Molecular graph generation method, device, storage medium, and program product

Abstract

The application discloses a molecular graph generation method and device, a storage medium and a program product, relates to the technical field of molecular graph processing, and comprises the following steps: predicting an input noise graph through a graph generation model to determine a probability adjacency matrix of the noise graph; performing power operation on the probability adjacency matrix and an original adjacency matrix of an original molecular graph multiple times to obtain a first high-order adjacency matrix of multiple noise graphs and a second high-order adjacency matrix of the original molecular graph; constructing a target loss function according to the first high-order adjacency matrices and the second high-order adjacency matrices; updating the graph generation model according to a back propagation algorithm and the target loss function, and performing the step of predicting the input noise graph through the graph generation model until the target loss function converges, so that a trained graph generation model is obtained. Through local-to-global structure consistency learning, the rationality of a molecular structure generated by the model is improved.

Description

Technical Field

[0001] This application relates to the field of molecular diagram processing technology, and in particular to molecular diagram generation methods, devices, storage media and program products. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning-based molecular generation models have become the mainstream technology for de novo molecular design. These models typically abstract molecules into graph structures composed of atoms (nodes) and chemical bonds (edges), and generate novel molecular graphs with expected physicochemical properties or biological activities by learning from the data distribution of massive amounts of known molecules.

[0003] Current common software implementations of autoregressive models, diffusion models, and flow-matching models predict molecular structures primarily by focusing on each atom, forecasting the atom type and its directly connected chemical bonds. However, this highly localized prediction process, revolving only around each atom, fails to explicitly model the global topological structure of the molecule. This makes the models inadequate for handling complex molecular structures, such as accurately grasping complex ring systems and long-range spatial relationships between atoms, resulting in low prediction accuracy and difficulty in generating structurally sound molecules.

[0004] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this application is to provide a molecular diagram generation method, apparatus, storage medium, and program product, aiming to solve the technical problem of how to improve the rationality of the molecular structure generated by the model.

[0006] To achieve the above objectives, this application proposes a method for generating molecular maps, the method comprising:

[0007] The input noisy graph is predicted using a graph generation model to determine the probability adjacency matrix of the noisy graph, wherein the noisy graph is obtained by injecting noise into the original molecular graph;

[0008] By performing multiple exponentiation operations on the probability adjacency matrix and the original adjacency matrix of the original molecular graph, a plurality of first higher-order adjacency matrices of the noise graph and a second higher-order adjacency matrix of the original molecular graph are obtained.

[0009] Based on each of the first higher-order adjacency matrices and each of the second higher-order adjacency matrices, construct the target loss function;

[0010] According to the back propagation algorithm and the target loss function, the graph generation model is updated, and the step of predicting the input noise graph by the graph generation model is performed until the target loss function converges, so as to obtain the trained graph generation model.

[0011] In an embodiment, the step of predicting the input noise graph by the graph generation model and determining the probability adjacency matrix of the noise graph comprises:

[0012] The prediction edge feature corresponding to the noise graph is determined by predicting the input noise graph by the graph generation model.

[0013] According to the preset classification activation function and the prediction edge feature, the connection probability between each atom of the noise graph is determined.

[0014] According to the connection probability between each atom of the noise graph, the probability adjacency matrix is determined.

[0015] In an embodiment, the target loss function comprises a local loss function and a topological loss function, and the step of predicting the input noise graph by the graph generation model and determining the prediction edge feature of the noise graph comprises:

[0016] The prediction node feature and the prediction edge feature of the noise graph are determined by predicting the input noise graph by the graph generation model.

[0017] The step of determining the target loss value according to each first high-order adjacency matrix and each second high-order adjacency matrix comprises:

[0018] The topological loss function is determined according to each first high-order adjacency matrix and each second high-order adjacency matrix.

[0019] The local loss function is determined according to the prediction node feature, the prediction edge feature, the atomic feature of the original molecular graph and the bond feature of the original molecular graph.

[0020] The target loss function is constructed according to the topological loss function and the local loss function.

[0021] In an embodiment, the step of determining the topological loss function according to each first high-order adjacency matrix and each second high-order adjacency matrix comprises:

[0022] Each first high-order adjacency matrix and each second high-order adjacency matrix is normalized.

[0023] According to the preset matrix norm, the difference between the first high-order adjacency matrix and the second high-order adjacency matrix with the same power is determined, and the loss value corresponding to each power is obtained.

[0024] According to the loss value corresponding to each power, the topological loss function is determined.

[0025] In an embodiment, before the step of predicting, by the graph generation model, an input noise graph and determining a probabilistic adjacency matrix of the noise graph, the method further comprises:

[0026] obtaining a bondless molecular graph corresponding to the original molecular graph, wherein the bondless molecular graph and the original molecular graph have the same number of atoms, and the bondless molecular graph has no chemical bond;

[0027] generating a noise graph according to a preset discrete flow matching rule, the original molecular graph, and the bondless molecular graph.

[0028] In an embodiment, the discrete flow matching rule includes a random time step, and the step of generating a noise graph according to a preset discrete flow matching rule, the original molecular graph, and the bondless molecular graph comprises:

[0029] determining a replacement probability of each atom and each chemical bond in the original molecular graph according to the random time step;

[0030] replacing atoms and chemical bonds in the original molecular graph with atoms and chemical bonds in the bondless molecular graph according to the replacement probability to obtain the noise graph.

[0031] In an embodiment, after the step of obtaining the trained graph generation model, the method further comprises:

[0032] processing a to-be-constructed molecular graph by the trained graph generation model to output an atom type prediction score and a chemical bond type prediction score;

[0033] determining a real atom type according to the atom type prediction score and a real chemical bond type according to the chemical bond type prediction score;

[0034] generating a target molecular graph according to the real atom type and the real chemical bond type.

