Protein mutation drug resistance prediction method and system based on multi-task deep learning

Abstract

The invention provides a protein mutation drug resistance prediction method and system based on multi-task deep learning, and the method comprises the steps: obtaining evidence entries from a CIViC / PharmGKB library, carrying out the preprocessing, and supplementing pathogenicity tags to build a double-tag corpus; the protein is characterized by using a mutation site neighborhood local one-bit code and an evolutionary comparison site specific scoring matrix PSSM dual path, the length is standardized, and the drug is mapped and spliced by using a PubChem881-dimensional fingerprint; constructing a Transform cascade model, firstly transfusing the pathogenicity probability, then splicing with the intermediate characterization to transfuse the drug resistance probability, and carrying out combined loss training; and dividing data set evaluation, and analyzing multiple mutations in combination with attention. The performance of the method is superior to or not lower than that of an existing tool, a high-resolution three-dimensional structure is not needed, accurate prediction can be achieved only through a small amount of input features, meanwhile, whether drug reaction is driven by mutation pathogenicity or not can be determined, differentiated medication strategies are formulated for different types of patients, and the method has good cross-center migration potential and is suitable for clinical application. The method can be widely applied to scenes such as accurate medication decision making, drug resistance early warning and drug redesign.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics, in particular to a protein mutation drug resistance prediction method and system based on multi-task deep learning. Especially, it relates to a multi-task deep learning method based on pathogenic signal guidance, which is used for accurate and interpretable prediction and attribution of protein mutation driven drug resistance, and is suitable for application scenarios such as precise medication decision, drug research and development, and variation interpretation in clinical genomics. BACKGROUND

[0002] Protein mutation driven drug resistance has become one of the core challenges faced by global public health and drug research and development, and widely appears in invasive tumors, severe bacterial infections and rapidly spreading viral diseases, continuously eroding the clinical efficacy of existing drugs and leading to treatment failure, recurrence and increased risk of death. Traditional understanding emphasizes that amino acid substitution occurring at the direct drug target is the main source of drug resistance, but a large amount of evidence shows that mutations located outside the drug-target system (such as proteins involved in drug metabolism, transmembrane transport and downstream signaling pathways) can also significantly affect the efficacy by changing drug exposure, adjusting network dynamics or triggering compensatory bypass. In the real world, a single patient often carries multiple mutations, and these mutations may exhibit additive, antagonistic or synergistic phenotypic effects in the same protein or across protein networks; at the same time, tissue-specific expression, heterogeneous microenvironment and selection pressure during treatment make the drug resistance phenotype exhibit highly complex and dynamic evolution characteristics. These factors together promote the research paradigm to move from the static judgment of "single target-single mutation" to the systematic prediction and attribution framework of "cross-network-multi-mutation-interpretable", so as to achieve earlier, more accurate and more stable drug resistance warning and intervention on the scale of clinical decision-making.

[0003] In terms of technical path, the identification and characterization of drug resistance mutations have long relied on the synergy of various experimental methods: upstream, directed evolution and mutant library screening in vitro are used to explore candidate sites that lead to decreased sensitivity, and downstream biochemical and biophysical experiments are used to evaluate the effects of mutations on drug binding, enzymatic kinetics and protein stability, and the effects are verified in cell or animal models. However, these methods naturally have the bottleneck of long cycle, high cost and limited throughput, which makes it difficult to cover the exponentially growing combination space of mutations. More importantly, there is still a migration gap between the conditions of experimental systems and the clinical phenotypes of the population, making it particularly difficult to carry out prospective scanning and mechanism hypothesis verification for potential drug resistance risks that have not yet appeared in the clinic. Therefore, developing a computational strategy that can quickly, systematically and interpretably evaluate the effects of mutations on drug efficacy in a large-scale mutation space has become an urgent need to crack the drug resistance problem and support precise medication and new drug design.

[0004] Despite the progress, these methods still have limitations in addressing complex clinical problems. The affinity change estimation approach, which is centered on the difference in protein-drug binding free energy, requires high-quality structural starting states and sufficient sampling, and is computationally expensive and difficult to scale up to cover multiple proteins, drugs, and mutations. More importantly, there is no universally accepted threshold or standard that can directly map a binding free energy difference value to clinical drug resistance, making the translation of molecular energy changes to clinically useful predictions uncertain. The structural perturbation and dynamics-oriented approach, while providing mechanistic clues, is limited in its generalization ability due to differences in systems and the availability of structures. The end-to-end sequence representation method, while alleviating the structural dependence and computational cost issues, generally lacks transparency in addressing mechanisms. When multiple sites coexist within the same protein, the model often has difficulty identifying which variation plays a dominant role in the drug resistance phenotype. In addition, most existing methods are primarily based on target-centered data construction and task definition, and do not adequately cover the clinical common situation of non-target site variations affecting drug efficacy through network-level effects, resulting in a methodological gap in real-world application scenarios.

