KR102617958B1 - Method and apparatus for cross attention mechanism based compound-protein interaction prediction - Google Patents

Abstract

The present invention relates to a method and device for predicting compound-protein interactions based on a cross-attention mechanism. The method for predicting compound-protein interactions based on a cross-attention mechanism of the present invention includes the steps of encoding compound information based on molecular graph data and molecular fingerprint data, encoding protein information based on protein sequence data, and encoding the encoded protein information based on protein sequence data. It may include inputting compound information and protein information into a first cross-attention block and predicting an interaction between a compound and a protein based on the output of the first cross-attention block.

Description

Method and device for predicting compound-protein interaction based on cross attention mechanism {METHOD AND APPARATUS FOR CROSS ATTENTION MECHANISM BASED COMPOUND-PROTEIN INTERACTION PREDICTION}

The present invention relates to a method and device for predicting compound-protein interactions based on a cross-attention mechanism. More specifically, it relates to a method and device for predicting compound-protein interactions using a cross-attention mechanism rather than a simple combination of compound and protein information while simultaneously using graph convolutional neural networks and molecular fingerprints data.

Compound-protein interaction prediction is an essential step in the new drug development process and is performed through expensive molecular docking simulations. Artificial intelligence-based approaches are being proposed to replace this. Recently, two types of deep learning models have been used to study compound data: a graph convolutional neural network model that learns using molecular structure as graph data, and an artificial neural network model that learns molecular fingerprints data that replaces the partial structure of the molecule in bits. It is being used for artificial intelligence. Most existing studies use only one of the two methodologies. In addition, recent research is actively conducting research on how to model the combination between data during learning, beyond simply combining and learning compound and protein data to study compound-protein interaction prediction. The present invention relates to a method and device for predicting compound-protein interactions using a cross-attention mechanism rather than a simple combination of compound and protein information while simultaneously using a graph convolutional neural network and molecular fingerprint data.

Republic of Korea Patent Publication No. 10-2020-0124923 (2020.11.04)

The purpose of the present invention is to propose a method for predicting compound-protein interactions that is effective and has improved performance compared to existing technologies that use only either molecular structure graph data or fingerprint data.

The technical problems of the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below.

A method for predicting compound-protein interactions based on a cross-attention mechanism according to an embodiment of the present invention to achieve the above-described technical problem includes encoding compound information based on molecular graph data and molecular fingerprint data, and protein sequence data. Encoding protein information based on, inputting the encoded compound information and protein information into a first cross attention block, and predicting the interaction of the compound and protein based on the output of the first cross attention block. may include.

According to one embodiment, the encoding of the compound information includes inputting the molecular graph data into a direct message passing neural network (D-MPNN), and inputting the molecular fingerprint data into a mulit layer perceptron (MLP) artificial neural network. and inputting the output of the D-MPNN and the output of the MLP artificial neural network to a second cross attention block.

According to one embodiment, the second cross attention block has inputs of Q (query), K (Key), and V (value), and the V and K are determined based on the output of the D-MPNN, Q can be determined based on the output of the MLP artificial neural network.

According to one embodiment, the output of the second cross attention block is based on the inputs of Q, K, and V, It is determined as follows, where C is the number of embedding dimensions and d may be the number of attention heads.

According to one embodiment, the step of encoding the compound information may further include inputting the output of the second cross attention block to a self-attention block.

According to one embodiment, the step of encoding protein information includes preprocessing protein data based on a TAPE (tasks assessing protein embeddings) tokenizer, and processing the protein sequence as one-dimensional data based on one-hot encoding. It may include generating data and inputting the protein sequence data into a 1D convolutional neural network (CNN).

According to one embodiment, the first cross attention block has inputs of Q (query), K (Key), and V (value), where Q is determined based on the encoded compound information, and the K and V Can be determined based on the encoded protein information.

According to one embodiment, the output of the first cross attention block is based on the inputs of Q, K, and V. It is determined as follows, where C is the number of embedding dimensions and d may be the number of attention heads.

