A method for predicting the activity cliff of task-specific drug molecules

Abstract

The present invention provides a task-specific drug molecule activity cliff prediction method, which relates to the field of drug molecule prediction. The MMP is represented as a pair of complete drug molecules, or a pair of molecules split into a core and two substituents. When the MMP represents a pair of complete drug molecules, transfer learning is used to update the initialization weights of an activity cliff prediction model for a molecular framework to predict the activity cliff of the drug molecule pair. When the MMP represents a molecule split into a core and two substituents, the direct relationship between the core and the substituents is learned based on a hyperconnected graph structure to predict the activity cliff of the drug molecule pair. The technical solution of the present invention overcomes the problem that the existing technology cannot consider the properties of MMPs in different scenarios to predict the activity cliff of drug molecules.

Description

Technical Field

[0001] The present invention relates to the field of drug molecule prediction, and in particular to a task-specific drug molecule activity cliff prediction method. Background Art

[0002] By analyzing the activity cliff of drugs, researchers can discover the sensitivity of drug molecules to changes in activity under minor structural changes, thereby avoiding further development of ineffective molecules and saving a lot of experimental resources and time. The identification of activity cliffs can help guide the optimization of drug structures, especially in the study of the "Structure-Activity Relationship" (SAR) of drugs, revealing which structural features are critical to biological activity. Secondly, through a deep understanding of the activity cliff, researchers can identify key structural features that affect drug activity and provide guidance for subsequent drug optimization. The prediction of drug activity cliffs can also reveal the complex interaction mechanism between drug molecules and biological targets, providing theoretical support for precision drug design.

[0003] Existing activity prediction techniques primarily rely on descriptors abstracted from molecular structure, which may not fully capture the complex interactions and fine spatial relationships between molecules. The limitations of this approach have given rise to structure-based prediction methods, particularly when understanding and predicting significant activity differences caused by minor structural changes. Heikamp et al. proposed the key concept that activity cliffs should be developed for pairs of molecules rather than for individual molecule classification, and successfully predicted activity cliffs using support vector machines. Park et al. proposed a method using graph convolutional networks to predict activity cliffs. This method directly utilizes the structural information of molecular graphs and, by learning the characteristics of nodes (atoms) and edges (chemical bonds), can reveal in detail the structural changes that lead to activity differences. Iqbal et al. used convolutional neural networks to predict activity cliffs from molecular images. This method differs from traditional graph-based methods in that it analyzes two-dimensional molecular images to identify pairs of molecules with similar structures but significantly different biological activities. Deep learning techniques are then used to distinguish activity cliffs, further demonstrating the potential of structure-based methods for processing complex molecular data.

[0004] However, existing methods for activity cliff prediction often treat matched molecular pairs (MMPs) in a simplistic manner, failing to thoroughly analyze the application of different MMP types in prediction tasks. Furthermore, data scarcity poses a challenge to model training. Finally, while the core and substituents together constitute a drug, how these two structural components jointly influence the overall properties of a molecule remains unclear.

[0005] Therefore, there is a need for a task-specific drug molecule activity cliff prediction method that considers the properties of MMPs in different scenarios. Summary of the Invention

[0006] The main purpose of the present invention is to provide a task-specific drug molecule activity cliff prediction method to solve the problem in the prior art that the drug molecule activity cliff prediction cannot be performed considering the MMP properties in different scenarios.

[0007] To achieve the above objectives, the present invention provides a task-specific drug molecule activity cliff prediction method, where MMP is represented as a pair of complete drug molecules, or a pair of molecules split into one core and two substituents. When MMP represents a pair of complete drug molecules, transfer learning is used to update the initialization weights of the activity cliff prediction model for the molecular framework to predict the activity cliff of drug molecules. When MMP represents a molecule split into one core and two substituents, the direct relationship between the core and the substituents is learned based on the hyperconnected graph structure to predict the activity cliff of drug molecules.

[0008] Furthermore, when MMP represents a pair of complete drug molecules, transfer learning is used to update the parameters of the activity cliff prediction model for the molecular framework to perform activity cliff prediction, which specifically includes the following steps:

[0009] S1, taking the first drug and the second drug as input, pre-trains the activity cliff prediction model for the molecular framework to obtain the initialization weights.

