Methods for predicting early drug resistance in ALK fusion gene-positive non-small cell lung cancer

Abstract

The present invention relates to the field of medical informatics and artificial intelligence technology, and more specifically, to a method for predicting early drug resistance in ALK fusion gene-positive non-small cell lung cancer, comprising the following steps: (1) data acquisition and preprocessing: collecting multi-omics data of ALK-positive non-small cell lung cancer patients, including gene mutation data, epigenetic data and transcriptome data, performing quality control, removing batch effects and standardizing the multi-omics data to form a unified feature matrix X; (2) multi-omics association analysis: using a canonical correlation analysis method to explore the association patterns between the multi-omics data and identify key features related to drug resistance; (3) deep learning prediction: inputting the feature matrix X into a deep learning model, and predicting the risk of early drug resistance through hierarchical progressive feature learning, drug structure adversarial learning and focused loss function optimization; and (4) drug selection recommendation.

Description

Technical Field

[0001] The present invention relates to the fields of medical informatics and artificial intelligence, and in particular to a method for predicting early drug resistance of ALK fusion gene-positive non-small cell lung cancer. Background Art

[0002] In recent years, patients with ALK fusion gene-positive non-small cell lung cancer (NSCLC) often show good efficacy in the early stage after receiving ALK inhibitor treatment, but as time goes by, some patients will gradually develop drug resistance, resulting in a decrease in treatment effect. Accurately predicting the early drug resistance risk of NSCLC patients is of great significance for guiding clinical medication and improving prognosis.

[0003] With the rapid development of precision medicine, the use of multi-omics data and artificial intelligence technology for disease diagnosis and prognosis prediction has become a hot topic in medical research. In particular, for patients with ALK fusion gene-positive non-small cell lung cancer (ALK+NSCLC), although ALK inhibitor treatment shows good efficacy in the early stage, some patients will gradually develop acquired resistance, leading to disease progression. Therefore, accurately predicting the early resistance risk of ALK+NSCLC patients is of great significance for guiding clinical decision-making and prolonging patient survival.

[0004] Existing NSCLC drug resistance prediction methods are mainly based on single-omics data such as gene mutations, such as the use of next-generation sequencing technology to detect gene mutations and evaluate the correlation between specific genes and drug resistance. However, such methods fail to fully consider information at different biological levels, such as gene expression regulation and protein interactions, and ignore the complexity of biological systems. Existing studies have shown that integrating multi-omics data such as genome, transcriptome, and proteome can more comprehensively characterize the occurrence and development mechanism of the disease and provide richer characteristic information for disease diagnosis and prognosis prediction. Therefore, there is an urgent need for a drug resistance prediction method that can effectively integrate multi-omics data.

[0005] In addition, existing integrated analysis methods based on multi-omics data, such as iCluster and SNF, mainly use linear models or probabilistic graphical models, which have limitations in modeling complex nonlinear relationships in multi-omics data. In recent years, deep learning technology has made breakthrough progress in many biomedical fields with its powerful feature learning and abstraction capabilities. However, existing deep learning models, such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), are mainly targeted at single-modality data such as images and sequences, and are difficult to directly apply to multi-omics data. How to design a deep learning framework suitable for multi-omics data and fully explore the collaborative information between different omics is an urgent problem to be solved.

[0006] Another thorny problem in the task of predicting drug resistance in NSCLC is that the number of early drug-resistant samples is far less than that of sensitive samples, that is, the class imbalance problem. Traditional machine learning algorithms, such as support vector machines and logistic regression, usually use accuracy as the optimization goal, which is prone to bias towards the majority class, resulting in poor prediction performance for minority class samples. Some improved methods, such as undersampling, oversampling, and cost-sensitive learning, have alleviated the class imbalance problem to a certain extent, but there are still defects such as insufficient information utilization and poor model generalization ability. Therefore, how to learn robust and effective feature representation under the condition of class imbalance is a bottleneck that needs to be broken through in the prediction of early drug resistance in NSCLC.

