Drug target prediction method and system based on credible routing and orthogonal residual expert

Abstract

The present application relates to the technical field of artificial intelligence assisted drug discovery, and discloses a drug target prediction method and system based on credible routing and orthogonal residual experts. The method comprises: obtaining drug and target protein input, extracting task-specific features from a backbone model; extracting word-level embeddings from a frozen pre-trained encoder and performing backbone conditional attention pooling to obtain sequence-level priors; calculating drug-side credibility, protein-side credibility and pairwise credibility after dimension reduction of the sequence-level priors, and constructing pairwise prior description vectors; generating orthogonal residual candidates through a low-rank residual expert; generating routing weights according to the pairwise prior description vectors and the pairwise credibility for credible residual injection to obtain enhanced representations; and outputting drug-target interaction probability or affinity value according to the enhanced representations. The present application reduces the negative transfer risk of low-credibility priors through double-scale credible calibration and orthogonal residual expert mechanism, and improves the adaptability and stability of the prediction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted drug discovery technology, and specifically relates to a drug target prediction method and system based on trusted routing and orthogonal residual experts. Background Technology

[0002] Drug-target prediction is a fundamental task in computational drug discovery. Drug-target interaction (DTI) prediction typically determines whether a given compound interacts with a target protein, while drug-target affinity (DTA) prediction typically estimates the strength of continuous binding between the two. Accurate drug-target prediction results can assist in virtual screening, target identification, drug relocation, and candidate drug prioritization.

[0003] In recent years, deep learning methods have been widely applied in this field. Typical drug-target prediction models usually include a drug encoder, a protein encoder, and an interaction module or prediction head. The drug side can use SMILES sequences, molecular graphs, or molecular fingerprints as input; the protein side can use amino acid sequences or their derived features as input; the interaction module models the relationship between drugs and proteins through methods such as splicing, attention, cross-attention, bilinear matching, gating fusion, or graph Transformers.

[0004] In existing solutions, some techniques set up parallel pre-trained sequence prior branches outside the original drug-target prediction backbone model, and concatenated the drug-side pre-trained representations and protein-side pre-trained representations with backbone features after pooling and dimensionality reduction. While these solutions can introduce external unlabeled sequence knowledge, their structures typically remain at the level of external branches, fixed pooling, low-dimensional compression, and simple fusion layers, failing to adequately address the following issues: First, the reliability of the pre-trained priors for the current sample is not computably determined; second, the modeling of whether the drug-side and protein-side priors constitute a credible pairwise relationship is lacking; third, prior injection easily leads to redundant coupling with backbone features or covers the original backbone representation; fourth, out-of-distribution or low-quality priors may cause negative transfer; fifth, different drug-target samples may require different prior injection substructures, and fixed concatenation or a single gating is insufficient to express this difference.

[0005] Therefore, there is an urgent need for a general technical solution that, while preserving the original backbone structure and training process as much as possible, not only introduces pre-trained drug sequence priors and pre-trained protein sequence priors, but also evaluates the credibility of unilateral and paired priors, dynamically selects residual experts, eliminates backbone parallel redundant components, and suppresses negative migration through counterfactual constraints.

[0006] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address or at least mitigate one or more of the above problems, a drug target prediction method based on trusted routing and orthogonal residual experts is provided. This method can reduce interference and negative migration risk caused by low-confidence priors while keeping the original backbone structure as unchanged as possible through dual-scale trusted calibration and orthogonal residual expert mechanism, thereby improving the adaptability and stability of drug-target interaction or affinity prediction.

[0008] To achieve the above objectives, according to a first aspect of the present invention, a drug target prediction method based on trusted routing and orthogonal residual experts is provided, comprising the following steps:

[0009] S1. Obtain drug input and target protein input, and further obtain corresponding labels during the training phase;

[0010] S2. The original drug-target prediction backbone model is used to extract features from the drug input and target protein input to obtain task-specific drug features and protein features.

[0011] S3. The drug sequence input and protein sequence input are encoded by the frozen drug-side pre-trained sequence encoder and protein-side pre-trained sequence encoder respectively, and the drug-side word-level context embedding and protein-side word-level context embedding are extracted.

[0012] S4. Using the task-specific drug features and protein features as query conditions, attention-weighted aggregation is performed on the drug-side terminology-level context embedding and the protein-side terminology-level context embedding to obtain drug sequence-level priors and protein sequence-level priors.

[0013] S5. Dimensionally reduce the drug sequence-level prior and the protein sequence-level prior, calculate the drug-side confidence, protein-side confidence and drug-target pairwise confidence based on the dimensionality-reduced prior, and construct the pairwise prior description vector.

[0014] S6. Multiple low-rank residual experts generate prior residual candidates based on the dimensionality-reduced priors, and remove redundant components parallel to the corresponding original backbone features from each prior residual candidate to obtain orthogonal residual candidates.

[0015] S7. Generate expert routing weights based on the pairwise prior description vectors and the pairwise confidence. After weighted fusion of each orthogonal residual candidate based on the pairwise confidence and the expert routing weights, inject the original backbone features to obtain enhanced drug representation and enhanced protein representation.

[0016] S8. Based on the enhanced drug representation and enhanced protein representation, output the drug-target interaction probability or affinity value.

[0017] Furthermore, the drug sequence-level prior and protein sequence-level prior are dimensionality-reduced. Based on the dimensionality-reduced priors, drug-side confidence, protein-side confidence, and drug-target pairwise confidence are calculated, and a pairwise prior description vector is constructed, including:

[0018] The drug sequence-level prior and protein sequence-level prior are mapped to a low-dimensional space by a low-dimensional projection matrix fitted by the training set, respectively, to obtain low-dimensional drug prior and low-dimensional protein prior.

[0019] Based on the positions of the low-dimensional drug prior and low-dimensional protein prior in the low-dimensional prior distribution of the training set, drug-side confidence, protein-side confidence, and drug-target pairwise confidence are calculated respectively. After mapping the low-dimensional drug prior and low-dimensional protein prior to the shared prior space, pairwise prior description vectors are constructed according to their combination terms, difference terms, element-wise product terms, and the drug-side confidence and protein-side confidence.

[0020] Furthermore, using the task-specific drug features and protein features as query conditions, attention-weighted aggregation is performed on the drug-side term meta-context embeddings and protein-side term meta-context embeddings, respectively, including:

[0021] After concatenating the task-specific drug features with the context embeddings of each drug-side lexical, the first learnable parameter matrix is ​​mapped and hyperbolic tangent activation is performed. Then, the attention score of each lexical is calculated through the first learnable weight vector. The attention score is then processed by a normalized exponential function along the drug lexical dimension to obtain the attention weight of each drug-side lexical. The attention weights are then used to perform a weighted summation of the drug-side lexical-level context embeddings to obtain the drug sequence-level prior.