[0010] S2 uses the initialization weights obtained from pre-training for transfer learning, updates the activity cliff prediction model for the molecular framework, and uses the activity cliff prediction model for the molecular framework to predict the activity cliff of drug molecule pairs.

[0011] Furthermore, the step S2 of using the molecular framework-oriented activity cliff prediction model to predict the activity cliff of the drug molecule pair specifically includes the following steps:

[0012] S2.1, taking the molecule pair as input, uses the graph convolutional neural network to update the features of the molecule to obtain the encoded molecular features, and performs pooling operation on the encoded molecular features to obtain the graph-level representation of the molecule.

[0013] S2.2, design a Transformer-based molecular feature enhancer to learn the relationship between molecular pairs and predict activity cliffs.

[0014] Furthermore, the molecule pair includes a first molecule and the second molecule , the feature update process is the same; the first molecule in the molecule pair Taking feature update as an example, step S2.1 specifically includes the following steps:

[0015] S2.1.1, use the RDKit tool to Convert to molecular graph ,The nodes in the molecular graph represent the atoms in the molecule, and the edges represent the chemical bonds between atoms.

[0016] S2.1.2, use a graph convolutional neural network to perform convolution operations on each atom and its neighboring atoms, update the atomic features and obtain the initial representation of the molecule:

[0017] (1);

[0018] in, is a node go through Features after layer convolution, is a node The set of neighbor nodes of and Represents nodes and nodes The degree, is a trainable weight matrix.

[0019] S2.1.3, use the pooling operation to aggregate all nodes of the molecule into a one-dimensional vector as the graph-level representation of the molecule:

[0020] (2);

[0021] in, Indicates that the molecule after pooling compression The feature representation of represents pooling, Represents molecules The number of nodes in the.

[0022] Furthermore, step S2.2 specifically includes the following steps:

[0023] S2.2.1, learning features between molecule pairs through attention mechanism:

[0024] (3);

[0025] (4);

[0026] (5);

[0027] in, , , and is the trainable weight matrix, represents a vector stacking operation, 、 and represent query vector, key vector and value vector respectively, represents the normalization operation, Indicates multi-head splicing operation, Indicates the number of heads, Represents molecular pairs and The final embedding.

[0028] S2.2.2, predict the activity cliff between drug molecule pairs using a multi-layer perceptron:

[0029] (6);

[0030] in, is a multi-layer perceptron, Represents the prediction probability score of the activity cliff prediction model for the molecular framework.

[0031] S2.2.3, using binary cross entropy loss function Training the model:

[0032] (7);

[0033] in, represents the total number of training samples, represents the true label of the training sample, indicates that an activity cliff is formed between the molecular pairs, which does not exist otherwise.

[0034] Furthermore, when the MMP represents a molecule that is split into a core and two substituents, the activity cliff prediction is performed by learning the direct relationship between the core and the substituents based on the hyperconnected graph structure. The specific steps include:

[0035] S3, using the graph attention network GAT to update the features of the core and substituents.

[0036] In S4, the core and substituent features are input into the hyperconnected graph structure, and pooling and splicing operations are performed in sequence. Finally, the activity cliff between drug molecule pairs is predicted through a multi-layer perceptron.

[0037] S5, using the binary cross entropy loss function to obtain the prediction loss , and use L2 normalization to quantify the constraint loss of the difference between molecular features and activity features. The prediction loss and the constraint loss are combined to obtain the final loss.

[0038] Furthermore, taking the feature update of the core and the first substituent as an example, step S3 specifically includes the following steps:

[0039] S3.1, the core of each MMP is represented as , the first substituent is represented by ; The graph structures of the core and the first substituent are represented as and ; Indicates Middle The initial embedding of nodes, Indicates Middle The initial embedding of the nodes.

[0040] S3.2, Update core and substituent features using GAT:

[0041] (8);

[0042] (9);

[0043] in, is the softmax activation function, is a learnable operator, and Respectively represent the core The node and the first substituent Nodes Features after layer GAT processing, is the activation function, and They represent the core of the hyperconnected graph. The node and the first substituent The associated set of nodes, Representation node go through Features after layer GAT processing.

[0044] Furthermore, step S4 specifically includes the following steps:

[0045] S4.1, using pooling operation to and All nodes in are compressed into one-dimensional vectors, and the final representation of the molecule consisting of the core and the first substituent is obtained using the concatenation operation:

[0046] (10);

[0047] in, Indicated by and Representation of the constituent molecules, Represents the core The number of nodes in Represents a substituent The number of nodes in Represents a splicing operation.