In order to achieve the technical object of the present invention, a compound-protein interaction prediction device based on a cross attention mechanism according to an embodiment of the present invention includes at least one processor, and loads a computer program executed by the at least one processor ( load) and a storage for storing the computer program, wherein the computer program operates by the at least one processor to encode compound information based on molecular graph data and molecular fingerprint data, based on protein sequence data. An operation of encoding protein information, an operation of inputting the encoded compound information and protein information into a first cross attention block, and an operation of predicting the interaction between a compound and a protein based on the output of the first cross attention block. It can be controlled to do so.

A computer program stored in a computer-readable recording medium according to an embodiment of the present invention for achieving the technical object of the present invention is combined with a computing device, and encodes compound information based on molecular graph data and molecular fingerprint data, An operation of encoding protein information based on protein sequence data, an operation of inputting the encoded compound information and protein information into a first cross attention block, and interaction of the compound and protein based on the output of the first cross attention block. Predictive actions can be executed.

According to the present invention, a method for predicting compound-protein interactions is disclosed that is effective and has improved performance compared to existing technologies that use only either molecular structure graph data or fingerprint data.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below.

Figure 1 is a diagram showing the structure of a compound-protein interaction prediction model according to an embodiment of the present invention.
Figure 2 is a diagram showing the detailed structure of a cross attention mechanism according to an embodiment of the present invention.
Figure 3 is a flowchart showing a method for predicting compound-protein interactions based on a cross-attention mechanism according to an embodiment of the present invention.
Figure 4 is a diagram showing the process of encoding compound information according to an embodiment of the present invention.
Figure 5 is a diagram showing the process of encoding protein information according to an embodiment of the present invention.
Figure 6 is a diagram showing the configuration of a compound-protein interaction prediction device based on a cross attention mechanism according to an embodiment of the present invention.

Details regarding the purpose and technical configuration of the present invention and its operational effects will be more clearly understood by the following detailed description based on the drawings attached to the specification of the present invention. Embodiments according to the present invention will be described in detail with reference to the attached drawings.

The embodiments disclosed in this specification should not be construed or used as limiting the scope of the present invention. It is obvious to those skilled in the art that the description, including embodiments, of this specification has various applications. Accordingly, any embodiments described in the detailed description of the present invention are illustrative to better explain the present invention and are not intended to limit the scope of the present invention to the embodiments.

The functional blocks shown in the drawings and described below are only examples of possible implementations. Other functional blocks may be used in other implementations without departing from the spirit and scope of the detailed description. Additionally, although one or more functional blocks of the present invention are shown as individual blocks, one or more of the functional blocks of the present invention may be a combination of various hardware and software components that perform the same function.

In addition, the expression including certain components is an “open” expression and simply refers to the presence of the corresponding components, and should not be understood as excluding additional components.

Furthermore, when a component is referred to as being “connected” or “connected” to another component, it should be understood that although it may be directly connected or connected to the other component, other components may exist in between. do.

Hereinafter, detailed embodiments of the present invention will be looked at with reference to the drawings.

Figure 1 is a diagram showing the structure of a compound-protein interaction prediction model according to an embodiment of the present invention.

Referring to Figure 1, the compound-protein interaction prediction model according to an embodiment of the present invention includes a direct message passing neural network (D-MPNN), a multi layer perceptrons (MLP) artificial neural network, and a 1D convolutional neural network (CNN). It can be composed of three input artificial neural networks, an attention block, and an output artificial neural network. D-MPNN and MLP artificial neural networks can be used to encode compound information, and 1D CNN can be used to encode protein information. Compound information and protein information encoded through D-MPNN, MLP, and 1D CNN are combined through a cross attention block, and the predicted Binding Score can be output through an output artificial neural network.

First, we describe how to encode compound information through an input neural network. Referring to Figure 1, among the input artificial neural networks, D-MPNN and MLP artificial neural networks can receive molecular structure graph data and molecular fingerprint data (Morgan fingerprint data), respectively.

Here, the molecular structure graph data and molecular fingerprint data can be expressed as follows.

- is a set of atoms and set of combinations It represents molecular graph data consisting of .

- Morgan fingerprint vector is a binary vector, expressing the presence or absence of a specific partial structure as 0 and 1. The Morgan algorithm searches all paths to all atoms in a molecule, and uses the Morgan fingerprint vector represents data obtained by replacing the unique path with bits.

molecular graph data is input to D-MPNN, and the Moragn fingerprint vector can be input to the MLP artificial neural network. The D-MPNN model is a model that learns relationships between atoms using directed edges rather than nodes in the molecular structure graph. can be output. MLP artificial neural network inputs By extracting nonlinear information from can be output.

two output values and Can be combined through a cross attention mechanism. The cross-attention mechanism has Q (Query), K (Key), and V (Value) as input values, is the output of the MLP artificial neural network from, and is the output of D-MPNN can be created from If this is expressed mathematically, it is as shown in Equation 1.

