Task-specific drug molecule activity cliff prediction method

Abstract

The invention provides a cliff prediction method for task-specific drug molecule activity, and relates to the field of drug molecule prediction, MMP represents a pair of complete drug molecules or a pair of molecules split into one core and two substituent groups; when the MMP represents a pair of complete drug molecules, updating an initialized weight of the molecular framework-oriented active cliff prediction model by using transfer learning, and performing active cliff prediction on the drug molecules; when MMP represents a molecule split into one core and two substituent groups, a direct relationship between the core and the substituent groups is learned based on a hyperliplex diagram structure, and activity cliff prediction of drug molecules is carried out. According to the technical scheme, the problem that drug molecule activity cliff prediction cannot be carried out by considering MMP properties in different scenes in the prior art is solved.

Description

Technical Field

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

[0002] Through the analysis of drug activity cliffs, researchers can discover the sensitivity of drug molecule activity changes under minor structural changes, thereby avoiding the further development of ineffective molecules and saving a large amount 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 crucial for biological activity. Secondly, through in-depth understanding of activity cliffs, researchers can identify the key structural features affecting drug activity and provide guidance for subsequent drug optimization. The prediction of drug activity cliffs can also reveal the complex interaction mechanisms between drug molecules and biological targets, providing theoretical support for precision drug design.

[0003] The existing activity predictions mainly rely on descriptors abstracted from molecular structures, which may not fully capture the complex interactions and fine spatial relationships between molecules. The limitations of this method have given rise to structure-based prediction methods, especially when it is necessary to understand and predict significant activity differences caused by minor structural changes. Heikamp et al. proposed a key concept that activity cliffs should be unfolded for a pair of molecular pairs rather than classifying individual molecules, and successfully predicted activity cliffs using support vector machines. Park et al. proposed a method using graph convolutional networks to predict activity cliffs, directly utilizing the structural information of molecular graphs. By learning the features of nodes (atoms) and edges (chemical bonds), it can detail the structural changes leading to activity differences. Iqbal et al. used convolutional neural networks to predict activity cliffs from molecular images. This method is different from traditional graph-based methods. It identifies pairs of molecules with similar structures but very different biological activities by analyzing the two-dimensional images of molecules and uses deep learning techniques to distinguish activity cliffs, further demonstrating the potential of structure-based methods in processing complex molecular data.

[0004] However, in the research of activity cliff prediction using existing methods, the handling of Matched Molecular Pairs (MMPs) is too single, and the application of different types of MMPs in the prediction task has not been deeply analyzed. In addition, the scarcity of data is also a challenge for model training. Finally, although the core and substituents together constitute a drug, it is not clear how these two structural components jointly affect the overall properties of a molecule.

[0005] Therefore, there is a need for a task-specific method for predicting the activity cliffs of drug molecules that takes into account the MMP properties in different scenarios. Summary of the Invention

[0006] The main object of the present invention is to provide a task-specific method for predicting the activity cliffs of drug molecules, so as to solve the problem that the prior art cannot consider the MMP properties in different scenarios for predicting the activity cliffs of drug molecules.

[0007] To achieve the above object, the present invention provides a task-specific method for predicting the activity cliffs of drug molecules. MMP is 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, and the activity cliff prediction of the drug molecule pair is performed. When MMP represents a molecule split into 1 core and 2 substituents, based on the hyper-connected graph structure, the direct relationship between the core and the substituents is learned, and the activity cliff prediction of the drug molecule pair is performed.

[0008] Further, 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, and the activity cliff prediction is performed. The specific steps are as follows: S1. Take the first drug and the second drug as inputs, and pre-train the activity cliff prediction model for the molecular framework to obtain the initial weights.

[0009] S2. Use the initial weights obtained by pre-training for transfer learning to update the activity cliff prediction model for the molecular framework, and use the activity cliff prediction model for the molecular framework to perform the activity cliff prediction on the drug molecule pair.

[0010] Further, the specific steps of using the activity cliff prediction model for the molecular framework to perform the activity cliff prediction on the drug molecule pair in step S2 are as follows: S2.1. Take the molecule pair as an input, use the graph convolutional neural network to update the features of the molecule to obtain the encoded molecular features, and perform pooling operations on the encoded molecular features to obtain the graph-level representation of the molecule.

