Drug interaction prediction method and system based on knowledge distillation

Abstract

This invention belongs to the technical field of drug interaction prediction. To address the inaccuracy of existing drug interaction prediction methods, a drug interaction prediction method and system based on knowledge distillation is proposed. This method utilizes a multi-scale modeling model to encode and characterize drugs separately, obtaining corresponding drug representations. These drug representations are then fused to obtain a joint representation of drug pairs, leading to the results of drug interactions. Specifically, multiple independent energy-inspired molecular encoders are used to extract multiple pharmacokinetic property representation vectors for each drug. These independent energy-inspired molecular encoders are pre-trained based on these multiple pharmacokinetic properties, enabling them to learn representations containing pharmacokinetic knowledge. Finally, the multi-scale modeling model is trained using cross-task knowledge distillation with these multiple pharmacokinetic property representation vectors, giving the model pharmacological causal reasoning capabilities and significantly improving the accuracy and generalization ability of drug interaction prediction.

Description

Technical Field

[0001] This invention belongs to the technical field of drug interaction prediction, and particularly relates to a drug interaction prediction method and system based on knowledge distillation. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Drug-drug interaction (DDI) refers to the phenomenon where one drug alters the pharmacokinetic or pharmacodynamic properties of another drug when two or more drugs are used simultaneously or sequentially. Drug-drug interactions are a significant cause of adverse drug reactions and clinical treatment failures, especially in modern medicine where multidrug combination therapy is increasingly prevalent (such as the comprehensive treatment of oncology, cardiovascular diseases, and mental illnesses). Accurate prediction of drug-drug interactions is of great clinical significance for ensuring patient medication safety, guiding rational combination drug use, and reducing the cost of new drug development.

[0004] Traditional drug-drug interaction prediction methods primarily rely on pharmacological experiments and clinical observations, which are costly, time-consuming, and unable to comprehensively cover all possible drug combinations. In recent years, computational methods based on graph neural networks have made significant progress because they can model drug molecules as graph structures with atoms as nodes and chemical bonds as edges, thereby learning molecular topological features. Drug-drug interactions are mainly mediated by pharmacokinetics (PK), involving interactions during drug absorption, distribution, metabolism, excretion, and toxicity. However, most existing methods operate within a single-task learning framework, treating drug-drug interaction prediction as an isolated end-to-end mapping from molecular structure to labels, failing to explicitly absorb and utilize pharmacokinetic mechanism knowledge, leading to inaccurate drug-drug interaction predictions. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a drug interaction prediction method and system based on knowledge distillation, which transfers pharmacokinetic knowledge to a multi-scale modeling model, enabling the multi-scale modeling model to have pharmacological causal reasoning ability, and significantly improving the accuracy, generalization ability and pharmacological interpretability of drug-drug interaction prediction.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a drug interaction prediction method based on knowledge distillation, comprising: The first drug and the second drug are characterized and encoded using a multi-scale modeling model. The first drug and the second drug are characterized by extracting local features, global representations and molecular structure semantic features of the first drug and the second drug and fusing them. The first drug characterization and the second drug characterization are fused to obtain a drug pair joint characterization. The drug pair joint characterization is then input into a prediction model to obtain the interaction results between the first drug and the second drug. Among them, multiple independent energy-inspired molecular encoders are used to extract multiple pharmacokinetic property representation vectors of the drug; multiple independent energy-inspired molecular encoders are pre-trained based on multiple pharmacokinetic properties of the drug, so that the independent energy-inspired molecular encoders learn representations containing pharmacokinetic knowledge. By using multiple pharmacokinetic property characterization vectors to train a multi-scale modeling model through cross-task knowledge distillation, the multi-scale modeling model is endowed with pharmacological causal reasoning ability.

[0007] Secondly, the present invention provides a drug interaction prediction system based on knowledge distillation, comprising: The multi-scale modeling module is configured to: use a multi-scale modeling model to encode the first drug and the second drug respectively, and extract and fuse the local features, global representations and molecular structure semantic features of the first drug and the second drug to obtain the characterization of the first drug and the second drug. Specifically, multiple independent energy-inspired molecular encoders are used to extract multiple pharmacokinetic property representation vectors of the drug; multiple independent energy-inspired molecular encoders are pre-trained based on multiple pharmacokinetic properties of the drug, so that the independent energy-inspired molecular encoders learn representations containing pharmacokinetic knowledge; and the multi-scale modeling model is trained by cross-task knowledge distillation using multiple pharmacokinetic property representation vectors, so that the multi-scale modeling model has the ability of pharmacological causal reasoning. The interaction prediction module is configured to: fuse the first drug characterization and the second drug characterization to obtain a joint drug pair characterization, input the joint drug pair characterization into a prediction model, and obtain the interaction results between the first drug and the second drug.

[0008] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0009] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0010] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0011] The above one or more technical solutions have the following beneficial effects: In this invention, local features, global representations, and molecular structural semantic features of a drug are extracted and fused through a multi-scale modeling model to obtain a comprehensive drug characterization. Multiple independent energy-inspired molecular encoders are obtained by pre-training multiple pharmacokinetic properties of the drug, enabling these independent energy-inspired molecular encoders to generate pharmacokinetic property characterization vectors rich in pharmacokinetic knowledge. Through cross-task knowledge distillation, pharmacokinetic knowledge is adaptively transferred to the multi-scale modeling model, giving the multi-scale modeling model a foundation for pharmacological causal reasoning, thereby significantly improving the accuracy, generalization ability, and pharmacological interpretability of drug interaction prediction.