[0022] After concatenating the task-specific protein features with the context embeddings of each lexical term on the protein side, the process is mapped by a second learnable parameter matrix and activated by hyperbolic tangent. Then, the attention score of each lexical term is calculated using a second learnable weight vector. The attention score is then processed by a normalized exponential function along the protein lexical dimension to obtain the attention weight of each lexical term on the protein side. Finally, the protein sequence-level prior is obtained by weighted summation of the protein lexical-level context embeddings using the attention weights.

[0023] Furthermore, the calculation of drug-side confidence includes:

[0024] Based on the low-dimensional projection matrix, the drug sequence-level prior is reconstructed by low-dimensional projection to obtain the drug-side reconstruction residual.

[0025] Calculate the Mahalanobis distance between the low-dimensional drug prior and the mean of the low-dimensional drug prior distribution in the training set;

[0026] The drug-side confidence level is obtained by multiplying the negative exponent of the drug-side reconstruction residual after adjustment with the first temperature coefficient and the negative exponent of the Mahalanobis distance after adjustment with the second temperature coefficient.

[0027] The protein-side confidence level is calculated in the same way from the protein-side reconstruction residual and the protein-side Mahalanobis distance.

[0028] The calculation of drug-target pairwise confidence includes:

[0029] The low-dimensional drug prior is mapped to the shared prior space through the first shared space mapping matrix, and the low-dimensional protein prior is mapped to the shared prior space through the second shared space mapping matrix.

[0030] The mapped drug prior, the mapped protein prior, the absolute difference between the two, the element-wise product of the two, the drug-side confidence and the protein-side confidence are concatenated to obtain a pairwise prior description vector.

[0031] Calculate the pairwise Mahalanobis distance between the pairwise prior description vectors and the mean of the distribution of the prior description vectors of the training ensemble;

[0032] The drug-target pairwise confidence is obtained by multiplying the drug-side confidence, the protein-side confidence, and the negative exponent of the paired Mahalanobis distance adjusted by the third temperature coefficient.

[0033] Furthermore, orthogonal residual candidates are generated through multiple low-rank residual experts, including:

[0034] K drug-side low-rank residual experts and K protein-side low-rank residual experts are set up. Each low-rank residual expert is composed of a first linear mapping matrix, a nonlinear activation function and a second linear mapping matrix in sequence. The first linear mapping matrix maps the dimensionality-reduced prior to the bottleneck dimension, and the second linear mapping matrix maps the bottleneck representation to the corresponding backbone feature dimension. The bottleneck dimension is smaller than the corresponding backbone feature dimension.

[0035] The dimensionality-reduced drug priors are input into K low-rank residual experts on the drug side to obtain K candidate quantities of drug-side prior residuals; the dimensionality-reduced protein priors are input into K low-rank residual experts on the protein side to obtain K candidate quantities of protein-side prior residuals.

[0036] For each drug-side prior residual candidate, calculate its inner product with the corresponding original backbone drug feature, and subtract the component parallel to the original backbone drug feature from the prior residual candidate based on the ratio of the inner product to the original backbone drug feature modulus to obtain the drug-side orthogonal residual candidate.

[0037] For each protein-side prior residual candidate, the component parallel to the corresponding original backbone protein feature is removed in the same way to obtain the protein-side orthogonal residual candidate.

[0038] Further, expert routing weights are generated based on the pairwise prior description vectors and the pairwise confidence levels. Then, the original backbone features are injected after weighted fusion of each orthogonal residual candidate based on the pairwise confidence levels and the expert routing weights. This includes:

[0039] The pairwise prior description vectors are concatenated with the pairwise confidence levels to obtain the routing input vector;

[0040] The routing input vector is mapped through a third learnable parameter matrix and then nonlinearly activated, followed by a fourth learnable parameter matrix mapping to obtain the K-dimensional original routing score.

[0041] Sparse probability mapping is performed on the K-dimensional original routing scores to make the output routing weights lie on the probability simplex with non-negative elements and a sum of one, where some elements can be zero, thus obtaining K-dimensional expert routing weights. The sparse probability mapping makes a single sample activate only some low-rank residual experts.

[0042] Based on the pairwise confidence and the expert routing weight, the original backbone features are injected after channel-level gated weighted fusion of each orthogonal residual candidate quantity.

[0043] Further, based on the enhanced drug representation and enhanced protein representation, the drug-target interaction probability or affinity value is output, including:

[0044] An interaction condition vector is generated based on the enhanced drug representation, enhanced protein representation, pairwise prior description vector, and pairwise confidence. The enhanced drug representation, enhanced protein representation, and interaction condition vector are then input into the prediction head or interaction module, and the drug-target interaction probability or affinity value is output.

[0045] Furthermore, the method also includes a counterfactual prior inhibition step during the training phase:

[0046] Constructing a path without prior knowledge: Retain the original backbone features but remove the prior condition vector to obtain a prediction result without prior knowledge;

[0047] Constructing mismatched prior paths: Replace the drug prior or protein prior of the current sample with the corresponding prior of other samples in the same training batch to obtain the mismatched prior prediction results;

[0048] Calculating the counterfactual prior suppression loss includes:

[0049] The first constraint term is to weight the square of the difference between the true prior path prediction result and the no prior path prediction result by the difference between the pairwise confidence of the current sample and 1, so that the true prior prediction of the low pairwise confidence sample approaches the no prior prediction.

[0050] The second constraint requires that the task loss of the true prior path be lower than the task loss of the mismatched prior path, and the difference be no less than a preset interval constant.

[0051] Furthermore, the method employs a joint loss during the training phase for model training, where the joint loss is a weighted sum of the following losses:

[0052] Task loss includes binary classification cross-entropy loss or mean squared error loss;

[0053] The low-confidence stability loss is weighted by the difference between the pairwise confidence of each sample and 1, which is used to weight the square of the magnitude of the difference between the enhanced representation and the original backbone feature representation, thus limiting the residual perturbation amplitude of low-confidence samples.

[0054] The backbone-prior redundancy removal loss is calculated from the squared Frobenius norm of the cross-correlation matrix between the original backbone feature matrix after batch standardization and the orthogonal residual matrix after batch standardization. It is used to reduce the linear correlation between backbone features and prior residual features.

[0055] The counterfactual prior suppresses loss;

[0056] The route sparsity loss, calculated from the entropy of the expert route weight vector for each sample, is used to encourage samples to choose more explicit expert combinations.

[0057] To achieve the above objectives, according to a second aspect of the present invention, a drug target prediction system based on trusted routing and orthogonal residual experts is provided, comprising:

[0058] The data input and preprocessing module is used to acquire drug input and target protein input, and further acquire corresponding labels during the training phase;

[0059] The original backbone feature extraction module is used to extract features from the drug input and target protein input by the original drug-target prediction backbone model to obtain task-specific drug features and protein features;

[0060] The drug pre-training prior module and the protein pre-training prior module are used to extract drug-side lexical-level context embeddings and protein-side lexical-level context embeddings from the frozen drug-side pre-trained sequence encoder and protein-side pre-trained sequence encoder, respectively.