[0005] Despite the progress, these methods still have limitations in addressing complex clinical problems. The affinity change estimation approach, which is centered on the difference in protein-drug binding free energy, requires high-quality structural starting states and sufficient sampling, and is computationally expensive and difficult to scale up to cover multiple proteins, drugs, and mutations. More importantly, there is no universally accepted threshold or standard that can directly map a binding free energy difference value to clinical drug resistance, making the translation of molecular energy changes to clinically useful predictions uncertain. The structural perturbation and dynamics-oriented approach, while providing mechanistic clues, is limited in its generalization ability due to differences in systems and the availability of structures. The end-to-end sequence representation method, while alleviating the structural dependence and computational cost issues, generally lacks transparency in addressing mechanisms. When multiple sites coexist within the same protein, the model often has difficulty identifying which variation plays a dominant role in the drug resistance phenotype. In addition, most existing methods are primarily based on target-centered data construction and task definition, and do not adequately cover the clinical common situation of non-target site variations affecting drug efficacy through network-level effects, resulting in a methodological gap in real-world application scenarios.

[0006] In summary, the prior art generally faces the following common limitations: First, the experimental approach has limited throughput and high resource consumption, making it difficult to cover the increasing number of mutations and combination mutations in actual clinical practice. Second, although the affinity estimation based on the difference in binding free energy can provide molecular-level clues, there is no widely accepted unified judgment standard directly targeting clinical drug resistance, making it difficult to convert the prediction results into clinical decisions. At the same time, such methods rely heavily on high-quality three-dimensional structures and complex computing processes, limiting their applicability in structural gaps or low-resolution scenarios. Third, models targeting structural perturbations, dynamics, and allosteric sites have varying degrees of generalization ability across different drug-protein systems, and the coverage of non-target mutations across protein networks is insufficient. Finally, although end-to-end learning solutions for sequences and molecular structures have improved the scope of application, they still struggle to explain whether the drug resistance phenotype is dominated by the intrinsic pathogenicity of the mutation, and they are also weak in causal dissection when multiple mutations coexist.

[0007] Therefore, there is an urgent need in the market for a multi-task deep learning-based protein mutation drug resistance prediction method and system that can explicitly introduce pathogenicity information, uniformly process targeted and non-targeted mutations, and achieve robust prediction and clear attribution in complex mutation backgrounds. SUMMARY

[0008] In view of the defects in the prior art, the purpose of the present application is to provide a multi-task deep learning-based protein mutation drug resistance prediction method and system.

[0009] According to the protein mutation drug resistance prediction method based on deep learning provided by the present application, the method comprises the following steps: Data acquisition and annotation step: obtaining evidence items of the influence of mutations on drug efficacy from a clinical drug-variation knowledge base, preprocessing the evidence items, and supplementing each mutation with a clinical pathogenicity label, only retaining two types of clear labels, pathogenic and benign, to form a training corpus with both drug resistance / sensitivity and pathogenicity / benignity labels; Feature engineering and input representation step: using a double-path representation for protein sequences and standardizing the length, and using a PubChem 881-dimensional structure fingerprint for drugs after full connection mapping and splicing with protein features to obtain a sequence tensor and a drug fingerprint that together constitute the model input; Cascade multi-task model training and reasoning step: constructing a cascade multi-task deep learning model with a Transformer as the core backbone, first learning the shared features of mutations and drugs and outputting the pathogenicity probability of mutations, then splicing the transformed and intermediate representations, and cascading the drug resistance probability; Model evaluation and interpretation analysis step: divide the double-labeled data into training set, validation set and independent test set, evaluate the performance by indicators, and realize sample joint typing based on double-task prediction results, and analyze the contribution of multiple mutations combined with attention weight.

[0010] Preferably, the clinical drug-variant knowledge base comprises CIViC knowledge base and PharmGKB knowledge base. The preprocessing includes deduplication, evidence level screening and field alignment.

[0011] Preferably, the double paths are local one-bit encoding of mutation site neighborhood and site-specific scoring matrix PSSM obtained by evolutionary alignment, respectively. The standardized length includes standardizing the sequence to a length L=1024, zero padding at the end of the sequence for sequences shorter than the length, and centering the mutation site to cut a fixed window.

[0012] Preferably, the cascaded multi-task model training and inference step comprises: Step S3.1: linearly map the drug fingerprint F and fuse it with the sequence tensor S into a unified input C,

[0013]

[0014] wherein is the sequence length, is the amino acid feature dimension, is the fingerprint vector dimension; Step S3.2: the unified input C is sent into a two-layer stacked Transformer encoder module, through a multi-head self-attention mechanism, to deeply capture the complex dependency between the residues within the sequence and between the sequence and the drug molecule, to generate a highly abstract shared context representation, which is shown as follows:

[0015] Step S3.2: flatten the shared context representation and send it into a fully connected layer to obtain an intermediate representation D1, which is shown as follows:

[0016] Step S3.3: based on the intermediate representation D1, through a dedicated multi-layer perceptron prediction head, and applying a Softmax function, output the posterior probability of mutation pathogenicity :

[0017] Step S3.4: introduce a cascading mechanism to combine P pathoThe feature vector that can characterize the pathogenicity conclusion is obtained by transforming the data through a task-specific fully connected layer. The formula is as follows:

[0018] Step S3.5: This conclusive feature The shared intermediate representation D1 from the upstream is spliced ​​together in a new dimension to construct an information-enhanced cascaded input. The formula is as follows:

[0019] Step S3.6: Cascaded Input The data is fed into a Transformer module to learn the correlation between the two types of information, and the final drug resistance / susceptibility probability is output by the subsequent MLP prediction head. .

