Method and system for drug target affinity prediction using multi-scale neural adaptive process

Abstract

The application discloses a drug target affinity prediction method and system of a multi-scale neural adaptive process, relates to the technical field of drug target affinity prediction, and comprises the following steps: acquiring a training sample set composed of a drug molecule sequence and a target amino acid sequence; constructing a multi-scale neural adaptive prediction model composed of four modules, i.e., a molecular representation learning module, an interaction mode learning module, a pharmacological mechanism learning module and an affinity prediction module, wherein the molecular representation module generates drug target interaction representation, the interaction mode module extracts interaction mode hidden variables, the pharmacological mechanism module extracts pharmacological representation, and the affinity prediction module fuses the above information to output a prediction value; training the model by using the training sample and a joint loss function containing an exponential weighted mean square error loss and a KL regularization loss; and inputting a drug-target pair to be tested into the trained model to obtain an affinity prediction result. Through multi-scale information fusion and probability modeling, accurate prediction and uncertainty estimation of drug target affinity in a few-sample scenario are realized.

Description

Technical Field

[0001] This application relates to the field of drug target affinity prediction technology, and in particular to a method and system for predicting drug target affinity in a multi-scale neural adaptive process. Background Technology

[0002] Drug-target affinity (DTA) prediction is a crucial step in drug discovery and retargeting. Early DTA prediction methods primarily relied on molecular docking, simulating binding modes using 3D structures such as GOLD and LeDock. With the rise of machine learning, methods such as KronRLS and Affinity2Vec have been explored for DTA prediction. Deep learning, in particular, has been widely used for DTA prediction because it can automatically learn efficient representations of compounds and proteins, reducing information loss during intermediate processes.

[0003] However, existing DTA prediction methods still face two core limitations: First, DTA data is inherently sparse and uncertain. Limited by high costs and technical bottlenecks, experimentally determined affinities only cover a small fraction of all possible drug-target combinations, making DTA prediction a few-shot learning task. Furthermore, measurement differences due to different experimental techniques, environmental perturbations, and instrument noise lead to fluctuations and biases in affinity values, which existing studies typically ignore, simplifying the problem to a deterministic regression task and failing to model prediction uncertainty. Second, mainstream methods usually focus on molecular structural features, neglecting the multi-scale biological information contained in DTA. DTA is essentially a comprehensive representation of multi-scale biological processes. The molecular structures of drugs and targets determine binding sites (molecular scale), different drug-target interaction modes jointly influence affinity (interaction scale), and the interactions between the same drug and different targets reveal pharmacological mechanisms (pharmacological scale). Integrating multi-scale information helps enhance the ability to capture intermolecular dynamics. Summary of the Invention

[0004] The purpose of this application is to provide a method and system for predicting drug target affinity in a multi-scale neural adaptive process, thereby solving the aforementioned problems in the prior art.

[0005] To achieve the above objectives, this application provides a method for predicting drug target affinity in a multi-scale neural adaptive process, comprising the following steps: S1: Obtain the drug SMILES sequence and target amino acid sequence to form a training sample set; S2: Construct a multi-scale neural adaptive prediction model, which includes a molecular representation learning module, an interaction pattern learning module, a pharmacological mechanism learning module, and an affinity prediction module. The molecular representation learning module is used to fuse drug-target correlation information and historical drug-target interaction information at the molecular scale to generate drug-target interaction representations. The interaction pattern learning module is used to extract latent variables of interaction patterns from drug-target pairs using an interaction pattern selector. The pharmacological mechanism learning module is used to extract pharmacological representations from a pharmacological prototype library. The affinity prediction module is used to fuse drug-target interaction representations and latent variables, and uses the pharmacological representations to modulate the parameters of the prediction network, outputting the predicted affinity value. S3: Train the multi-scale neural adaptive prediction model using training samples, with the optimization objective being a joint loss function that includes exponentially weighted mean squared error loss and KL regularization loss; S4: Input the drug-target pair to be tested into the trained multi-scale neural adaptive prediction model to obtain the drug-target affinity prediction results.

[0006] Preferably, in step S1, the training sample set is organized into multiple training tasks according to the meta-learning paradigm, and each training task includes a support set and a query set.