[0012] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0013] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0014] Figure 1 This is an overall flowchart of the drug interaction prediction method based on knowledge distillation in Embodiment 1 of the present invention; Figure 2 This is a visualization of the molecular characterization embedded in UMAP obtained in the drug-drug interaction prediction task in the DrugBank dataset in Embodiment 1 of the present invention; Figure 3 This is a visualization of the molecular characterization embedding UMAP obtained in the drug-drug interaction prediction task of the Twosides dataset in Embodiment 1 of the present invention. Detailed Implementation

[0015] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0016] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0017] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0018] Example 1 like Figure 1 As shown, this embodiment provides a drug interaction prediction method based on knowledge distillation, including: A multi-scale modeling model is used to encode the first and second drugs separately. By extracting and fusing the local features, global representations, and molecular structural semantic features of the first and second drugs, the characterizations of the first and second drugs are obtained. Specifically, multiple independent energy-inspired molecular encoders are used to extract multiple pharmacokinetic property representation vectors of the drugs. These independent energy-inspired molecular encoders are pre-trained with multiple pharmacokinetic properties of the drugs, enabling them to learn representations containing pharmacokinetic knowledge. The multi-scale modeling model is then trained using cross-task knowledge distillation using these multiple pharmacokinetic property representation vectors, giving the model pharmacological causal reasoning capabilities. The characterization of the first drug and the characterization of the second drug are fused to obtain the joint characterization of the drug pair. The joint characterization of the drug pair is then input into the prediction model to obtain the interaction results between the first drug and the second drug.

[0019] This embodiment extracts and fuses local features, global representations, and molecular structural semantic features of a drug through a multi-scale modeling model to obtain a comprehensive drug characterization. Multiple independent energy-inspired molecular encoders are obtained by pre-training multiple pharmacokinetic properties of the drug, enabling these independent energy-inspired molecular encoders to generate pharmacokinetic property characterization vectors rich in pharmacokinetic knowledge. Through cross-task knowledge distillation, pharmacokinetic knowledge is adaptively transferred to the multi-scale modeling model, giving the model a foundation for pharmacological causal reasoning and significantly improving the accuracy, generalization ability, and pharmacological interpretability of drug interaction prediction.

[0020] The following is a detailed description of the drug interaction prediction method based on knowledge distillation proposed in this embodiment: S1. Obtain the molecular structure data and atomic three-dimensional coordinates of the first and second drugs.

[0021] Specifically, the SMILES strings of the first and second drugs are obtained and converted into a molecular graph using the RDKit toolkit. The molecular graph is constructed with atoms as nodes and chemical bonds as edges. A three-dimensional conformation is generated using the ETKDGv3 method to obtain the three-dimensional coordinates of the atoms, which are used for subsequent energy calculations and distance biasing in the Transformer.

[0022] S2. The first drug and the second drug are characterized and encoded using a multi-scale modeling model. The local features, global representations and molecular structure semantic features of the first drug and the second drug are extracted and fused to obtain the characterization of the first drug and the second drug.

[0023] In this embodiment, a KAN-driven multi-scale modeling model is used to encode the molecular characteristics of the first drug and the second drug, respectively, to obtain the first drug characterization and the second drug characterization. The multi-scale modeling model includes an improved graph convolutional network, a Transformer encoder, and a language model. The improved graph convolutional network takes a molecular graph as input, and the Transformer encoder takes the drug's SMILES sequence as input. The improved graph convolutional network and the Transformer encoder encode the molecular characteristics of the drug.

[0024] In this embodiment, the improved graph convolutional network is: the Kolmogorov-Arnold network, or KAN, is used to replace the linear transformation in the standard graph convolutional network. KAN places the activation function on the edge, and each univariate function is composed of learnable B-spline basis functions.

[0025] The node update formula for each graph convolutional layer is:

[0026] in, For the first l Layer nodes v i Features For nodes v i The set of neighboring nodes, where ⊕ represents the concatenation operation. KAN For the Kolmogorov-Arnold network layer, after Drug delivery via layer message passing The final local features are denoted as After propagation through 3 layers, local features are obtained through atomic average pooling. ;u represents a node v i The neighboring nodes.

[0027] In this embodiment, the Transformer encoder employs a multi-head self-attention mechanism, introducing a three-dimensional distance bias between atoms in the attention calculation to output a global representation. :

[0028] .

[0029] in, , , Through initial characterization of the drug By performing linear projection, It is a learnable mapping matrix. It is the dimension of K. Represents atoms and atoms The three-dimensional distance between them. Initial drug characterization. It is a fixed or learnable embedding matrix constructed based on prior knowledge of molecular structure (atomic properties).

[0030] To extract rich semantic features of molecular structure, a pre-trained language model is used for encoding: ,in, Textual information representing molecular structure or properties, such as molecular descriptions obtained from literature or databases; The features representing the text description are encoded by BERT and then processed by a multilayer perceptron. Feature vectors mapped to appropriate dimensions; It is a multilayer perceptron, which maps the output of a pre-trained language model to the appropriate dimension.

[0031] In this embodiment, local features, global representation, and molecular structure semantic features are integrated to combine local features. Global representation and molecular structure semantic features By piecing them together, the drug can be obtained. Final drug characterization : The superscript L indicates the Lth floor. Indicates splicing.