[0011] S2.2. Design a molecular feature enhancer based on Transformer to learn the relationship between molecule pairs and predict the activity cliffs.

[0012] Further, the molecule pair includes a first molecule and a second molecule , and the feature update processes are the same. Taking the feature update of the first molecule in the molecule pair as an example, step S2.1 specifically includes the following steps: S2.1.1, convert to a molecular graph through the RDKit tool where the nodes in the molecular graph represent atoms in the molecule, and the edges represent chemical bonds between atoms.

[0013] S2.1.2, use a graph convolutional neural network to perform convolutional operations on each atom and its adjacent atoms, update the features of the atoms, and obtain the initial representation of the molecule: (1); where is the feature of node after layers of convolution, is the set of neighbor nodes of node , and represent the degrees of node and node respectively, is a trainable weight matrix.

[0014] S2.1.3, use pooling operations to aggregate all nodes of the molecule into a one-dimensional vector as the graph-level representation of the molecule: (2); where represents the feature representation of molecule after pooling compression, represents pooling, represents molecule and

[0015] Furthermore, step S2.2 specifically includes the following steps: S2.2.1, learn the features between molecule pairs through an attention mechanism: (3); (4); (5); where , , and are trainable weight matrices, represents vector stacking operation, , and represent query vector, key vector and value vector respectively, represents normalization operation, represents multi-head concatenation operation, represents the number of heads. Represents a molecular pair and of the final embedding.

[0016] S2.2.2, predicting the activity cliff activity between drug molecule pairs through a multi-layer perceptron: (6); Among them, is a multi-layer perceptron, represents the predicted probability score of the activity cliff prediction model of the molecular framework.

[0017] S2.2.3, using the binary cross-entropy loss function to train the model: (7); Among them, represents the total number of training samples, represents the true label of the training sample, represents the formation of an activity cliff between molecular pairs, otherwise it does not exist.

[0018] Furthermore, when MMP represents a molecule split into 1 core and 2 substituents, based on the hyper-connected graph structure, learn the direct relationship between the core and the substituents, and perform activity cliff prediction, which specifically includes the following steps:

[0019] S3, using the graph attention network GAT to update the features of the core and substituents.

[0020] S4, input the core and substituent features into the hyper-connected graph structure, perform pooling operations and concatenation operations in sequence, and finally predict the activity cliff between drug molecule pairs through a multi-layer perceptron.

[0021] S5, using the binary cross-entropy loss function to obtain the prediction loss , and using L2 normalization to quantify the constraint loss of the difference between the molecular feature and the activity feature, and merge the prediction loss and the constraint loss to obtain the final loss.

[0022] Furthermore, taking the feature update of the core and the first substituent as an example, step S3 specifically includes the following steps: S3.1, represent the core of each MMP as , and the first substituent as ; the graph structures of the core and the first substituent are respectively represented as and ; represents the initial embedding of the th node in , represents in The initial embedding of the th node.

[0023] S3.2. Update the features of the core and substituents using GAT: (8); (9); Where is the softmax activation function, is a learnable operator, and respectively represent the features of the th node in the core and the th node in the first substituent after being processed by layers of GAT, is the activation function, and respectively represent the associated sets of the th node of the core and the th node of the first substituent in the hyper-connected graph, represents the feature of node after being processed by layers of GAT.

[0024] Furthermore, step S4 specifically includes the following steps: S4.1. Use pooling operation to compress all nodes in and into one-dimensional vectors, and use concatenation operation to obtain the final representation of the molecule composed of the core and the first substituent: (10); Where, represents the representation of the molecule composed of and , represents the number of nodes in the core , represents the number of nodes in the substituent , represents the concatenation operation.

[0025] S4.2. Denote the molecular representation composed of the core and the second substituent as . Use concatenation operation to combine and into the final embedding of the two molecules, and use MLP to obtain the prediction result : (11).

[0026] Further, in step S5, the expression of the constraint loss for quantifying the difference between the molecular feature and the activity feature using L2 normalization is as follows: (12); (13); where is the feature of the molecule after quantification using L2 normalization, is the feature of the molecular activity after quantification using L2 normalization, is the matrix transpose, is the activity information of the molecule composed of the core and the first substituent, is the constraint loss of the molecule composed of the core and the first substituent, is the Frobenius norm, is the constraint loss of the molecule composed of the core and the second substituent , The calculation method is the same as that of .