[0035] In addition, to achieve the above object, the application further provides a molecular graph generation device, which comprises:

[0036] a model prediction module configured to predict, by a graph generation model, an input noise graph and determine a probabilistic adjacency matrix of the noise graph, wherein the noise graph is obtained by noise injection on an original molecular graph;

[0037] a power operation module configured to perform multiple power operations on the probabilistic adjacency matrix and an original adjacency matrix of the original molecular graph to obtain a first high-order adjacency matrix of a plurality of noise graphs and a second high-order adjacency matrix of the original molecular graph.

[0038] The loss construction module is used to construct the target loss function based on each of the first higher-order adjacency matrices and each of the second higher-order adjacency matrices;

[0039] The model update module is used to update the graph generation model according to the backpropagation algorithm and the target loss function, and to perform the step of predicting the input noisy graph through the graph generation model until the target loss function converges to obtain the trained graph generation model.

[0040] In addition, to achieve the above objectives, this application also proposes an electronic device, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the molecular map generation method as described above.

[0041] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the molecular diagram generation method described above.

[0042] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the molecular map generation method described above.

[0043] The one or more technical solutions provided in the application have at least the following technical effects: first, the input noise graph is predicted by the graph generation model to determine the probability adjacency matrix of the noise graph, so that the model learns the distribution of the atomic nodes and chemical bonds of the original molecular graph from the noise, laying a foundation for subsequent structure recovery; then, the power operation is performed on the probability adjacency matrix and the original adjacency matrix of the original molecular graph multiple times to obtain a first high-order adjacency matrix of multiple noise graphs and a second high-order adjacency matrix of the original molecular graph, the power operation of the adjacency matrix learns the connection between the atomic nodes through the multi-hop path, and the global topological features of the multi-step path in the molecular graph are explicitly captured; then, the target loss function is constructed according to each first high-order adjacency matrix and each second high-order adjacency matrix, the similarity of the generated graph and the real graph in the topological structure is more accurately and comprehensively measured by comparing the global structure difference, the structure distortion problem that may occur when only the first-order connection is compared is avoided, and the model is guided to generate a more reasonable molecular graph in the overall topology; then, the graph generation model is updated according to the back propagation algorithm and the target loss function until the target loss function converges, the trained graph generation model is obtained, and an efficient closed loop of the model training process is formed, so that the model can be closer to the structure characteristics of the real molecule when generating the molecular graph. The application improves the modeling ability of the graph generation model for the global topological structure of the molecular graph through the structure consistency learning from the local to the global, enhances the prediction accuracy of the model in the complex molecular environment, and thus improves the rationality of the molecular structure generated by the model. BRIEF DESCRIPTION OF DRAWINGS

[0044] The accompanying drawings, which are incorporated into and form part of the specification, illustrate embodiments consistent with the present application and, together with the specification, serve to explain the principles of the application.

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0046] Figure 1 The flowchart provided for the molecular graph generation method embodiment one of the application;

[0047] Figure 2 The construction flowchart of the topological loss function provided for the embodiment one of the application;

[0048] Figure 3 The overall flowchart of the molecular graph generation method provided for the embodiment two of the application;

[0049] Figure 4 The module structure diagram of the molecular graph generation device provided for the embodiment of the application;

[0050] Figure 5 The device structure diagram of the hardware running environment involved in the molecule graph generation method in the embodiments of the present application.

[0051] The object implementation, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0052] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application, and are not used to limit the present application.

[0053] In order to better understand the technical solutions of the present application, the specific embodiments will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] The common software implementation schemes such as autoregressive model, diffusion model and flow matching model at present mainly take each atom as the center to predict the type of the atom and the chemical bond directly connected with the atom when predicting the molecular structure. However, this prediction process is highly localized, only develops around each atom, cannot explicitly model the global topological structure of the molecule, so that the model is not satisfactory when dealing with complex molecular structure, for example, it is difficult to accurately grasp the complex ring system structure and the long-range spatial relationship between atoms, thereby resulting in low prediction accuracy and being difficult to produce a molecular structure with reasonable structure.

[0055] The application provides a solution. First, a graph generation model is used to predict an input noise graph to determine a probability adjacency matrix of the noise graph, so that the model learns the distribution of atomic nodes and chemical bonds of an original molecular graph from noise, laying a foundation for subsequent structure recovery. Then, the probability adjacency matrix and an original adjacency matrix of the original molecular graph are subjected to multiple power operations to obtain a first high-order adjacency matrix of the noise graph and a second high-order adjacency matrix of the original molecular graph. The power operation of the adjacency matrix learns the connection between atomic nodes through a multi-hop path, and explicitly captures the global topological features of the multi-step path in the molecular graph. Then, a target loss function is constructed according to each first high-order adjacency matrix and each second high-order adjacency matrix. By comparing the global structural differences, the similarity of the generated graph and the real graph in the topological structure is more accurately and comprehensively measured, the structural distortion problem that may occur when only the first-order connection is compared is avoided, and the model is guided to generate a more reasonable molecular graph in the overall topology. Then, the graph generation model is updated according to the back propagation algorithm and the target loss function until the target loss function converges, the trained graph generation model is obtained, and an efficient closed loop of the model training process is formed, so that the model can be closer to the structural features of the real molecule when generating the molecular graph. Through the structure consistency learning from the local to the global, the modeling ability of the graph generation model for the global topological structure of the molecular graph is improved, the prediction accuracy of the model in the complex molecular environment is enhanced, and the rationality of the molecular structure generated by the model is improved.

[0056] It should be noted that the execution subject of the embodiment can be an electronic device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone and the like.