[0032] This embodiment employs a multi-scale architecture combining a KAN-based improved graph convolutional network, Transformer, and semantic encoding. The KAN-based improved graph convolutional network uses learnable spline basis functions to replace fixed activation functions, enabling adaptive modeling of complex nonlinear interactions between atoms. The Transformer introduces a three-dimensional distance bias to effectively capture global conformational dependencies, and semantic encoding further supplements structural textual knowledge, together achieving a comprehensive representation of molecules from atoms to the whole.

[0033] This embodiment trains the multi-scale modeling model through cross-task knowledge distillation, enabling the multi-scale modeling model to have pharmacological causal reasoning capabilities.

[0034] First, multiple independent energy-inspired molecular encoders are constructed and pre-trained on multiple pharmacokinetic property prediction tasks. This step aims to enable the independent energy-inspired molecular encoders to learn representations containing pharmacokinetic knowledge. Independent energy-inspired molecular encoders are pre-trained on five pharmacokinetic prediction tasks: absorption, distribution, metabolism, excretion, and toxicity. Each energy-inspired molecular encoder has the same network structure but different model parameters.

[0035] Multiple independent energy-inspired molecular encoders were pre-trained for five pharmacokinetic prediction tasks: ESOL (water solubility, absorption), BBBP (blood-brain barrier permeability, distribution), HLM Clearance (human liver microsomal clearance, metabolism), OCT2 Inhibition (organic cation transporter 2 inhibition, excretion), and Clintox (clinical toxicity).

[0036] Each pharmacokinetic prediction task is trained for 100 epochs. The regression task uses mean squared error loss, and the classification task uses binary cross-entropy loss. Five pharmacokinetic representation vectors are obtained after pre-training. .

[0037] Since molecular interactions are inherently governed by the energy landscape, the stability, reactivity, and binding affinity of drugs to targets are determined by various intramolecular and intermolecular interaction energies, such as bond stretching energy, bond bending energy, dihedral torsion energy, van der Waals energy, and electrostatic interaction energy. Existing methods primarily learn from two-dimensional topological structures, failing to explicitly incorporate these fundamental physical energy constraints into the model. This results in molecular representations learned by the model that, while statistically fitting the training data, exhibit inherent geometric-energy relationships that contradict real physical laws, severely limiting the model's ability to generalize to out-of-distribution samples.

[0038] This embodiment enables an independent energy-inspired molecular encoder to generate molecular characterizations that incorporate both physical energy constraints and pharmacokinetic knowledge by explicitly embedding physical energy constraints.

[0039] The following explains the processing procedure of the energy-inspired molecular encoder: First, the bond stretching energy is calculated based on the CHARMM force field. Key angle bending energy Dihedral torsional energy Abnormal torsional energy And Lennard-Jones van der Waals and electrostatic coulombic energy .

[0040] The energy-inspired molecular encoder in this embodiment includes three hierarchical modules: atomic-level module, bond-level module, and molecular-level module.

[0041] At the atomic level, the representation of each atom is adjusted by aggregating the multi-hop neighborhood information of the atom itself and weighting it according to the interaction strength with the central atom. When an atom node is updated, the one-hop neighbors, i.e., directly connected atoms, are adjusted by bond stretching energy. As a aggregation weight, it reflects the stability of local covalent bonds; two-hop neighbors, that is, atoms connected through an intermediate atom, have bond angle bending energy. Lennard-Jones van der Waals and electrostatic coulombic energy The sum, as the aggregation weight, reflects geometric distortion and weak interaction; three-hop neighbors are atoms connected through two intermediate atoms with dihedral torsional energy. Lennard-Jones van der Waals and electrostatic coulombic energy The sum is used as the aggregation weight to reflect long-range stereoelectronic effects and weak interactions; at the same time, abnormal torsional energy is introduced. To perceive planar distortion / chirality information.

[0042] The updated formula is as follows:

[0043]

[0044]

[0045]

[0046]

[0047] in, , and Each represents an atom v i The 1-hop, 2-hop, and 3-hop neighbor sets; This represents the aggregated message from a 1-hop neighbor atom to the target atom, while and Similarly, this represents the aggregated message from 2-hop neighbor atoms and 3-hop neighbor atoms to the target atom; For atoms v i With atoms The bond stretching energy; For atoms v i ,atom and atoms The bending energy of the bond angle between them; For atoms v i and atoms Lennard-Jones van der Waals between; For atoms v i and atoms The electrostatic coulomb energy between them; For atoms v i ,atom ,atom and atoms Torsional energy of the dihedral angle between them; For atoms v i and atoms Lennard-Jones van der Waals between; For atoms v i and atoms The electrostatic coulomb energy between them; For atoms v i The central atom, with three neighboring atoms , , The abnormal torsional energy formed; Represents the tensor product; Indicates l-1 After layer message transmission, drug molecules Graph-level features; Indicates l-1 After layer message passing, the node v i eigenvectors; Indicates l After layer message passing, atoms v i eigenvectors; This indicates that after message passing through layer l-1, the atom v i With atoms Edge characteristics between them; ; They represent respectively l-1 After layer message passing, atoms ,atom eigenvectors; , , They represent respectively l-1 After layer message passing, atoms v i Neighboring atoms , , The feature vectors; MLP represents a multilayer perceptron.

[0048] In the bond-level module, the representation of each chemical bond is updated by aggregating its multi-hop neighborhood information. During bond-level updates, the aggregation weight of a one-hop neighbor bond is based on the bond angle bending energy between that neighbor bond and the target bond. The aggregation weight of two-hop adjacent bonds is based on the dihedral torsional energy of the adjacent bond relative to the target bond. .