[0027] Further, in step S5, the classification loss and the constraint losses with different weights are combined to obtain the final loss The expression of is as follows: (14); where is the hyperparameter for balancing the loss.

[0028] The present invention has the following beneficial effects: 1) The present invention sets two modeling ideas according to the data nature for different activity cliff scenarios, reasonably utilizes the data under different scenarios, and improves the prediction accuracy.

[0029] 2) The present invention uses transfer learning to fully utilize the knowledge in other related tasks to reduce the impact of data scarcity and improve the prediction accuracy.

[0030] 3) The present invention designs a hyper-connected graph structure to establish a fully connected relationship between the core and each substituent, enhancing information integration and the adaptability to the internal structure changes of drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. In the drawings: Figure 1 Shows the flow chart for predicting activity cliffs when MMP represents a pair of complete drug molecules in the present invention.

[0032] Figure 2 Shows the flow chart for predicting activity cliffs when MMP represents a molecule split into 1 core and 2 substituents in the present invention.

[0033] Figure 3 Shows the drug molecule structures of the first group of MMPs (MMP - nonCliff) that do not form activity cliffs when MMP represents a pair of complete drug molecules.

[0034] Figure 4 Shows the drug molecule structures of the second group of MMPs (MMP - nonCliff) that do not form activity cliffs when MMP represents a pair of complete drug molecules.

[0035] Figure 5 Shows the drug molecule structures of the first group of MMPs (MMP - Cliff) that form activity cliffs when MMP represents a pair of complete drug molecules.

[0036] Figure 6 Shows the drug molecule structures of the second group of MMPs (MMP - Cliff) that form activity cliffs when MMP represents a pair of complete drug molecules.

[0037] Figure 7 Shows the drug molecule structures of the first group of MMPs (MMP - nonCliff) that do not form activity cliffs when MMP represents a molecule split into 1 core and 2 substituents.

[0038] Figure 8 Shows the drug molecule structures of the second group of MMPs (MMP - nonCliff) that do not form activity cliffs when MMP represents a molecule split into 1 core and 2 substituents.

[0039] Figure 9 Shows the drug molecule structures of the first group of MMPs (MMP - Cliff) that form activity cliffs when MMP represents a molecule split into 1 core and 2 substituents.

[0040] Figure 10 Shows the drug molecule structures of the second group of MMPs (MMP - Cliff) that form activity cliffs when MMP represents a molecule split into 1 core and 2 substituents. Detailed implementation mode

[0041] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0042] A method for predicting activity cliffs of task-specific drug molecules. MMP represents 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 for activity cliff prediction. When MMP represents a molecule split into 1 core and 2 substituents, based on the hyper-connected graph structure, the direct relationship between the core and the substituents is learned for activity cliff prediction.

[0043] Most of the existing activity cliff data is currently collected from the ChEMBL database. The generation of MMPs is achieved by swapping a pair of unique substructures (such as end groups or central fragments). The algorithm proposed by Hussain and Rea is commonly used for this task, and this algorithm focuses on the fragmentation and indexing of molecules in MMP analysis. Taking the single-cut example (i.e., MMPs with only a single terminal fragment different), in the first stage of the algorithm, all molecules in the input dataset are fragmented, and all possible single-cut cases at the labeled bonds in the molecules are listed, and then the generated fragments are indexed. Therefore, MMPs generally have two expressions, namely 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 composed of the core and the first substituent. is the activity information of the molecule composed of the core and the second substituent. When the activity difference between a pair of molecules is at least 100 times or 2 orders of magnitude (i.e., ), then this pair of molecules is regarded as having an activity cliff. When the difference is 10 times or 1 order of magnitude (i.e., ), it is regarded as having no activity cliff. When the MMPs between 1 and 2 are excluded.