[0057] Based on this, the embodiment of the application provides a molecular graph generation method, referring to Figure 1 , Figure 1 The flowchart of the first embodiment of the molecular graph generation method of the application is shown.

[0058] In the embodiment, the molecular graph generation method includes steps S10-S40:

[0059] In step S10, a graph generation model is used to predict an input noise graph to determine a probability adjacency matrix of the noise graph, wherein the noise graph is obtained by noise injection on an original molecular graph.

[0060] The graph generation model refers to a deep learning-based computing model, usually adopting a Generative Adversarial Network (GAN), a Variational Auto-Encoders (VAE) architecture, etc., for learning distribution from input graph data and generating new graph structures; it can capture local and global patterns of a graph by learning multi-layer node and edge feature transformations, and the trained graph generation model can be used for molecular graph generation.

[0061] The noise graph refers to a degenerate graph representation generated by applying random perturbations (such as randomly changing atom types or removing chemical bonds) to the adjacency matrix or node features of the original molecular graph, for simulating data incompleteness or enhancing training diversity. The original molecular graph refers to a graph structure representation of a real molecule, composed of nodes (atoms) and edges (chemical bonds), which can be represented in the form of G= (X, A, E), where G represents the molecular graph; X is a node feature matrix, storing information such as the type of each atom; A is an adjacency matrix, where the value of each element is 0 or 1, used to represent whether there is a chemical bond between atoms, for example represents that there is a chemical bond between atom j and atom k; E is an edge feature matrix, storing the specific type of each chemical bond (such as single bond, double bond, etc.).

[0062] The probability adjacency matrix refers to a matrix used to describe the connection probability between nodes in a molecular graph, where each element represents the probability of connection between the corresponding two nodes.

[0063] Exemplarily, the original molecular graph can be loaded from a database, and random noise can be added to the adjacency matrix or node features of the original molecular graph by multiplying a noise mask or adding Gaussian distribution random values, etc., to generate a noise graph; then, the noise graph is taken as the input of the graph generation model, triggering the forward calculation process of the model; then, the graph generation model can process the noise graph through internal layers (such as graph convolution layers or attention mechanisms), outputting a preliminary probability adjacency matrix for subsequent processing.

[0064] In a feasible implementation, step S10 includes:

[0065] Step S11, predicting the input noise graph through the graph generation model to determine the predicted edge features corresponding to the noise graph.

[0066] The predicted edge features refer to a high-dimensional vector or matrix encoding the probability of edges (chemical bonds) in the noise graph belonging to various types of chemical bonds, obtained through forward propagation calculation of the graph generation model; the element values in the vector or matrix can represent the likelihood of a predicted edge being a single bond, double bond, or no bond, etc.

[0067] Step S12, according to the preset classification activation function and the predicted edge feature, the connection probability between each atom of the noise graph is determined;

[0068] The classification activation function refers to a mathematical function used to map the predicted edge feature to the probability space, such as Softmax, Sigmoid function, etc., which is used to output the probability of each edge belonging to a specific category (such as single bond, double bond or no bond, etc.), ensuring that the result is within the range of [0, 1] and conforms to the probability distribution; and the probability of each edge belonging to a specific category other than no bond is added, that is, the probability of the existence of a chemical bond between each atom (connection probability) is obtained.

[0069] Step S13, according to the connection probability between each atom of the noise graph, a probability adjacency matrix is determined.

[0070] Exemplarily, the molecule graph (noise graph) subjected to noise injection is input into the graph generation model to trigger the forward calculation process of the model; then, the graph generation model processes the noise graph through internal layers to output the feature vector of each edge between the atom nodes, obtaining the edge prediction feature, wherein the feature vector of each edge encodes the feature attributes of the edge (chemical bond), such as bond type, bond strength, etc.; then, the predicted edge feature is processed through the preset classification activation function (such as Softmax) to calculate the probability value of each edge corresponding to different chemical bond types, and the probabilities of each edge belonging to other categories other than no bond are added to obtain the connection probability between each atom; then, the connection probability between each atom is organized into a matrix form, the rows and columns of the matrix correspond to the atom nodes in the noise graph, and the matrix element value is the connection probability between the corresponding atoms.

[0071] In this embodiment, the graph generation model is used to predict the edge feature, capture the feature attributes of the chemical bond, and improve the learning accuracy of the local structure; then, the classification activation function is introduced to construct a probability adjacency matrix according to the predicted edge feature output by the model, so that the model can consider the connection relationship between all atoms at the same time, rather than judging the isolated atom pair, thereby enhancing the model's understanding ability of the overall topology of the molecule, so as to evaluate and further optimize the model's ability to construct the overall topology of the molecule graph during the training process.

[0072] Step S20, the probability adjacency matrix and the original adjacency matrix of the original molecule graph are subjected to multiple power operations respectively to obtain a first high-order adjacency matrix of multiple noise graphs and a second high-order adjacency matrix of the original molecule graph;

[0073] The original adjacency matrix refers to the binary connection matrix of the original molecule graph, which is used as the target benchmark for model training; the element value thereof reflects whether there is a real connection between each atom node, and in general cases, the element value of 1 indicates the existence of a chemical bond connection, and the element value of 0 indicates no connection.

[0074] A higher-order adjacency matrix is ​​an adjacency matrix obtained after exponentiation. It contains information about the higher-order connections between atomic nodes in the molecular graph, reflecting the connection possibilities and path data after multiple jumps. For example, an m-order adjacency matrix obtained by exponentiation m times. In, elements This represents the number of paths from atom i to atom j that require m jumps (the number of paths of length m from atom i to j). To distinguish them, the higher-order adjacency matrix obtained by exponentiation of the predicted probability adjacency matrix is ​​called the first higher-order adjacency matrix, and the higher-order adjacency matrix obtained by exponentiation of the original adjacency matrix is ​​called the second higher-order adjacency matrix.