[0049] The updated formula is as follows:

[0050]

[0051]

[0052]

[0053] in, This represents the aggregated message from the 1-hop neighbor key to the target key, while This represents the aggregated message from the 2-hop neighbor key to the target key; This indicates that after message passing through layer l-1, the atom v i With atoms Edge characteristics between them; Indicates l-1 After layer message passing, the node v i eigenvectors; Indicates l After layer message passing, atoms v i eigenvectors; Indicates l-1 After layer message transmission, drug molecules Graph-level features; , Each represents an atom The set of 1-hop and 2-hop neighbors; For atoms v i ,atom and atoms The bending energy of the bond angle between them; For atoms v i ,atom ,atom and atoms Torsional energy of the dihedral angle between them; It is an atom v i ,atom and atoms Interaction terms are used to supplement bond angle bending energies to explain vibrational coupling between bonded atoms.

[0054] In the molecular-level module, global molecular features are obtained by merging atomic features and chemical bond features (edge ​​features) after performing mean pooling on each feature separately.

[0055] The specific update formula is as follows:

[0056] Where Mean() represents the average pooling operation, l Indicates the level number, Represents a node v i The sum of characteristics Represents atoms v i With atoms The sum of edge features; V represents the set of nodes; Indicates l After layer message passing, atoms v i With atoms Edge characteristics between them; E represents the set of edges; Indicates l-1 After layer message transmission, drug molecules Graph-level features; superscript l Representation layer index.

[0057] go through After layer message passing, drug molecules are obtained. The final graph-level feature is denoted as .

[0058] In this embodiment, the bond stretching energy, bond angle bending energy, dihedral angle torsion energy, and non-bonded interaction energy in the molecular force field are used as aggregation weights for multi-hop message passing, so that the molecular characterization learning process follows the laws of physics and enhances the physical realism and distributional out-of-distribution generalization ability of the model.

[0059] This embodiment uses cross-task knowledge distillation to adaptively transfer five pre-trained pharmacokinetic knowledge sets to a multi-scale modeling model, specifically: Obtain multiple pharmacokinetic property representation vectors generated by a pre-trained energy-inspired molecular encoder for the same drug. The pharmacokinetic properties are characterized in five dimensions: absorption, distribution, metabolism, excretion, and toxicity. The five pharmacokinetic property vectors are then compared with the drug characterization vector generated by a multi-scale modeling module for the same drug. Perform L2 normalization to obtain and The adaptive weights of each pharmacokinetic characterization vector are calculated using an attention mechanism.

[0060] No. Attention score of each ADMET attribute (abbreviation of Absorption, Distribution, Metabolism, Excretion, Toxicity). calculate:

[0061]

[0062] in, , and These are trainable parameters. It is a temperature parameter that controls the smoothness of attention distribution; the superscript T indicates transpose. This represents the tangent function.

[0063] Based on adaptive weights Calculate drugs Cross-task knowledge distillation loss :

[0064] in, This represents the square of the L2 norm.

[0065] Drug effect Total distillation loss of the multi-scale model:

[0066] in, This represents the cross-task knowledge distillation loss for drug A; This represents the cross-task knowledge distillation loss of drug B.

[0067] S3: The first drug characterization and the second drug characterization are fused to obtain the drug pair joint characterization. The drug pair joint characterization is input into the prediction model to obtain the interaction probability between the first drug and the second drug.

[0068] In this embodiment, the first drug characterization γ A With the second drug characterization γ B spliced ​​into γ AB = [γ A || γ BThe input to the prediction model is first processed by a three-layer multilayer perceptron, and then the interaction probability is output by the Sigmoid function. AB .

[0069] Optionally, the hidden layer dimensions of the three-layer multilayer perceptron are 256, 128, and 1, respectively.

[0070] The binary cross-entropy loss is used as the prediction loss of the prediction model. The total loss function is:

[0071] Wherein, λ is the balance parameter, with a value ranging from 0.1 to 0.5, and λ=0.3 is preferred through grid search.

[0072] End-to-end training was performed using the Adam optimizer with a learning rate of 0.001 and a batch size of 32.

[0073] This example illustrates the implementation process using the drug pair "ketoconazole-simvastatin" known to interact with each other in the DrugBank database.

[0074] S101, Data Acquisition: Acquire ketoconazole (C 26 H 28 Cl2N4O4) and simvastatin (C 25 H 38 SMILES of O5 were converted into molecular diagrams using RDKit.

[0075] Ketoconazole has 43 atoms and 46 bonds; simvastatin has 68 atoms and 73 bonds. Atomic and bond features were extracted, and a three-dimensional conformation was generated using ETKDGv3. The atomic features include approximately 80 dimensions such as atom type, degree, formal charge, hybridization state, and intra-ring characteristics; the bond features include approximately 10 dimensions such as bond type, conjugation, and intra-ring characteristics.

[0076] S102, Energy-Inspired Multi-Source Pre-training: Calculating Energy Priors Based on CHARMM Force Fields.

[0077] Taking the ketoconazole imidazole ring as an example: CN bond E bond Approximately 0.015 kcal / mol; CNC bond angle E angle Approximately 0.017 kcal / mol; the dihedral angle E connecting the imidazole ring and the benzene ring. dihedral Approximately 0.4 kcal / mol; the distance between the two oxygen atoms in the non-bonded interaction is 5.2 Å, E nonbond It is approximately 19.1 kcal / mol.