[0044] Deep learning methods usually require sufficient data to perform well, but the molecular pairs of MMPs are relatively difficult to process and belong to high-quality labeled data. Therefore, the model will encounter the challenge of data scarcity. For this reason, the present invention uses a large number of molecules from DrugBank and drug data with rich bioactivity information, and pre-trains the model using the relatively simple drug interaction task in data labeling, and then uses transfer learning to update the model parameters, which greatly shortens the training time and improves the ability to identify the activity differences between new molecular pairs, thus achieving high-precision and efficient activity cliff prediction.

[0045] 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, as Figure 1 shown, which specifically includes the following steps:

[0046] S1. Take the first drug and the second drug as inputs, construct a drug interaction prediction task, pre-train the activity cliff prediction model for the molecular framework, and obtain the initial weights.

[0047] The pre-training stage of step S1 is specifically as follows: Input the drugs, encode the drug features using a graph convolutional neural network, pool the features of the two drugs to obtain two drug graph-level representations, splice the two drug-level representations, and then learn the detailed interaction between the molecule pairs through a molecular feature enhancer, and subsequently perform drug-drug interaction (DDI) prediction.

[0048] S2. Use the initial weights obtained from pre-training for transfer learning to update the activity cliff prediction model for the molecular framework, and use the activity cliff prediction model for the molecular framework to perform activity cliff prediction on drug molecule pairs.

[0049] Specifically, the specific steps of using the activity cliff prediction model for the molecular framework to perform activity cliff prediction in step S2 include the following:

[0050] S2.1. Take the molecule pair as input, use a graph convolutional neural network to update the molecular features, obtain the encoded molecular features, and perform a pooling operation on the encoded molecular features to obtain the graph-level representation of the molecule.

[0051] S2.2. Design a molecular feature enhancer based on Transformer to learn the relationship between molecule pairs and predict the activity cliff.

[0052] Specifically, the molecule pair includes a first molecule and a second molecule , and the feature update processes are the same; taking the feature update of the first molecule in the molecule pair as an example, step S2.1 specifically includes the following steps:

[0053] S2.1.1. Convert to a molecular graph through the RDKit tool, where the nodes in the molecular graph represent the atoms in the molecule and the edges represent the chemical bonds between the atoms.

[0054] S2.1.2. Use a graph convolutional neural network to perform convolutional operations on each atom and its adjacent atoms, update the atom features, and obtain the initial representation of the molecule: (1); Among them, is the feature after node passes through layers of convolution, is the set of neighbor nodes of node , and respectively represent the degrees of node and node , is a trainable weight matrix.

[0055] S2.1.3. Use pooling operation to aggregate all nodes of the molecule into a one-dimensional vector as the graph-level representation of the molecule: (2); Among them, represents the feature representation of the molecule after pooling compression, represents pooling, represents the molecule the number of nodes in.

[0056] Although the graph convolutional neural network GCN shows excellent performance in capturing molecular structures, when dealing with complex molecular pair interactions, a single GCN may not be able to fully capture the subtle differences between molecules, and these differences play a crucial role in the process of the model analyzing the communication between molecular pairs. To solve this problem, the present invention designs a molecular feature enhancer based on Transformer, aiming to learn the detailed communication between molecular pairs through the self-attention mechanism.

[0057] Specifically, step S2.2 specifically includes the following steps: S2.2.1. Learn the features between molecular pairs through the attention mechanism: (3); (4); (5); Among them, , , and are trainable weight matrices, represents the vector stacking operation, , and respectively represent the query vector, key vector and value vector, represents the normalization operation, represents the multi-head concatenation operation, Indicates the number of heads, Indicates the molecular pair and The final embedding.

[0058] S2.2.2. Predict the activity cliff activity between drug molecule pairs through a multi-layer perceptron: (6); Wherein, Is a multi-layer perceptron, Represents the predicted probability score of the activity cliff prediction model of the molecular framework.

[0059] S2.2.3. Regard the activity cliff prediction task as a binary classification task and use the binary cross-entropy loss function Train the model: (7); Wherein, 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, otherwise it does not exist.