[0075] For example, it can be based on a preset maximum number of hops k, for For each integer in the matrix, the probability adjacency matrix output by the graph generation model and the original adjacency matrix are raised to the power of m to obtain the first m-order adjacency matrix and the second m-order adjacency matrix. As the value of m increases, each power operation enhances the representation of long-range connectivity information in the matrix. Lower orders (e.g., m=2) capture local loop features, while higher orders (e.g., m=5) capture the global topology of the molecule. The resulting set of high-order matrices is then fed into the loss calculation module, where the differences between the generated results and the real molecules are compared at various topological levels.

[0076] Understandably, by directly capturing multi-step path information through exponentiation, higher-order adjacency matrices of different orders provide multi-level supervision signals from local to global, enabling the model to simultaneously optimize structural accuracy at different scales during training. This allows the model to perceive global structural relationships beyond local neighbors, improving the rationality and chemical effectiveness of the molecular structures generated by the model.

[0077] Step S30: Construct the target loss function based on each first higher-order adjacency matrix and each second higher-order adjacency matrix;

[0078] The target loss function is a function used to measure the difference between the model's predicted output and the true label. By minimizing the target loss function, the model can continuously adjust its parameters to make the prediction result closer to the true label. In this embodiment, the target loss function can be constructed by integrating and weighting the differences between the multiple first higher-order adjacency matrices corresponding to the prediction result (probability adjacency matrix) and the multiple second higher-order adjacency matrices corresponding to the true label (original adjacency matrix).

[0079] In one feasible implementation, the target loss function includes a local loss function and a topological loss function, and step S11 includes:

[0080] Step S111, predicting the input noise graph through the graph generation model to determine predicted node features and predicted edge features of the noise graph;

[0081] The predicted node features refer to predicted values of attributes or features possessed by each node in the noise graph, such as atomic types, which are predicted by the graph generation model according to the input noise graph and the graph structure knowledge learned by the model itself.

[0082] For example, in the process of predicting the input noise graph through the graph generation model, first, the node information and edge information in the noise graph are extracted through the model; then, the node information and edge information extracted at different levels are transmitted and fused through the graph attention network in the model, and these information is used for prediction, and finally the predicted node features and predicted edge features are output.

[0083] Step S30 includes:

[0084] Step S31, determining a topological loss function according to each first high-order adjacency matrix and each second high-order adjacency matrix;

[0085] The topological loss function refers to a loss function for measuring the difference between the first and second high-order adjacency matrices, reflecting the gap between the model prediction results and the original molecular graph in the global topological structure, and is used to guide the graph generation model to optimize the topological structure.

[0086] For example, the element difference between each first high-order adjacency matrix and the corresponding second high-order adjacency matrix of the power can be calculated to obtain a plurality of difference matrices; then, each element in the difference matrix is squared or taken the absolute value, etc. to obtain an error matrix at the element level; then, according to the relative importance of each power, the error matrices corresponding to different powers are weighted and summed to construct the topological loss function.

[0087] It can be understood that through the supervision of the high-order adjacency matrix, the model can learn to maintain the overall topological structure of the molecule, solve the topological distortion problem caused by the local view of the traditional method, and thus improve the structural rationality and chemical effectiveness of the generated molecule.

[0088] In a feasible implementation, step S31 includes:

[0089] Step S311, normalizing each first high-order adjacency matrix and each second high-order adjacency matrix;

[0090] Normalization refers to a numerical preprocessing method that maps matrix elements to a uniform numerical range through linear or nonlinear transformation, which is used to eliminate the problem of dimension difference and numerical imbalance; common normalization methods include minimum-maximum normalization, symmetric normalization, logarithmic normalization, etc., and the normalization method used in the present embodiment is not specifically limited.

[0091] Optionally, the normalization processing method can be determined according to the order of each high-order adjacency matrix.

[0092] For example, for a second-order adjacency matrix, symmetric normalization can be used to eliminate the influence of node degree on path count; for a third-order or higher-order adjacency matrix, logarithmic normalization can be used to reduce the exponential growth of the number of paths of the same length and reduce the numerical explosion problem.

[0093] It can be understood that, through normalization processing, the dimensional difference between matrices of different orders can be eliminated, the contribution of each order topological feature in the loss function is balanced, and the low-order features are not overwhelmed by the high-order features; at the same time, the numerical explosion problem of high-order matrices can be effectively suppressed, the oscillation phenomenon in the reverse gradient optimization process is reduced, and the stable progress of the model training process is ensured.

[0094] In step S312, the difference between the first high-order adjacency matrix and the second high-order adjacency matrix of the same power is determined according to the preset matrix norm, and a loss value corresponding to each power is obtained.

[0095] The matrix norm refers to a mathematical operator used to measure the overall size or distance of the matrix elements. Common matrix norms include Frobenius norm (square root of the sum of the squares of the matrix elements), spectral norm (maximum singular value), and L1 norm (sum of the absolute values of the elements).

[0096] For example, for each order (power) of the first high-order adjacency matrix and the second high-order adjacency matrix, the square of the Frobenius norm can be used to measure the difference between the two, which specifically includes: calculating the difference matrix between the first high-order adjacency matrix and the second high-order adjacency matrix, and calculating the sum of the squares of the elements in the difference matrix to obtain the square value of the norm; then, the square value of the norm is divided by the number of elements in the difference matrix, and the loss value under the order is obtained.

[0097] In step S313, a topological loss function is determined according to the loss value corresponding to each power.