[0078] A four-layer multi-scale model (128 hidden dimensions) was constructed and updated at the atomic, bond, and molecular levels as described above. The model was independently pre-trained for 100 rounds on five tasks: ESOL (water solubility, absorption), BBBP (blood-brain barrier permeability, distribution), HLM Clearance (human liver microsomal clearance, metabolism), OCT2 Inhibition (organic cation transporter 2 inhibition, excretion), and Clintox (clinical toxicity), extracting five 128-dimensional pharmacokinetic characteristics of ketoconazole.

[0079] S103, KAN-driven multi-scale modeling: Construct a 3-layer KAN-GCN (256 hidden dimensions), where each univariate function in the KAN layer uses 5 B-spline basis functions. Propagation is performed according to the node update formula, and after 3 layers, 256-dimensional local features are obtained through atomic mean pooling.

[0080] Construct a 3-head attention Transformer (256-dimensional), introduce a 3D distance bias into the attention, and output 256-dimensional global features.

[0081] The 512-dimensional ketoconazole drug characterization is obtained by splicing local features and global representations. keto Simvastatin drug characterization γ simva .

[0082] S104, Cross-task knowledge distillation: Characterizing ketoconazole drug γ keto Projected to 128 dimensions, and L2 normalized to the five pre-trained representations respectively, the attention score and weight β are calculated. i .

[0083] Ketoconazole had the highest weighting for metabolic characterization, β≈0.45.

[0084] Calculate the cross-task knowledge distillation loss:

[0085] Similarly:

[0086] in, For cross-task knowledge distillation loss of ketoconazole; This represents the cross-task knowledge distillation loss of simvastatin; For adaptive weights; , These are the drug characterization vectors for ketoconazole and simvastatin, respectively. This represents a vector characterizing pharmacokinetic properties.

[0087] S105, Prediction and Optimization: Characterizing Ketoconazole Drugs with γ keto Synvastatin drug characterization γ simvaThe concatenation results in a 1024-dimensional array, which is input into a three-layer MLP (1024→256→128→1) and outputs probabilities via a Sigmoid function. True label y=1. Total loss: .

[0088] Adam optimization, learning rate 0.001, batch size 32, training for 100 epochs. AUROC of 99.21% was achieved on the transductive test set.

[0089] Table 1 shows a performance comparison between the method of this embodiment and the baseline model on the inductive drug-drug interaction prediction task.

[0090] Table 1:

[0091] Table 2 compares the performance of the method in this embodiment with that of the baseline model on the inductive drug-drug interaction prediction task.

[0092] Table 2:

[0093] Inductive setup: On the DrugBank dataset (1,706 drugs), the dataset is divided by drug. 20% of the drugs (approximately 342) are randomly selected as the test set, and the remaining 80% (approximately 1,364) are used as the training set. The drugs and their interaction pairs in the training set are used for model training, while the drugs in the test set are completely hidden during the training phase.

[0094] The prediction task is to predict the interaction probability between each new drug in the test set and all drugs in the training set.

[0095] Evaluation metrics included ACC (accuracy), AUROC (area under the ROC curve), Precision, and F1-score. Each experiment was independently repeated five times, and the mean ± standard deviation was used. Results showed that this embodiment achieved an accuracy of 95.39% ± 0.31%, an AUROC of 97.87% ± 0.22%, a precision of 96.09% ± 0.35%, and an F1 score of 95.23% ± 0.28%. Compared to the closest baseline LSA-DDI (85.41% / 93.21% / 90.35% / 86.21%), the AUROC improved by approximately 4.7 percentage points, and the F1 score improved by approximately 9 percentage points.

[0096] Case Study: Two clinically relevant drugs were selected: Lomitapide (a microsomal triglyceride transfer protein inhibitor used for homozygous familial hypercholesterolemia) and Escitalopram (a selective serotonin reuptake inhibitor used for major depressive disorder) for a new drug generalization test. All known interactions involving Lomitapide and Escitalopram were completely removed from the training set to simulate the scenario of their introduction as "new drugs." After training, the model was used to predict the probability of interactions between these two drugs and all other drugs in DrugBank, and the 10 candidate interactions with the highest confidence were selected for each. Cross-validation was performed using a drug interaction checker.

[0097] The results showed that for Lomitapide (a microsomal triglyceride transfer protein inhibitor), the Top-10 predicted risks included abemaciclib (cumulative risk of hepatotoxicity), stavudine (hepatotoxicity), erlotinib (hepatotoxicity), pralsetinib (hepatotoxicity), colestipol (decreased absorption), armodafinil (increased blood drug concentration), lomustine (hepatotoxicity), tocilizumab (hepatotoxicity), vemurafenib (hepatotoxicity), and oxandrolone (hepatotoxicity). All 10 predictions were validated by Drugs.com, with a validation rate of 100%. For escitalopram (a selective serotonin reuptake inhibitor), the Top-10 predictions include apalutamide (risk of arrhythmia), tazemetostat (risk of bleeding), fluconazole (risk of arrhythmia), zaleplon (overlay of central nervous system depression), eszopiclone (central nervous system depression), tenecteplase (risk of bleeding), remimazolam (risk of central nervous system depression), valerian (risk of central nervous system depression), offloxacin (risk of arrhythmia), and iloprost (risk of bleeding), all of which have been validated through clinical databases.

