Comogolov GCN protein ligand affinity prediction method

Through the Kolmogolov GCN protein ligand affinity prediction method, combining self-attention mechanism feature fusion and KAN layer to replace traditional flat layer, the data processing problem and explanatory problem in the prediction of drug target affinity in the prior art is solved and the data processing problem and explanationable value in the prediction of drug target affinity in the prior art is solved. sexual problems achieve more efficient protein-ligand interaction prediction and virtual screening performance improvement in drug discovery tasks.

CN118969062APending Publication Date: 2024-11-15NANJING UNIV OF POSTS & TELECOMM
View PDF 0 Cites 6 Cited by

Patent Information

Application Number
CN202410842887.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-06-27
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

Existing machine learning methods have problems such as overfitting, difficulty in dealing with high-dimensional data, limited interpretability and inability to capture complex molecular interactions in drug target affinity prediction, and traditional neural network structures are also difficult to be effective. Processing two-dimensional graph structure data of proteins and ligands.

Method used

Kolmogolov GCN protein ligand affinity prediction method is used to extract abstract information of ligands and proteins through GCN, and combine self-attention mechanism characteristics fusion, using KAN instead of traditional flat layers for task-oriented learning, thereby improving Characterization efficiency of protein and ligand interactions.

Benefits of technology

It improves the performance of virtual screening technology in drug discovery tasks, improves the interpretability and flexibility of the model, and can more efficiently predict the interaction relationship between protein and ligands.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969062A_ABST
    Figure CN118969062A_ABST
Patent Text Reader

Abstract

The invention discloses a method for predicting affinity of a Comogolov GCN protein ligand, which is characterized in that a graph neural network is associated with the interaction between a ligand and a protein (DTI) to predict the affinity in the DTI. When the affinity is predicted, a Kolmogorov network (KAN) is introduced after GCN stacking so as to effectively judge which subsequences in the protein are more likely to bind to the ligand. According to the method, GCN is used for extracting abstract information of ligands and protein molecules, so that characterization of interaction of the protein and the ligands is more efficient. The KAN places the activation function on the edge (weight) and makes it learnable. Thus, the calculation of the KAN is reexpressed by processing the input by using different basis functions or activating the input, and then linearly combining the output. For DTI, the recoding based on KAN ensures the expressive force of a GCN model during virtual screening, and can also improve the interpretability of a prediction model and enhance the flexibility of the model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a Kolmogorov GCN protein ligand affinity prediction method, belonging to the technical field of computer networks. Background Art

[0002] Drug-Target Binding Affinity (DTA) prediction is a key step in computer-aided drug design, and its goal is to accurately predict the binding affinity between proteins and drugs. Molecular docking virtual screening using artificial intelligence can show that drug design time can be saved. The recently popular Kolmogorov Arnold Network (KAN) is a new neural network structure that is expected to replace the multilayer perceptron (MLP), which is providing opportunities for improving deep learning models. The present invention attempts to explore various combinations of KAN and graph convolutional neural network (GCN) to search for a new neural network architecture that can effectively process protein and ligand two-dimensional graph structure data, and bring new virtual screening technology solutions to drug discovery tasks based on deep learning. In order to facilitate comparison with existing terms in the field, the terms drug and ligand are interchangeable, and the terms protein and drug target are interchangeable below.

[0003] The traditional machine learning methods currently used for DTA prediction are summarized as: Linear Regression (LR), Support Vector Machine (SVM), Artificial Neural Network (ANN), Random Forest, K-nearest neighbor (K-NN), etc. These traditional machine learning methods have the following limitations in DTA prediction, such as overfitting, difficulty in handling high-dimensional data, limited interpretability, and inability to capture complex molecular interactions.

[0004] In order to overcome these shortcomings of traditional machine learning, researchers have explored new machine learning and deep learning methods in the past five years, such as deep neural networks (DNN), graph convolutional neural networks, and attention models. These deep learning methods have shown good results in DTA prediction, but there is still room for improvement, such as using new MLP components, deep neural networks, graph convolutional neural networks (GCN), and attention models.

[0005] Let's expand on the previous paragraph. MLP is the predecessor of DNN. It can be viewed as a directed graph, consisting of multiple node layers, each of which is fully connected to the next layer. In addition to the input node, each node is a neuron (or processing unit) with a nonlinear activation function. Its advantage is that it has strong expressive power and can handle nonlinear problems and high-dimensional data. Relevant mathematical proofs reveal that MLP is a universal approximator, that is, they can approximate any function, similar to how all functions can be represented by Taylor series. DNN is a more general deep feedforward neural network. It is a network composed of multiple layers, usually including at least an input layer, an output layer, and at least one hidden layer. If there are fewer hidden layers, DNN is called a shallow neural network. When the hidden layer increases and an appropriate activation function is used, it is called a deep neural network.