[0007] Preferably, the molecular representation learning module is used to fuse inter-drug correlation information and drug target historical interaction information at the molecular scale to generate drug target interaction representations, specifically including: The embedding representations of the drug and the target are obtained by using the drug encoder and the target encoder, respectively, based on the drug SMILES sequence and the target amino acid sequence. Based on the similarity between drug embedding representations, an auxiliary drug is selected for the current drug. The embedding representation of the selected auxiliary drug is pooled to obtain the drug auxiliary representation. The auxiliary interaction representation is obtained by calculating the embedding representation of the current target and the embedding representation of the historical interaction targets through an attention mechanism; The drug embedding representation, target embedding representation, drug-assisted representation, and interaction-assisted representation are fused to generate a drug-target interaction representation.

[0008] Preferably, the similarity between drug embedding representations is represented as follows: ; in, For drugs Embedded representation, yes Drug embedding representation sorted by similarity yes In addition to Drug embedding representation outside of drug; The expression for drug adjuvant is: ; in, Indicates pooling operation; The auxiliary interaction expression is: ; in, Dimensions for queries, keys, and values; The expression for drug-target interaction is: ; in, Target Embedded representation.

[0009] Preferably, the interaction mode selector is a gated hybrid expert network, and the process of extracting latent variables of the interaction mode includes: The drug target interaction representation is concatenated with the corresponding affinity value and then input into the gating network to calculate the weights of each expert network. The outputs of each expert network are weighted and summed to obtain a mixed representation of the samples. The sample hybrid representations in the support set and query set are exchanged and aggregated to obtain the support set aggregate representation and the task aggregate representation; Based on support set aggregation representation and task aggregation representation, the mean and log-variance are obtained through linear mapping, respectively, and the latent variables of interaction patterns are obtained by sampling using reparameterization techniques.

[0010] Preferably, the formula for calculating the weights in a gated network is: ; ; in, This indicates the affinity value. It is the unnormalized weight of the gating network. and These are the weights and biases of the gated network. For the first Normalized weights corresponding to each expert; The mean and log-variance are expressed as follows: ; ; in, To support aggregate representation, The weights of the mean-mapped network, This is the bias term for the mean-mapped network. The weights of the log-variance mapping network, For the bias term of the log-variance mapping network; The implicit variables of the interaction mode are represented as follows: ; Here, ⊙ represents element-wise multiplication.

[0011] Preferably, the process by which the pharmacological mechanism learning module obtains a pharmacological representation specifically includes: The target interaction representation and affinity value supporting the set are input into the feedforward neural network to obtain the initial representation of the task: ; in, To support the set; A pharmacological prototype library was introduced. Each of them This represents a learnable pharmacological prototype, using the Student's t-distribution as the kernel function for measurement. and Normalized similarity between them: ; The pharmacological representation is obtained by weighted summation based on similarity: .

[0012] Preferably, the parameter modulation process of the affinity prediction module includes: The drug target interaction representation is concatenated with the latent variables of the interaction pattern and used as the input to the prediction network; A characteristic linear modulation mechanism is used to generate scaling and offset coefficients for each layer of the prediction network based on pharmacological representation.

[0013] Preferably, the joint loss function consists of three sub-losses, expressed as: ; ; ; ; in, yes The weight, Represents the training set, It is used to emphasize the importance of high-affinity samples. , This is the weighting growth factor. For high affinity threshold, For the task posterior distribution, For the prior distribution based on the support set, For the target distribution, This represents the normalized similarity distribution between the initial pharmacological representation and the pharmacological prototype.

[0014] A drug target affinity prediction system based on a multi-scale neural adaptive process, comprising: The data acquisition module is used to acquire drug SMILES sequences and target amino acid sequences to form a training sample set; The model building module is used to construct a multi-scale neural adaptive prediction model, which includes a molecular representation learning module, an interaction pattern learning module, a pharmacological mechanism learning module, and an affinity prediction module. Specifically, the molecular representation learning module fuses drug-target correlation information with historical drug-target interactions at the molecular scale to generate drug-target interaction representations; the interaction pattern learning module extracts latent variables of interaction patterns from drug-target pairs using an interaction pattern selector; the pharmacological mechanism learning module extracts pharmacological representations from a pharmacological prototype library; and the affinity prediction module fuses drug-target interaction representations and latent variables, and uses the pharmacological representations to modulate the parameters of the prediction network, outputting predicted affinity values. The model training module is used to train a multi-scale neural adaptive prediction model using training samples. The optimization objective is a joint loss function that includes exponentially weighted mean squared error loss and KL regularization loss. The prediction module is used to input the drug-target pair to be tested into a trained multi-scale neural adaptive prediction model to obtain the drug-target affinity prediction results.