[0098] This case study demonstrates that this embodiment successfully learned transferable pharmacokinetic principles rather than memorizing specific drug pairs, thus possessing the ability to make forward-looking predictions about new drugs.

[0099] Using a pushover setup on the DrugBank dataset, with the complete model as the baseline, the following components were removed or replaced: the entire energy-inspired multi-source pre-trained module was removed; the molecular semantic encoder was removed; the energy-inspired encoder was replaced with a standard MPNN; and the KAN-GCN was replaced with a standard GCN.

[0100] The experimental results, as shown in Table 3, indicate that the complete model achieved an accuracy of 99.82%, an AUROC of 99.21%, a precision of 99.15%, and an F1 score of 99.38%. Removing the energy-inspired multi-source pre-training module resulted in a significant drop in accuracy to 93.28%, a decrease of 6.54 percentage points, and an AUROC to 94.05%, a decrease of 5.16 percentage points, demonstrating that pharmacokinetic pre-training provided the necessary theoretical foundation for the model. Removing the semantic encoder resulted in a slight decrease in performance (accuracy 97.05%), indicating that semantic knowledge played an auxiliary enhancement role. Replacing the energy-inspired encoder with an MPNN restored the accuracy to 95.83%, but it was still significantly lower than the complete model (99.82%), confirming the necessity of the energy-aware design. Replacing KAN-GCN with GCN reduced the AUROC from 99.21% to 97.13% and the F1 score from 99.38% to 97.49%, indicating that the adaptive nonlinearity of KAN is crucial for high-precision prediction.

[0101] Table 3:

[0102] This embodiment examines the impact of key hyperparameters, providing empirical guidance for model configuration.

[0103] Energy-inspired multi-source pre-training module depth: Performance was tested on DrugBank and Twosides datasets with depths of 1–6 layers. Results show that all metrics peak at a depth of 4 layers in both transductive and inductive settings. With depths less than 4 layers, the model underfits (both training and validation losses are high); with depths greater than 4 layers, the validation loss shows an upward trend (overfitting).

[0104] KAN-GCN Depth: The impact of 1-5 layers on AUROC was tested. Results consistently showed that 3 layers resulted in the highest AUROC (99.21% for direct projection from DrugBank and 97.87% for inductive projection). Increasing the number of layers increased the risk of gradient vanishing and overfitting.

[0105] Number of attention heads: Test 1-5 attention heads. Optimal performance is achieved with 3 attention heads (AUROC 99.21%). Too few (1 head) results in insufficient expressive power (98.45%), while too many (5 heads) introduce redundancy and make training unstable (98.92%).

[0106] Balance coefficient λ: Grid search is performed in the [0,1] interval with a step size of 0.1. When λ=0 (no distillation), the AUROC is 98.01%; performance gradually improves as λ increases, reaching a peak of 99.21% when λ=0.3; performance begins to decline after λ>0.3 (98.56% when λ=0.5), indicating that excessively strong distillation signals can interfere with the main task optimization.

[0107] Visualization of pre-trained task molecular embeddings, such as Figure 2 As shown, UMAP was used to project the molecular embeddings of the model outputs from five pharmacokinetic property prediction (PPP) tasks into a two-dimensional space. For the pre-training task, blood-brain barrier-penetrating molecules (positive, purple) and non-penetrating molecules (negative, blue) formed two clearly separated clusters (one clustered in the lower left region, and the other in the upper right region), with clear boundaries. Clear classification boundaries were also observed for the OCT2Inhibition and Clintox tasks. For the ESOL and HLM Clearance regression tasks, the diagonal plots of predicted and true values ​​showed that most points were closely distributed around the y=x diagonal. The R² for ESOL was 0.94, and for HLM Clearance, it was 0.88, demonstrating that the model can accurately quantify continuous pharmacokinetic properties.

[0108] Drug-drug interaction embedding UMAP visualization, such as Figure 3 The final joint representation of drug pairs is projected into a two-dimensional space. Interacting pairs (green) and non-interacting pairs (orange) form two distinct clusters, with clear manifold structures observed in both DrugBank and Twosides datasets, indicating that the model successfully encodes the interaction relationships into a linearly separable representation space.

[0109] Attention heatmaps for drug-drug interaction tasks: Attention heatmaps for four representative drug pairs are generated using a gradient-based attention visualization method. Absorption Case (Yantoraprazole-Dasatinib): The model focuses on the benzimidazole ring, pyridine ring, and sulfoxide moiety of yantoprazole, and the piperazine ring, pyrimidine ring, and thiazole ring of dasatinib, consistent with the mechanism by which proton pump inhibitors increase gastric pH and affect the absorption of weakly basic drugs.

[0110] Distribution Case (Warfarin-Aspirin): The model focuses on the coumarin ring system and β-diketone moiety of warfarin, and the salicylic acid core and carboxylic acid group of aspirin, which is consistent with the pharmacological mechanism of the two competing for the albumin binding site (Sudlow site I).

[0111] Metabolic case (ketoconazole-simvastatin): The model focuses on the imidazole ring, dichlorobenzene ring, and piperazine side chain of ketoconazole (imidazole nitrogen coordinates with CYP3A4 heme iron), and on the δ-lactone ring, β-hydroxy acid moiety, and alkyl side chain of simvastatin, consistent with the mechanism by which CYP3A4 inhibition leads to reduced metabolism of statins.