[0006] In recent years, deep learning methods have made significant progress in DTA prediction. Sequence-based methods have always been an important method for DTA prediction in computational biology. There have been many papers studying various neural networks to extract features of protein sequences and molecular SMILES (Simplified Molecular Input Line Specification). Sequence-based methods can be further divided into two types according to different needs, including compound-protein interaction (CPI) prediction and drug-target affinity (DTA) prediction (Mingjian Jiang, https: / / doi.org / 10.1186 / s12864-022-08648-9). The present invention mainly considers sequence-based, two-dimensional data-oriented DTA prediction methods.

[0007] The latest progress includes prediction of drug-target affinity based on graph neural network and protein contact graph, prediction of drug-target binding affinity through cross-scale graph comparison learning, etc. (Rohan Gorantla, https: / / doi.org / 10.1021 / acs.jcim.3c01208, Jingru Wang, https: / / doi.org / 10.1093 / bib / bbad516). The present invention focuses on the prediction of protein-ligand interaction based on graph neural network in the field of DTA, and especially considers the application design of the latest KAN+GCN structure in DTA prediction when the two-dimensional graph structure data of protein and ligand is input.

[0008] KAN is a new type of neural network structure that appeared in April 2024. Here we compare KAN with MLP (Ziming Liu, https: / / doi.org / 10.48550 / arXiv.2404.19756) from the perspective of mathematical principles, structure, interpretability, and mathematical expressions. First, we discuss the interpretability of KAN from a mathematical perspective. The design of KAN was inspired by the insight of physicists into the discovery of natural laws, and the mathematical principle behind it is the Kolmogorov-Arnold (KA) representation theorem. This theorem states that if f is any multivariable continuous function defined on a bounded domain, then the function f can be expressed as a two-layer nested addition of a finite number of single-variable, continuous functions, that is, any multivariable function can be expressed by the addition of one-dimensional nonlinear functions. The KA theorem is very similar to the Fourier series, which is a continuous periodic function generated by the sum of harmonically related sine functions. Note that this theorem only discusses a two-layer superposition structure. These one-dimensional functions are parameterized as B-Spline (Basic Spline, B-Spline). This means that each weight parameter is no longer a single value, but a function.

[0009] For each deep learning task, a function is actually learned, and the usual end-to-end deep learning requires the stacking of multiple components (or layers), each layer usually consisting of linear transformations and nonlinearities. Therefore, recent researchers are paying more attention to deeper KAN networks rather than the two-layer KAN structure directly discussed in the KA representation theorem.

[0010] Let's continue discussing the interpretability of KAN. Unlike MLP, which uses activation processing on nodes, KAN uses activation functions on edges. Its training process can be visualized intuitively. Since visualization makes it easier for people to understand the decision-making process of machine learning models, this improves the interpretability of the model. Existing literature shows that KAN can outperform MLP with large parameters using only a small number of parameters (Ziming Liu), which may help improve the efficiency of deep learning training in mathematical research, physical research, and drug discovery. The visualization and interactivity of KAN make it potentially applicable in many fields. Summary of the invention

[0011] The purpose of the present invention is to address the defects and shortcomings of the above-mentioned prior art and propose a Kolmogorov GCN protein ligand affinity prediction method, which uses GCN to extract abstract information of ligand and protein molecules respectively, and then proposes a task-oriented learning through a combination of feature fusion including a self-attention mechanism, so as to make the characterization of protein-ligand interaction more efficient.

[0012] The technical solution adopted by the present invention to solve the technical problem is: a method for predicting Kolmogorov GCN protein ligand affinity, the method comprising the following steps:

[0013] Obtain two ligand molecule graphs and protein graphs to be predicted;

[0014] Use GCN to encode the two predicted graphs and obtain the corresponding embedding vectors;

[0015] Using KAN to replace the traditional flat layer to obtain a new flat layer;

[0016] The proposed feature fusion strategy is used to fuse the ligand flat layer output and the protein flat layer output to obtain an updated encoding representation vector;

[0017] The encoded representation is input into multiple task-related KAN linear layers, and the final output loss is used as a predictor indicator to evaluate the interaction relationship between the protein ligand to be predicted;

[0018] The design of a protein-ligand affinity prediction system integrating a KAN sublayer requires weighing multiple metrics to obtain a preferred KAN-based GCN scheme.