[0007] Finally, the current clinical drug selection for ALK+NSCLC patients mainly relies on the experience of doctors and lacks personalized guidance. A series of ALK inhibitors, such as crizotinib, ceritinib, alectinib, etc., have significant differences in efficacy and toxic side effects. To match the best drug combination for each patient, it is necessary to comprehensively consider multiple factors such as the patient's molecular characteristics, medical history, and tolerance. Most of the existing drug selection systems are based on rules or random traversal, lack personalized considerations, and are difficult to adapt to the increasingly sophisticated diagnosis and treatment needs. Therefore, developing an intelligent drug recommendation algorithm that makes full use of the results of multi-omics data and prediction models is of great significance to improving the treatment effect and quality of life of ALK+NSCLC patients.

[0008] In summary, the existing NSCLC early drug resistance prediction and drug selection methods still have deficiencies in terms of multi-omics data integration, deep learning algorithm design, class imbalance problems and personalized decision-making. Therefore, it is urgent to develop a new intelligent solution to break through the limitations of traditional methods and maximize the synergistic effect of multi-omics data, thereby significantly improving the accuracy of early drug resistance prediction and drug selection for ALK+NSCLC patients and benefiting more patients. This is exactly the research motivation and significance of this invention. Summary of the invention

[0009] The purpose of the present invention is to provide a method for predicting early drug resistance in ALK-positive NSCLC. This method constructs a full-process drug resistance prediction solution through multi-omics data integration and deep learning technology, which can significantly improve the accuracy and robustness of early drug resistance prediction.

[0010] The present invention provides a method for predicting early drug resistance of ALK fusion gene-positive non-small cell lung cancer, comprising the following steps:

[0011] The method for predicting early drug resistance of ALK fusion gene-positive non-small cell lung cancer includes the following steps:

[0012] (1) Data collection and preprocessing: Collect multi-omics data of ALK-positive non-small cell lung cancer patients, including gene mutation data, epigenetic data, and transcriptome data, perform quality control, remove batch effects, and standardize the multi-omics data to form a unified feature matrix X;

[0013] (2) Multi-omics association analysis: canonical correlation analysis was used to explore the association patterns among the multi-omics data and identify key features associated with drug resistance;

[0014] (3) Deep learning prediction: The feature matrix X is input into the deep learning model, and the early drug resistance risk is predicted through hierarchical progressive feature learning, drug structure adversarial learning and focused loss function optimization;

[0015] (4) Drug selection recommendation: Based on the early drug resistance risk prediction results and combined with the patient's molecular characteristics, the drug knowledge base is retrieved and recommended for subsequent treatment drugs suitable for the patient.

[0016] Specifically, the hierarchical progressive feature learning in step (3) includes:

[0017] Step 1: Extract low-level features within a single group:

[0018] Use the autoencoder to encode the i-th omics data X i Extract low-dimensional feature representation V i , the formula is:

[0019] V i =σ(W 1i ·X i +b 1i )

[0020] Among them, W 1i and b 1i are the weight matrix and bias term of the encoder respectively, and σ is the activation function;

[0021] Step 2: Attention mechanism learning omics interaction features:

[0022] Use the attention mechanism to calculate the attention weight α of omics discussion omics j ij , the formula is:

[0023] Among them, Q i =W 2i ·V i ,K i =W 3i ·V i ,W 2i and W 3i is the learnable weight matrix, d is Q i , K jThe dimension of the interaction feature representation C of omics i i By weighted summation we get:

[0024]

[0025] Step 3: Fusion of multi-head self-attention mechanism:

[0026] Multi-scale features use a multi-head self-attention mechanism for C1, C2, ..., C m After fusion, we get the multi-omics feature representation MF, the formula is:

[0027] MF=W5·Concat(C i1 ,C i2 ,...,C ih )+b5

[0028] Among them, W5 and b5 are the learnable weight matrix and bias term.

[0029] Specifically, the drug structure adversarial learning in step (3) includes: assuming that the drug molecular structure feature is FS, the drug-resistant and sensitive cell lines are Sr and Ss respectively, and by minimizing the adversarial loss L adv , so that the feature representation MF can distinguish Sr and Ss, enhancing its drug resistance prediction ability.