[0112] Excretion case (probenecid-benzylpenicillin): The model focuses on the aromatic ring, sulfonamide group, and carboxylic acid moiety of probenecid (OAT inhibitor pharmacophore), and on the β-lactam-thiazolidinyl core and benzyl side chain of benzylpenicillin (OAT substrate characteristics), consistent with the mechanism of competitive inhibition of renal tubular secretion mediated by OAT transporters.

[0113] Experiments on multiple benchmark datasets such as DrugBank and Twosides show that this embodiment achieves optimal performance in both inductive and inductive settings, and in particular, its generalization ability for new drugs is significantly better than existing methods in the inductive setting.

[0114] Attention heatmaps show that the molecular structures and pharmacological mechanisms of interest identified by the model in this embodiment are highly consistent with those of absorption (benzimidazole ring, piperazine ring), distribution (coumarin ring, salicylic acid core), metabolism (imidazolium ring, δ-lactone ring), and excretion (sulfonamide group, β-lactam ring). Case studies show that the Top-10 predictions for the novel drugs Lomitapide and Escitalopram were validated by the Drugs.com clinical database with a 100% validation rate. UMAP visualizations show that the drug pairs are embedded in two clearly separated clusters between interacting and non-interacting categories, indicating that the model has learned meaningful discriminative features.

[0115] Example 2 The purpose of this embodiment is to provide a drug interaction prediction system based on knowledge distillation, including: The multi-scale modeling module is configured to: encode the first drug and the second drug using a multi-scale modeling model; extract and fuse the local features, global representations, and molecular structure semantic features of the first and second drugs to obtain the drug characterizations; wherein, multiple independent energy-inspired molecular encoders are used to extract multiple pharmacokinetic property characterization vectors of the drug; the multiple independent energy-inspired molecular encoders are pre-trained from multiple pharmacokinetic properties of the drug; and the multi-scale modeling model is trained using cross-task knowledge distillation using multiple pharmacokinetic property characterization vectors to enable the multi-scale modeling model to have pharmacological causal reasoning capabilities; The interaction prediction module is configured to: fuse the characterization of the first drug and the characterization of the second drug to obtain a joint characterization of the drug pair, input the joint characterization of the drug pair into the prediction model, and obtain the interaction results between the first drug and the second drug.

[0116] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0117] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0118] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0119] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0120] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0121] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0122] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0123] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0124] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0125] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0126] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A drug interaction prediction method based on knowledge distillation, characterized in that, include: The first drug and the second drug are characterized and encoded using a multi-scale modeling model. The first drug and the second drug are characterized by extracting local features, global representations and molecular structure semantic features of the first drug and the second drug and fusing them. The first drug characterization and the second drug characterization are fused to obtain a drug pair joint characterization. The drug pair joint characterization is then input into a prediction model to obtain the interaction results between the first drug and the second drug. Specifically, multiple independent energy-inspired molecular encoders are used to extract multiple pharmacokinetic property representation vectors of the drug; multiple independent energy-inspired molecular encoders are pre-trained based on multiple pharmacokinetic properties of the drug, so that the independent energy-inspired molecular encoders learn representations containing pharmacokinetic knowledge; and the multi-scale modeling model is trained by cross-task knowledge distillation using multiple pharmacokinetic property representation vectors, so that the multi-scale modeling model has the ability of pharmacological causal reasoning. The energy-inspired molecular encoder includes an atomic-level module, a bond-level module, and a molecular-level module. In the atomic-level module, the representation of each atom is obtained by aggregating multi-hop neighborhood information of the atom itself and adjusting it with energy weighting based on the interaction strength with the central atom. In the bond-level module, the representation of each chemical bond is updated by aggregating multi-hop neighborhood information to obtain chemical bond features. In the molecular-level module, the global molecular features are obtained by fusing the atomic features and chemical bond features after performing mean pooling on them respectively. In the atomic-level module, the updated formula is as follows: in, , and Each represents an atom v i The 1-hop, 2-hop, and 3-hop neighbor sets; This represents the aggregated message from a 1-hop neighbor atom to the target atom. This represents the aggregated message from a 2-hop neighbor atom to the target atom; This represents the aggregated message from a 3-hop neighbor atom to the target atom; For atoms v i With atoms The bond stretching energy; For atoms v i ,atom and atoms The bending energy of the bond angle between them; For atoms v i and atoms Lennard-Jones van der Waals between; For atoms v i and atoms The electrostatic coulomb energy between them; For atoms v i ,atom ,atom and atoms Torsional energy of the dihedral angle between them; For atoms v i and atoms Lennard-Jones van der Waals between; For atoms v i and atoms The electrostatic coulomb energy between them; For atoms v i The central atom, with three neighboring atoms , , The abnormal torsional energy formed; Represents the tensor product; Indicates l-1 After layer message transmission, drug molecules Graph-level features; Indicates l-1 After layer message passing, the node v i eigenvectors; Indicates l-1 After layer message passing, the node eigenvectors; Indicates l After layer message passing, atoms v i eigenvectors; This indicates that after message passing through layer l-1, the atom v i With atoms Edge characteristics between them; , They represent respectively l-1 After layer message passing, atoms ,atom eigenvectors; , , They represent respectively l-1 After layer message passing, atoms v i Neighboring atoms , , The feature vectors; MLP represents a multilayer perceptron.