[0019] Furthermore, the method of the present invention obtains two ligand molecular graphs and protein graphs to be predicted, including: considering each small molecule ligand molecule as a graph, and the chemical bonds between ligand atoms as edges to obtain an atom list and an adjacency matrix in the molecular graph, defining amino acids in the protein as nodes, and connecting non-continuous amino acids through edges, and generating a two-dimensional protein graph from a contact graph, that is, converting the two-dimensional structure of the protein into a graph representation.

[0020] Furthermore, the method of the present invention utilizes the first part of representation learning, namely, respectively encoding the two-dimensional ligand molecule graph and the two-dimensional protein graph using GCN stacking to obtain corresponding embedding vectors.

[0021] Furthermore, the present invention uses a new flat layer, i.e., the second part of the representation learning, which respectively uses the ligand embedding vector output by the GCN stack and the protein embedding vector output by the GCN stack to perform KAN stacking. The KAN stacking can use KAN linear, fastKAN and traditional fully connected layers or dense layers and their combinations. The fastKAN is a fast implementation of KANlinear. The second part entity will output the corresponding, updated representation vector matrices X and Xt.

[0022] Furthermore, the method of the present invention utilizes the proposed feature fusion strategy to achieve task- and data-related feature fusion, that is, a combination of feature fusion schemes including an attention mechanism is introduced and the second one is selected to fuse the ligand flat layer output X and the protein flat layer output Xt to obtain a feature-fused, updated encoding representation vector.

[0023] Furthermore, the method of the present invention utilizes task-related learning to process the representation vector after feature fusion, and inputs the encoded representation after feature fusion into multiple task-related KAN linear layers. Finally, the output loss will be used as a predictor indicator to evaluate the interaction relationship between the protein ligands to be predicted. The multiple KAN linear layers are a combination of KAN linear layers and traditional fully connected layers.

[0024] Furthermore, the method described in the present invention utilizes the proposed protein ligand affinity prediction system to obtain the preferred KAN-based GCN scheme. The performance of the affinity predictor integrating the KAN sublayer is related to the data set, data reconstruction, representation learning and task-related learning. The predictor indicators for different combinations will be different, such as RMSE, the number of rounds to reach the minimum RMSE and the training time.

[0025] Beneficial effects:

[0026] 1. The present invention uses GCN to extract abstract information of ligand and protein molecules respectively, and then proposes a task-oriented learning method through a combination of feature fusion including self-attention mechanism, so as to make the representation of protein-ligand interaction more efficient.

[0027] 2. The KAN-based recoding of the present invention can ensure the expressiveness of the GCN model during virtual screening while also improving the interpretability of the prediction model and enhancing the flexibility of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of the mathematical expression of the network model of the present invention.

[0029] Figure 2 Flowchart of the method for protein ligand affinity prediction based on the combination of KAN and GCN proposed in the present invention.

[0030] Figure 3 This is a schematic diagram of the affinity prediction model proposed in the present invention.

[0031] Figure 4 This is a flow chart of the feature fusion mechanism of the self-attention mechanism of the present invention.

[0032] Figure 5 It is a schematic diagram of KAN combination and optimization according to task requirements in the affinity prediction system of the present invention. DETAILED DESCRIPTION

[0033] The invention will be further described in detail below in conjunction with the accompanying drawings.

[0034] This paper will explore how to use the KAN layer to replace the original MLP layer in a specific GCN virtual screening task.

[0035] like Figure 1 As shown, the mathematical expression of a KAN network model when L = 2 is formula (1),

[0036]

[0037] In such Figure 1 In the above equation, 1=0 represents the KAN layer closest to the input vector, i.e. Φ0. The input vector x=[x 0.0 , x 0.1 ] T , the superscript "T" indicates the transpose operation, It represents the stacking operation of KAN layers. Assume that there is a KAN composite network with L+1 layers (here L=2), and the number of nodes in each layer is [n0,…, n1,…, n L ]( Figure 1 The number of nodes is [2, 3, 3, 1]); (i, j) represents the jth neuron in the layer, and x(i, j) represents the value of this neuron. The external function is defined as Φ q (Here Φ0, Φ1, Φ2), the internal function is defined as Φ q.p (Here the internal function Φ of layer 0 0,0 , Φ 0,1 , Φ 0,2 ). Due to space limitations, Figure 1 Internal functions are not directly exposed.