[0048] S4.2, core and the second substituent The molecular structure is represented by , use the splicing operation to and Merge the final embedding of the two molecules and use MLP to obtain the prediction results :

[0049] (11).

[0050] Furthermore, the expression of the constraint loss for quantifying the difference between the molecular features and the activity features using L2 normalization in step S5 is:

[0051] (12);

[0052] (13);

[0053] in, To use L2 normalization to quantize the molecular features, is the feature of molecular activity after quantification using L2 normalization, is the matrix transpose, is the activity information of the molecule consisting of the core and the first substituent, Constraint loss for the molecule consisting of the core and the first substituent, is the Frobenius norm, Constraint loss for the molecule consisting of the core and the second substituent , Calculation method and same.

[0054] Furthermore, in step S5, the classification loss and the constraint loss of different weights are combined to obtain the final loss The expression is:

[0055] (14);

[0056] in, is a hyperparameter of the balance loss.

[0057] The present invention has the following beneficial effects:

[0058] 1) The present invention sets two modeling approaches for different active cliff scenarios according to their data properties, rationally utilizes data from different scenarios, and improves prediction accuracy.

[0059] 2) This paper uses transfer learning to fully utilize knowledge from other related tasks to reduce the impact of data scarcity and improve prediction accuracy.

[0060] 3) The present invention designs a hyperconnected graph structure to establish a fully connected relationship between the core and each substituent, thereby enhancing information integration and adaptability to changes in the internal structure of the drug. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0062] Figure 1 The flowchart of the present invention for predicting activity cliff when MMP represents a pair of complete drug molecules is shown.

[0063] Figure 2 The flowchart of the present invention for predicting activity cliff when MMP represents a molecule separated into one core and two substituents is shown.

[0064] Figure 3 The drug molecular structure of the first group of MMPs is shown, in which no activity cliff is formed when the MMP represents a pair of complete drug molecules (MMP-nonCliff).

[0065] Figure 4 The drug molecular structures of the second group of MMPs are shown, in which MMPs do not form an activity cliff (MMP-nonCliff) when a pair of complete drug molecules is represented.

[0066] Figure 5 The molecular structure of the first group of MMP drugs showing that an MMP forms an activity cliff (MMP-Cliff) when a pair of intact drug molecules is represented.

[0067] Figure 6 The drug molecular structure of the second group of MMPs is shown, in which an MMP represents an MMP activity cliff (MMP-Cliff) when paired with a pair of intact drug molecules.

[0068] Figure 7The drug molecular structure of the first group of MMPs is shown, which represents an MMP that does not form an activity cliff (MMP-nonCliff) when the MMP is decomposed into a core and two substituent molecules.

[0069] Figure 8 The drug molecular structure of the second group of MMPs is shown, which represents an MMP that does not form an activity cliff (MMP-nonCliff) when the MMP is decomposed into a core and two substituent molecules.

[0070] Figure 9 The drug molecular structure of the first group of MMPs is shown, which represents the MMP-Cliff that forms an activity cliff when the MMP molecule is split into a core and two substituents.

[0071] Figure 10 The drug molecular structure of the second group of MMPs is shown, which represents the MMP-cliff that forms an activity cliff when the MMP molecule is split into a core and two substituents (MMP-Cliff). DETAILED DESCRIPTION

[0072] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0073] A task-specific drug molecule activity cliff prediction method is proposed. The MMP is represented as a pair of complete drug molecules or a pair of molecules split into a core and two substituents. When the MMP represents a pair of complete drug molecules, transfer learning is used to update the initialization weights of the activity cliff prediction model for the molecular framework to perform activity cliff prediction. When the MMP represents a molecule split into a core and two substituents, the direct relationship between the core and the substituents is learned based on the hyperconnected graph structure to perform activity cliff prediction.