[0061] The backbone conditionalized trust pooling module is used to perform attention-weighted aggregation on the word-level context embedding using the task-specific drug features and protein features as query conditions to obtain drug sequence-level priors and protein sequence-level priors.

[0062] The dimensionality reduction and credibility calibration module is used to reduce the dimensionality of the drug sequence-level prior and the protein sequence-level prior, calculate the drug-side credibility, protein-side credibility and pairwise credibility based on the dimensionality-reduced prior, and construct the pairwise prior description vector.

[0063] The low-rank orthogonal residual expert adaptation module is used to generate prior residual candidates with the same dimension as the corresponding backbone features based on the dimensionality reduction of multiple low-rank residual experts, and remove redundant components that are parallel to the corresponding original backbone features to obtain orthogonal residual candidates.

[0064] A sparse pairwise interactive routing module is used to generate expert routing weights based on the pairwise prior description vectors and pairwise confidence, and to perform reliable residual injection based on the pairwise confidence and expert routing weights to obtain enhanced drug representation and enhanced protein representation.

[0065] The prediction output module is used to output the drug-target interaction probability or affinity value based on the enhanced drug representation and enhanced protein representation.

[0066] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art:

[0067] This invention utilizes the synergistic effects of backbone conditional pooling, dimensionality reduction and reliable calibration, orthogonal residual expert adaptation, and sparse pairwise routing to inject the prior knowledge of bilateral pre-trained sequences into the original backbone model in a controlled manner as redundant orthogonal residuals after reliability evaluation. This achieves reliable coupling of prior knowledge while preserving the original backbone structure, reduces the negative transfer risk of low-reliability priors, and improves the accuracy and cross-backbone reusability of drug-target prediction.

[0068] Specifically, this invention controls the residual injection intensity by using both one-sided confidence and paired confidence, thereby automatically reducing the influence of external priors on low-confidence samples and making prior injection more reliable.

[0069] Specifically, this invention removes redundant components parallel to the backbone features by using orthogonal residual experts, so that the prior mainly supplements the information not covered by the backbone, thereby reducing feature fusion redundancy.

[0070] Specifically, this invention enhances the ability to express pairs of objects through the combined influence of the difference term, product term, and routing weight between drug priors and protein priors on prediction.

[0071] Specifically, this invention constrains low-confidence samples to approach no-priority paths by using a counterfactual prior suppression module, and requires that true priors are superior to mismatched priors, thereby reducing the risk of negative migration.

[0072] Specifically, the pre-trained encoder of the present invention can be frozen, the word-level context embedding, low-dimensional projection and distribution statistics can be pre-computed or fixed, the low-rank expert and routing module have a small number of parameters, and are suitable for different backbones such as CNN, GNN, Transformer, graph Transformer and multimodal gating models, with good cross-backbone reusability.

[0073] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0074] The accompanying drawings, as part of this invention, are provided to further illustrate the invention. The illustrative embodiments and descriptions of the invention are used to explain the invention, but do not constitute an undue limitation thereof. Clearly, the drawings described below are merely some embodiments; those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0075] In the attached diagram:

[0076] Figure 1 This is a schematic diagram of the overall process of the drug target prediction method based on trusted routing and orthogonal residual experts in this specific embodiment;

[0077] Figure 2 This is a schematic diagram of the framework for extracting the original backbone features in this specific embodiment;

[0078] Figure 3 This is a schematic diagram of the framework for paired trusted routing and orthogonal residual experts in this specific implementation.

[0079] Figure 4 This is a schematic diagram of the structure of the drug target prediction system based on trusted routing and orthogonal residual experts in this specific embodiment. Detailed Implementation

[0080] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments will be clearly and completely described below with reference to the accompanying drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0081] In the following description, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0082] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of the present invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0083] Please see Figure 1 This invention provides a drug target prediction method based on trusted routing and orthogonal residual experts, comprising the following steps:

[0084] S1. Obtain drug input and target protein input, and further obtain corresponding labels during the training phase;

[0085] S2. The original drug-target prediction backbone model is used to extract features from the drug input and target protein input to obtain task-specific drug features and protein features.

[0086] S3. The drug sequence input and protein sequence input are encoded by the frozen drug-side pre-trained sequence encoder and protein-side pre-trained sequence encoder respectively, and the drug-side word-level context embedding and protein-side word-level context embedding are extracted.

[0087] S4. Using the task-specific drug features and protein features as query conditions, attention-weighted aggregation is performed on the drug-side terminology-level context embedding and the protein-side terminology-level context embedding to obtain drug sequence-level priors and protein sequence-level priors.

[0088] S5. Dimensionally reduce the drug sequence-level prior and the protein sequence-level prior, calculate the drug-side confidence, protein-side confidence and drug-target pairwise confidence based on the dimensionality-reduced prior, and construct the pairwise prior description vector.

[0089] S6. Multiple low-rank residual experts generate prior residual candidates based on the dimensionality-reduced priors, and remove redundant components parallel to the corresponding original backbone features from each prior residual candidate to obtain orthogonal residual candidates.

[0090] S7. Generate expert routing weights based on the pairwise prior description vectors and the pairwise confidence. After weighted fusion of each orthogonal residual candidate based on the pairwise confidence and the expert routing weights, inject the original backbone features to obtain enhanced drug representation and enhanced protein representation.

[0091] S8. Based on the enhanced drug representation and enhanced protein representation, output the drug-target interaction probability or affinity value.

[0092] Data input and original backbone feature extraction

[0093] Obtain the drug set D, the target protein set T, and drug-target sample pairs. Each sample contains drug inputs. Protein input And the label y. For DTI classification tasks, y indicates whether an interaction exists, with a value of 0 or 1; for DTA regression tasks, y represents a continuous affinity value, such as binding affinity, inhibition constant, etc.

[0094] Let the original drug encoder be The original protein encoder is The original prediction head or interaction module is The original backbone path is calculated using the following formula:

[0095] (1a);

[0096] (1b);

[0097] (1c);

[0098] in, and These are the task-specific drug features and protein features learned from the original backbone model, respectively, both of which are fixed-dimensional real vectors; This is the original backbone prediction result without the injection of external priors. This embodiment maintains... , and The main structure remains unchanged, that is, the number of network layers, connection methods and parameter initialization strategies within the original backbone model are not modified.

[0099] The drug inputs to the original backbone model include SMILES (Simplified Molecular Input Line Entry System) strings, molecular graphs, molecular fingerprints, or combinations thereof; protein inputs include amino acid sequences or protein features derived from amino acid sequences. The original backbone model is a DTI classification model or a DTA regression model that includes a drug encoder, a protein encoder, explicit interaction modules, or a prediction head.