[0030] L adv =-E[logD(FS|Sr)]-E[log(1-D(FS|Ss))]

[0031] Where D is the discriminator, which is used to distinguish whether FS comes from Sr or Ss.

[0032] Specifically, the optimization of the focus loss function in step (3) includes: using the focus loss function to deal with the sample imbalance problem and strengthen the model's attention to minority class samples. The loss function is defined as:

[0033] L focal =-(1-p) γ ylog(p)-(p) γ (1-y)log(1-p)

[0034] Among them, y∈{0,1} is the true label of the sample, p is the probability that the model predicts that the sample is a positive class, γ is the focusing factor, γ≥0.

[0035] Specifically, the multi-omics data in step (1) include: gene mutation data, epigenetic data and transcriptome data, with a sample size of no less than 100 cases and including samples of early resistant and sensitive patients.

[0036] Specifically, the quality control of multi-omics data in step (1) includes: using FastQC to evaluate the quality of sequencing data and using Trimmomatic to remove adapters and low-quality sequences.

[0037] Specifically, the removal of batch effects of multi-omics data in step (1) includes: performing variation detection and annotation on gene mutation data, calculating methylation levels on epigenetic data, performing expression quantification on transcriptome data, and performing batch correction.

[0038] Specifically, the canonical correlation analysis in step (2) identifies the features that contribute most to drug resistance by maximizing the correlation coefficient of the linear combination of variables between different omics data. The optimization goal is:

[0039]

[0040]

[0041] Specifically, the deep learning model in step (3) is trained using stochastic gradient descent, and early stopping and L2 regularization are used to avoid overfitting. The loss function is the weighted sum of focus loss and adversarial loss:

[0042] L total =L focal +λ·L adv

[0043] Where λ is the balancing factor, which controls the weight of the two losses.

[0044] Specifically, the drug selection recommendation in step (4) includes: dividing patients into high-risk and low-risk categories according to the predicted probability of early drug resistance risk, and for high-risk patients, retrieving drugs with similar molecular characteristics as candidates from a pre-constructed drug knowledge base; the drug knowledge base contains the molecular structure, mechanism of action, toxicology and clinical data information of marketed and clinical trial drugs, and uses chemical informatics methods to achieve drug comparison and retrieval.

[0045] The present invention has the following beneficial effects:

[0046] 1. Synergistic effect of multi-omics data: Different omics data reflect different levels of biological systems, such as genome reflects genetic information, epigenome reflects gene expression regulation, and transcriptome reflects gene expression level. There are complex regulatory relationships between these omics, which jointly affect the occurrence and development of diseases. The present invention uses methods such as canonical correlation analysis to quantify the correlation between different omics from a mathematical perspective, and use it to guide feature selection and prediction model construction. Compared with a single omics method, the multi-omics integration strategy can more comprehensively characterize the state of the biological system, and the complementary information of different omics can significantly improve the prediction performance, resulting in a synergistic effect of "1+1>2".

[0047] 2. Information extraction and transmission of end-to-end deep learning model: The present invention constructs an end-to-end deep learning model, which gradually extracts and fuses multi-omics features of different levels and scales through hierarchical progressive feature learning. Among them, the autoencoder extracts the essential features within the omics, the attention mechanism learns the interaction mode between different omics, and the multi-head self-attention mechanism further mines high-order associations. This hierarchical feature learning method can make full use of the information in the original data while avoiding the deviation caused by artificially designed features. In addition, the present invention introduces drug structure adversarial learning in model training, and uses the drug molecular structure to guide feature learning, so that the final feature representation can well distinguish between resistant and sensitive samples. This adversarial learning strategy that integrates multiple prior knowledge enhances the generalization performance and robustness of the model.