[0074] Most of the currently available activity cliff data are collected from the ChEMBL database, and the generation of MMP is achieved by exchanging a pair of unique substructures (such as terminal groups or central fragments). The algorithm proposed by Hussain and Rea is often used for this task, which focuses on the fragmentation and indexing of molecules in MMP analysis. Taking the single-cut example (i.e., MMP with only a single terminal fragment different) as an example, the first stage of the algorithm fragments all molecules in the input data set, lists all possible single-cut situations at the annotated bonds in the molecules, and then indexes the generated fragments. Therefore, MMP generally contains two expressions, divided into a pair of complete drug molecules and a pair of molecules split into 1 core and 2 substituents, is the activity information of the molecule consisting of the core and the first substituent, It is the activity information of the molecule composed of the core and the second substituent. When the activity difference of a pair of molecules is at least 100 times or 2 orders of magnitude (i.e. ), then the pair of molecules is considered to have an activity cliff, when the difference is 10 times or 1 order of magnitude (i.e. ), then there is no activity cliff. MMPs between 1 and 2 were eliminated.

[0075] Deep learning methods typically require sufficient data to perform well, but MMP molecular pairs are difficult to process and require high-quality labeled data, so models face the challenge of data scarcity. To address this, the present invention uses a large amount of molecular and bioactivity-rich drug data from DrugBank, and pre-trains the model using relatively simple drug interaction tasks from the labeled data. Transfer learning is then used to update the model parameters. This significantly shortens training time and improves the ability to identify activity differences between new molecular pairs, thereby achieving highly accurate and efficient activity cliff predictions.

[0076] Specifically, when MMP represents a pair of complete drug molecules, transfer learning is used to update the parameters of the activity cliff prediction model for the molecular framework to perform activity cliff prediction, such as Figure 1 As shown, the specific steps include:

[0077] S1, taking the first drug and the second drug as input, constructs a drug interaction prediction task, pre-trains the activity cliff prediction model for the molecular framework, and obtains the initialization weights.

[0078] The pre-training stage of step S1 is as follows: the drugs are input, the drug features are encoded using a graph convolutional neural network, and the features of the two drugs are pooled to obtain two drug graph-level representations. After splicing the two drug-level representations, the detailed communication between the molecular pairs is learned through a molecular feature enhancer, and then drug interaction DDI prediction is performed.

[0079] S2 uses the initialization weights obtained from pre-training for transfer learning, updates the activity cliff prediction model for the molecular framework, and uses the activity cliff prediction model for the molecular framework to predict the activity cliff of drug molecule pairs.

[0080] Specifically, the step S2 of using the molecular framework-oriented activity cliff prediction model to predict the activity cliff of the drug molecule pair specifically includes the following steps:

[0081] S2.1, taking the molecule pair as input, uses the graph convolutional neural network to update the features of the molecule to obtain the encoded molecular features, and performs pooling operation on the encoded molecular features to obtain the graph-level representation of the molecule.

[0082] S2.2, design a Transformer-based molecular feature enhancer to learn the relationship between molecular pairs and predict activity cliffs.

[0083] Specifically, the molecule pair includes a first molecule and the second molecule , the feature update process is the same; the first molecule in the molecule pair Taking feature update as an example, step S2.1 specifically includes the following steps:

[0084] S2.1.1, use the RDKit tool to Convert to molecular graph ,The nodes in the molecular graph represent the atoms in the molecule, and the edges represent the chemical bonds between atoms.

[0085] S2.1.2, use a graph convolutional neural network to perform convolution operations on each atom and its neighboring atoms, update the atomic features and obtain the initial representation of the molecule:

[0086] (1);

[0087] in, is a node go through Features after layer convolution, is a node The set of neighbor nodes of and Represents nodes and nodes The degree, is a trainable weight matrix.

[0088] S2.1.3, use the pooling operation to aggregate all nodes of the molecule into a one-dimensional vector as the graph-level representation of the molecule:

[0089] (2);

[0090] in, Indicates that the molecule after pooling compression The feature representation of represents pooling, Represents molecules The number of nodes in the.

[0091] Although graph convolutional neural networks (GCNs) have demonstrated excellent performance in capturing molecular structures, when dealing with complex molecular pair interactions, a single GCN may not fully capture the subtle differences between molecules, which play a crucial role in the model's interpretation of the communication between molecular pairs. To address this issue, we designed a Transformer-based molecular feature enhancer that aims to learn the detailed communication between molecular pairs through a self-attention mechanism.

[0092] Specifically, step S2.2 includes the following steps:

[0093] S2.2.1, learning features between molecule pairs through attention mechanism:

[0094] (3);

[0095] (4);

[0096] (5);

[0097] in, , , and is the trainable weight matrix, represents a vector stacking operation, 、 and represent query vector, key vector and value vector respectively, represents the normalization operation, Indicates multi-head splicing operation, Indicates the number of heads, Represents molecular pairs and The final embedding.