[0015] Therefore, the above-mentioned method and system for predicting drug target affinity based on a multi-scale neural adaptive process has the following beneficial effects: (1) This application adopts a probabilistic learning architecture, which introduces Gaussian random variables to model the uncertainty in the prediction of drug target affinity in a small number of samples, thereby improving the generalization ability and prediction robustness of the model in low-resource scenarios.

[0016] (2) At the molecular scale, this application introduces auxiliary drugs and historical interaction targets to supplement potential association information based on the principle of "guilt-by-association"; at the interaction scale, an interaction mode selector is designed to dynamically identify different types of molecular interaction modes; at the pharmacological scale, a pharmacological prototype library is constructed to learn the potential pharmacological mechanisms of drugs, thereby realizing multi-level and multi-perspective molecular relationship modeling.

[0017] (3) The exponentially weighted mean square error loss function of this application, by giving higher affinity samples a greater weight, makes the model pay more attention to key active samples, thereby enhancing its guiding value in actual new drug discovery and drug optimization.

[0018] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a block diagram of the multi-scale neural adaptive prediction model in the embodiments of this application. Detailed Implementation

[0020] The following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning as understood by a person of ordinary skill in the art to which this application pertains.

[0022] The terms "comprising" or "including," as used in this application, mean that the element preceding the term encompasses the element listed after the term, and do not exclude the possibility of encompassing other elements as well. The terms "inner," "outer," "upper," and "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In this application, unless otherwise expressly specified and limited, the term "attached," etc., should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can refer to a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication of two elements or the interaction relationship between two elements. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0023] Example 1: This application draws on the ideas of neural processes (NP), assuming that each task Corresponding to a stochastic process Therefore, the "observed target affinity" can be viewed as the sampled values ​​of a random function at discrete input points. NP approximates the random process as a deep neural network and uses random vectors. By parameterizing the process, under the condition of independent and identically distributed distribution, the process can be represented as: ; in Drugs and target Interactive representation, To represent the affinity value between the two, for ease of representation, we will use... and Abbreviated as and .from Sampling This can be viewed as sampling from the function space. .

[0024] Due to the true posterior distribution Difficult to estimate, this application uses amortized variational inference to learn it. A variational posterior distribution is introduced. and conditional priors Establish a variational lower bound (ELBO) for predictions on the query set: ; in Task The support set in Task The query set in Expressing expectations, This represents the KL divergence between two distributions. The first term in the formula corresponds to the reconstruction loss, and the second term corresponds to the KL regularization.

[0025] Specifically, a method for predicting drug target affinity in a multi-scale neural adaptive process includes: S1: Obtain the drug SMILES sequence and target amino acid sequence to form a training sample set; The training sample set is organized into multiple training tasks according to the meta-learning paradigm, and each training task contains a support set and a query set.

[0026] S2: Construct a multi-scale neural adaptive prediction model, such as Figure 1 As shown, the multi-scale neural adaptive prediction model includes a molecular representation learning module, an interaction pattern learning module, a pharmacological mechanism learning module, and an affinity prediction module. The molecular representation learning module is used to fuse drug-target association information and historical drug-target interaction information at the molecular scale to generate drug-target interaction representations. The interaction pattern learning module is used to extract latent variables of interaction patterns from drug-target pairs using an interaction pattern selector. The pharmacological mechanism learning module is used to extract pharmacological representations from a pharmacological prototype library. The affinity prediction module is used to fuse drug-target interaction representations and latent variables, and uses the pharmacological representations to modulate the parameters of the prediction network, outputting the predicted affinity value. The molecular representation learning module is used to generate interactive representations for each drug target pair. The SMILE sequence of the drug amino acid sequence of the target After being input into this module, the data first passes through the drug encoder ChemBERTa and the target encoder ESM2 to obtain the embedding representations of each drug and target. and Embedding representation for a given drug target pair and The molecular representation learning module uses drug embedding enhancement and historical interaction representation operations to generate drug-aided representations. and interactive assistance representation .

[0027] Drug embedding enhancement: represented by drugs As input, generate drug-aided representations. Regarding medication In other words, it is necessary to support other drugs from the central drug supply chain. Search for auxiliary drugs within the matrix. Calculate the similarity between the current drug and other drug embeddings: ; in yes Embeddings sorted by similarity yes In addition to Drug embeddings other than those mentioned above. Sort by similarity from highest to lowest, and select the top... The embedding of an auxiliary drug is obtained through pooling. ; in This indicates a pooling operation.