[0098] For example, according to the preset topological loss function construction rule (such as weighted summation, mean calculation, etc.), the loss values corresponding to each power can be integrated into a topological loss function through mathematical operations; wherein the determination of the weight coefficient can be based on prior knowledge, experimental experience or automatically learned through an optimization algorithm.

[0099] For example, please refer to Figure 2 , Figure 2A101, normalizing each high-order adjacency matrix (including the first high-order adjacency matrix and the second high-order adjacency matrix); A102, calculating the difference between the first high-order adjacency matrix and the second high-order adjacency matrix of the same order by using matrix norm, obtaining the matrix loss value of each order (power), and eliminating the dimension effect and the influence of abnormal values by using the matrix norm; and A103, weighting and summing the matrix loss values of all orders to combine the topological loss function.

[0100] In this embodiment, by normalization and matrix norm processing, the topological difference between the model prediction result and the real situation can be more accurately measured, the model can more accurately capture the global structural features of the molecule, and the chemical effectiveness and structural stability of the generated molecule can be improved. In addition, by calculating the loss value corresponding to each power to construct the topological loss function, the graph generation model can comprehensively consider the connection relationship of different orders in the graph structure. The connection relationship of low power reflects the direct or close interaction between nodes, and the connection relationship of high power reflects the indirect or long-distance influence between nodes. By comprehensively considering the loss values of different powers, the topological loss function can more comprehensively describe the topological features of the graph structure, and improve the global understanding ability of the model for the molecular structure.

[0101] In step S32, a local loss function is determined according to the predicted node features, the predicted edge features, the atomic features of the original molecular graph, and the bond features of the original molecular graph.

[0102] The atomic features and the bond features of the original molecular graph refer to the reference attribute information about the atomic nodes and the chemical bonds directly extracted from the original molecular graph, which is used to provide accurate supervision signals for model training.

[0103] The local loss function refers to a function for measuring the difference between the prediction result of the graph generation model and the original molecular graph in the local structure of the molecule, which is used to guide the model to more accurately predict the local structure of the molecule.

[0104] Optionally, the node loss can be determined according to the predicted node features and the atomic features of the original molecular graph, and the edge loss can be determined according to the predicted edge features and the bond features of the original molecular graph. Then, the local loss function is determined according to the node loss and the edge loss.

[0105] For example, the difference between the predicted node features and the atomic features of the original molecular graph, and the difference between the predicted edge features and the bond features of the original molecular graph can be calculated by using the cross-entropy function to construct the local loss function. Please refer to the following formula:

[0106]

[0107] wherein, denotes a local loss function, denotes a node loss, denotes an edge loss; N denotes the number of atomic nodes; denotes the value corresponding to the i-th atomic node in the atomic feature, denotes the value corresponding to the i-th atomic node in the predicted node feature generated by the model; denotes the value corresponding to the edge from the i-th atomic node to the j-th atomic node in the bond feature, denotes the value corresponding to the edge from the i-th atomic node to the j-th atomic node in the predicted edge feature generated by the model.

[0108] Step S33, constructing a target loss function according to the topological loss function and the local loss function.

[0109] Exemplarily, the target loss function is obtained by weighted sum of the topological loss function and the local loss function wherein, λ is a hyperparameter balancing the importance of the topological loss and the local loss, which is manually set by the user and remains unchanged during the model training process.

[0110] In the present embodiment, the target loss function is constructed by combining the topological loss function and the local loss function, so that the graph generation model pays attention to both the overall structure and the local details of the data during the training process, which helps the model to learn the features of the data more comprehensively and improve the prediction accuracy of the model in various aspects (overall and local), thereby improving the rationality of the molecular structure generated by the model.

[0111] Step S40, updating the graph generation model according to the backpropagation algorithm and the target loss function, and performing the step of predicting the input noise graph by the graph generation model until the target loss function converges, obtaining the trained graph generation model.

[0112] The backpropagation algorithm refers to an optimization algorithm for training neural networks, which calculates the gradient of the target loss function with respect to the parameters of the neural network model (corresponding to the graph generation model in the present embodiment), and then uses an optimization algorithm to update the model parameters, thereby minimizing the loss function.

[0113] Exemplarily, the partial derivative of the target loss function can be calculated to calculate the gradient of each network layer in the graph generation model; then, according to the pre-set optimization algorithm (such as Adam algorithm, gradient descent method) etc., the update amount of each parameter in each network layer is calculated, and the parameters of the graph generation model are updated accordingly; then the noise graph is predicted again by the updated graph generation model until the target loss function converges, obtaining the trained graph generation model.

[0114] The embodiment provides a molecular graph generation method. Through multi-scale topological supervision of a high-order adjacency matrix, the model can capture complex ring systems and long-range dependency relationships, so as to guide the model to generate a more reasonable molecular graph in the overall topology. Through the combination of node feature supervision, edge feature supervision and high-order topological constraints, double supervision of the local structure and the global topology of the molecular graph in the graph generation model training process is realized, the structure consistency learning from the local to the global is realized, the modeling capability of the graph generation model for the global topological structure of the molecular graph is improved, the prediction accuracy of the model in the complex molecular environment is enhanced, and the rationality of the molecular structure generated by the model is improved.

[0115] Based on the first embodiment of the present application, the same or similar contents as the above embodiment one can be referred to the above introduction, and will not be repeated hereinafter. On this basis, before step S10, it also includes:

[0116] Step S01, obtaining a bondless molecular graph corresponding to an original molecular graph, wherein the number of atoms of the bondless molecular graph is the same as that of the original molecular graph, and the bondless molecular graph has no chemical bond;

[0117] The bondless molecular graph refers to a degenerate graph structure that only retains the atom node number information of the original molecular graph and removes all chemical bonds, so that its adjacency feature matrix is a zero matrix, and all entries in its edge feature matrix are in the state of “no chemical bond”. The types of all atoms can be initialized to a unified basic type, usually the most common carbon atom in chemistry; or part of the atom nodes in the original molecular graph can be retained, and the other atoms are initialized to carbon atoms, which is not limited in the present embodiment.