[0020] Preferably, step S3.1 includes: Drug fingerprinting via a fully connected layer Perform a linear mapping so that its dimension corresponds to the sequence feature dimension. Matching is performed, and a non-linear activation is introduced, as shown in the following formula:

[0021] in, and It consists of a learnable weight matrix and a bias vector; Transformed drug characteristics Reshaped into The tensor, and the sequence feature tensor Concatenate the sequences along their length to form a unified combined input tensor. : .

[0022] Preferably, the model employs a multi-task joint loss function for end-to-end training, using cross-entropy as the loss function, as follows:

[0023] in, For batch size, For real labels; Total loss It is a weighted sum of the losses from the two tasks, and the formula is as follows:

[0024] Weighting coefficient It is an adjustable hyperparameter used to balance the importance of the two tasks during training.

[0025] Preferably, all learnable parameters in the network are updated by a backpropagation algorithm with AdamW adaptive learning rate optimizer, including learning rate, batch size, Dropout rate and loss weight .

[0026] According to the protein mutation-based prediction system for the influence of drug resistance provided by the application, the system comprises: A data acquisition and annotation module: evidence items of the influence of mutations on drug efficacy are obtained from a clinical drug-variation knowledge base, the evidence items are preprocessed, and a clinical pathogenicity label is supplemented for each mutation, only two types of clear labels of pathogenicity and benignity are retained, thereby forming training corpus with double labels of drug resistance / sensitivity and pathogenicity / benignity; A feature engineering and input representation module: a double-path representation is used for protein sequences and the length is standardized, and PubChem 881-dimensional structure fingerprints are used for drugs, which are mapped by full connection and spliced with protein features, and the obtained sequence tensor and drug fingerprints jointly constitute model input; A cascaded multi-task model training and reasoning module: a cascaded multi-task deep learning model with a Transformer core backbone is constructed, the shared features of mutations on drugs are learned first, and the pathogenicity probability of mutations is output, then the transformed and intermediate representation are spliced, and the drug resistance probability is output in cascade; A model evaluation and explanation analysis module: double-label data is divided into a training set, a validation set and an independent test set, performance is evaluated by indicators, joint typing of samples is realized based on double-task prediction results, and the contribution of multiple mutations is analyzed combined with attention weight.

[0027] Preferably, the clinical drug-variation knowledge base comprises a CIViC knowledge base and a PharmGKB knowledge base; The preprocessing comprises deduplication, evidence level screening and field alignment; The double paths are local one-bit encoding of mutation site neighborhood and site-specific scoring matrix PSSM obtained by evolutionary alignment; The standardized length comprises standardizing the sequence to a length L=1024, zero padding at the end of the sequence for sequences shorter than the length, and centering the mutation site to cut a fixed window for sequences longer than the length.

[0028] Preferably, the cascaded multi-task model training and reasoning module comprises: Module M3.1: linearly mapping drug fingerprints F and fusing them with sequence tensor S into unified input C,

[0029]

[0030] where is the sequence length, is the amino acid feature dimension, is the fingerprint vector dimension; Module M3.2: The unified input C is sent into a two-layer stacked Transformer encoder module, which captures the complex dependencies between residues within the sequence and between the sequence and the drug molecule through multi-head self-attention mechanisms, and generates a highly abstract shared context representation, which is shown as follows:

[0031] Module M3.2: The shared context representation is flattened and sent into a fully connected layer to obtain the intermediate representation D1, which is shown as follows:

[0032] Module M3.3: Based on the intermediate representation D1, a dedicated multi-layer perceptron prediction head is used, and a Softmax function is applied, to output the posterior probability of mutation pathogenicity .

[0033] Module M3.4: A cascade mechanism is introduced to combine P patho with the task-specific fully connected layer to obtain a feature vector that can represent the pathogenicity conclusion , which is shown as follows:

[0034] Module M3.5: This conclusion feature is spliced with the upstream shared intermediate representation D1 in a new dimension to construct an information-enhanced cascade input , which is shown as follows:

[0035] Module M3.6: The cascade input is sent into a Transformer module to learn the association between the two types of information, and the final drug resistance / sensitivity probability is output by the subsequent MLP prediction head .

[0036] Compared with the prior art, the present application has the following beneficial effects: 1. The present application jointly models "mutation pathogenicity" and "drug response" in the same framework, explicitly capturing the internal relationship and difference between the two in the representation space, overcoming the limitations of traditional methods that simply equate or separate the two.

[0037] 2、The application gives more accurate numerical prediction for both tasks on the independent test set (pathogenicity AUC=0.96, drug resistance AUC=0.94), which can accurately distinguish mutations and drug response status; compared with representative tools such as ESM-1v and AlphaMissense, the application is superior to or not lower than the prior art in F1, MCC, Accuracy and other indicators; and compared with similar systems (such as Emden), the application can obtain higher discrimination with less input features, reduce the data dependence threshold, and improve the efficiency of practical application.