[0038] Next, we will explain the connection between the first layer and the l+1th layer. In KAN, each weight is a learnable activation function. The neurons between the two layers are interconnected and coexist (n1*n 1+1 ) activation function. The jth neuron (l, j) in the ,th layer is connected to the ith neuron (l+1, i) in the l+1th layer through an activation function. The activation function is represented as a set of Φ(l, i, j) and is expressed in matrix form. The multi-layer KAN uses a structure similar to MLP. The layers are fully connected. Each element x in the next layer (l+1,j) All elements x in the previous layer (l,i) Combination, the vector x of the l+1th layer l+1 for

[0039] x l+1 =Φ l x l (2)

[0040] In formula (2), Φ l =[Φ l,i,j ],i=1,...,(n L +1), j=1, ..., n L In short, KAN puts the activation function on the edge, so that each element in the previous layer corresponds to an activation function.

[0041] B. Combination of KAN and GCN

[0042] Graph neural networks (GNNs) are a class of neural networks that operate directly on graphs, leveraging the structural and semantic information of nodes and edges. GNNs are able to learn vector representations (or embeddings) of nodes that capture their features and context in the graph. This representation can be used for a variety of tasks, such as node classification, link prediction, graph generation, spatial and temporal reasoning, etc.

[0043] Existing domain knowledge shows that for end-to-end deep learning tasks, the specific combination of two or more different structures (or neural network layers) will depend on the specific problem domain and dataset. Existing mature graph encoders usually use GNN, such as GCN, Graph Attention Network (GAT) and GraphSAGE. After encoding the input graph data, GNN will generate node-level representations to capture local and global graph properties. The structure of GCN is usually composed of the following main components: graph convolution layer (GCNConv), linear layer (Linear), ReLU activation function, Dropout layer, etc. The combination and stacking of these components form different types of graph neural network models. Here GCNConv is a key layer (entity) in GCN in the Pytorch framework, which is used to implement graph convolution operations.

[0044] The question here is whether the graph convolution layer can directly use a mechanism similar to KAN, such as the KAN network places learnable activation functions on the edges connecting nodes, rather than the nodes themselves. This means that the KAN network does not have a linear weight matrix, but uses learnable spline functions or other (mathematical) symbolic functions to replace each weight parameter. Since KAN is still in its early stages, after limited literature searches, it was found that no research work has considered directly introducing learnable nonlinear activation functions on the edges in a single graph convolution layer, although the latest literature reports the progress of the KAN convolution layer. In this regard, the present invention first conducted its own verification experiments and tried to make such a simple change to the graph convolution layer. However, one result of the own verification experiment shows that the output value of the KAN graph convolution layer is "nan", that is, directly introducing nonlinear transformations on the edges will lead to instability in numerical calculations, which is also a future research point.

[0045] After considering the work of others and our own experiments, this paper will introduce a feasible fusion idea of ​​KAN and GCN, and will provide an effective deep learning design in the subsequent part of this paper. The idea is briefly described as follows:

[0046] 1) Integrate KAN into GCN application, that is, first apply GCN to the constructed 2D graph, and then use KAN linear layer to replace the traditional linear layer;

[0047] 2) Introducing a task- and data-dependent feature fusion scheme, and then using the KAN linear layer to replace the subsequent traditional linear layer;

[0048] 3) Screen various affinity prediction models containing KAN through experiments and optimization, because in the early stage, the KAN scheme only provides many possibilities for parameter tuning, but does not indicate how to do it.

[0049] Figure 2 This is the overall flow chart of the proposed protein-ligand affinity prediction based on the combination of KAN and GCN. The input for the protein is the protein sequence (ASCII characters), and then the contact map is used to construct a two-dimensional protein graph. The input for the ligand is smiles encoding, which is then converted into a ligand molecular graph. The protein graph and the ligand molecular graph are respectively output with two representation vectors through KAN-based GCN representation learning. Secondly, these representation vectors are first fused with task-related features, and then passed through a single or multi-layer KAN linear layer, and finally the protein-ligand affinity value is estimated. The proposed encoding and feature fusion with attention mechanism help to reveal how the KAN component contributes to the protein-ligand interaction mechanism.