[0098] S2.2.2, predict the activity cliff between drug molecule pairs using a multi-layer perceptron:

[0099] (6);

[0100] in, is a multi-layer perceptron, Represents the prediction probability score of the activity cliff prediction model for the molecular framework.

[0101] S2.2.3, treat the activity cliff prediction task as a binary classification task and use the binary cross entropy loss function Training the model:

[0102] (7);

[0103] in, represents the total number of training samples, represents the true label of the training sample, indicates that an activity cliff is formed between the molecular pairs, which does not exist otherwise.

[0104] Specifically, when the MMP represents a molecule that is split into one core and two substituents, the direct relationship between the core and the substituents is learned based on the hyperconnected graph structure to predict the activity cliff. In this scenario, simply directly analyzing the split structure often ignores the key interaction between the core and the substituents, which is crucial for the accurate prediction of the activity cliff. This is especially true when the activity differences caused by the substituents on the molecule are very subtle, because traditional models find it difficult to accurately capture how these changes affect biological activity, resulting in inaccurate predictions. To overcome this limitation, the present invention proposes a hyperconnected graph structure for this scenario, such as Figure 2 As shown, this approach aims to improve the ability to identify significant differences in activity between the core and different substituents. In the hyperconnected graph, atomic-level connections are established between the core and the two substituents, enabling the model to calculate attention scores for all nodes between the core and the substituents and use them to update node features. This provides the model with a deeper understanding of the dynamic relationship between the core and substituents, allowing it to more accurately identify the impact of these subtle changes on molecular activity.

[0105] like Figure 2 As shown, the specific steps include:

[0106] S3, using the graph attention network GAT to update the features of the core and substituents.

[0107] In S4, the core and substituent features are input into the hyperconnected graph structure, and pooling and splicing operations are performed in sequence. Finally, the activity cliff between drug molecule pairs is predicted through a multi-layer perceptron.

[0108] S5, using the binary cross entropy loss function to obtain the prediction loss And use L2 normalization to quantify the constraint loss of the difference between molecular features and activity features. The prediction loss and constraint loss are combined to obtain the final loss.

[0109] Specifically, the core and the first substituent (i.e. Figure 2 Taking the substituent 1) in the figure as an example, step S3 specifically includes the following steps:

[0110] S3.1, the core of each MMP is represented as , the first substituent is represented by ; The graph structures of the core and the first substituent are represented as and ; Indicates Middle The initial embedding of nodes, Indicates Middle The initial embedding of the nodes.

[0111] S3.2, Update core and substituent features using GAT:

[0112] (8);

[0113] (9);

[0114] in, is the softmax activation function, is a learnable operator, and Respectively represent the core The node and the first substituent Nodes Features after layer GAT processing, is the activation function, and They represent the core of the hyperconnected graph. The node and the first substituent The associated set of nodes, Representation node go through Features after layer GAT processing.

[0115] The core and the second substituent (i.e. Figure 2 The process of updating the characteristics of the substituents in 2) is the same as the above process.

[0116] Specifically, step S4 includes the following steps:

[0117] S4.1, using pooling operation to and All nodes in are compressed into one-dimensional vectors, and the final representation of the molecule consisting of the core and the first substituent is obtained using the concatenation operation:

[0118] (10);

[0119] in, Indicated by and Representation of the constituent molecules, Represents the core The number of nodes in Represents a substituent The number of nodes in Represents a splicing operation.

[0120] S4.2, core and the second substituent The molecular structure is represented by , use the splicing operation to and Merge the final embedding of the two molecules and use MLP to obtain the prediction results :

[0121] (11).

[0122] To further improve the model's predictive performance, the present invention incorporates information about the activity of both molecules. These values are mapped into a high-dimensional space using a linear layer. To effectively utilize these embeddings, a constrained loss is introduced that uses L2 normalization to quantify the difference between molecular and activity features, thereby improving the ability to identify differences in molecular activity.