As in equation 1, Is from It is created as, Is from each , It can be created as . At this time is a projection function and can be defined as follows (meaning weight and bias, respectively).

Figure 2 is a diagram showing the detailed structure of a cross attention mechanism according to an embodiment of the present invention.

Output through D-MPNN and MLP artificial neural networks and Is Can be combined through a cross-attention mechanism with as an input value. If this is expressed mathematically, it is as shown in Equation 2.

here and means the number of embedding dimensions and the number of attention heads, respectively. This model uses a single-head cross-attention mechanism ( ) has been used, but note that this is for illustrative purposes only and does not limit the scope of the present invention. According to one embodiment of the present invention, in order to improve the compound phenotype, self-attention is additionally applied to obtain the final phenotype of the compound Feautre as shown in Equation 2. can be created.

Next, we will explain how to encode protein information. Referring to Figure 1, among the input artificial neural networks, 1-D CNN can receive protein sequence data. Protein data is preprocessed using TAPE (Tasks Assessing Protein Embeddings) tokenizer, which can express residues as unique numbers assigned by UniRep Vocabulary, and then one-hot encoding is used to construct the protein sequence into one-dimensional data and input into a 1-D CNN model. It can be. Through this, the protein feature can be created.

To predict accurate compound-protein interactions, compound features and protein features obtained through input artificial neural networks ( The way to combine them is important. The present invention proposes a new asymmetric artificial intelligence model that combines various data on the structure of a compound and finally combines them with protein data using a cross-attention module to predict interactions. The cross-attention module is trained to identify patterns of influence of compound data on various protein data. In other words, it models how proteins react with compounds. In this model and The combination formula is shown in Equation 3 below.

is input to the output artificial neural network to generate a binding score between proteins and compounds, and based on this, interactions between compounds and proteins can be predicted. For example, the output of the cross attention block is is input to the MLP artificial neural network, and the MLP artificial neural network can generate a Binding Score as an output. Based on the Binding Score, the final interaction between the compound and protein can be predicted.

Figure 3 is a flowchart showing a method for predicting compound-protein interactions based on a cross-attention mechanism according to an embodiment of the present invention. It will be understood that each step of the cross-attention mechanism-based compound-protein interaction prediction method according to an embodiment of the present invention, which will be described below with reference to the drawings, is performed by a cross-attention mechanism-based compound-protein interaction prediction device. You can. Additionally, in describing the present invention below, unless otherwise defined, “device” may be understood to mean a device for predicting compound-protein interactions based on a cross-attention mechanism according to various embodiments of the present invention.

Referring to FIG. 3, the device can first encode compound information based on the molecular graph and molecular fingerprint data (Morgan fingerprint data) (S310). Molecular graph data is a collection of atoms and set of combinations representing molecular graph data consisting of It can be expressed as Molecular fingerprint data is a binary vector, the Morgan fingerprint vector expressing the presence or absence of a specific partial structure as 0 and 1. It can be expressed as The process for encoding compound information in step S310 will be described below with reference to FIG. 4.

Figure 4 is a diagram showing the process of encoding compound information according to an embodiment of the present invention.

Referring to Figure 4, molecular graph data Can be input to D-MPNN (direct message passing neural network) (S410). Here, D-MPNN is a model that learns relationships between atoms using directed edges rather than nodes in the molecular structure graph. can be output. Also, Morgan fingerprint Vector molecular fingerprint data expressed as can be input to a multi layer perceptron (MLP) artificial neural network (S420). MLP artificial neural network inputs By extracting nonlinear information from can be output.

Output of D-MPNN and the output of the MLP artificial neural network. is input to the cross attention block and can be combined through the cross attention mechanism (S430). The cross-attention block used in step S430 has Q (Query), K (Key), and V (Value) as input values, is the output of the MLP artificial neural network from, and is the output of D-MPNN can be created from and from For the process of generating a value, refer to Equation 1 described above.