[0060] Specifically, when MMP represents a molecule split into 1 core and 2 substituents, based on the hyper-connected graph structure, learn the direct relationship between the core and the substituents and perform activity cliff prediction. In this scenario, simply directly analyzing the split structure often ignores the key interaction associations between the core and the substituents, which is crucial for accurate activity cliff prediction. Especially when the activity differences caused by the substituents on the molecule are very subtle, the problem is even more serious because traditional models are difficult to accurately capture how these changes affect biological activity, resulting in inaccurate predictions. To overcome this limitation, the present invention proposes a hyper-connected graph structure for this scenario, as Figure 2 Shown, aiming to improve the ability to identify significant activity differences formed by the combination of the core and different substituents. In the hyper-connected graph, connection relationships are established between the core and the two substituents at the atomic level respectively, so that the model can calculate the attention scores of all nodes between the core and the substituents for updating the node features. Thus, the model has a deeper understanding of the dynamic relationship between the core and the substituents and can more accurately identify the impact of such subtle changes on molecular activity.

[0061] As Figure 2 Shown, specifically including the following steps: S3. Use the graph attention network GAT to update the features of the core and the substituents.

[0062] S4. Input the core and substituent features into the hyper-connected graph structure, perform pooling operations and concatenation operations in sequence, and finally predict the activity cliffs between drug molecule pairs through a multi-layer perceptron.

[0063] S5. Obtain the prediction loss using the binary cross-entropy loss function , and use L2 normalization to quantify the constraint loss of the difference between molecular features and activity features, and combine the prediction loss and the constraint loss to obtain the final loss.

[0064] Specifically, taking the feature update of the core and the first substituent (i.e., Figure 2 substituent 1 in S3.1 represents the core of each MMP as , and the first substituent as ; the graph structures of the core and the first substituent are respectively represented as and ; represents the initial embedding of the th node in , represents the initial embedding of the th node in .

[0065] S3.2 Update the features of the core and the substituent using GAT: (8); (9); where is the softmax activation function, is a learnable operator, and respectively represent the features of the th node in the core and the th node in the first substituent after being processed by layers of GAT, is the activation function, and respectively represent the association sets of the th node of the core and the th node of the first substituent in the hyper-connected graph, represents the feature of node after being processed by layers of GAT.

[0066] The process of feature update for the core and the second substituent (i.e., Figure 2 substituent 2 in

[0067] Specifically, step S4 specifically includes the following steps: S4.1, use pooling operation to and compress all nodes in into a one-dimensional vector, and use concatenation operation to obtain the final representation of the molecule composed of the core and the first substituent: (10); Among them, represents the representation of the molecule composed of and , represents the core the number of nodes in, represents the substituent the number of nodes in, represents the concatenation operation.

[0068] S4.2, the molecular representation composed of the core and the second substituent is denoted as , use concatenation operation to and merge into the final embedding of the two molecules, and use MLP to obtain the prediction result : (11).

[0069] To further improve the prediction performance of the model, the present invention introduces the activity information of two kinds of molecules. These values are mapped into a high-dimensional space through a linear layer. To effectively utilize these embeddings, a constraint loss that uses L2 normalization to quantify the difference between molecular features and activity features is also introduced, thereby improving the ability to identify the differences in molecular activities.

[0070] Specifically, the expression of the constraint loss that uses L2 normalization to quantify the difference between molecular features and activity features in step S5 is: (12); (13); Among them, is the feature of the molecule after quantization using L2 normalization, is the feature of the molecular activity after quantization using L2 normalization, is the matrix transpose, perform L2 normalization operation on the embedding, so that the L2 norm of each embedding is equal to 1, is the activity information of the molecule composed of the core and the first substituent, is the constraint loss of the molecule composed of the core and the first substituent, is the Frobenius norm, and the loss is calculated by computing the difference in the Frobenius norm. is the constraint loss of the molecule composed of the core and the second substituent , The calculation method is the same as .

[0071] Specifically, in step S5, the classification loss and the constraint losses with different weights are combined to obtain the final loss The expression of which is: (14); where is the hyperparameter for balancing the loss.

[0072] As shown in Table 1, the performance comparison of the model (TS-AC) proposed in the present invention with the current advanced methods in two tasks in the Thrombin dataset shows that the model achieves the optimal performance in five out of six evaluation metrics, verifying its application potential in different scenarios of activity cliffs. Among them, CC represents the scenario for the complete molecular pair, and CS represents the scenario for the molecule decomposed into the core and substituents. TPR is used to measure the ability of the model to correctly identify positive samples; TNR is used to measure the ability of the model to correctly identify negative samples; BA is the average of TPR and TNR, used to measure the ability of the model to address class imbalance; F1 takes into account the precision and recall of the model and is the harmonic mean of the two; MCC is an evaluation metric that comprehensively considers true positives, false positives, true negatives, and false negatives; AUC is used to measure the performance of the model at all possible classification thresholds.