[0050] The overall flowchart contains three parts: data reconstruction, ligand representation learning / protein representation learning, and task-related learning. In most previous works, the input representation of protein sequence and ligand SMILES has been shown to be an effective method for predicting protein-ligand binding affinity. Data reconstruction contains the input of simple protein sequence and ligand SMILES. Data reconstruction contains two outputs, one is the two-dimensional ligand molecular graph reconstructed from the ligand SMILES data, and the other is the use of protein sequence to contact map and then to two-dimensional protein graph structure data. Since the residues interact to form a protein, the residue contact map can reflect the spatial structure of the entire protein. Therefore, if the contact map can be quickly obtained based on the sequence, fine-grained structure and interaction information can be quickly captured, which is very useful for affinity prediction. Two-dimensional graph structure data can express more complete information about the drug, so this method can obtain better prediction performance than one-dimensional data. It should be mentioned that most previous works represent the drug or the ligand in the present invention as a one-dimensional string during data reconstruction, but this is not the only way to represent molecules. The present invention represents the drug as a two-dimensional graph.

[0051] Ligand representation learning / protein representation learning and task-related learning are elaborated in detail in the subsequent specific implementation.

[0052] The present invention comprises 4 key components.

[0053] 1) GCN-based embedding vector extractor

[0054] We use the existing graph convolutional layer GCNconv to enable each node in the 2D graph to learn an embedding that contains information about its neighborhood.

[0055] 2) KAN-based feature extractor

[0056] KAN* further extracts topological features from the GCN-based embedding vector extractor, including node features, edge features, and graph-level features;

[0057] 3) Task-related feature fusion

[0058] The proposed feature fusion strategy involves the selection of multiple feature fusions and their fusion schemes. Here, a self-attention mechanism fusion scheme is particularly considered, and the details are shown in the subsequent specific implementation;

[0059] 4) Task-dependent linear merging

[0060] The feature vector output by feature fusion can be used for subsequent task-related learning. Usually, this can be processed by traditional fully connected layers (including activation functions at nodes), but this solution introduces multiple KAN linear layers. For details, please refer to the subsequent specific implementation.

[0061] The Davis dataset used for the verification of the present invention (Davis, MI; Comprehensive analysis of kinase inhibitor selectivity. Nat. Biotechnol. 2011, 29, 1046-1051). The Davis dataset and the KIBA dataset have been widely used as benchmarks for binding affinity prediction and evaluation. The former contains a dataset of affinity values ​​corresponding to 422 compounds and 60 target proteins, and the latter contains a dataset of affinity values ​​corresponding to kinase inhibitors and 242 protein kinases. Common datasets for protein ligand affinity prediction also include the PDBbind database and the ChEMBL database. These datasets / databases can be divided into two categories: sequence data and structural data. The KIBA database is currently the best sequence-based database. The PDBbind database is currently the most widely used database due to its large sample size and strict filtering conditions. Different datasets usually differ in size, molecular type, and target protein, allowing researchers to evaluate the performance of graph neural networks in different tasks and scenarios. The binding affinity of the Davis dataset is the corresponding dissociation constant Kd, and a high Kd value indicates a high affinity. Figure 2 and Figure 3 The protein 2D coding is generated from the protein map. The first protein contact map data used for the protein map is "pconsc4".

[0062] Affinity prediction models such as Figure 3 shown. Figure 1 In the figure, (J, K) indicates that the input dimension and output dimension of a neural network layer are J and K respectively; the node dimension is marked with bold symbols, such as [K, K] represents the dimension of the ligand representation vector X and the protein representation vector Xt before the row feature fusion. This dimension can also be extended to other values ​​to facilitate the introduction of more feature fusion entities, because the KAN* stacking (corresponding to the traditional flatten layer) after the GCN stacking can convert high-dimensional features into low-dimensional features.

[0063] B. Technical factors

[0064] In the previous introduction of the general flow chart, the present invention has already outlined data reconstruction and described in detail the technical elements related to representation learning and tasks. Specifically, it includes:

[0065] 1) GCN stacking in ligand representation learning and protein representation learning

[0066] Ligand representation learning and protein representation learning can use the same stack (also called multi-layer GCN components), but the feature dimensions of the two are different because the dimensions of the ligand graph and protein graph input to the representation learning are different. Public knowledge in this field shows that a single GCN layer directly implements information propagation of 1st-order neighbors. After the initial input passes through multiple layers of GCN, the information can propagate farther in the graph, helping to learn more complex graph structures. More precisely, on the one hand, the more GCN layers there are, the more it can cover multi-hop neighbor aggregation information; on the other hand, too many GCN layers may lead to over-smoothing, that is, the features of different nodes become more and more similar, which will lead to loss of discrimination. Therefore, this model uses 3 layers of GCN. Multi-layer GCN is convenient for capturing the interaction between protein binding pockets and ligands. The ligand only binds at favorable binding sites (also called protein substructures) in the protein structure. The following description is mainly based on ligand representation learning, and the protein representation learning process is similar to it;