[0048] 3. Focus loss function solves the problem of class imbalance: In clinical data, the proportion of early drug-resistant patients is usually much lower than that of sensitive patients, which leads to serious class imbalance problems and poses challenges to the training and optimization of prediction models. The traditional cross entropy loss function tends to favor the majority class and ignore minority class samples. To address this problem, the present invention introduces a focused loss function, which dynamically adjusts the loss weights so that the model pays more attention to minority class samples that are difficult to classify. Mathematically, the focused loss function amplifies the loss contribution of mispredicted samples through the term (1-p)^γ, where p is the probability that the sample is predicted to be the correct class, and γ is the focusing factor. The larger the γ value, the greater the penalty for misclassified samples. This strategy effectively alleviates the negative impact of class imbalance on model training and improves the prediction sensitivity for early drug-resistant patients.

[0049] 4. Drug selection recommendation reflects the concept of precision medicine: ALK-positive NSCLC patients have very different responses to different ALK inhibitors. Selecting the best treatment plan for each patient is the core goal of precision medicine. Traditional drug selection mainly relies on the doctor's experience and lacks personalized considerations. Based on the multi-omics characteristics of patients, the present invention recommends potentially effective drugs from a large number of candidate drugs through similarity search. Here, technologies such as molecular fingerprints of chemical informatics are used to comprehensively consider the multi-dimensional properties of drugs such as chemical structure, mechanism of action, and side effects to match the best drug for patients. This recommendation algorithm makes full use of the multi-omics prediction results, embodies the concept of "tailor-made" precision medicine, and provides a practical reference for subsequent clinical treatment.

[0050] In summary, the present invention skillfully combines multi-omics data fusion, deep learning algorithms, and domain knowledge, achieving complementary advantages and conflict resolution in multiple links, and finally forming a full-process early drug resistance prediction scheme. This scheme not only has significantly better prediction performance than existing methods, but also provides precise guidance for subsequent clinical treatment. It has high application value and promotion potential, and has opened up a new research direction for intelligent medical auxiliary diagnosis and treatment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The present invention is a flow chart of the method for predicting early drug resistance of ALK fusion gene-positive non-small cell lung cancer. DETAILED DESCRIPTION

[0052] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the method for predicting early drug resistance of ALK fusion gene-positive non-small cell lung cancer proposed by the present invention, its specific implementation method, structure, characteristics and efficacy in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0053] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0054] The specific scheme of the method for predicting early drug resistance of ALK fusion gene positive non-small cell lung cancer provided by the present invention is described in detail below with reference to the accompanying drawings.

[0055] Example 1

[0056] See also Figure 1 The present invention discloses a method for predicting early drug resistance of ALK fusion gene positive non-small cell lung cancer, comprising the following steps:

[0057] (1) Data collection and preprocessing: Collect multi-omics data of ALK-positive non-small cell lung cancer patients, including but not limited to gene mutation data, epigenetic data, and transcriptome data. To ensure the robustness of the model, the sample size should be no less than 100 cases, and both early drug-resistant patients and treatment-sensitive patients should be included.

[0058] The obtained multi-omics data were subjected to comprehensive quality control, and the sequencing data quality was evaluated using tools such as FastQC, and adapters and low-quality sequences were removed using tools such as Trimmomatic to ensure the reliability of the data.

[0059] Specific preprocessing is performed for different types of omics data. Gene mutation data are subjected to variation detection and functional annotation; epigenetic data such as DNA methylation chips or sequencing data are used to calculate methylation levels; transcriptome data such as RNA-seq are used to calculate gene expression. At the same time, batch correction is performed on each type of data to eliminate the impact of batch effects.

[0060] After the above preprocessing steps, the multi-omics data from different sources are integrated into a unified feature matrix X, ready for subsequent analysis.

[0061] (2) Multi-omics association analysis: In order to fully explore the association information between different omics, canonical correlation analysis (CCA) was used to explore the association patterns of multi-omics data. CCA maximizes the correlation coefficient between linear combinations of different omics data to identify feature sets that contribute most to drug resistance.

[0062] Suppose there are m omics data types, the optimization goal of CCA is:

[0063]

[0064] The constraints ensure that the transformed features have equal variance, so that the correlations obtained across different omics are comparable:

[0065]

[0066] Among them, a i For omics data X i The linear transformation vector of . The alternating optimization algorithm can be used to solve CCA by iteratively updating a i , until convergence. The linear combination of variables obtained by CCA reflects the contribution of different omics to drug resistance, thus providing key feature information and biological prior knowledge for the construction of drug resistance prediction models.