[0028] Historical interaction representation: represented by target points Generate interactive auxiliary representations based on the input. During training, historical interaction targets are continuously iterated, for In other words, the former Targets in each sample These are historical interaction targets. During testing, for the support set, historical interaction targets are obtained iteratively, while for the query set, the last one is directly selected from the support set. Each target is used as a historical interaction target. The current target and historical target are embedded and fed into a linear transformation to obtain a query, key, and value vector: ; in For query, For key, For value, , and These are their mapping matrices, where || denotes vector concatenation. Then, auxiliary interaction representations are calculated based on the attention mechanism. ; in Dimensions for queries, keys, and values.

[0029] get , , and Then, we concatenate them and perform feature fusion using a feedforward neural network (FNN) with multiple hidden layers to generate the final drug-target interaction representation. ; The purpose of the interactive mode learning module is to provide tasks Generative variational approximation and Then, latent variables containing interaction information are sampled. To capture different drug target interaction patterns, the interaction pattern selector calculates expert weights through a gating network. Process each sample: ; ; in It is the unnormalized weight of the gating network. and These are the weights and biases of the gating network. The outputs of each expert are then weighted and summed according to the gating weights to obtain a mixed representation of the samples: ; get Subsequently, to ensure independence from the order of samples, the interaction pattern learning module will... Aggregation is supported by set aggregation representation and task aggregation representation: ; ; in For modeling , For simulation ⊕ represents the exchange operation, and the model in this application implements this operation through vector addition to improve computational efficiency.

[0030] To derive the hidden variables of a stochastic process As extensively studied in Variational Auto-Encoders (VAEs), this application uses a Gaussian distribution to represent... That is, assumption The mean and log-variance are obtained through a linear layer. The mapping yields: ; Subsequently, reparameterization techniques were used to achieve differentiability, based on the task representation. generate : ; Where ⊙ represents element-wise multiplication.

[0031] In order to capture pharmacological characteristics This application introduces a pharmacological mechanism learning module, which aims to learn from the support set Learning in China This is to assist in parameter modulation of the affinity prediction module. First, [the following is done]... The input is fed into the FNN to obtain the representation. : ; To further explore the pharmacological mechanism of action, a pharmacological prototype library was introduced. Each of them This represents a learnable pharmacological prototype. The Student's t-distribution is used as the kernel function for measurement. and Normalized similarity between them: ; in These are the degrees of freedom of the student's t-distribution. Finally, the pharmacological representation... Depend on The weighted sum of the original and the prototype constitutes: ; Due to the lack of real pharmacological prototype data, an auxiliary target distribution is introduced. To improve the quality of soft clustering assignment and facilitate task prototype learning, the target distribution is designed with reference to deep clustering methods. The definition is as follows: ; The pharmacological prototype library can then be optimized by minimizing the KL divergence.

[0032] After obtaining the molecular representation Interaction mode representation and pharmacological expression Then, the affinity prediction module learns the conditional likelihood. ,according to In affinity To make predictions, this application employs Feature-wise Linear Modulation (FiLM) to construct a multi-scale fusion scorer in order to effectively integrate multi-scale information.

[0033] FiLM is a differentiable modulation mechanism that is highly effective in many fields. Its core idea is to adaptively influence the model's predictions within a neural network by applying feature-level affine transformations to the intermediate features of each layer. Formally, for any intermediate representation... The FiLM modulation operation can be described as follows: ; in and These are the scaling and offset coefficients, respectively. In this model, FiLM is in the... modulation parameters of the layer and Represented by global task The decision is made using the following formula: ; in Used to constrain the modulation coefficients in the interval [-1, 1] and They are respectively and The mapping matrix.

[0034] When performing affinity prediction, and This is concatenated as input for the affinity prediction module: ; Subsequently, the multi-scale fusion scorer executes affine modulation and activation functions layer by layer to form a deep modulation network: ; go through After the modulation network of the layers, the final output prediction value is: ; S3: Train the multi-scale neural adaptive prediction model using training samples, with the optimization objective being a joint loss function that includes exponentially weighted mean squared error loss and KL regularization loss; The optimization objective of this model consists of three sub-losses, and the overall loss function is as follows: ; in yes The weight.