[0100] Prior extraction of pre-trained sequences

[0101] A frozen pre-trained sequence prior branch is constructed outside the original backbone model. A pre-trained molecular sequence encoder is used on the drug side. SMILES sequence of the drug Encoding is performed; a pre-trained protein sequence encoder is used on the protein side. protein amino acid sequence Encode the code using the following formula:

[0102] (2a);

[0103] (2b);

[0104] in, and These are the word-level context embedding matrices for the drug side and the protein side, respectively; and These are the sequence lengths of the drug and the protein, respectively. and These represent the hidden dimensions of the pre-trained encoders on the drug side and the protein side, respectively. The drug-side pre-trained sequence encoder can use Transformer models pre-trained based on molecular sequences, such as SMILES-BERT and ChemBERTa; the protein-side pre-trained sequence encoder can use Transformer models pre-trained based on amino acid sequences, such as ProtBERT and ESM-2. Both pre-trained encoders maintain parameter freeze during training and inference and do not participate in gradient updates.

[0105] Backbone Conditional Trust Pooling

[0106] Please see Figure 2 Unlike fixed mean pooling or CLS pooling, this embodiment uses the original backbone features as query conditions to calculate sample-related attention weights for word-level pre-trained embeddings.

[0107] Drug side pooling is calculated using the following formula:

[0108] (3a);

[0109] (3b);

[0110] Protein side pooling is calculated using the following formula:

[0111] (4a);

[0112] (4b);

[0113] in, and These are the attention weights for the i-th word on the drug side and the j-th word on the protein side, respectively, with values ​​ranging from [0,1] and the sum of the weights on each side being 1; , These represent the context embeddings of the i-th word on the drug side and the j-th word on the protein side, respectively. The first learnable parameter matrix, The second learnable parameter matrix is ​​used to perform linear mapping on the word embeddings of the drug side and the protein side, respectively. , It is a linear mapping matrix of the backbone features; This is the first learnable weight vector. This is the second learnable weight vector, used to calculate the attention scores of each word on the drug side and the protein side, respectively. , is the bias vector; tanh is the hyperbolic tangent activation function; the subscripts i and j of the softmax function represent normalization along the drug lexicon dimension and the protein lexicon dimension, respectively; and These are drug sequence-level priors and protein sequence-level priors, respectively. This design allows the aggregation of pre-trained priors to be correlated with current backbone features, highlighting lexical segments that are more useful to the current drug-target sample, rather than applying the same fixed pooling strategy to all samples.

[0114] In some embodiments, the drug-side pre-trained sequence encoder and the protein-side pre-trained sequence encoder are kept frozen, and their output lexical-level context embeddings can be pre-computed and cached in a storage medium, and can be directly read during inference without repeated encoding, thereby reducing inference latency.

[0115] Low-dimensional projection and one-sided confidence estimation

[0116] To reduce the risk of dimensionality mismatch and redundancy between high-dimensional pre-trained representations and backbone features, this invention fits principal component analysis (PCA) or other low-dimensional projections without supervision labels onto the training set.

[0117] Taking the drug side as an example, low-dimensional projection and reconstruction are calculated according to the following formula:

[0118] (5a);

[0119] (5b);

[0120] in, The PCA center vector of the training set; To preserve the principal component matrix, its column vectors are the directions of the first few principal components; This serves as a priori for drug-side sequence-level pre-training before dimensionality reduction; A priori for low-dimensional drugs; The high-dimensional approximate representation of the drug side is obtained by reconstructing the low-dimensional drug prior. These two representations correspond to the original representation before dimensionality reduction and the reconstructed representation after dimensionality reduction, respectively. The low-dimensional projection matrix and PCA center of each data partition (such as each fold in five-fold cross-validation) are fitted only by the training set of that data partition and do not involve the validation or test set data.

[0121] The normalized reconstruction residuals and the low-dimensional Mahalanobis distance are calculated using the following formulas:

[0122] (6a);

[0123] (6b);

[0124] in, To normalize the reconstructed residuals, measure the proportion of information about the current drug prior outside the PCA subspace; tr(·) is a diagonal matrix consisting of the set of residual eigenvalues ​​corresponding to the principal components not retained by PCA, and tr(·) represents the trace of the matrix; and These are the mean vector and covariance matrix of the low-dimensional drug priors in the training set, respectively; To prevent extremely small positive numbers with denominators of zero or singular matrix, It is the identity matrix; This is a low-dimensional Mahalanobis distance, which measures the standardized distance between the current low-dimensional prior and the prior distribution center of the training set.

[0125] The prior confidence level of the drug is calculated using the following formula:

[0126] (7);

[0127] in, The first temperature coefficient, The second temperature coefficient is a pre-set normal value, which controls the sensitivity of the confidence level to the reconstruction residual and the Mahalanobis distance, respectively. This represents the prior confidence level on the drug side, with a value range of (0, 1). The smaller the value, the more accurate the reconstruction. A smaller value indicates that the current prior is closer to the center of the training set distribution, and a higher negative exponent of the product indicates higher confidence. The protein side is obtained in the same way. , , and , respectively, are low-dimensional protein prior, protein-side normalized reconstruction residual, protein-side low-dimensional Mahalanobis distance, and protein-side prior confidence.

[0128] Dual-scale pairwise reliable calibration

[0129] Please see Figure 2 One-sided confidence can only determine whether the prior on the drug side or the protein side is reliable on its own, but it cannot determine whether the two, as a drug-target sample pair, are in a reliable paired region. Therefore, this invention further maps the drug prior and the protein prior to a shared space and constructs paired prior description vectors.

[0130] The shared space mapping and the construction of pairwise prior description vectors are calculated according to the following formula:

[0131] (8a);

[0132] (8b);

[0133] (8c);

[0134] in, This is the first shared space mapping matrix. The second shared space mapping matrix is ​​used to map the low-dimensional priors on both sides to the shared prior space, and its dimension is such that... and Dimensions are consistent, therefore the differences are... (Absolute value of each element) and element-wise product terms The calculation of (element-wise multiplication) is legal and has the same dimension; [·;·] represents vector concatenation operation; For paired prior description vectors, their dimension is 4 times the dimension of the shared prior space plus 2.

[0135] Pairwise distribution distance and pairwise confidence are calculated using the following formulas:

[0136] (9a);

[0137] (9b);

[0138] in, and These are the mean vector and covariance matrix of the training ensemble for the prior description vector, respectively, which are fitted only by the training set; For paired Mahalanobis distances; The third temperature coefficient controls the sensitivity of the confidence level to the pairwise Mahalanobis distance; The pairwise confidence level is influenced by the drug-side confidence level, the protein-side confidence level, and the location of the pairwise prior distribution. It is used for subsequent residual injection and routing strength control. High pairwise confidence level indicates that both priors of the drug-target sample pair are reliable and the pairwise relationship is located in a common region of the training set.