[0067] Deep learning prediction: This step is the core of the present invention. It uses advanced deep learning technology to build a full-process drug resistance prediction model, which mainly includes three modules: hierarchical progressive feature learning, drug structure adversarial learning, and focused loss function optimization. The hierarchical progressive feature learning module uses technologies such as autoencoders, attention mechanisms, and multi-head self-attention mechanisms to adaptively extract and fuse features from multi-omics data. This module is divided into three key steps:

[0068] Step 1: Extract low-level features within a single omics dataset X i , use autoencoder for dimensionality reduction and feature extraction, and learn a low-dimensional feature representation V i :

[0069] V i =σ(W 1i ·X i +b 1i )

[0070] Among them, W 1i and b 1i are the weight matrix and bias vector of the encoder, respectively, and σ is a nonlinear activation function such as ReLU. As an unsupervised feature learning method, the autoencoder can adaptively extract important feature patterns in omics.

[0071] Step 2: Attention mechanism learns the interactive features of omics. In Step 1, the low-level feature representation V of each omics is obtained. i Finally, the attention mechanism is used to model the interaction between different groups, and the attention weight α of group j on group j is learned. ij :

[0072]

[0073] Among them, Q i =W 2i ·V i ,K i =W 3i ·V i denote the query vector and key vector respectively, W 2i and W 3i is the learnable mapping matrix, d is Q i and K j The softmax function normalizes the attention coefficient to between 0 and 1. The interactive feature representation C of group i i By weighted summation of attention, we get:

[0074] Step 3: Multi-head self-attention mechanism integrates multi-scale features in Step 2, omics interaction features C i Only paired omics relationships are considered. To further explore high-order omics correlations, a multi-head self-attention mechanism is introduced to perform self-attention on C1 to C m The multi-head self-attention mechanism will i Divide into h subspaces, perform the attention calculation of Step 2 in parallel in each subspace, and obtain h groups of interactive feature representations C i1 ,C i2 ,...,C ih This mechanism is equivalent to modeling omics associations from different perspectives, enhancing the diversity and robustness of feature representation.

[0075] Finally, concatenation and linear transformation are used to fuse the h sub-feature spaces into a unified multi-omics feature representation MF:

[0076] MF=W5·Concat(C i1 ,C i2 ,...,C ih )+b5

[0077] Among them, W5 and b5 are the weight matrix and bias vector respectively, and Concat represents the concatenation operation.

[0078] The above-mentioned hierarchical progressive feature learning module makes full use of the complementary information of multi-omics data, and obtains a comprehensive and effective multi-omics feature representation MF through a hierarchical feature extraction method from local to global and shallow to deep. The drug structure adversarial learning module uses the molecular structure information of NSCLC-related drugs (such as ALK inhibitors) to enhance the drug sensitivity prediction ability of MF through adversarial learning. Suppose the drug molecular structure feature is FS, and the resistant and sensitive cell lines are Sr and Ss, respectively. A discriminator network D is introduced, whose purpose is to distinguish whether FS comes from Sr or Ss. By minimizing the following adversarial loss, MF can distinguish Sr and Ss as much as possible, thereby improving its drug resistance prediction effect:

[0079] L adv =-E[logD(FS|Sr)]-E[log(1-D(FS|Ss))]

[0080] Where E represents the mathematical expectation. Intuitively, the first term in the above formula indicates that for drug-resistant samples, D should output a larger probability; the second term indicates that for sensitive samples, D should output a smaller probability. Therefore, minimizing L adv This means making it difficult for D to distinguish the source of FS, while allowing MF to identify resistance / sensitive characteristics to the greatest extent.