[0038] 3、The application not only outputs improved prediction accuracy, but also provides mechanism-level explanation and clinically usable joint typing, which can clearly answer the key question of "whether drug response is driven by mutation pathogenicity", and develop differentiated drug strategies for different types of patients; at the same time, due to the characteristics of "weak structure dependence and minimal input", the application has good cross-center migration potential and can be quickly landed in scenarios such as precision medicine, variation interpretation, drug resistance warning, and drug redesign. BRIEF DESCRIPTION OF DRAWINGS

[0039] Other features, objects and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings: Figure 1 The following is a schematic diagram of the working method flow framework of the application; Figure 2 The following is a schematic diagram of the data flow of the cascaded multi-task deep learning model with Transformer as the core backbone in the application. DETAILED DESCRIPTION

[0040] The application will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the application. These all belong to the protection scope of the application.

[0041] Example 1 According to the protein mutation affecting drug resistance prediction method based on deep learning provided by the application, as shown in the following formula: Figure 1 The method comprises the following steps: Data acquisition and annotation step: obtaining evidence items of the effect of mutations on drug efficacy from a clinical drug-variation knowledge base, pre-processing the evidence items, and supplementing a clinical pathogenicity label for each mutation, only retaining two types of clear labels of pathogenicity and benignity, thereby forming a training corpus with both drug resistance / sensitivity and pathogenicity / benignity labels. The data acquisition and annotation step comprises: Step S1.1: Obtain evidence entries of mutation impact on drug efficacy from expert curated clinical drug-variant knowledge bases, and download a fixed version at the same time point to ensure reproducibility. The clinical drug-variant knowledge bases include CIViC knowledge base and PharmGKB knowledge base.

[0042] Step S1.2: After deduplication, evidence level screening and field alignment of the knowledge bases, a set of mutation-drug entries covering multiple disease fields is formed (e.g. 710 entries are obtained).

[0043] Step S1.3: On this basis, further supplement the clinical pathogenicity label for each mutation, preferably from the integrated human variation annotation of UniProt (including sources such as ClinVar, Ensembl VEP, etc.), and only keep the two types of clear labels “pathogenic” and “benign”, so as to form a training corpus with both “drug resistance / sensitivity” and “pathogenicity / benignity” labels (e.g. 570 entries are obtained, of which 409 are pathogenic, 161 are benign; 149 are drug resistant, and 421 are sensitive).

[0044] To support the hypothesis that “non-target mutations may also affect drug efficacy through pathological pathways”, entries are optionally docked with pChEMBL affinity data, and the maximum pChEMBL value represents the strength of drug-protein interaction, which is used to distinguish “directly acting targets” and “network layer indirect action”, for example, 308 entries with IC50 records are obtained from the 570 entries.

[0045] Feature engineering and input representation step: Local one-bit encoding and site-specific scoring matrix PSSM are used to represent protein sequences in two paths and standardized to length 1024; PubChem 881-dimensional structure fingerprint is used for drugs, which is mapped by full connection and spliced with protein features to form model input. Specifically, the protein sequence is represented by a double path, which is local one-bit encoding of the mutation site neighborhood, and site-specific scoring matrix (PSSM) obtained by evolutionary alignment. To unify the input dimension, the sequence is standardized to a length L=1024, and if it is shorter than this length, it is zero-padded at the end of the sequence, and if it is longer than this length, it is truncated to a fixed window centered on the mutation site. Wild-type and mutant sequences generate features according to the same process. Drug molecules are represented by PubChem's 881-dimensional structure fingerprint, and mapped to an embedding space that can be spliced with sequence channels by full connection. The sequence tensor obtained by the above processing and the drug fingerprint together constitute the model input. This design takes into account the local mutation environment and long-term evolutionary clues without relying on high-resolution three-dimensional structures, and provides sufficient context capacity for multiple mutations coexisting.

[0046] Training and Inference Steps of the Cascaded Multi-Task Model: A cascaded multi-task deep learning model with Transformer as its core backbone is constructed. First, the model learns the shared features of mutations with drugs and outputs the probability of mutation pathogenicity. Then, the transformed representation is concatenated with the intermediate representation to output the drug resistance probability. The cascaded multi-task deep learning model first fuses the drug fingerprint F with the sequence tensor S as a unified input C after linear mapping. It then learns the shared representation through two layers of self-attention Transformer modules. Subsequently, the shared representation is flattened and fed into a fully connected layer to obtain the intermediate representation D1, while simultaneously outputting the probability P of mutation pathogenicity. patho To enable pathogenicity tasks to explicitly guide drug resistance prediction, this invention introduces a cascade mechanism, which... patho The task-specific vector Ft1 is obtained through a dedicated fully connected transformation and then concatenated with D1 to form a cascaded input C. cascade Ultimately, C cascade After entering the dedicated Transformer module, the final drug resistance probability is output. Specifically, such as Figure 2 As shown, this model employs a novel explicit information flow guidance mechanism, using the output of the pathogenicity prediction task as prior knowledge for the drug resistance prediction task, thereby improving the model's predictive performance and interpretability in complex biological scenarios. The model initially receives two heterogeneous inputs: protein sequence feature tensors. (in For sequence length, (Amino acid feature dimension) and chemical fingerprint vector of drug molecules (in (For fingerprint vector dimensions). To fuse these two types of information, the model first processes the drug fingerprint through a fully connected layer (FC layer). Perform a linear mapping so that its dimension corresponds to the sequence feature dimension. Matching and introducing non-linear activation:

[0047] in, and These are learnable weight matrices and bias vectors. Subsequently, the transformed drug features... Reshape into The tensor, and the sequence feature tensor Concatenate the sequences along their length to form a unified combined input tensor. :

[0048] This combined tensor Subsequently, it is fed into two stacked Transformer encoder modules, which capture complex dependencies between residues within the sequence and between the sequence and the drug molecule through multi-head self-attention mechanisms, and generate highly abstracted shared context representations

[0049] After obtaining the shared representations , the model first performs the pathogenicity prediction task (task one). The core intermediate representation is flattened and passed through a fully connected layer :

[0050] The intermediate representation D1 contains a deep encoding of the combined effects of the mutation and the drug. Based on D1, the model predicts the posterior probability of the pathogenicity of the mutation through a dedicated multi-layer perceptron (MLP) prediction head and applies a softmax function :

[0051] To enable the pathogenicity prediction to explicitly guide the drug resistance prediction, we design a cascade mechanism. Specifically, the output probability of task one is not the training endpoint, but rather serves as an information carrier, which is transformed through a task-specific fully connected layer to generate a feature vector that can represent the pathogenicity conclusion wherein:

[0052] Subsequently, we concatenate this conclusive feature with the upstream shared intermediate representation D1 in a new dimension to construct an information-enhanced cascade input . This input can be viewed as a new sequence of length 2, with the first element being the combined context information and the second element being the pathogenicity judgment information.

[0053]

[0054] Finally, the cascade input is fed into a Transformer module to learn the association between the two types of information, and the final drug resistance / sensitivity probability is output by the subsequent MLP prediction head .

[0055] The model is trained end-to-end by using a joint loss, the weight coefficient is set to balance the pathogenicity and drug resistance tasks, the training hyperparameters can be selected by using an adaptive optimizer and a cross-entropy loss function, and parameters such as learning rate and batch size can be optimized according to the performance of the verification set. In the inference stage, the model can simultaneously give pathogenicity and drug resistance prediction based on the minimum input features, and the drug resistance prediction and the pathogenicity branch share optimization, so as to realize multi-task learning and mechanism inference. Specifically, the model training method is as follows: in order to realize the unified optimization of the entire network, the present application adopts a multi-task joint loss function for end-to-end training. Both tasks adopt cross-entropy as the loss function:

[0056] wherein, is the batch size (batch size), is the true label. The total loss is the weighted sum of the loss of the two tasks:

[0057] The weight coefficient is an adjustable hyperparameter, which is used to balance the importance of the two tasks in the training process, and ensures that the model optimizes the drug resistance prediction while not ignoring the accuracy of the pathogenicity prediction. In addition, the AdamW adaptive learning rate optimizer is used to update all learnable parameters in the network through the backpropagation algorithm. The learning rate, batch size, dropout rate and loss weight and other key hyperparameters.

[0058] The model evaluation and explanation analysis steps are as follows: the double-label data is divided into a training set, a verification set and an independent test set, the performance is evaluated by using AUC, F1, MCC and the like, and the existing tools are compared; based on the double-task prediction results, the sample joint typing is realized, and the contribution of multiple mutations is analyzed based on the attention weight. The double-label data is divided into a training set, a verification set and an independent test set (for example, 450 / 50 / 70), the classification performance of the pathogenicity and drug resistance tasks in the test set is evaluated, and the representational learning is used to analyze the separability and clustering structure of different labels in the embedding space. The advantages of the present application in accuracy and interpretability are verified compared with the existing computing tools through external methods and case-level case analysis. In the multi-mutation analysis, the contribution of multiple sites of the same protein is finely divided by combining methods such as attention weight, so as to support the differentiation of subtypes such as “pathogenic but sensitive” and “pathogenic and resistant”, and to locate the dominant site and possible synergistic / antagonistic relationship in the coexistence of multiple mutations. In order to meet the clinical application, threshold calibration and uncertainty quantification can be selected to make the output risk score and “drug resistance / sensitivity” judgment flexible.

[0059] The application aims to propose a cascade multi-task deep learning method and device with pathogenicity signal as prior, which is used for accurate and interpretable prediction and attribution of mutation-driven drug resistance. The scheme overcomes the problems of low throughput, strong dependence on high resolution structure and poor threshold dependence on binding free energy, and insufficient coverage of non-target and multi-mutation scenarios in the prior art, and unifies the processing of in-target and out-target situations under the minimum input features, and outputs clinically usable drug resistance risk assessment, supporting precise drug use and drug redesign.