[0067] 2) Workflow of a single GCN layer in ligand representation learning

[0068] As a commonly used graph neural network, GCN can apply convolutional neural network (CNN) to graph structured data. In GCN, node features are aggregated to their neighbor nodes, and then the representation of the node is updated through some nonlinear transformations. The core idea is to use the information of neighbor nodes to update the representation of each node. The workflow of GCN usually includes the following three steps, namely neighbor aggregation, feature update and multi-layer GCN. During the neighbor aggregation process, GCN collects the features of its neighbor nodes. During the feature update process, GCN combines the features of the current node and the aggregated features of the neighbor nodes to update the representation of the current node. The results of some existing literature show that GCN can learn the interaction information of protein-ligand in detail. A single GCN layer can effectively capture the local relationship between nodes, and GCN stacking facilitates the acquisition of more global information. Figure 1 After each layer of GCN, a nonlinear activation function (such as ReLU) operation is also included.

[0069] 3) Representation Learning KAN Stacking in Ligand Representation Learning

[0070] In the representation learning of the present invention, after the GCN stacking, there is the KAN stacking (also called the multi-layer KAN* component). KAN* means that it can be used in components such as KAN linear and fastKAN. KAN stacking has two functions. One is to replace the traditional torch.nn.Linear with KAN Linear (ZimingLiu), and the other is to perform linear transformation. It should be mentioned here that the flatten layer compresses the two-dimensional or higher-dimensional data output by the convolution layer and converts it into one-dimensional data. It is often used in the transition from the convolution layer to the fully connected layer. Its operation can increase the width and depth of the data, so that the data can be better utilized by the fully connected layer and subsequent neural network layers. In the design of the present invention, the traditional flatten layer is replaced by the KAN stacking. After balancing the input data dimension, computational efficiency and memory consumption, this design considers a 2-layer substructure. Use KAN to replace the traditional flat layer to obtain a new flat layer (i.e., KAN stacking).

[0071] 4) fastKAN in ligand representation learning

[0072] FastKAN represents a fast implementation of the KAN algorithm. The experiments of the present invention refer to the open source project https: / / gitcode.com / ZiyaoLi / fast-kan, but are not limited to the source code. The core contribution of FastKAN is to use Gaussian radial basis functions (RBF) to approximate the B-spline basis, which can run faster than the KAN linear algorithm while maintaining the accuracy of the model.

[0073] 5) KAN stacking in ligand representation learning

[0074] The representation learning KAN stack (also called the multi-layer KAN* component) processes the embedding vectors of the GCN stack output 2D graph to obtain the ligand feature vector x and the protein feature vector xt.

[0075] 6) Feature Fusion Schemes in Task-Related Learning

[0076] The input of GCN is a graph structure, and the output is also a graph structure, but the nodes in the output graph are embedded with semantic information. The output structure of the GCN stack is low-dimensional, continuous, and dense. Figure 2The representation learning has two output features, corresponding to the ligand GCN stacking and the protein GCN stacking. How to merge the two features to effectively discover the local and global information in the two graphs? Existing protein ligand affinity prediction schemes usually adopt the following three feature fusion methods, namely "concat", "sum" and "⊙", which respectively represent the concatenation of two feature vectors, the addition of elements of two feature vector matrices and the product of elements of two feature vector matrices, but there are few works on attention fusion for protein ligand affinity prediction tasks. The present invention introduces a self-attention mechanism into the task under study, and proposes a strategy for selecting the optimal fusion among multiple feature fusion schemes, that is, for different protein-ligand interaction tasks, different feature fusion schemes need to be selected;

[0077] 7) Attention mechanism feature fusion mechanism flow chart

[0078] Figure 4 This is the proposed self-attention mechanism feature fusion mechanism flow chart. The self-attention mechanism is a kind of attention mechanism, which can pay more attention to the key information in the input. Figure 4 The upper middle branch. The model structure contains three matrices Q, K, and V, which are obtained by linearly transforming the ligand abstract information x. Q is used to match other units, K is used to be matched by other units, and V is the information that needs to be extracted. Secondly, the similarity α is calculated by Q and K, and then α_ is calculated by softmax and normalized (scale). Thirdly, the comprehensive information Y1 of each unit after attention is calculated. For the protein input x t The subsequent operation is the fusion of the two comprehensive information. Fusion 1, fusion 2 and fusion 3 in Table 4 can be selected. Note that fusion 3 is a traditional direct fusion that does not consider the attention mechanism but has the lowest computational complexity. The present invention does not consider the traditional "concat";