[0081] During the model training process, the discriminator D and the feature representation MF adopt an alternating optimization strategy, and improve each other through mini-max game, so that MF can finally learn the drug feature information related to drug resistance. The focus loss function optimization module is used to deal with the common class imbalance problem in the NSCLC drug resistance prediction task, that is, the number of drug-resistant samples is far less than that of sensitive samples. In order to increase the model's attention to minority class samples, the focus loss function (FocalLoss) is introduced as the optimization target of the model:

[0082] L focal =-(1-p) γ ylog(p)-p γ (1-y)log(1-p)

[0083] Among them, y∈{0,1} is the true label of the sample (0 means sensitive, 1 means resistant), p is the probability that the model predicts that the sample is resistant, and γ≥0 is the focusing factor.

[0084] When γ=0, the focus loss degenerates to the standard cross entropy loss. When γ>0, the loss of the wrongly predicted samples is amplified by (1-p) compared to the correctly predicted samples. γ times, so that the model pays more attention to samples that are difficult to classify correctly. The larger the value of γ, the more attention is paid to samples that are difficult to classify. Through this strategy, the focus loss function effectively alleviates the negative impact of class imbalance.

[0085] When training deep learning models, stochastic gradient descent is used as the optimization algorithm, and early stopping and L2 weight regularization are used to prevent overfitting. The final loss function of the model is the weighted sum of the focus loss and the adversarial loss:

[0086] L total =L focal +λ·L adv

[0087] Among them, λ is the balance factor, which controls the relative importance of the two losses.

[0088] (4) Drug selection recommendation: Based on the risk probability of early drug resistance predicted by the deep learning model, patients are divided into high-risk and low-risk categories. For high-risk patients, appropriate follow-up treatment plans are recommended.

[0089] Specifically, a knowledge base containing marketed drugs and clinical trial drugs is constructed to record the molecular structure, mechanism of action, toxicology, clinical data, etc. of each drug. For each high-risk patient, similar drug molecules are retrieved from the knowledge base based on their multi-omics characteristics as potential drug candidates.

[0090] The calculation of drug similarity can adopt common methods in the field of chemical informatics, such as molecular fingerprint, chemical substructure search, etc. Here, taking molecular fingerprint as an example, the cosine similarity of the patient feature vector v and the drug molecular fingerprint f is calculated:

[0091]

[0092] Candidate drugs are ranked according to similarity scores, and the Top-N are taken as the recommendation results to assist doctors in making treatment decisions. This similarity-based recommendation strategy takes into account the multi-dimensional attribute information of drugs and can provide personalized and accurate medication guidance.

[0093] Compared with the prior art, the present invention has the following advantages and effects:

[0094] (1) Comprehensively utilize multi-dimensional data such as genome, epigenome and transcriptome to characterize disease status from multiple biological levels, overcoming the limitation of existing methods that only focus on a single group; (2) Use methods such as canonical correlation analysis to reveal the association patterns between different groups from a statistical perspective, and use them to guide feature selection and drug resistance prediction model construction, thereby improving the interpretability and robustness of the model; (3) Construct an end-to-end deep learning model, and through innovative strategies such as hierarchical progressive feature learning, adversarial learning and focused loss function optimization, while improving prediction accuracy, effectively alleviate problems such as high-dimensional small samples and class imbalance; (4) For patients predicted to be at high risk, intelligently recommend personalized treatment plans from a large number of candidate drugs to provide decision support for precision medicine use.

[0095] In summary, this invention uses artificial intelligence technology to build a full-process early drug resistance prediction scheme based on multi-omics data fusion. This scheme not only improves the prediction level, but also provides practical guidance for subsequent treatment. It has significant clinical application value and represents an important development direction of intelligent medical assisted diagnosis and treatment.