[0139] In some embodiments, the residual injection intensity is determined by drug-side confidence. Protein side credibility Pairwise credibility The joint regulation automatically reduces the residual injection intensity of low-confidence drug inputs, low-confidence target protein inputs, or low-pair confidence samples, and makes the corresponding predictions mainly rely on the original backbone features, thereby avoiding interference from low-quality priors.

[0140] Low-rank orthogonal residual expert adaptation

[0141] Please see Figure 2 After obtaining low-dimensional priors and pairwise confidence, this invention does not directly splice... and Instead, it generates prior residual candidates with the same dimension as the backbone features by using K low-rank residual experts.

[0142] The k-th drug-side low-rank residual expert and the protein-side low-rank residual expert are calculated according to the following formula (10):

[0143] (10a);

[0144] (10b);

[0145] in, Let the first linear mapping matrix of the k-th expert on the drug side be the low-dimensional prior. Mapping to the bottleneck dimension ; Let be the second linear mapping matrix of the k-th expert on the drug side, which maps the bottleneck representation to the drug backbone feature dimension; and This is the protein-side correspondence matrix; For nonlinear activation functions, ReLU, GELU, or SiLU can be used; and These are the k-th prior residual candidates on the drug side and the protein side, respectively, and their dimensions are consistent with the corresponding backbone feature dimensions. Bottleneck dimension The parameters are smaller than the corresponding core feature dimensions, so the number of parameters for each expert is much smaller than the number of parameters for a complete linear mapping, and the total number of parameters for K experts is controlled within an acceptable range.

[0146] To avoid the prior residuals repeatedly expressing the existing information of the original backbone, this invention performs sample-level orthogonalization on each residual candidate quantity, and calculates it according to the following formula (11):

[0147] (11a);

[0148] (11b);

[0149] in, and These are the candidate orthogonal residuals for the drug-side and protein-side after removing the parallel components of the backbone; ) represents the original backbone drug feature vector. With residual candidate quantity The inner product, for The squared L2 norm of the L2 norm. This orthogonal projection operation will include the candidate residuals with... Subtracting components with parallel directions ensures that the retained components primarily contain supplementary information orthogonal to the backbone features. If a candidate prior residual is primarily related to... or If the parallel component is highly redundant, it will be significantly weakened, thus making the injected information more inclined to supplement the original backbone representation rather than cover it.

[0150] Sparse Pairwise Interaction Routing and Trusted Residual Injection

[0151] Please see Figure 3 This invention concatenates paired prior description vectors with paired confidence levels to form routing inputs, and generates K-dimensional routing weights accordingly, which are used to dynamically select or weight low-rank orthogonal residual experts.

[0152] Route weights are calculated using the following formula:

[0153] (12a);

[0154] (12b);

[0155] (12c);

[0156] in, The routing input vector is composed of paired prior description vectors. Paired credibility obtained by piecing together; For the third learnable parameter matrix, This is the fourth learnable parameter matrix; and It is the bias vector; It is a non-linear activation function; It is a sparse probability mapping function that maps any real vector to a probability simplex, and the output may contain zero elements; Let K be the K-dimensional expert routing weight vector for the current drug-target sample, where each element is non-negative and the sum is 1. Due to the use of sparse mapping, a single sample typically activates only a small number of residual experts, and the routing weights corresponding to the inactive experts are zero, thereby improving structural selectivity and reducing irrelevant prior interference.

[0157] Based on routing weights and pairwise trust levels, bilateral trusted residual injection is calculated using the following formula:

[0158] Channel-level gating vector generation:

[0159] (13a);

[0160] (13b);

[0161] Enhanced drug representation:

[0162] (14);

[0163] Enhanced protein representation:

[0164] (15);

[0165] in, and These are the channel-level gating vectors for the k-th experts on the drug side and the protein side, respectively. They are concatenated from the corresponding backbone features, orthogonal residual candidates, and pairwise confidence scores, and then linearly mapped and processed by the Sigmoid function. Each element has a value range of (0, 1), which is used to selectively scale the components of the residual candidate quantity in the channel dimension. For element-wise multiplication; LN is for level normalization; and These are enhanced drug representation and protein representation, respectively. When When the value is low, the entire residual injection term is proportionally reduced; when When the routing weights are high and concentrated on a few experts, the model can select the residual expert combination that best matches the current sample for prior injection.

[0166] Paired interaction condition generation and prediction output

[0167] In order to influence the interactive prediction of drug-side priors and protein-side priors on a sample pair basis, this invention constructs a pair of interactive condition vectors in front of the prediction head.

[0168] The interaction conditions are generated according to the following formulas (16) to (18):

[0169] Shared interactive space mapping:

[0170] (16a);

[0171] (16b);

[0172] Paired interaction condition vectors:

[0173] (17);

[0174] Final prediction input and prediction output:

[0175] (18a);

[0176] (18b);

[0177] (18c);

[0178] in, and To map the enhanced bilateral backbone features to a linear mapping matrix in the shared interaction space, such that... and Dimensional consistency, thus element-wise product terms and differences Dimensional validity; For learnable parameter matrix, It is the bias vector; These are paired interaction condition vectors; For learnable weight vectors, It is a biased scalar; The interaction condition strength scalar obtained by the Sigmoid function has a value range of (0, 1). The final prediction input vector is composed of an enhanced drug representation, an enhanced protein representation, and an interaction condition vector scaled by conditional intensity. This is the final predicted value; The final prediction head is obtained by adapting the input dimensions based on the original prediction head or interactive module, and its task output format is consistent with the original prediction head.

[0179] For the DTI binary classification task, Converted to interaction probabilities using the Sigmoid or Softmax function; for DTA regression tasks, These are predictions of continuous affinity (such as binding affinity, inhibition constant, etc.).

[0180] In some embodiments, the reliable residual injection is positioned at the encoder-level fusion location, immediately following the outputs of the drug encoder and protein encoder; the sparse pairwise interaction routing module is positioned at the interaction level or before the prediction head. When the original backbone model has cross-attention, bilinear matching, gated fusion, or graph Transformer interaction structures, the interaction condition vector... It can be used as a global conditional lexical, auxiliary lexical, or gating condition input into the interaction structure to enhance the original interaction module's ability to perceive prior information.

[0181] Model training phase

[0182] Please see Figure 3 The following describes a specific implementation of the present invention during the training phase.

[0183] Counterfactual prior suppression and joint training

[0184] To further reduce negative transfer caused by low-credibility priors and mismatched priors, this invention constructs two types of counterfactual inputs: no prior input. and mismatched prior input No prior input preserves the original backbone features but removes prior conditions (i.e., the interaction condition vector is set to zero or a prediction generated only by backbone features is used). Mismatched prior input replaces the current drug or protein prior with the corresponding prior of other samples in the same training batch.