[0123] Specifically, the expression of the constraint loss for quantifying the difference between the molecular features and the activity features using L2 normalization in step S5 is:

[0124] (12);

[0125] (13);

[0126] in, To use L2 normalization to quantize the molecular features, is the feature of molecular activity after quantification using L2 normalization, is the matrix transpose, and the embedding is L2 normalized so that the L2 norm of each embedding is equal to 1. is the activity information of the molecule consisting of the core and the first substituent, Constraint loss for the molecule consisting of the core and the first substituent, is the Frobenius norm, and the Frobenius norm difference is calculated as the loss. Constraint loss for the molecule consisting of the core and the second substituent , Calculation method and same.

[0127] Specifically, in step S5, the classification loss and the constraint loss of different weights are combined to obtain the final loss The expression is:

[0128] (14);

[0129] in, is a hyperparameter of the balance loss.

[0130] As shown in Table 1, the proposed model (TS-AC) achieves optimal performance in two tasks on the Thrombin dataset compared to state-of-the-art methods in five of six evaluation metrics, validating its potential for application in different activity cliff scenarios. CC represents the scenario for intact molecular pairs, and CS represents the scenario for molecules decomposed into core and substituent groups. TPR measures the model's ability to correctly identify positive samples; TNR measures the model's ability to correctly identify negative samples; BA, the average of TPR and TNR, measures the model's ability to address class imbalance; F1 considers the model's precision and recall, being the harmonic average of the two; MCC is an evaluation metric that comprehensively considers true positives, false positives, true negatives, and false negatives; and AUC measures the model's performance at all possible classification thresholds.

[0131] Table 1 Performance comparison table

[0132]

[0133] like Figure 3-Figure 6 Figure 2 shows a visualization of the proposed model in a CC scenario. Two groups of MMPs without activity cliffs were selected. In the first group of MMPs without activity cliffs (MMP-nonCliff), the labeled nodes for both drugs were identical, and none of these nodes fell on substituents responsible for structural changes. In the first group of MMPs with activity cliffs (MMP-Cliff), drug ID 11471 exhibited higher activity, and the substituents responsible for the structural changes were labeled. In contrast, drug ID 11464 lacked nodes related to the substituents, indicating that the proposed model identifies substituent changes in this group as key to the activity cliff. This finding highlights the model's sensitivity and predictive power for detecting activity changes caused by key structural differences.

[0134] like Figure 7-10As shown, in the first group of MMPs in MMP-nonCliff, it can be observed that the activity values of the two drugs are relatively close. The model of the present invention concentrates the labeled nodes on the core without marking any substituents, which indicates that changes in substituents have little effect on the activity of the drug. Similarly, in the second group of MMPs, all important labels are also concentrated on the core. This consistent result further verifies that in this case, changes in substituents do not significantly affect the activity of the drug, reflecting that the model can identify the impact of changes in substituents on drug activity. In the two groups of MMPs in MMP-Cliff, the model successfully marked the changed substituents as key nodes. These nodes are considered to be the driving factors of activity changes. It is worth noting that in both groups, it can be clearly observed that the model of the present invention assigns higher attention scores to substituents that cause high activity.

[0135] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A task-specific drug molecule activity cliff prediction method, characterized in that: MMPs are represented as a pair of complete drug molecules or a pair of molecules split into 1 core and 2 substituents; When MMP represents a pair of complete drug molecules, transfer learning is used to update the initial weights of the activity cliff prediction model for the molecular framework to predict the activity cliff of drug molecules. When MMP represents a molecule that is split into a core and two substituents, the direct relationship between the core and the substituents is learned based on the hyperconnected graph structure to predict the activity cliff of drug molecules. When MMP represents a pair of complete drug molecules, transfer learning is used to update the parameters of the activity cliff prediction model for the molecular framework to perform activity cliff prediction. The specific steps include the following: S1, taking the first drug and the second drug as input, pre-training the activity cliff prediction model for the molecular framework to obtain the initial weights; S2, using the initialization weights obtained from pre-training to perform transfer learning, update the activity cliff prediction model for the molecular framework, and use the activity cliff prediction model for the molecular framework to predict the activity cliff of drug molecule pairs; The process of predicting the activity cliff of drug molecules using the molecular framework-oriented activity cliff prediction model in step S2 specifically includes the following steps: S2.1, taking the molecule pair as input, uses a graph convolutional neural network to update the features of the molecule to obtain the encoded molecular features, and performs a pooling operation on the encoded molecular features to obtain a graph-level representation of the molecule; S2.2, design a Transformer-based molecular feature enhancer to learn the relationship between molecular pairs and predict activity cliffs; When the MMP represents a molecule that is split into a core and two substituents, the activity cliff prediction is performed by learning the direct relationship between the core and the substituents based on the hyperconnected graph structure. The specific steps include: S3, using the graph attention network GAT to update the features of the core and substituents; S4, the core and substituent features are input into the hyperconnected graph structure, and pooling and concatenation operations are performed in sequence. Finally, the activity cliff between drug molecule pairs is predicted by a multi-layer perceptron. S5, using the binary cross entropy loss function to obtain the prediction loss , and use L2 normalization to quantify the constraint loss of the difference between molecular features and activity features. The prediction loss and the constraint loss are combined to obtain the final loss.