[0079] 8) KAN stacking in task-related learning

[0080] In the traditional GCN scheme, the abstract information output by GCN will be sent to the task-related learning entity, that is, multiple fully connected layers. These fully connected layers (FC) will implement task-related classification, regression, etc. For protein ligand affinity prediction, the present invention specifically considers using KAN to process the feature vector output by the feature fusion entity. The KAN stack will estimate the loss function, which will be used to evaluate the downstream regression task. Note that the impact of KAN on the encoding used for representation learning and task-related learning has yet to be evaluated.

[0081] Table 1

[0082]

[0083]

[0084] 4 Performance Evaluation

[0085] Table 1 above is a description of the feature fusion scheme, which contains 4 different schemes. Table 2 is a benchmark test. Note that the training time and RMSE are related to the specific simulation environment. The 200 rounds of training are only used for performance evaluation. The number of training and test batches is 128. The data sets here are all Davis. Unless otherwise specified, the task-related learning in Cases 1-7 uses the traditional Fully Connected Layer (FC), which is also called the Dense Layer.

[0086] In the regression problem involved in predicting the affinity of protein-ligand interaction, Table 2 shows that the RMSE performance of 2D graph representation learning is significantly better than that of 1D representation learning cases (cases 1, 2, and 3). Secondly, cases 1-3 use 1D protein CNN encoding and 2D GCN ligand encoding. The 2D GCN ligand encoding of case 1 uses traditional torch.nn.Linear and considers fusion 1. KAN linear and fastKAN are used after GCN stacking in the representation learning of cases 2 and 3, respectively.

[0087] In cases 4-8, after GCN stacking in representation learning, the solutions of the present invention all use KAN linear to replace the traditional torch.nn.Linear; in case 8, FC1, FC2 and FC3 all use KAN linear; in case 9, fastKAN is used after GCN stacking, but FC1 uses KAN linear, and FC2 and FC3 use fastKAN.

[0088] The comparison results show that it is feasible to replace MLP with KAN layer, although the time complexity of KAN linear is slightly higher than that of basic linear processing such as MLP. Case 4 shows the best RMSE performance, but the number of rounds to reach the minimum RMSE during training is the largest, which partly explains that it takes a long time to train. Comparing Case 3 and Case 7, it can be seen that the time complexity problem of KAN linear can be solved with the help of fast KAN, although the fast KAN algorithm is still in its early stages. Third, we further reveal the advantages of KAN. Comparing Cases 5-9, it can be seen that after replacing the original linear layer with KAN, the number of rounds to reach the minimum RMSE will be significantly advanced, which partly shows the embedding of KAN. Fourth, different feature fusion schemes have different effects on the performance of GCN based on KAN. For actual tasks and data, the designers of algorithm prediction systems need to carefully consider the compromise between encoding and model and their fusion schemes.

[0089] Table 2 Benchmark test results

[0090]

[0091]

[0092] exist Figure 2 , Figure 3 and Figure 4 In this paper, a method and algorithm prediction system for calculating protein ligand binding affinity are proposed, and its model is called KANDTA. This method aims to reduce the difficulty of drug development, reduce the time and cost of discovering the interaction between drug small molecules and proteins, and shorten the drug development cycle.

[0093] The KAN-based GCN protein ligand binding affinity prediction scheme of the present invention not only presents a new deep learning model, but also improves the interpretability of the model and prediction results. The ability to process graph structured data, interpretability, and the combination of multiple KANs and GCNs will make the proposed prediction system have more application potential, such as Western medicine drug discovery and the construction of early Chinese medicine drug-target networks, which can help understand the interaction between Chinese medicine ingredients and targets, because the work of graph neural networks in virtual screening and reuse of Chinese medicine drugs is still in its early stages.

[0094] Figure 5 The KAN combination and the optimization according to task requirements in the affinity prediction system are shown, that is, the benchmark test results in Table 2 are Figure 5 The data visualization shown, where RMSE, number of rounds, and time are multiplied by the scaling factors 200, 1, and 20 for the values ​​in Table 2, respectively. Figure 5The case number indicates the specific encoding proposed from feature to encoding, and the fusion sequence number indicates the specific feature fusion proposed. The proposed KAN-based GCN protein ligand affinity predictor is a protein ligand affinity prediction system in the art, which uses a combination of linear KAN (such as KAN linear) and fast KAN (such as fastKAN) and the concept of optimization according to task requirements can be further generalized to the GNN protein ligand affinity prediction design. The concept of combination and optimization according to task requirements can enhance the training of GNN, because KAN improves the training quality and interpretability of traditional GNN predictors in terms of achieving the minimum RMSE rounds; the GCN used in the present invention can be extended to other GCN variants.