2. The drug interaction prediction method based on knowledge distillation as described in claim 1, characterized in that, A multi-scale modeling model is used to characterize and encode the first and second drugs respectively. By extracting and fusing the local features, global representations, and molecular structural semantic features of the first and second drugs, the characterizations of the first and second drugs are obtained, specifically as follows: Local features of the first and second drugs are extracted using an improved graph convolutional network; wherein the improved graph convolutional network is: a KAN network is used to replace the multilayer perceptron in the standard graph convolutional network. A global representation of the first and second drugs is extracted using a Transformer encoder; where the atomic 3D distance is introduced as a bias in the attention calculation. Based on a pre-trained language model, feature extraction is performed on the textual information of the molecular structure or properties of the first and second drugs to obtain the corresponding molecular structure semantic features. The local features, global representations, and molecular structure semantic features corresponding to the first drug and the second drug are fused to obtain the characterization of the first drug and the characterization of the second drug.

3. The drug interaction prediction method based on knowledge distillation as described in claim 1, characterized in that, The formula is updated in the key-level module as follows: in, This represents the aggregated message from the 1-hop neighbor key to the target key. This represents the aggregated message from the 2-hop neighbor key to the target key; This indicates that after message passing through layer l-1, the atom v i With atoms Edge characteristics between them; Indicates l- 1 After layer message passing, the node v i eigenvectors; Indicates l-1 After layer message transmission, drug molecules Graph-level features; , Each represents an atom The set of 1-hop neighbors and the set of 2-hop neighbors; For atoms v i ,atom and atoms The bending energy of the bond angle between them; For atoms v i ,atom ,atom and atoms Torsional energy of the dihedral angle between them; It is an atom v i ,atom and atoms Interaction terms; MLP stands for Multilayer Perceptron; Indicates l-1 After layer message passing, the node eigenvectors; This indicates that after message passing through layer l-1, the atom With atoms Edge features between them.

4. The drug interaction prediction method based on knowledge distillation as described in claim 1, characterized in that, The multi-scale modeling model is trained using cross-task knowledge distillation by utilizing multiple pharmacokinetic property characterization vectors, specifically as follows: Based on multiple independent energy-inspired molecular encoders, multiple pharmacokinetic property characterization vectors of drugs are extracted; The multiple pharmacokinetic property characterization vectors of a drug are normalized to the drug characterization of the same drug. The adaptive weights of each pharmacokinetic characterization vector are calculated using an attention mechanism; Based on the calculated adaptive weights, and combined with each pharmacokinetic characterization vector and the corresponding drug characterization, a cross-task knowledge distillation loss function is constructed. Cross-task knowledge distillation loss function is used to train multi-scale modeling models for cross-task knowledge distillation.

5. A drug interaction prediction system based on knowledge distillation, characterized in that, include: The multi-scale modeling module is configured to: encode the first drug and the second drug using a multi-scale modeling model; extract and fuse the local features, global representations, and molecular structural semantic features of the first and second drugs to obtain the drug representations; wherein, multiple independent energy-inspired molecular encoders are used to extract multiple pharmacokinetic property representation vectors of the drugs; multiple independent energy-inspired molecular encoders are pre-trained based on multiple pharmacokinetic properties of the drugs, enabling the independent energy-inspired molecular encoders to learn representations containing pharmacokinetic knowledge; and the multi-scale modeling model is trained by cross-task knowledge distillation using multiple pharmacokinetic property representation vectors, enabling the multi-scale modeling model to have pharmacological causal reasoning ability; The energy-inspired molecular encoder includes an atomic-level module, a bond-level module, and a molecular-level module. In the atomic-level module, the representation of each atom is obtained by aggregating multi-hop neighborhood information of the atom itself and adjusting it with energy weighting based on the interaction strength with the central atom. In the bond-level module, the representation of each chemical bond is updated by aggregating multi-hop neighborhood information to obtain chemical bond features. In the molecular-level module, the global molecular features are obtained by fusing the atomic features and chemical bond features after performing mean pooling on them respectively. In the atomic-level module, the updated formula is as follows: in, , and Each represents an atom v i The 1-hop, 2-hop, and 3-hop neighbor sets; This represents the aggregated message from a 1-hop neighbor atom to the target atom. This represents the aggregated message from a 2-hop neighbor atom to the target atom; This represents the aggregated message from a 3-hop neighbor atom to the target atom; For atoms v i With atoms The bond stretching energy; For atoms v i ,atom and atoms The bending energy of the bond angle between them; For atoms v i and atoms Lennard-Jones van der Waals between; For atoms v i and atoms The electrostatic coulomb energy between them; For atoms v i ,atom ,atom and atoms Torsional energy of the dihedral angle between them; For atoms v i and atoms Lennard-Jones van der Waals between; For atoms v i and atoms The electrostatic coulomb energy between them; For atoms v i The central atom, with three neighboring atoms , , The abnormal torsional energy formed; Represents the tensor product; Indicates l-1 After layer message transmission, drug molecules Graph-level features; Indicates l-1 After layer message passing, the node v i eigenvectors; Indicates l-1 After layer message passing, the node eigenvectors; Indicates l After layer message passing, atoms v i eigenvectors; This indicates that after message passing through layer l-1, the atom v i With atoms Edge characteristics between them; , They represent respectively l-1 After layer message passing, atoms ,atom eigenvectors; , , They represent respectively l-1 After layer message passing, atoms v i Neighboring atoms , , The feature vectors; MLP represents a multilayer perceptron; The interaction prediction module is configured to: fuse the first drug characterization and the second drug characterization to obtain a joint drug pair characterization, input the joint drug pair characterization into a prediction model, and obtain the interaction results between the first drug and the second drug.

6. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-4.

8. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-4.

Images