[0185] The counterfactual prior suppression loss is calculated according to the following formula (19):

[0186] (19);

[0187] Where B is the training batch size; Let be the pairwise confidence level of the i-th sample; , and These are the prediction results for the true prior path, the path without prior knowledge, and the mismatched prior path, respectively. For classification or regression tasks, the loss function is the binary cross-entropy for classification tasks and the mean squared error for regression tasks. The preset counterfactual interval is a normal number (e.g., 0.1 or 0.2). is the non-negative weight coefficient; max(0, ·) is the hinge function.

[0188] The first constraint makes the true prior predictions for low-pair confidence samples approach the predictions without priors. That is, when the prior is unreliable, the model automatically degenerates into a prediction relying solely on the original backbone features, thus suppressing interference from unreliable priors. The second constraint requires that the true prior input has a lower task loss than the mismatched prior input (and the difference is not less than the margin). This ensures that the prior information injected by the correct drug-target pairing is superior to the prior information of random mismatches.

[0189] The overall training loss is calculated according to the following formula (20):

[0190] (20);

[0191] The definitions of each loss are as follows:

[0192] (1) Mission loss : This can be the cross-entropy loss for binary classification, the mean squared error loss, or the task loss function built into the original backbone model, depending on the specific task type.

[0193] (2) Low reliability stability loss Calculate according to the following formula (21):

[0194] (twenty one);

[0195] The loss is weighted by the difference between the pairwise confidence of each sample and 1, which makes the residual perturbation amplitude of low confidence samples more constrained and prevents unreliable priors from causing excessive perturbation to the backbone features.

[0196] (3) Backbone-Prior Redundancy Removal Loss Calculate according to the following formulas (22) and (23):

[0197] (22a);

[0198] (22b);

[0199] (23a);

[0200] (24b);

[0201] in, and These are the original backbone drug feature matrix and protein feature matrix after being centered and standardized within the batch, respectively. Each row is the backbone feature vector of a sample minus the within-batch mean and then divided by the within-batch standard deviation. and These are the orthogonal residual matrices of the drug side and protein side after in-batch centering and standardization (the weighted sum of all expert orthogonal residuals). and This is the cross-correlation matrix between backbone features and prior orthogonal residuals; Let Frobenius be the square of the matrix's norm. By minimizing this loss, the linear correlation between backbone features and prior residual features can be reduced, allowing the prior residuals to primarily provide supplementary information not captured by the original backbone rather than repeatedly expressing existing information.

[0202] (4) Counterfactual prior suppression of loss As defined in formula (19) above.

[0203] (5) Routing sparsity loss , which is the negative entropy regularization term defined in the second part of formula (23), is used to encourage the selection of more specific expert combinations for samples and avoid the model from degenerating to using only a single expert.

[0204] The loss is the intra-batch mean of the negative entropy of the routing weight vector for each sample. Minimizing the negative entropy is equivalent to maximizing the entropy of the routing weights, encouraging that the routing weights of the samples are not too concentrated and preventing the model from degenerating into using only a single expert. In the early stages of training, this loss helps each expert obtain a balanced training signal; in the later stages of training, combined with the sparsity property of sparsemax, the samples can naturally form a preference for a few experts.

[0205] The above loss weights , , and These are non-negative hyperparameters, which can be determined on the validation set through grid search or Bayesian optimization. Typical values ​​are: α ∈ [0.01, 0.5], β ∈ [0.001, 0.1], χ ∈ [0.01, 1.0], δ ∈ [0.0001, 0.01].

[0206] During training, both the drug-side and protein-side pre-trained sequence encoders maintain their parameters frozen and do not participate in gradient backpropagation. The low-dimensional projection matrix, PCA centers, low-dimensional prior distribution statistics, and pairwise prior distribution statistics for each data partition (e.g., each fold in five-fold cross-validation) are fitted only to the training set of that partition, without involving the validation or test set data. During training, the word-level context embeddings output by the frozen pre-trained encoders can be pre-computed and cached, avoiding repetitive encoding in each training epoch and thus accelerating training.

[0207] Experimental comparison and verification

[0208] To verify the effectiveness of the method of the present invention, three publicly available benchmark datasets, Davis, KIBA and Human, were used to conduct DTI classification and DTA regression experiments according to the backbone model and dataset combinations listed in Table 1 below.

[0209] Table 1: Experimental Dataset Size and Task Purpose

[0210]

[0211] GraphormerDTI, EviDTI, MAARDTI, and MGNDTI are used to report DTI classification results, while GK-BertDTA is used to report DTA regression results on the Davis and KIBA datasets.

[0212] All results were based on five-fold cross-validation (n=5) and are expressed as mean ± standard deviation of the five-fold results. AUROC, AUPR, Accuracy, F1-score, MCC, and CI are all expressed as percentages, with higher values ​​indicating better performance; MSE is expressed as the original regression error, with lower values ​​indicating better performance. Accuracy, F1-score, and MCC are threshold-dependent classification metrics.

[0213] The drug-side pre-trained encoder used SMILES-BERT, and the protein-side pre-trained encoder used ProtBERT. Both pre-trained encoders were frozen during training. The PCA projection of each fold was fitted only using the training set for that fold before being applied to the corresponding validation and test sets. The PCA output dimensions for the drug-side and protein-side models were consistent with the corresponding backbone representation dimensions. The training hardware was an NVIDIA GeForce RTX 4090 GPU. The optimizer, learning rate, batch size, and number of training epochs for each backbone model are shown in Table 2 below.

[0214] Table 2: Main Training Configurations for Each Backbone Model and Dataset Combination

[0215]

[0216] In Tables 3 and 4 below, the "Method of this Invention (PlugDTI)" simultaneously enables the frozen bilateral pre-trained sequence encoder, backbone conditionalized trusted pooling, low-dimensional projection, dual-scale prior trusted calibration, low-rank orthogonal residual expert, sparse pairwise routing, trusted residual injection, interactive condition generation, counterfactual prior suppression, and joint training constraints; "Original Backbone Model (Baseline)" refers to the backbone model without enabling the above prior coupling structures. All comparisons employ the same data preprocessing, cross-validation partitioning, and evaluation process, and the low-dimensional projection of each fold is fitted only by the training set of that fold.

[0217] Table 3: Comparison of classification and regression performance between the original backbone model and the method of this invention (five-fold cross-validation)

[0218]

[0219] Table 4: Threshold-dependent classification indices of the original backbone model and the method of this invention (five-fold cross-validation)

[0220]

[0221] As shown in Tables 3 and 4, the method of this invention improves AUROC, AUPR, and threshold-dependent metrics (Accuracy and F1-score) on multiple classification backbone models. In the DTA regression task, the complete method reduces MSE and improves CI on the Davis and KIBA datasets. These results demonstrate that, under the same five-fold cross-validation, data preprocessing, and training configuration, the method of this invention can effectively improve the predictive performance of drug-target interaction classification and drug-target affinity regression.