[0095] The specific implementation scheme described above further describes in detail the purpose, technical scheme and beneficial effects of the present invention. It should be understood that the above is only a specific implementation scheme of the present invention and is not intended to limit the scope of the present invention. Any equivalent changes and modifications made by any technician in the field without departing from the concept and principle of the present invention should fall within the scope of protection of the present invention.

Claims

1. A Kolmogorov GCN protein ligand affinity prediction method, characterized in that: The method comprises the following steps: Obtain two ligand molecule graphs and protein graphs to be predicted; Use GCN to encode the two predicted graphs and obtain the corresponding embedding vectors; Using KAN to replace the traditional flat layer to obtain a new flat layer; The proposed feature fusion strategy is used to fuse the ligand flat layer output and the protein flat layer output to obtain an updated encoding representation vector; The encoded representation is input into multiple task-related KAN linear layers, and the final output loss is used as a predictor indicator to evaluate the interaction relationship between the protein ligand to be predicted; The design of a protein-ligand affinity prediction system integrating a KAN sublayer requires weighing multiple metrics to obtain a preferred KAN-based GCN scheme.

2. A Kolmogorov GCN protein ligand affinity prediction method as claimed in claim 1, characterized in that: The method obtains two ligand molecular graphs and protein graphs to be predicted, including: considering each small molecule ligand molecule as a graph, considering the chemical bonds between ligand atoms as edges, so as to obtain the atom list and adjacency matrix in the molecular graph, defining the amino acids in the protein as nodes, and connecting non-continuous amino acids through edges, and generating a two-dimensional protein graph from the contact graph, that is, converting the two-dimensional structure of the protein into a graph representation.

3. A Kolmogorov GCN protein ligand affinity prediction method as claimed in claim 1, characterized in that: The method uses the first part of representation learning, i.e. encoding the 2D ligand molecule graph and the 2D protein graph respectively using GCN stacking to obtain the corresponding embedding vectors.

4. A Kolmogorov GCN protein ligand affinity prediction method as claimed in claim 1, characterized in that: The use of the new flat layer, i.e. the second part of the representation learning, respectively utilizes the ligand embedding vector output by the GCN stack and the protein embedding vector output by the GCN stack to perform KAN stacking. The KAN stacking uses KAN linear, fastKAN and traditional fully connected layers or dense layers and their combinations. The fastKAN is a fast implementation of KAN linear. The second part entity will output the corresponding, updated representation vector matrices X and Xt.

5. A Kolmogorov GCN protein ligand affinity prediction method as claimed in claim 1, characterized in that: The method utilizes the proposed feature fusion strategy to achieve task- and data-related feature fusion, that is, introducing a combination of feature fusion schemes including an attention mechanism and selecting the next one to fuse the ligand flat layer output X and the protein flat layer output Xt to obtain a feature-fused, updated encoding representation vector.

6. A Kolmogorov GCN protein ligand affinity prediction method as claimed in claim 1, characterized in that: The method uses task-related learning to process the representation vector after feature fusion, and inputs the encoded representation after feature fusion into multiple task-related KAN linear layers. Finally, the output loss will be used as a predictor indicator to evaluate the interaction relationship between the protein ligands to be predicted. The multiple KAN linear layers are a combination of KAN linear layers and traditional fully connected layers.

7. A Kolmogorov GCN protein ligand affinity prediction method as claimed in claim 1, characterized in that: The method uses the proposed protein ligand affinity prediction system to obtain the preferred KAN-based GCN scheme. The performance of the affinity predictor integrating the KAN sublayer is related to the dataset, data reconstruction, representation learning and task-related learning. The predictor indicators for different combinations will be different, such as RMSE, the number of rounds to reach the minimum RMSE and the training time.

Citation Information

Cited By

  • Drug-drug interaction prediction method based on multi-channel feature fusion

    CN119446314A

  • Tumor position monitoring method and system

    CN119868823A

  • Drug target affinity prediction method based on state space model and time-frequency transformation

    CN119964634A

  • Drug and target effect prediction method based on Floyd algorithm network

    CN120375913A

  • A method for predicting drug-target interactions based on Freud algorithm network

    CN120375913B