[0118] Step S02, generating a noise graph according to a preset discrete flow matching rule, the original molecular graph and the bondless molecular graph.

[0119] The discrete flow matching (DFM) rule refers to a mathematical framework based on probabilistic path interpolation, which generates a noise graph by discretely introducing noise information in the bondless molecular graph into the original molecular graph, so that the generated noise graph not only retains part of the structural features of the original molecular graph, but also has certain randomness and diversity, so as to simulate the possible state of the molecular graph under different noise disturbances.

[0120] In a feasible implementation manner, the discrete flow matching rule includes a random time step, and step S02 includes:

[0121] Step S021, determining the replacement probability of each atom and each chemical bond in the original molecular graph according to the random time step;

[0122] The random time step refers to a randomly selected time interval in the process of generating the noise graph; in each time step, a certain degree of replacement operation is performed on the original molecular graph according to the discrete flow matching rule, the closer the time step t is to 1, the more noise is introduced, and the stronger the noise of the generated noise graph is; the closer t is to 0, the weaker the noise of the generated noise graph is. In this embodiment, the introduction of the random time step makes the generation process of the noise graph uncertain, and can simulate the changes of the molecular structure under different interference degrees.

[0123] In step S022, according to the replacement probability, the atoms and chemical bonds in the original molecular graph are replaced with the atoms and chemical bonds in the bondless molecular graph to obtain a noise graph.

[0124] Exemplarily, a time step t can be randomly sampled in the continuous interval of [0, 1]; then for each atom and chemical bond in the original molecular graph, the original feature is kept unchanged with a probability of (1-t), and the feature is replaced with the corresponding feature (such as a carbon atom or a bondless chemical bond) in the bondless molecular graph with a replacement probability of t.

[0125] Exemplarily, in order to assist in understanding the implementation process of the molecular graph generation method obtained after combining the above-mentioned embodiment one, please refer to Figure 3 , Figure 3 A general flowchart of a molecular graph generation method is provided, which first obtains an original molecular graph in a training data set (B101), and samples a noise graph (B102) therefrom. The sampling process can be realized based on the discrete flow matching rule; then B103 is performed, the noise graph is input into a graph generation model for processing, the model outputs the predicted edge features and predicted node features corresponding to the noise graph (B104); then B105 is performed, the composite loss is calculated (including local loss calculation and topological loss calculation), the cross-entropy loss calculation is performed on the predicted node features and the atom features of the original molecular graph, and the predicted edge features and the chemical bond features of the original molecular graph, respectively, to obtain the local loss; the probability adjacency matrix is generated according to the predicted edge features, the power operation is performed on the probability adjacency matrix and the original adjacency matrix of the original molecular graph respectively, and the single-order loss between the high-order adjacency matrices of each power is calculated to obtain the topological loss of the noise graph; then after the loss calculation is completed, the model parameters can be updated by back propagation (B106); then after the model update is completed, the original molecular graph in the training data set can be reacquired, and the above-mentioned process can be repeated until the loss function converges.

[0126] In this embodiment, by introducing a discrete flow matching rule, determining the replacement probability according to the random time step, and generating a noise graph according to the replacement probability, diversified data samples are provided for the training of the graph generation model, which helps the model to learn the feature representation of the molecular structure under different replacement disturbances, enhances the adaptability of the model to the change of the molecular structure, so that the graph generation model can more accurately predict the situation after the atoms or chemical bonds are partially replaced, and improve the quality and structural rationality of the generated molecules.

[0127] In a feasible implementation, after step S40, the method further comprises:

[0128] Step S50, processing the to-be-constructed molecular graph by the trained graph generation model to output atom type prediction scores and chemical bond type prediction scores;

[0129] The to-be-constructed molecular graph refers to the initial input of the molecular generation model, which can be a noise graph, a simple skeleton (such as a bondless molecular graph) or a partial structure, serving as the starting point of molecular construction.

[0130] The atom type prediction score refers to the probability matrix of each atom belonging to each element type (C, N, O, etc.) output by the trained graph generation model. The higher the score of each row of atoms, the greater the possibility that the atom belongs to the corresponding score type.

[0131] The chemical bond type prediction score refers to the probability tensor of each pair of atoms forming various types of chemical bonds (single bond, double bond or no bond, etc.) output by the trained graph generation model. The higher the score of each row of chemical bonds, the greater the possibility that the chemical bond belongs to the corresponding score type.

[0132] Illustratively, the to-be-constructed molecular graph is input into the trained model, forward propagation calculation is performed, and the probability distribution of all atoms and chemical bonds is output to obtain the atom type prediction score and the chemical bond type prediction score.

[0133] Step S60, determining the real atom type according to the atom type prediction score, and determining the real chemical bond type according to the chemical bond type prediction score;

[0134] The real atom type refers to the real atom category to which the discrete atom belongs based on the prediction score; the real chemical bond type refers to the real chemical bond category to which the chemical bond belongs based on the prediction score.

[0135] Illustratively, the argmax function can be applied to each row of the atom type prediction score to find the index with the highest score, thereby determining the final type of each atom; similarly, the argmax function can be applied to the score vector of each atom pair in the chemical bond prediction score to find the index with the highest score, thereby determining the final chemical bond type (which can be "no chemical bond") between them.

[0136] Step S70, generating a target molecule graph according to the real atomic types and the real bond types.