[0222] The method of this invention can be used in various application scenarios such as drug relocation, candidate target discovery, virtual screening of compounds, affinity ranking, cold-start drug-protein association discovery, and cross-dataset drug-target prediction migration.

[0223] Please see Figure 4 Based on the same inventive concept, this invention also provides a drug target prediction system based on trusted routing and orthogonal residual experts, comprising:

[0224] The data input and preprocessing module is used to acquire drug input and target protein input, and further acquire corresponding labels during the training phase;

[0225] The original backbone feature extraction module is used to extract features from the drug input and target protein input by the original drug-target prediction backbone model to obtain task-specific drug features and protein features;

[0226] Drug pre-training prior branches and protein pre-training prior branches are used to extract drug-side lexical-level context embeddings and protein-side lexical-level context embeddings from the frozen drug-side pre-trained sequence encoder and protein-side pre-trained sequence encoder, respectively.

[0227] The backbone conditionalized trust pooling module is used to perform attention-weighted aggregation on the word-level context embedding using the task-specific drug features and protein features as query conditions to obtain drug sequence-level priors and protein sequence-level priors.

[0228] The dimensionality reduction and credibility calibration module is used to reduce the dimensionality of the drug sequence-level prior and the protein sequence-level prior, calculate the drug-side credibility, protein-side credibility and pairwise credibility based on the dimensionality-reduced prior, and construct the pairwise prior description vector.

[0229] The low-rank orthogonal residual expert adaptation module is used to generate prior residual candidates with the same dimension as the corresponding backbone features based on the dimensionality reduction of multiple low-rank residual experts, and remove redundant components that are parallel to the corresponding original backbone features to obtain orthogonal residual candidates.

[0230] A sparse pairwise interactive routing module is used to generate expert routing weights based on the pairwise prior description vectors and pairwise confidence, and to perform reliable residual injection based on the pairwise confidence and expert routing weights to obtain enhanced drug representation and enhanced protein representation.

[0231] The prediction output module is used to output the drug-target interaction probability or affinity value based on the enhanced drug representation and enhanced protein representation.

[0232] Unless otherwise defined, the technical or scientific terms used in this invention shall have the general meaning understood by one of ordinary skill in the art to which this invention pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this invention do not indicate quantitative limitations and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this invention are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “A plurality” used in this invention refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. In this invention, the terms "first," "second," "third," etc., are used only to distinguish similar objects and do not represent a specific order of objects.

[0233] Obviously, the accompanying drawings are merely some examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, certain design, manufacturing, or production modifications made according to the technical content disclosed in the present invention are merely conventional technical means and should not be considered as insufficient disclosure of the present invention.

[0234] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-described technical content to create equivalent embodiments without departing from the scope of the present invention. The implementation schemes in the above embodiments can also be further combined or replaced. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A drug target prediction method based on trusted routing and orthogonal residual experts, characterized in that, Includes the following steps: S1. Obtain drug input and target protein input, and further obtain corresponding labels during the training phase; S2. The original drug-target prediction backbone model is used to extract features from the drug input and target protein input to obtain task-specific drug features and protein features. S3. The drug sequence input and protein sequence input are encoded by the frozen drug-side pre-trained sequence encoder and protein-side pre-trained sequence encoder respectively, and the drug-side word-level context embedding and protein-side word-level context embedding are extracted. S4. Using the task-specific drug features and protein features as query conditions, attention-weighted aggregation is performed on the drug-side terminology-level context embedding and the protein-side terminology-level context embedding to obtain drug sequence-level priors and protein sequence-level priors. S5. Dimensionally reduce the drug sequence-level prior and the protein sequence-level prior, calculate the drug-side confidence, protein-side confidence and drug-target pairwise confidence based on the dimensionality-reduced prior, and construct the pairwise prior description vector. S6. Multiple low-rank residual experts generate prior residual candidates based on the dimensionality-reduced priors, and remove redundant components parallel to the corresponding original backbone features from each prior residual candidate to obtain orthogonal residual candidates. S7. Generate expert routing weights based on the pairwise prior description vectors and the pairwise confidence. After weighted fusion of each orthogonal residual candidate based on the pairwise confidence and the expert routing weights, inject the original backbone features to obtain enhanced drug representation and enhanced protein representation. S8. Based on the enhanced drug representation and enhanced protein representation, output the drug-target interaction probability or affinity value.

2. The method according to claim 1, characterized in that, The drug sequence-level and protein sequence-level priors are dimensionality-reduced. Based on the dimensionality-reduced priors, drug-side confidence, protein-side confidence, and drug-target pairwise confidence are calculated, and a pairwise prior description vector is constructed, including: The drug sequence-level prior and protein sequence-level prior are mapped to a low-dimensional space by a low-dimensional projection matrix fitted by the training set, respectively, to obtain low-dimensional drug prior and low-dimensional protein prior. Based on the positions of the low-dimensional drug prior and low-dimensional protein prior in the low-dimensional prior distribution of the training set, drug-side confidence, protein-side confidence, and drug-target pairwise confidence are calculated respectively. After mapping the low-dimensional drug prior and low-dimensional protein prior to the shared prior space, pairwise prior description vectors are constructed according to their combination terms, difference terms, element-wise product terms, and the drug-side confidence and protein-side confidence.

3. The method according to claim 1, characterized in that, Using the task-specific drug features and protein features as query conditions, attention-weighted aggregation is performed on the drug-side term meta-context embeddings and protein-side term meta-context embeddings, respectively, including: After concatenating the task-specific drug features with the context embeddings of each drug-side lexical, the first learnable parameter matrix is ​​mapped and hyperbolic tangent activation is performed. Then, the attention score of each lexical is calculated through the first learnable weight vector. The attention score is then processed by a normalized exponential function along the drug lexical dimension to obtain the attention weight of each drug-side lexical. The attention weights are then used to perform a weighted summation of the drug-side lexical-level context embeddings to obtain the drug sequence-level prior. After concatenating the task-specific protein features with the context embeddings of each lexical term on the protein side, the process is mapped by a second learnable parameter matrix and activated by hyperbolic tangent. Then, the attention score of each lexical term is calculated using a second learnable weight vector. The attention score is then processed by a normalized exponential function along the protein lexical dimension to obtain the attention weight of each lexical term on the protein side. Finally, the protein sequence-level prior is obtained by weighted summation of the protein lexical-level context embeddings using the attention weights.