[0060] The method of the application can be realized by program instructions in the system. Without departing from the spirit of the application, the data source, feature dimension, network layer number and hyperparameter can be equivalently replaced, and three-dimensional structure features and other molecular representations can also be included as optional channels. The distribution analysis of double-label data shows that a large number of "pathogenic but still sensitive" and "non-pathogenic but can change drug efficacy" samples objectively exist, confirming that pathogenicity and drug response are not linearly dependent, but they may share some molecular clues and pathway backgrounds. This fact provides a solid basis for the pathogenicity-guided multi-task learning of the application. Objective evaluation on an independent test set shows that the application achieves high accuracy and good robustness on both tasks: the AUC of pathogenicity prediction reaches 0.96, and the AUC of drug resistance prediction reaches 0.94. In the external method comparison (such as ESM-1v, AlphaMissense and other representative tools), the application performs better than or not lower than the prior art in F1, MCC and Accuracy and other indicators, indicating that the cascade information flow mediated by pathogenicity can effectively improve the quality of drug resistance determination. At the same time, compared with the Emden system, the application still obtains higher discriminability under the condition of using fewer input features, which embodies the significant gain of "pathogenicity guidance + cascade multi-task".

[0061] More importantly, the application not only gives more accurate numerical prediction (pathogenicity AUC=0.96, drug resistance AUC=0.94 on an independent test set), but also provides operable mechanism-level explanation and clinically usable joint typing capability. In an independent test set (n=70), the joint prediction of the model naturally divides the samples into 25 cases of "benign & sensitive" (about 35.7%), 27 cases of "pathogenic & sensitive" (about 38.6%) and 18 cases of "pathogenic & resistant" (about 25.7%). The distribution clearly shows that the model does not simply equate "pathogenic" with "resistant", but can distinguish whether a pathogenic mutation directly interferes with the mechanism of a specific drug, thereby answering the key question of "whether drug response is mainly driven by the intrinsic pathogenicity of the mutation", and providing quantitative basis for clinical interpretation.

[0062] In summary, under the premise of weak structure dependence and minimal input, the application realizes high-accuracy prediction of mutation-driven drug resistance, interpretable reasoning of mechanism, and robust handling of complex situations. The model output can be used to differentiate the interpretation and intervention strategy for different joint genotypes, has good cross-center migration potential, and can directly serve precision medicine decision-making, variant interpretation, drug resistance risk early warning, and drug re-designing applications.

[0063] Embodiment 2 The application also provides a protein mutation drug resistance prediction system based on multi-task deep learning, which can be realized by executing the process steps of the protein mutation drug resistance prediction method based on multi-task deep learning, that is, the protein mutation drug resistance prediction method based on multi-task deep learning can be understood by those skilled in the art as the preferred embodiment of the protein mutation drug resistance prediction system based on multi-task deep learning.

[0064] According to the protein mutation drug resistance prediction system based on deep learning provided by the application, the system comprises: The data acquisition and annotation module obtains evidence items of the influence of mutations on drug efficacy from a clinical drug-variant knowledge base, pre-processes the evidence items, and supplements each mutation with a clinical pathogenicity label, only retaining two types of clear labels, pathogenicity and benignity, thereby forming training corpus with both drug resistance / sensitivity and pathogenicity / benignity labels. The clinical drug-variant knowledge base includes CIViC knowledge base and PharmGKB knowledge base. The pre-processing includes deduplication, evidence level screening and field alignment.

[0065] The feature engineering and input representation module adopts double-path representation for protein sequences and standardizes the length, and adopts PubChem 881-dimensional structure fingerprint for drugs after full connection mapping and splicing with protein features, to obtain sequence tensors and drug fingerprints which together constitute the model input. The double paths are local one-bit encoding of mutation site neighborhood and site-specific scoring matrix PSSM obtained through evolutionary alignment. Standardizing the length includes standardizing the sequence to a length of L=1024, zero-padding the sequence tail for sequences shorter than the length, and centering the mutation site to obtain a fixed window.

[0066] The cascaded multi-task model training and reasoning module constructs a cascaded multi-task deep learning model with Transformer as the core backbone, first learns the shared features of mutations and drugs, and outputs the mutation pathogenicity probability, then splices the transformed and intermediate representations, and cascades to output the drug resistance probability. The cascaded multi-task model training and reasoning module comprises: module M3.1: fusing the drug fingerprint F after linear mapping with the sequence tensor S into a unified input C,

[0067]

[0068] where is the sequence length, is the amino acid feature dimension, is the fingerprint vector dimension. Module M3.1 includes: linear mapping of drug fingerprints through a fully connected layer, so that its dimension matches the sequence feature dimension , and introduces a nonlinear activation function, as follows: where, and are learnable weight matrices and bias vectors. The transformed drug features are reshaped into a tensor and concatenated with the sequence feature tensor in the sequence length dimension to form a unified combined input tensor : .