[0137] The target molecule graph refers to a complete graph structure composed of determined atomic nodes and bond edges, representing a generated new molecule, and can be used for molecular design evaluation and application.

[0138] Exemplarily, the real atomic types and the real bond types obtained by the above analysis are combined with the to-be-constructed molecule graph, so that the finally generated discrete molecule graph is obtained.

[0139] In this embodiment, by one forward propagation of the trained graph generation model, a molecule graph with local chemical correctness and global topological consistency can be generated, which shortens the time of molecular design and improves the research and development efficiency of molecular structure.

[0140] It should be noted that the above examples are only used to understand the present application and do not constitute a limitation on the molecular graph generation method of the present application. More forms of simple transformation based on this technical concept are within the protection scope of the present application.

[0141] The present application also provides a molecular graph generation device, please refer to Figure 4 The molecular graph generation device comprises:

[0142] The model prediction module 10 is configured to predict the input noise graph by the graph generation model to determine the probability adjacency matrix of the noise graph, wherein the noise graph is obtained by noise injection on the original molecule graph.

[0143] The power operation module 20 is configured to perform multiple power operations on the probability adjacency matrix and the original adjacency matrix of the original molecule graph respectively, to obtain a first high-order adjacency matrix of the multiple noise graphs and a second high-order adjacency matrix of the original molecule graph.

[0144] The loss construction module 30 is configured to construct a target loss function according to each first high-order adjacency matrix and each second high-order adjacency matrix.

[0145] The model update module 40 is configured to update the graph generation model according to the back propagation algorithm and the target loss function, and perform the step of predicting the input noise graph by the graph generation model until the target loss function converges, to obtain the trained graph generation model.

[0146] The molecular graph generation device provided in the embodiments of the present application adopts the molecular graph generation method in the above embodiments, and can solve the technical problem of how to improve the rationality of the molecular structure generated by the model. Compared with the prior art, the beneficial effects of the molecular graph generation device provided in the present application are the same as those of the molecular graph generation method provided in the above embodiments, and other technical features in the molecular graph generation device are the same as those disclosed in the above embodiments, and thus will not be described herein.

[0147] The embodiments of the present application provide an electronic device, which comprises at least one processor and a memory connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the molecular graph generation method in the above embodiment one.

[0148] Reference will be made to the following description of the embodiments of the present application, taken in conjunction with the accompanying drawings, in which Figure 5 which shows a structural schematic diagram of an electronic device suitable for implementing the embodiments of the present application. The electronic device in the embodiments of the present application can include, but is not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 5 The electronic device shown is only an example, and should not bring any limitation to the functions and use range of the embodiments of the present application.

[0149] As Figure 5As shown, the electronic device can include a processing device 1001 (e.g., a central processor, a graphics processor, etc.) that can perform various appropriate actions and processes according to programs stored in a read-only memory 1002 or loaded from a storage device 1003 into a random access memory 1004. Various programs and data required for operation of the electronic device are also stored in the random access memory 1004. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are connected to each other through a bus 1005. An input / output interface 1006 is also connected to the bus. Generally, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; the storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the electronic device to communicate wirelessly or wired with other devices to exchange data. Although the electronic device with various systems is shown in the figure, it should be understood that all the shown systems are not required to be implemented or possessed. More or less systems can be alternatively implemented or possessed.

[0150] In particular, according to embodiments of the present disclosure, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, embodiments of the present disclosure include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network through the communication device, or installed from the storage device 1003, or installed from the read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the methods of the embodiments of the present disclosure are performed.

[0151] The electronic device provided by the embodiments of the present disclosure adopts the molecular graph generation method in the above-mentioned embodiments, and can solve the technical problem of how to improve the rationality of the molecular structure generated by the model. Compared with the prior art, the electronic device provided by the present disclosure has the same beneficial effects as the molecular graph generation method provided by the above-mentioned embodiments, and other technical features in the electronic device are the same as the features disclosed in the previous embodiment method, which will not be repeated here.

[0152] It should be understood that parts of the present disclosure can be realized by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0153] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0154] The embodiment of the present application provides a computer readable storage medium having computer readable program instructions (i.e. computer programs) stored thereon, and the computer readable program instructions are used to execute the molecular graph generation method in the above embodiment.

[0155] The computer readable storage medium provided by the embodiment of the present application may, for example, be a U disk, but is not limited to an electric, magnetic, optical, electromagnetic, infrared, or semiconductor system or device, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to: an electric connection with one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM: Random Access Memory), a read-only memory (ROM: Read Only Memory), an erasable programmable read-only memory (EPROM: Erasable Programmable Read Only Memory or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM: CD-Read Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the embodiment, the computer readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system or device. The program code contained on the computer readable storage medium can be transmitted by any suitable medium, including but not limited to: an electric wire, an optical cable, an RF (Radio Frequency: Radio Frequency), etc., or any suitable combination of the above.

[0156] The above computer readable storage medium can be contained in an electronic device; or can exist separately without being assembled into an electronic device.

[0157] The computer readable storage medium described above carries one or more programs, when the one or more programs are executed by the electronic device, cause the electronic device to: predict an input noise graph through a graph generation model, determine a probability adjacency matrix of the noise graph, wherein the noise graph is obtained by noise injection on an original molecular graph; perform power operation on the probability adjacency matrix and an original adjacency matrix of the original molecular graph respectively to obtain a first high-order adjacency matrix of a plurality of noise graphs and a second high-order adjacency matrix of the original molecular graph; construct a target loss function according to each first high-order adjacency matrix and each second high-order adjacency matrix; update the graph generation model according to a back propagation algorithm and the target loss function, and perform the step of predicting the input noise graph through the graph generation model until the target loss function converges, to obtain a trained graph generation model.