4. The method according to claim 2, characterized in that, The calculation of drug-side confidence includes: Based on the low-dimensional projection matrix, the drug sequence-level prior is reconstructed by low-dimensional projection to obtain the drug-side reconstruction residual. Calculate the Mahalanobis distance between the low-dimensional drug prior and the mean of the low-dimensional drug prior distribution in the training set; The drug-side confidence level is obtained by multiplying the negative exponent of the drug-side reconstruction residual after adjustment with the first temperature coefficient and the negative exponent of the Mahalanobis distance after adjustment with the second temperature coefficient. The protein-side confidence level is calculated in the same way from the protein-side reconstruction residual and the protein-side Mahalanobis distance. The calculation of drug-target pairwise confidence includes: The low-dimensional drug prior is mapped to the shared prior space through the first shared space mapping matrix, and the low-dimensional protein prior is mapped to the shared prior space through the second shared space mapping matrix. The mapped drug prior, the mapped protein prior, the absolute difference between the two, the element-wise product of the two, the drug-side confidence and the protein-side confidence are concatenated to obtain a pairwise prior description vector. Calculate the pairwise Mahalanobis distance between the pairwise prior description vectors and the mean of the distribution of the prior description vectors of the training ensemble; The drug-target pairwise confidence is obtained by multiplying the drug-side confidence, the protein-side confidence, and the negative exponent of the paired Mahalanobis distance adjusted by the third temperature coefficient.

5. The method according to claim 1, characterized in that, Orthogonal residual candidates are generated through multiple low-rank residual experts, including: K drug-side low-rank residual experts and K protein-side low-rank residual experts are set up. Each low-rank residual expert is composed of a first linear mapping matrix, a nonlinear activation function and a second linear mapping matrix in sequence. The first linear mapping matrix maps the dimensionality-reduced prior to the bottleneck dimension, and the second linear mapping matrix maps the bottleneck representation to the corresponding backbone feature dimension. The bottleneck dimension is smaller than the corresponding backbone feature dimension. The dimensionality-reduced drug priors are input into K low-rank residual experts on the drug side to obtain K candidate quantities of drug-side prior residuals; the dimensionality-reduced protein priors are input into K low-rank residual experts on the protein side to obtain K candidate quantities of protein-side prior residuals. For each drug-side prior residual candidate, calculate its inner product with the corresponding original backbone drug feature, and subtract the component parallel to the original backbone drug feature from the prior residual candidate based on the ratio of the inner product to the original backbone drug feature modulus to obtain the drug-side orthogonal residual candidate. For each protein-side prior residual candidate, the component parallel to the corresponding original backbone protein feature is removed in the same way to obtain the protein-side orthogonal residual candidate.

6. The method according to claim 1, characterized in that, Expert routing weights are generated based on the pairwise prior description vectors and the pairwise confidence levels. After weighted fusion of each orthogonal residual candidate variable based on the pairwise confidence levels and the expert routing weights, the original backbone features are injected, including: The pairwise prior description vectors are concatenated with the pairwise confidence levels to obtain the routing input vector; The routing input vector is mapped through a third learnable parameter matrix and then nonlinearly activated, followed by a fourth learnable parameter matrix mapping to obtain the K-dimensional original routing score. Sparse probability mapping is performed on the K-dimensional original routing scores to make the output routing weights lie on the probability simplex with non-negative elements and a sum of one, where some elements can be zero, thus obtaining K-dimensional expert routing weights. The sparse probability mapping makes a single sample activate only some low-rank residual experts. Based on the pairwise confidence and the expert routing weight, the original backbone features are injected after channel-level gated weighted fusion of each orthogonal residual candidate.

7. The method according to claim 1, characterized in that, Based on the enhanced drug representation and enhanced protein representation, output drug-target interaction probability or affinity values, including: An interaction condition vector is generated based on the enhanced drug representation, enhanced protein representation, pairwise prior description vector, and pairwise confidence. The enhanced drug representation, enhanced protein representation, and interaction condition vector are then input into the prediction head or interaction module, and the drug-target interaction probability or affinity value is output.

8. The method according to claim 1, characterized in that, The method also includes a counterfactual prior inhibition step during the training phase: Constructing a path without prior knowledge: Retain the original backbone features but remove the prior condition vector to obtain a prediction result without prior knowledge; Constructing mismatched prior paths: Replace the drug prior or protein prior of the current sample with the corresponding prior of other samples in the same training batch to obtain the mismatched prior prediction results; Calculating the counterfactual prior suppression loss includes: The first constraint term is to weight the square of the difference between the true prior path prediction result and the no prior path prediction result by the difference between the pairwise confidence of the current sample and 1, so that the true prior prediction of the low pairwise confidence sample approaches the no prior prediction. The second constraint requires that the task loss of the true prior path be lower than the task loss of the mismatched prior path, and the difference be no less than a preset interval constant.

9. The method according to claim 1, characterized in that, The method employs a joint loss during the training phase for model training, where the joint loss is a weighted sum of the following losses: Task loss includes binary classification cross-entropy loss or mean squared error loss; The low-confidence stability loss is weighted by the difference between the pairwise confidence of each sample and 1, which is used to weight the square of the magnitude of the difference between the enhanced representation and the original backbone feature representation, thus limiting the residual perturbation amplitude of low-confidence samples. The backbone-prior redundancy removal loss is calculated from the squared Frobenius norm of the cross-correlation matrix between the original backbone feature matrix after batch standardization and the orthogonal residual matrix after batch standardization. It is used to reduce the linear correlation between backbone features and prior residual features. The counterfactual prior suppresses loss; The route sparsity loss, calculated from the entropy of the expert route weight vector for each sample, is used to encourage samples to choose more explicit expert combinations.

10. A drug target prediction system based on trusted routing and orthogonal residual experts, characterized in that, include: The data input and preprocessing module is used to acquire drug input and target protein input, and further acquire corresponding labels during the training phase; The original backbone feature extraction module is used to extract features from the drug input and target protein input by the original drug-target prediction backbone model to obtain task-specific drug features and protein features; The drug pre-training prior module and the protein pre-training prior module are used to extract drug-side lexical-level context embeddings and protein-side lexical-level context embeddings from the frozen drug-side pre-trained sequence encoder and protein-side pre-trained sequence encoder, respectively. The backbone conditionalized trust pooling module is used to perform attention-weighted aggregation on the word-level context embedding using the task-specific drug features and protein features as query conditions to obtain drug sequence-level priors and protein sequence-level priors. The dimensionality reduction and credibility calibration module is used to reduce the dimensionality of the drug sequence-level prior and the protein sequence-level prior, calculate the drug-side credibility, protein-side credibility and pairwise credibility based on the dimensionality-reduced prior, and construct the pairwise prior description vector. The low-rank orthogonal residual expert adaptation module is used to generate prior residual candidates with the same dimension as the corresponding backbone features based on the dimensionality reduction of multiple low-rank residual experts, and remove redundant components that are parallel to the corresponding original backbone features to obtain orthogonal residual candidates. A sparse pairwise interactive routing module is used to generate expert routing weights based on the pairwise prior description vectors and pairwise confidence, and to perform reliable residual injection based on the pairwise confidence and expert routing weights to obtain enhanced drug representation and enhanced protein representation. The prediction output module is used to output the drug-target interaction probability or affinity value based on the enhanced drug representation and enhanced protein representation.

Images