[0069] Module M3.2: The unified input C is fed into a two-layer stacked Transformer encoder module, which captures complex dependencies between residues within the sequence and between the sequence and the drug molecule through a multi-head self-attention mechanism, generating a highly abstract shared context representation, as follows:

[0070] Module M3.2: The shared context representation is flattened and fed into a fully connected layer to obtain the intermediate representation D1, as follows:

[0071] Module M3.3: Based on the intermediate representation , a dedicated multi-layer perceptron prediction head is used, and the Softmax function is applied to output the posterior probability of mutation pathogenicity :

[0072] Module M3.4: A cascade mechanism is introduced to transform P patho into a feature vector that can represent the pathogenicity conclusion , as follows:

[0073] Module M3.5: This conclusive feature is combined with the upstream shared intermediate representation Concatenate in new dimensions, build an information-enhanced cascade input , as follows:

[0074] Module M3.6: Cascade input is fed into a Transformer module to learn the association between the two kinds of information, and the final drug resistance / sensitivity probability is output by the subsequent MLP prediction head .

[0075] The model is trained end-to-end using a multi-task joint loss function, with cross-entropy as the loss function, as follows:

[0076] where, is the batch size, is the true label. The total loss is the weighted sum of the losses of the two tasks, as follows:

[0077] The weight coefficient is an adjustable hyperparameter that balances the importance of the two tasks during training. The AdamW adaptive learning rate optimizer is used to update all learnable parameters in the network, including learning rate, batch size, Dropout rate, and loss weight, through the backpropagation algorithm .

[0078] Model evaluation and interpretation analysis module: divide the double-label data into training set, validation set and independent test set, evaluate the performance through indicators, and realize sample joint typing based on the double-task prediction results, and analyze the contribution of multiple mutations combined with attention weight.

[0079] Those skilled in the art know that in addition to implementing the system provided by the present application and each device, module, unit thereof in pure computer readable program code, the system provided by the present application and each device, module, unit thereof can also be realized in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers and embedded microcontrollers, etc. by logically programming the method steps to achieve the same function. Therefore, the system provided by the present application and each device, module, unit thereof can be considered as a hardware component, and the devices, modules, units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, units for implementing various functions can also be considered as both software modules implementing methods and structures within hardware components.

[0080] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be combined with each other at will without conflict.

Claims

1. A deep learning-based method for predicting the impact of protein mutations on drug resistance, characterized in that, include: Data collection and annotation steps: Obtain evidence entries on the impact of mutations on drug efficacy from the clinical drug-variant knowledge base, preprocess the evidence entries, and add clinical pathogenicity labels to each mutation, retaining only the two categories of explicit labels: pathogenic and benign, thereby forming training corpus with dual labels of drug resistance / sensitivity and pathogenicity / benignity. Feature engineering and input representation steps: The protein sequence is characterized by dual paths and its length is normalized, and the drug is represented by PubChem881-dimensional structural fingerprint, which is then mapped by a fully connected layer and concatenated with the protein features. The resulting sequence tensor and the drug fingerprint together constitute the model input. Training and inference steps of cascaded multi-task model: Construct a cascaded multi-task deep learning model with Transformer as the core backbone. First, learn the shared features of mutations with drugs and output the pathogenicity probability of mutations. Then, concatenate the transformed model with the intermediate representation and output the drug resistance probability in a cascaded manner. Model evaluation and interpretation analysis steps: Divide the dual-label data into training set, validation set and independent test set, evaluate performance through indicators, and achieve joint genotyping of samples based on dual-task prediction results, and analyze the contribution of multiple mutations by combining attention weights.

2. The method for predicting the impact of protein mutations on drug resistance based on deep learning according to claim 1, characterized in that, The clinical drug-variant knowledge base includes the CIViC knowledge base and the PharmGKB knowledge base; The preprocessing includes deduplication, evidence level filtering, and field alignment.

3. The method for predicting the impact of protein mutations on drug resistance based on deep learning according to claim 1, characterized in that, The two paths are the local one-bit encoding of the neighborhood of the mutation site and the site-specific scoring matrix (PSSM) obtained through evolutionary alignment, respectively. The normalized length includes normalizing the sequence to a length of L=1024. For sequences shorter than this length, zero padding is performed at the end of the sequence. For sequences longer than this length, a fixed window is truncated with the mutation site as the center.

4. The method for predicting the impact of protein mutations on drug resistance based on deep learning according to claim 1, characterized in that, The training and inference steps of the cascaded multi-task model include: Step S3.1: After linear mapping, the drug fingerprint F is fused with the sequence tensor S into a unified input C. in For sequence length, For amino acid characteristics, The dimension of the fingerprint vector; Step S3.2: The unified input C is fed into a two-layer stacked Transformer encoder module. Through a multi-head self-attention mechanism, it deeply captures the complex dependencies between residues within the sequence and between the sequence and the drug molecule, generating a highly abstract shared context representation, as shown in the following formula: Step S3.2: Flatten the shared context representation and feed it into the fully connected layer to obtain the intermediate representation D1, as shown in the following formula: Step S3.3: Based on intermediate representations It uses a dedicated multilayer perceptron prediction head and applies the Softmax function to output the posterior probability of mutation pathogenicity. : Step S3.4: Introduce a cascading mechanism to P patho The feature vector that can characterize the pathogenicity conclusion is obtained by transforming the feature vector through a task-specific fully connected layer. The formula is as follows: Step S3.5: Present this conclusive feature Shared intermediate representation with upstream By splicing together elements in a new dimension, an information-enhanced cascading input can be constructed. The formula is as follows: Step S3.6: Cascaded Input The data is fed into a Transformer module to learn the correlation between the two types of information, and the final drug resistance / susceptibility probability is output by the subsequent MLP prediction head. .