[0158] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0159] The flow and block diagrams in the drawings show architectural, functional, and operational representations of possible implementations of systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flow and block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may

[0160] The modules described in the embodiments of the present application can be implemented in the form of software or in the form of hardware. In some cases, the name of the module does not constitute a limitation on the module itself.

[0161] The readable storage medium provided by the embodiments of the present application is a computer readable storage medium, which stores computer readable program instructions (i.e., a computer program) for executing the above-mentioned molecular graph generation method, and can solve the technical problem of how to improve the rationality of the molecular structure generated by the model. Compared with the prior art, the computer readable storage medium provided by the present application has the same beneficial effects as the molecular graph generation method provided by the above-mentioned embodiments, and will not be repeated here.

[0162] The embodiments of the present application also provide a computer program product, which includes a computer program. When the computer program is executed by a processor, the steps of the above-mentioned molecular graph generation method are implemented.

[0163] The computer program product provided by the embodiments of the present application can solve the technical problem of how to improve the rationality of the molecular structure generated by the model. Compared with the prior art, the computer program product provided by the present application has the same beneficial effects as the molecular graph generation method provided by the above-mentioned embodiments, and will not be repeated here.

[0164] The above only describes some embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation, direct / indirect application in other related technical fields based on the technical concept of the present application, and the content of the specification and drawings are included in the patent protection scope of the present application.

Claims

1. A method of generating a molecular graph, the method comprising: receiving a plurality of molecular fragments; and generating a molecular graph by combining the plurality of molecular fragments. The molecular graph generation method comprises: predicting the input noise graph by a graph generation model to determine a probability adjacency matrix of the noise graph, wherein the noise graph is obtained by noise injection on an original molecular graph; performing power operation on the probability adjacency matrix and an original adjacency matrix of the original molecular graph respectively to obtain a plurality of first high-order adjacency matrices of the noise graph and a second high-order adjacency matrix of the original molecular graph; constructing a target loss function according to each of the first high-order adjacency matrices and each of the second high-order adjacency matrices; updating the graph generation model according to a back propagation algorithm and the target loss function, and performing the step of predicting the input noise graph by the graph generation model until the target loss function converges to obtain a trained graph generation model.

2. The molecular graph generation method of claim 1, wherein, The step of predicting the input noise graph by the graph generation model to determine the probability adjacency matrix of the noise graph comprises: predicting the input noise graph by the graph generation model to determine a predicted edge feature corresponding to the noise graph; determining a connection probability between atoms of the noise graph according to a preset classification activation function and the predicted edge feature; determining the probability adjacency matrix according to the connection probability between the atoms of the noise graph.

3. The molecular graph generation method of claim 2, wherein, The target loss function comprises a local loss function and a topological loss function, and the step of predicting the input noise graph by the graph generation model to determine the predicted edge feature of the noise graph comprises: predicting the input noise graph by the graph generation model to determine a predicted node feature and a predicted edge feature of the noise graph; The step of determining the target loss value according to each of the first high-order adjacency matrices and each of the second high-order adjacency matrices comprises: determining a topological loss function according to each of the first high-order adjacency matrices and each of the second high-order adjacency matrices; determining a local loss function according to the predicted node feature, the predicted edge feature, an atomic feature of the original molecular graph and a chemical bond feature of the original molecular graph; constructing a target loss function according to the topological loss function and the local loss function.

4. The molecular graph generation method of claim 3, wherein, The step of determining the topological loss function according to each of the first high-order adjacency matrices and each of the second high-order adjacency matrices comprises: normalizing each of the first high-order adjacency matrices and each of the second high-order adjacency matrices; determining a difference between the first high-order adjacency matrices and the second high-order adjacency matrices with the same power according to a preset matrix norm to obtain a loss value corresponding to each power; determining the topological loss function according to the loss value corresponding to each power.

5. The molecular graph generation method of claim 1, wherein, Before the step of predicting the input noise graph by the graph generation model to determine the probability adjacency matrix of the noise graph, the method further comprises: obtaining a bond-free molecular graph corresponding to the original molecular graph, wherein the bond-free molecular graph and the original molecular graph have the same number of atoms, and the bond-free molecular graph has no chemical bond; generating a noise graph according to a preset discrete flow matching rule, the original molecular graph and the bond-free molecular graph.

6. The molecular graph generation method of claim 5, wherein, The discrete flow matching rule comprises a random time step, and the step of generating a noise graph according to the preset discrete flow matching rule, the original molecular graph and the bondless molecular graph comprises: According to the random time step, determining the replacement probability of each atom and each chemical bond in the original molecular graph; According to the replacement probability, the atoms and chemical bonds in the original molecular graph are replaced with the atoms and chemical bonds in the bondless molecular graph to obtain the noise graph.

7. The molecular graph generation method of claim 1, wherein, After the step of obtaining the trained graph generation model, further comprising: processing a to-be-constructed molecular graph through the trained graph generation model to output atom type prediction scores and chemical bond type prediction scores; According to the atom type prediction scores, determining the real atom type, and according to the chemical bond type prediction scores, determining the real chemical bond type; According to the real atom type and the real chemical bond type, generating a target molecular graph.

8. An electronic device, comprising: The device comprises a memory, a processor and a computer program stored on the memory and executable on the processor, and the computer program is configured to implement the steps of the molecular graph generation method according to any one of claims 1 to 7.

9. A storage medium, characterized by The storage medium is a computer readable storage medium, and the storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the molecular graph generation method according to any one of claims 1 to 7.

10. A computer program product, characterised in that, The computer program product comprises a computer program, and the computer program is executed by the processor to implement the steps of the molecular graph generation method according to any one of claims 1 to 7.

Images