According to one embodiment, in order to improve the compound phenotype, a self-attention mechanism is additionally applied by using the output of the cross attention block as the input of a self-attention block to obtain the final phenotype of the compound. can be created.

Going back to the description of FIG. 3, the device can encode protein information based on protein sequence data (S320). The process for encoding protein information in step S320 will be described below with reference to FIG. 5.

Figure 5 is a diagram showing the process of encoding protein information according to an embodiment of the present invention.

Referring to Figure 5, protein data is preprocessed using TAPE (Tasks Assessing Protein Embeddings) tokenizer, which can express residues as unique numbers assigned by UniRep Vocabulary (S510), and then protein data, which is one-dimensional data, is generated based on one-hot encoding. Sequence data may be generated (S520). The generated protein sequence data is input to a 1-D CNN (convolutional neural network), and the final protein feature is generated as the output of the 1D CNN. may be created (S530).

The compound information and protein information encoded through steps S310 and S320 are input to the cross attention block and can be combined through a cross attention mechanism (S330). The cross-attention block used in step S330 has Q (Query), K (Key), and V (Value) as input values, is the encoded compound information from, and is the encoded protein information can be created from and from For the process of generating , refer to Equation 3 described above.

Finally, the interaction between chemicals and proteins can be predicted based on the output of the cross-attention block used in the device S330 step (S340). As an example, as shown in Figure 1, the output of the cross attention block in step S330 can generate a Binding Score through an MLP artificial neural network, and the final interaction between the compound and the protein can be predicted based on the Binding Score. there is.

Figure 6 is a diagram showing the configuration of a compound-protein interaction prediction device based on a cross attention mechanism according to an embodiment of the present invention.

The cross-attention mechanism-based compound-protein interaction prediction device 600 according to an embodiment of the present invention includes a processor 610, a network interface 620, a memory 630, a storage 640, and a data bus connecting them. It may include (650), and of course, it may also include other additional components required to achieve the purpose of the present invention.

The processor 610 controls the overall operation of each component. The processor 610 may be a central processing unit (CPU), a micro processor unit (MPU), a micro controller unit (MCU), or any of the types of processors widely known in the technical field to which the present invention pertains. In addition, the processor 610 may perform operations on at least one application or program to perform a method for predicting compound-protein interactions based on a cross-attention mechanism according to an embodiment of the present invention.

The network interface 620 supports wired and wireless Internet communication of the cross-attention mechanism-based compound-protein interaction prediction device 600 according to an embodiment of the present invention, and may also support other known communication methods. Accordingly, the network interface 620 may be configured to include a corresponding communication module.

The memory 630 stores various information, commands, and/or information, and stores one or more computer programs from the storage 640 to perform the cross-attention mechanism-based compound-protein interaction prediction method according to an embodiment of the present invention. (641) can be loaded. In FIG. 1, RAM is shown as one of the memories 630, but of course, various storage media can also be used as the memory 630.

Storage 640 may non-temporarily store one or more computer programs 641 and large-capacity network information. This storage 640 includes non-volatile memory such as Read Only Memory (ROM), Erasable Programmable ROM (EPROM), Electrically Erasable Programmable ROM (EEPROM), flash memory, hard disk (HDD), secondary storage match (SSD), and removable memory. It may be a disk or any type of computer-readable recording medium widely known in the technical field to which the present invention pertains.

A computer program 641 is loaded into memory 630 and operates by one or more processors 610 to encode compound information based on molecular graph data and molecular fingerprint data, and encode protein information based on protein sequence data. An operation of inputting the encoded compound information and protein information into the cross attention block, and an operation of predicting the interaction between the compound and the protein based on the output of the cross attention block can be performed.

The data bus 650 serves as a path for moving commands and/or information between the processor 610, network interface 620, memory 630, and storage 640 described above.

Although embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention can be implemented in other specific forms without changing its technical idea or essential features. You will be able to understand it. Therefore, the embodiments described above should be understood in all respects as illustrative and not restrictive.

