function of the molecular fingerprint representation of the test compound converted and output by the SMILE module to the molecular feature vector of the test compound output by the self-attention mechanism module. A control method for an attention-based neural network system for predicting the stability of a compound in human plasma, comprising: a fourth step of simulating with fully connected layers to output a result value; and a fifth step in which, after the fourth step, a PredPS neural network control module analyzes the result value obtained from the simulation module to predict human plasma stability through binary classification of stability or instability, wherein in the fifth step, the PredPS neural network control module uses a dense layer of the neural network as a non-linear function for message transmission convolution. Claim 13 delete Claim 14 A method for controlling an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in claim 12, the PredPS neural network control module uses a dense layer of the neural network as a non-linear function for message transmission convolution. Claim 15 A method for controlling an attention-based neural network system for predicting the stability of compounds in human plasma, characterized in that, in the first to fifth steps, the PredPS neural network control module further includes a test step for testing four ML and DL algorithms—RF (Random Forest), SVM (Support Vector Machine), DMPNN (Directed Message Passing Neural Network) (Yang et al., 2019), and CMPNN (Communicative message passing neural network) (Song et al., 2020)—to predict human plasma stability and evaluate performance. Claim 16 A control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in the first step, the PredPS neural network control module collects human plasma stability data through a SMILE module, and the collected information provides two types of human plasma stability data, namely the concentration (%) of the compound remaining after a certain period of time and the half-life (t1 / 2). Claim 17 A method for controlling an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in claim 12, the above-mentioned fifth step further includes a step in which the PredPS neural network control module considers a compound with a half-life of 10 hours or more in human plasma as stable, while classifying a compound with a half-life of less than 10 hours as unstable. Claim 18 A method for controlling an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in the 12th paragraph, the above 5th step further includes a step in which the PredPS neural network control module sets a number of tree structures proposed in metabolic stability for the RF model to predict plasma stability, and an SVM model with a radial basis function (RBF) and a kernel is used for plasma stability. Claim 19 A method for controlling an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in claim 12, the above-mentioned 5th step further includes a step in which the PredPS neural network control module evaluates both RF and SVM using 5-fold cross-validation. Claim 20 A method for controlling an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in claim 19, the structural information of the compound is encoded by an MPNN (Gilmer et al., 2017) to predict molecular characteristics during the evaluation step. Claim 21 A control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in the 5th step, the PredPS neural network control module updates an edge message vector by subtracting the previous hidden state of the backconnection from the hidden state of a node, and then updates the hidden state of the edge by supplying it to a fully connected layer using the initial hidden state as a bias and a ReLU activation function. Claim 22 A control method for an attention-based neural network system for predicting the stability of a compound in human plasma, characterized in that, in the 12th paragraph, the above 5th step further includes a step of predicting the stability of human plasma by binary classification of stability or instability by connecting the representation of a molecular graph to a molecular fingerprint and training a fully connected layer under the functional control of a PredPS neural network control module.

Images