5. The method for predicting the impact of protein mutations on drug resistance based on deep learning according to claim 4, characterized in that, Step S3.1 includes: Drug fingerprinting via a fully connected layer Perform a linear mapping so that its dimension corresponds to the sequence feature dimension. Matching is performed, and a non-linear activation is introduced, as shown in the following formula: in, and It consists of a learnable weight matrix and a bias vector; Transformed drug characteristics Reshaped into The tensor, and the sequence feature tensor Concatenate the sequences along their length to form a unified combined input tensor. : 。 6. The method for predicting the impact of protein mutations on drug resistance based on deep learning according to claim 1, characterized in that, The model employs a multi-task joint loss function for end-to-end training, using cross-entropy as the loss function, as shown in the following formula: in, For batch size, For real labels; Total loss It is a weighted sum of the losses from the two tasks, as shown in the following formula: Weighting coefficient It is an adjustable hyperparameter used to balance the importance of the two tasks during training.

7. The method for predicting the impact of protein mutations on drug resistance based on deep learning according to claim 1, characterized in that, The AdamW adaptive learning rate optimizer is used to update all learnable parameters in the network, including the learning rate, batch size, dropout rate, and loss weights, through the backpropagation algorithm. .

8. A deep learning-based predictive system for the impact of protein mutations on drug resistance, characterized in that, include: Data acquisition and annotation module: Obtain evidence entries on the impact of mutations on drug efficacy from the clinical drug-variant knowledge base, preprocess the evidence entries, and add clinical pathogenicity labels to each mutation, retaining only the two clear labels of pathogenicity and benignity, thereby forming training corpus with dual labels of drug resistance / sensitivity and pathogenicity / benignity. Feature engineering and input representation module: Protein sequences are characterized using a dual-path approach and their lengths are standardized. Drug fingerprints are generated using PubChem881-dimensional structure fingerprints and then fully connected and concatenated with protein features. The resulting sequence tensor and drug fingerprint together constitute the model input. Cascaded multi-task model training and inference module: Construct a cascaded multi-task deep learning model with Transformer as the core backbone. First, learn the shared features of mutations with drugs and output the pathogenicity probability of mutations. Then, concatenate the transformed model with intermediate representations and output the drug resistance probability in a cascaded manner. Model evaluation and interpretation analysis module: Divide the dual-label data into training set, validation set and independent test set, evaluate performance through indicators, and realize joint genotyping of samples based on dual-task prediction results, and analyze the contribution of multiple mutations by combining attention weights.

9. The deep learning-based predictive system for the impact of protein mutations on drug resistance according to claim 8, characterized in that, The clinical drug-variant knowledge base includes the CIViC knowledge base and the PharmGKB knowledge base; The preprocessing includes deduplication, evidence level filtering, and field alignment. The two paths are the local one-bit encoding of the neighborhood of the mutation site and the site-specific scoring matrix (PSSM) obtained through evolutionary alignment, respectively. The normalized length includes normalizing the sequence to a length of L=1024. For sequences shorter than this length, zero padding is performed at the end of the sequence. For sequences longer than this length, a fixed window is truncated with the mutation site as the center.

10. The deep learning-based predictive system for the impact of protein mutations on drug resistance according to claim 8, characterized in that, The cascaded multi-task model training and inference module includes: Module M3.1: After linear mapping, the drug fingerprint F is fused with the sequence tensor S into a unified input C. in For sequence length, For amino acid characteristics, The dimension of the fingerprint vector; Module M3.2: The unified input C is fed into a two-layer stacked Transformer encoder module. Through a multi-head self-attention mechanism, it deeply captures the complex dependencies between residues within the sequence and between the sequence and drug molecules, generating a highly abstract shared context representation, as shown in the following formula: Module M3.2: Flatten the shared context representation and feed it into the fully connected layer to obtain the intermediate representation D1, as shown in the following formula: Module M3.3: Based on the intermediate representation D1, it uses a dedicated multilayer perceptron prediction head and applies the Softmax function to output the posterior probability of mutation pathogenicity. : Module M3.4: Introduces a cascading mechanism to connect P patho The feature vector that can characterize the pathogenicity conclusion is obtained by transforming the feature vector through a task-specific fully connected layer. The formula is as follows: Module M3.5: This conclusive feature The shared intermediate representation D1 from the upstream is spliced ​​together in a new dimension to construct an information-enhanced cascaded input. The formula is as follows: Module M3.6: Cascade Input The data is fed into a Transformer module to learn the correlation between the two types of information, and the final drug resistance / susceptibility probability is output by the subsequent MLP prediction head. .

Images