600: Compound-protein interaction prediction device based on cross attention mechanism
610: processor
620: Network interface
630: memory
640: Storage
641: computer program
650: data bus

Claims (10)

In the compound-protein interaction prediction method based on the cross-attention mechanism,
Encoding compound information based on molecular graph data and molecular fingerprint data, respectively;
Encoding protein information based on protein sequence data;
Inputting the encoded compound information and protein information into a first cross attention block; and
In the cross-attention mechanism-based compound-protein interaction prediction method, which includes predicting the interaction between a compound and a protein based on the output of the first cross-attention block,
The first cross attention block has inputs of Q (query), K (Key), and V (value), where Q is determined based on the encoded compound information, and K and V are encoded protein information. It is decided based on
The step of encoding the compound information is,
Inputting the molecular graph data into a direct message passing neural network (D-MPNN);
Inputting the molecular fingerprint data into a multi layer perceptron (MLP) artificial neural network; and
Inputting the output of the D-MPNN and the output of the MLP artificial neural network into a second cross-attention block to output the encoded compound information and Q input to the first cross-attention block;
Including,
Compound-protein interaction prediction method based on cross-attention mechanism. delete According to paragraph 1,
The second cross attention block has inputs of Q (query), K (Key), and V (value), where V and K are determined based on the output of the D-MPNN, and Q is the MLP artificial neural network. A compound-protein interaction prediction method based on a cross-attention mechanism determined based on the output of . According to paragraph 3,
The output of the second cross attention block is based on the inputs of Q, K, and V.

It is determined as follows, where C is the number of embedding dimensions and d is the number of attention heads. A compound-protein interaction prediction method based on a cross attention mechanism. According to paragraph 1,
The step of encoding the compound information is,
A method for predicting compound-protein interactions based on a cross-attention mechanism, further comprising inputting the output of the second cross-attention block to a self-attention block. According to paragraph 1,
The step of encoding protein information is,
Preprocessing protein data based on TAPE (tasks assessing protein embeddings) tokenizer;
Generating the protein sequence data as one-dimensional data based on one-hot encoding; and
A method for predicting compound-protein interactions based on a cross-attention mechanism, including the step of inputting the protein sequence data into a 1D convolutional neural network (CNN). delete According to paragraph 1,
The output of the first cross attention block is based on the inputs of Q, K, and V.

It is determined as follows, where C is the number of embedding dimensions and d is the number of attention heads. A compound-protein interaction prediction method based on a cross attention mechanism. at least one processor;
a memory that loads a computer program executed by the at least one processor; and
Includes storage for storing the computer program,
The computer program is operated by the at least one processor,
An operation to encode compound information based on molecular graph data and molecular fingerprint data, respectively;
An operation to encode protein information based on protein sequence data;
Inputting the encoded compound information and protein information into a first cross attention block; and
In the cross-attention mechanism-based compound-protein interaction prediction device controlled to perform an operation to predict the interaction between a compound and a protein based on the output of the first cross-attention block,
The first cross attention block has inputs of Q (query), K (Key), and V (value), where Q is determined based on the encoded compound information, and K and V are encoded protein information. It is decided based on
The operation of encoding the compound information is,
Inputting the molecular graph data into a direct message passing neural network (D-MPNN);
Inputting the molecular fingerprint data into a multi layer perceptron (MLP) artificial neural network; and
An operation of inputting the output of the D-MPNN and the output of the MLP artificial neural network to a second cross-attention block to output the encoded compound information and Q input to the first cross-attention block;
Including,
Compound-protein interaction prediction device based on cross-attention mechanism. Combined with a computing device,
An operation to encode compound information based on molecular graph data and molecular fingerprint data, respectively;
An operation to encode protein information based on protein sequence data;
Inputting the encoded compound information and protein information into a first cross attention block; and
In the computer program stored in a computer-readable recording medium that performs an operation of predicting an interaction between a compound and a protein based on the output of the first cross attention block,
The first cross attention block has inputs of Q (query), K (Key), and V (value), where Q is determined based on the encoded compound information, and K and V are encoded protein information. It is decided based on
The operation of encoding the compound information is,
Inputting the molecular graph data into a direct message passing neural network (D-MPNN);
Inputting the molecular fingerprint data into a multi layer perceptron (MLP) artificial neural network; and
An operation of inputting the output of the D-MPNN and the output of the MLP artificial neural network into a second cross-attention block to output the encoded compound information and Q input to the first cross-attention block;
Including,
A computer program stored on a computer-readable recording medium.