[0096] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, replacement, and improvement made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting early drug resistance in ALK fusion gene-positive non-small cell lung cancer, characterized in that: The following steps are involved: (1) Data collection and preprocessing: Collect multi-omics data of ALK-positive non-small cell lung cancer patients, including gene mutation data, epigenetic data, and transcriptome data, perform quality control, remove batch effects, and standardize the multi-omics data to form a unified feature matrix X; (2) Multi-omics association analysis: canonical correlation analysis was used to explore the association patterns among the multi-omics data and identify key features associated with drug resistance; (3) Deep learning prediction: The feature matrix X is input into the deep learning model, and the early drug resistance risk is predicted through hierarchical progressive feature learning, drug structure adversarial learning and focused loss function optimization; (4) Drug selection recommendation: Based on the early drug resistance risk prediction results and combined with the patient's molecular characteristics, the drug knowledge base is retrieved to recommend subsequent treatment drugs suitable for the patient; The hierarchical progressive feature learning in step (3) includes: Step 1: Extract low-level features within a single group: Use the autoencoder to train the i-th omics data X i Extract low-dimensional feature representation V i , the formula is: V i =σ(W 1i ·X i +b 1i ) Among them, W 1i and b 1i are the weight matrix and bias term of the encoder respectively, and σ is the activation function; Step 2: Attention mechanism learning omics interaction features: Use the attention mechanism to calculate the attention weight α of omics discussion omics j ij , the formula is: Among them, Q i =W 2i ·V i ,K i =W 3i ·V i ,W 2i and W 3i is the learnable weight matrix, d is Q i , K j The dimension of the interaction feature representation C of omics i i By weighted summation we get: Step 3: Fusion of multi-head self-attention mechanism: Multi-scale features use a multi-head self-attention mechanism for C1, C2, ..., C m After fusion, we get the multi-omics feature representation MF, the formula is: MF=W5·Concat(C i1 ,C i2 ,...,C ih )+b5 Among them, W5 and b5 are learnable weight matrices and bias terms; Assume that the molecular structure of the drug is FS, the resistant and sensitive cell lines are Sr and Ss respectively, and minimize the loss of resistance L adv , so that the feature representation MF can distinguish Sr and Ss, enhancing its drug resistance prediction ability. L adv =-E[logD(FS|Sr)]-E[log(1-D(FS|Ss))] Where D is the discriminator, which is used to distinguish whether FS comes from Sr or Ss.

2. The method according to claim 1, characterized in that: The optimization of the focus loss function in step (3) includes: using the focus loss function to deal with the sample imbalance problem and strengthen the model's attention to minority class samples. The loss function is defined as: L focal =-(1-p) γ ·ylog(p)-(p) γ ·(1-y)log(1-p) Among them, y∈{0,1} is the true label of the sample, p is the probability that the model predicts that the sample is a positive class, γ is the focusing factor, γ≥0.

3. The method according to claim 1, characterized in that ,The multi-omics data in step (1) include: gene mutation data, epigenetic data and transcriptome data, with a sample size of no less than 100 cases, and including early resistant and sensitive patient samples.

4. The method according to claim 1, characterized in that ,The quality control of multi-omics data in step (1) includes: using FastQC to assess the ,quality of sequencing data and using Trimmomatic to remove adapters and ,low-quality sequences.

5. The method according to claim 1, characterized in that ,The removal of batch effects of multi-omics data in step (1) includes: variant detection and annotation of ,gene mutation data, methylation level calculation of epigenetic data, ,expression quantification of transcriptome data, and batch correction.

6. The method according to claim 1, characterized in that The canonical correlation analysis in step (2) identifies the features that contribute most to drug resistance by maximizing the correlation coefficient of the linear combination of variables between different omics data. The optimization goal is:

7. The method according to claim 1, characterized in that The deep learning model in step (3) is trained using stochastic gradient descent, and early stopping and L2 regularization are used to avoid overfitting. The loss function is the weighted sum of focus loss and adversarial loss: L total =L focal +λ·L adv Where λ is the balancing factor, which controls the weight of the two losses.

8. The method according to claim 1, characterized in that ,The drug selection recommendation in step (4) includes: dividing the patients into high-risk and low-risk categories according to the predicted probability of early drug resistance, and retrieving drugs with similar molecular characteristics as candidates from a pre-constructed drug knowledge base ; the drug knowledge base contains the molecular structure, mechanism of action, toxicology and clinical data information of marketed and clinical trial drugs, and using chemical informatics methods to achieve drug comparison and retrieval.

Images