[0073] Table 1 Performance comparison table

[0074] As Figures 3 - 6 shown, the visualization scheme of the model proposed in the present invention in the CC scenario. Two groups of MMPs that do not form activity cliffs are selected. In the first group of MMPs of MMP-nonCliff that do not form activity cliffs, the labeled nodes in the two drugs are the same, and none of these nodes fall on the substituents that cause structural changes in the two drugs. In the first group of MMPs of MMP-Cliff that form activity cliffs, the drug with ID 11471 has higher activity, and the substituents that cause structural changes are labeled, while the nodes at the substituent positions in the drug with ID 11464 are not concerned, indicating that the model of the present invention believes that the change in substituents in this group of MMPs is the key to causing the activity cliff. This finding highlights the sensitivity and predictive ability of the model to the activity changes caused by key structural differences.

[0075] As Figures 7 - 10As shown, in the first group of MMPs of MMP-nonCliff, the activity values of the two drugs can be observed to be relatively close. The model of the present invention focuses the labeled nodes on the core and does not label any substituents, indicating that the change of substituents has little effect on the drug activity. Similarly, in the second group of MMPs, all the importance labels are also concentrated on the core. This consistent result further verifies that in this case, the change of substituents does not significantly affect the drug activity, reflecting that the model can distinguish the influence of substituent changes on drug activity. In the two groups of MMPs of MMP-Cliff, the model successfully labels the changed substituents as key nodes, and these nodes are considered as the driving factors for 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 the substituents that cause high activity.

[0076] 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 those skilled in the art within the scope of the essence of the present invention should also fall within the protection scope 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 split into 1 core and 2 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.

2. A task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: 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: 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 initialization weights; S2, use the initialization weights obtained from pre-training for 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.

3. A task-specific drug molecule activity cliff prediction method according to claim 2, characterized in that: The process of predicting the activity cliff of drug molecule pairs using the molecular framework-oriented activity cliff prediction model in step S2 specifically includes the following steps: S2.1, taking the molecular pair as input, using graph convolutional neural network to update the features of the molecule to obtain the encoded molecular features, and performing pooling operation on the encoded molecular features to obtain the 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.

4. A task-specific drug molecule activity cliff prediction method according to claim 3, characterized in that: The first molecule in the pair 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 features of the atom 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 after pooling compression, the molecule The characteristic representation of represents pooling, Represents molecules The number of nodes in .

5. A task-specific drug molecule activity cliff prediction method according to claim 3, 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 denote the query vector, key vector and value vector respectively, represents the normalization operation, Indicates a multi-head splicing operation. Indicates the number of heads, Represents molecular pair and The final embedding of S2.2.2, predict the activity cliff between drug molecule pairs through 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 Train 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.

6. A task-specific drug molecule activity cliff prediction method according to claim 1, characterized in that: When 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, which specifically includes the following steps: S3, using the graph attention network GAT to update the features of the core and substituents; S4, input the core and substituent features into the hyperconnected graph structure, perform pooling operations and splicing operations in sequence, and finally predict the activity cliff between drug molecule pairs through 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.

7. A task-specific drug molecule activity cliff prediction method according to claim 6, 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 ; Indicated in Middle The initial embedding of nodes, Indicated in 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 The 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.

8. A task-specific drug molecule activity cliff prediction method according to claim 6, characterized in that: Step S4 specifically includes the following steps: S4.1, using pooling operation 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 The representation of the molecules that make up the Indicates 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 into the final embedding of the two molecules and use MLP to obtain the prediction results : （11）。 9. A task-specific drug molecule activity cliff prediction method according to claim 6, characterized in that: The expression of the constraint loss in step S5 that quantifies the difference between the molecular features and the activity features using L2 normalization is: （12）； （13）； in, To quantize the features of the molecule using L2 normalization, 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 molecules consisting of the core and the second substituent , Calculation method and same.

10. A task-specific drug molecule activity cliff prediction method according to claim 6, 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