[0035] The first part is: ; in Represents the training set, It is a weight used to emphasize the importance of high-affinity samples, defined as: ; in This is the weighting growth factor. This represents the high affinity threshold. This loss term corresponds to the reconstruction term in neural process modeling. By minimizing the weighted mean square error between the predicted value and the true label, it effectively improves the model's fitting ability in high affinity regions.

[0036] The second part is: ; This term is a KL regularization term, used to constrain the posterior distribution of the task. Compared with prior distributions based on support sets Consistency between them ensures that the model infers It exhibits stability and generalization.

[0037] The third part is: ; This loss encourages the model to pay more attention to high-confidence task-prototype matching, thereby improving the discriminativeness and purity of the prototype.

[0038] S4: Input the drug-target pair to be tested into the trained multi-scale neural adaptive prediction model to obtain the drug-target affinity prediction results.

[0039] Example 2: A drug target affinity prediction system based on a multi-scale neural adaptive process, comprising: The data acquisition module is used to acquire drug SMILES sequences and target amino acid sequences to form a training sample set; The model building module is used to construct a multi-scale neural adaptive prediction model, which includes a molecular representation learning module, an interaction pattern learning module, a pharmacological mechanism learning module, and an affinity prediction module. Specifically, the molecular representation learning module fuses drug-target correlation information with historical drug-target interactions at the molecular scale to generate drug-target interaction representations; the interaction pattern learning module extracts latent variables of interaction patterns from drug-target pairs using an interaction pattern selector; the pharmacological mechanism learning module extracts pharmacological representations from a pharmacological prototype library; and the affinity prediction module fuses drug-target interaction representations and latent variables, and uses the pharmacological representations to modulate the parameters of the prediction network, outputting predicted affinity values. The model training module is used to train the multi-scale neural adaptive prediction model using training samples. The optimization objective is a joint loss function that includes exponentially weighted mean squared error loss and KL regularization loss. The prediction module is used to input the drug-target pair to be tested into a trained multi-scale neural adaptive prediction model to obtain the drug-target affinity prediction results.

[0040] Therefore, this application adopts the above-mentioned multi-scale neural adaptive process drug target affinity prediction method and system, which improves the generalization ability and robustness of the model in low-resource scenarios by using a probabilistic element learning architecture and Gaussian random variable modeling to predict uncertainty; it also enhances the ability to characterize complex interaction patterns and potential pharmacological mechanisms by co-modeling the drug-target relationship at the molecular, interaction, and pharmacological scales; and by increasing the training weight of high affinity samples, the model pays more attention to key active samples, thereby improving its predictive value and application effect in new drug discovery and drug optimization.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of this application, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of this application.

Claims

1. A method for predicting drug target affinity in a multi-scale neural adaptive process, characterized in that, Includes the following steps: S1: Obtain the drug SMILES sequence and target amino acid sequence to form a training sample set; S2: Construct a multi-scale neural adaptive prediction model, which includes a molecular representation learning module, an interaction pattern learning module, a pharmacological mechanism learning module, and an affinity prediction module; wherein, the molecular representation learning module is used to fuse drug-target association information and drug-target historical interaction information at the molecular scale to generate drug-target interaction representation; the interaction pattern learning module is used to extract latent variables of interaction patterns from drug-target pairs using an interaction pattern selector. The pharmacological mechanism learning module is used to extract pharmacological representations from the pharmacological prototype library; the affinity prediction module is used to fuse drug-target interaction representations and latent variables, and use pharmacological representations to modulate the parameters of the prediction network to output the affinity prediction value. S3: Train the multi-scale neural adaptive prediction model using training samples, with the optimization objective being a joint loss function that includes exponentially weighted mean squared error loss and KL regularization loss; S4: Input the drug-target pair to be tested into the trained multi-scale neural adaptive prediction model to obtain the drug-target affinity prediction results.

2. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 1, characterized in that, In step S1, the training sample set is organized into multiple training tasks according to the meta-learning paradigm, and each training task includes a support set and a query set.

3. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 2, characterized in that, The molecular representation learning module is used to fuse inter-drug correlation information and drug target historical interaction information at the molecular scale to generate drug target interaction representations, specifically including: The embedding representations of the drug and the target are obtained by using the drug encoder and the target encoder, respectively, based on the drug SMILES sequence and the target amino acid sequence. Based on the similarity between drug embedding representations, an auxiliary drug is selected for the current drug. The embedding representation of the selected auxiliary drug is pooled to obtain the drug auxiliary representation. The auxiliary interaction representation is obtained by calculating the embedding representation of the current target and the embedding representation of the historical interaction targets through an attention mechanism; The drug embedding representation, target embedding representation, drug-assisted representation, and interaction-assisted representation are fused to generate a drug-target interaction representation.

4. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 3, characterized in that, The similarity between drug embeddings is represented as follows: ； in, For drugs Embedded representation, yes Drug embedding representation sorted by similarity yes In addition to Drug embedding representation outside of drug; The expression for drug adjuvant is: ； in, Indicates pooling operation; The auxiliary interaction expression is: ； in, Dimensions for queries, keys, and values; The expression for drug-target interaction is: ； in, Target Embedded representation.

5. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 4, characterized in that, The interaction mode selector is a gated hybrid expert network. The process of extracting the latent variables of the interaction mode includes: The drug target interaction representation is concatenated with the corresponding affinity value and then input into the gating network to calculate the weights of each expert network. The outputs of each expert network are weighted and summed to obtain a mixed representation of the samples. The sample hybrid representations in the support set and query set are exchanged and aggregated to obtain the support set aggregate representation and the task aggregate representation. Based on support set aggregation representation and task aggregation representation, the mean and log-variance are obtained through linear mapping, respectively, and the latent variables of interaction patterns are obtained by sampling using reparameterization techniques.

6. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 5, characterized in that, The formula for calculating weights in a gating network is: ； ； in, This indicates the affinity value. It is the unnormalized weight of the gating network. and These are the weights and biases of the gated network. For the first Normalized weights corresponding to each expert; The mean and log-variance are expressed as follows: ； ； in, To support aggregate representation, The weights of the mean-mapped network, This is the bias term for the mean-mapped network. The weights of the log-variance mapping network are... For the bias term of the log-variance mapping network; The implicit variables of the interaction mode are represented as follows: ； Here, ⊙ represents element-wise multiplication.

7. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 6, characterized in that, The process by which the pharmacological mechanism learning module obtains pharmacological representations specifically includes: The target interaction representation and affinity value supporting the set are input into the feedforward neural network to obtain the initial representation of the task: ； in, To support the set; A pharmacological prototype library was introduced. Each of them This represents a learnable pharmacological prototype, using the Student's t-distribution as the kernel function for measurement. and Normalized similarity between them: ； The pharmacological representation is obtained by weighted summation based on similarity: 。 8. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 7, characterized in that, The parameter modulation process of the affinity prediction module includes: The drug target interaction representation is concatenated with the latent variables of the interaction pattern and used as the input to the prediction network; A characteristic linear modulation mechanism is used to generate scaling and offset coefficients for each layer of the prediction network based on pharmacological representation.

9. The method for predicting drug target affinity in a multi-scale neural adaptive process according to claim 8, characterized in that, The joint loss function consists of three sub-losses, expressed as: ； ； ； ； in, yes The weight, Represents the training set, It is used to emphasize the importance of high-affinity samples. , This is the weighting growth factor. For high affinity threshold, For the task posterior distribution, For the prior distribution based on the support set, For the target distribution, This represents the normalized similarity distribution between the initial pharmacological representation and the pharmacological prototype.

10. A drug target affinity prediction system for a multi-scale neural adaptive process, employing the drug target affinity prediction method for a multi-scale neural adaptive process as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire drug SMILES sequences and target amino acid sequences to form a training sample set; The model building module is used to construct a multi-scale neural adaptive prediction model, which includes a molecular representation learning module, an interaction pattern learning module, a pharmacological mechanism learning module, and an affinity prediction module. Specifically, the molecular representation learning module fuses drug-target correlation information with historical drug-target interactions at the molecular scale to generate drug-target interaction representations; the interaction pattern learning module extracts latent variables of interaction patterns from drug-target pairs using an interaction pattern selector; the pharmacological mechanism learning module extracts pharmacological representations from a pharmacological prototype library; and the affinity prediction module fuses drug-target interaction representations and latent variables, and uses the pharmacological representations to modulate the parameters of the prediction network, outputting predicted affinity values. The model training module is used to train a multi-scale neural adaptive prediction model using training samples. The optimization objective is a joint loss function that includes exponentially weighted mean squared error loss and KL regularization loss. The prediction module is used to input the drug-target pair to be tested into a trained multi-scale neural adaptive prediction model to obtain the drug-target affinity prediction results.

Images