2. A task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: The molecule pair includes the first molecule and the second molecule , the feature update process is the same; the first molecule in the molecule pair Taking feature update as an example, step S2.1 specifically includes the following steps: S2.1.1, through the RDKit tool Convert to molecular graph ,The nodes in the molecular graph represent the atoms in the molecule, and the edges represent the chemical bonds between atoms; S2.1.2, use a graph convolutional neural network to perform convolution operations on each atom and its neighboring atoms, update the atomic features and obtain the initial representation of the molecule: （1）； in, is a node go through Features after layer convolution, is a node The set of neighbor nodes of and Represents nodes and nodes The degree, is a trainable weight matrix; S2.1.3, use the pooling operation to aggregate all nodes of the molecule into a one-dimensional vector as the graph-level representation of the molecule: （2）； in, Indicates that the molecule after pooling compression The feature representation of represents pooling, Represents molecules The number of nodes in the.

3. The task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: Step S2.2 specifically includes the following steps: S2.2.1, learning features between molecule pairs through attention mechanism: （3）； （4）； （5）； in, , , and is the trainable weight matrix, represents a vector stacking operation, 、 and represent query vector, key vector and value vector respectively, represents the normalization operation, Indicates multi-head splicing operation, Indicates the number of heads, Represents molecular pairs and The final embedding of S2.2.2, predict the activity cliff between drug molecule pairs using a multi-layer perceptron: （6）； in, is a multi-layer perceptron, represents the prediction probability score of the activity cliff prediction model of the molecular framework; S2.2.3, using binary cross entropy loss function Training the model: （7）； in, represents the total number of training samples, represents the true label of the training sample, indicates that an activity cliff is formed between the molecular pairs, which does not exist otherwise.

4. The task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: Taking the feature update of the core and the first substituent as an example, step S3 specifically includes the following steps: S3.1, the core of each MMP is represented as , the first substituent is represented by ; The graph structures of the core and the first substituent are represented as and ; Indicates Middle The initial embedding of nodes, Indicates Middle The initial embedding of nodes; S3.2, Update core and substituent features using GAT: （8）； （9）； in, is the softmax activation function, is a learnable operator, and Respectively represent the core The node and the first substituent Nodes Features after layer GAT processing, is the activation function, and They represent the core of the hyperconnected graph. The node and the first substituent The associated set of nodes, Representation node go through Features after layer GAT processing.

5. The task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: Step S4 specifically includes the following steps: S4.1, using pooling operation to and All nodes in are compressed into one-dimensional vectors, and the final representation of the molecule consisting of the core and the first substituent is obtained using the concatenation operation: （10）； in, Indicated by and Representation of the constituent molecules, Represents the core The number of nodes in Represents a substituent The number of nodes in Represents a splicing operation; S4.2, core and the second substituent The molecular structure is represented by , use the splicing operation to and Merge the final embedding of the two molecules and use MLP to obtain the prediction results : （11）。 6. The task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: The expression of the constraint loss for quantifying the difference between molecular features and activity features using L2 normalization in step S5 is: （12）； （13）； in, To use L2 normalization to quantize the molecular features, is the feature of molecular activity after quantification using L2 normalization, is the matrix transpose, is the activity information of the molecule consisting of the core and the first substituent, Constraint loss for the molecule consisting of the core and the first substituent, is the Frobenius norm, Constraint loss for the molecule consisting of the core and the second substituent , Calculation method and same.

7. The task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: In step S5, the classification loss and the constraint loss of different weights are combined to obtain the final loss The expression is: （14）； in, is a hyperparameter of the balance loss.

Images