A method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection

Abstract

The present application relates to the technical field of mass spectrometry detection, and specifically discloses a method for screening tuberculosis-resistant substances based on mass spectrometry detection, which divides tuberculosis culture samples into a control group and a drug-treated group, extracts samples at different time points and performs rapid mass spectrometry detection to obtain multi-stage original mass spectrometry data. It is converted into structured visual coding features through feature extraction to reflect the specific response pattern of the strain under the action of drugs. In order to eliminate the baseline fluctuation interference caused by non-drug factors, a baseline fluctuation semantic correction strategy is introduced to quantify the systematic fluctuations based on the time change of the control group, and perform feature domain correction on the drug-treated group data, thereby stripping off background noise and highlighting the real drug response signal. Finally, the corrected features are used to distinguish the drug resistance category and output the confidence level, thereby improving the detection accuracy and reliability and providing a more solid technical support for the assessment of drug resistance.

Description

Technical Field

[0001] The present application relates to the technical field of mass spectrometry detection, and more specifically, to a method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection. Background Art

[0002] Tuberculosis (TB) is a chronic infectious disease caused by Mycobacterium tuberculosis (Mtb), a serious threat to human health. Despite the availability of a variety of anti-TB drugs for clinical treatment, the emergence and spread of drug-resistant TB, particularly multidrug-resistant TB (MDR-TB) and extensively drug-resistant TB (XDR-TB), has become a significant challenge to global TB prevention and control. The spread of drug-resistant strains not only increases treatment difficulty and failure rates, prolongs the infectious period, but also significantly increases treatment costs and the risk of mortality. Therefore, rapid and accurate screening and characterization of Mtb susceptibility to various anti-TB drugs, particularly identifying drug-resistant substances or resistant phenotypes, is crucial for guiding rational clinical drug use, curbing the spread of drug resistance, and improving patient outcomes. Therefore, the development of efficient and reliable screening protocols for drug-resistant substances in Mtb is urgently needed.

[0003] Currently, clinical methods for detecting Mtb drug resistance primarily include traditional phenotypic drug susceptibility testing (DST) and molecular biology-based genotyping. Phenotypic DSTs assess drug resistance by culturing Mtb in drug-containing culture media and observing its growth. However, their primary limitation is the slow growth of Mtb, with the entire testing cycle typically taking weeks or even months, making them inadequate for rapid early diagnosis and treatment adjustments in the clinic. Genotypic methods, such as gene chips and real-time PCR (e.g., GeneXpert MTB / RIF), assess drug resistance by detecting specific genetic mutations associated with drug resistance. These methods offer the advantage of rapidity, providing results within hours to days. However, these methods rely primarily on known resistance genes and their mutational profiles, and may fail to detect resistance caused by unknown mutations or non-genetic mechanisms (such as overexpression of drug efflux pumps). Furthermore, incomplete concordance between genotype and phenotype can lead to limitations in interpreting results.

[0004] Therefore, an optimized screening strategy for drug-resistant substances in M. tuberculosis is desired. Summary of the Invention

[0005] The present application is proposed to solve the above technical problems. The embodiments of the present application provide a method for screening drug-resistant substances in Mycobacterium tuberculosis based on mass spectrometry detection, which can use mass spectrometry technology to quickly capture the molecular changes in the strain's response to drugs, thereby achieving efficient and accurate drug resistance screening.

[0006] According to one aspect of the present application, a method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection is provided, comprising: dividing an Mtb culture into a control group and a drug-treated group; extracting a sample from the control group and processing and detecting the sample to obtain original mass spectrometry baseline sample data; adding an anti-tuberculosis drug to be tested to the drug-treated group, and after a preset time, extracting a sample from each of the control group and the drug-treated group and rapidly processing and detecting the sample to obtain original mass spectrometry baseline transformation sample data and reaction mass spectrometry sample data; determining original mass spectrometry baseline fluctuation visual features based on the original mass spectrometry baseline sample data and the original mass spectrometry baseline transformation sample data; performing feature domain-based baseline fluctuation semantic correction on the reaction mass spectrometry sample data based on the original mass spectrometry baseline fluctuation visual features to obtain baseline fluctuation semantic correction reaction mass spectrometry visual features; and determining the drug resistance category and its confidence level based on the baseline fluctuation semantic correction reaction mass spectrometry visual features.

[0007] In one possible implementation, the original mass spectrum baseline fluctuation visual feature is determined based on the original mass spectrum baseline sample data and the original mass spectrum baseline transformation sample data, including: extracting sample mass spectrum visual features from the original mass spectrum baseline sample data and the original mass spectrum baseline transformation sample data respectively to obtain an original mass spectrum baseline sample visual feature coding vector and an original mass spectrum baseline transformation visual feature coding vector; calculating a baseline fluctuation visual feature coding vector between the original mass spectrum baseline sample visual feature coding vector and the original mass spectrum baseline transformation visual feature coding vector as the original mass spectrum baseline fluctuation visual feature.

[0008] In one possible implementation, sample mass spectrum visual features are extracted from the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data to obtain an original mass spectrum baseline sample visual feature encoding vector and an original mass spectrum baseline transformed visual feature encoding vector, including: passing the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data through a sample mass spectrum visual feature extractor based on a void convolutional neural network to obtain the original mass spectrum baseline sample visual feature encoding vector and the original mass spectrum baseline transformed visual feature encoding vector.

[0009] In one possible implementation, based on the original mass spectrum baseline fluctuation visual feature, the reaction mass spectrum sample data is subjected to a feature domain-based baseline fluctuation semantic correction to obtain a baseline fluctuation semantically corrected reaction mass spectrum visual feature, including: extracting sample mass spectrum visual features from the reaction mass spectrum sample data to obtain a reaction mass spectrum visual feature encoding vector; based on the baseline fluctuation visual feature encoding vector, the reaction mass spectrum visual feature encoding vector is subjected to a baseline fluctuation semantic correction to obtain a baseline fluctuation semantically corrected reaction mass spectrum visual feature encoding vector as the baseline fluctuation semantically corrected reaction mass spectrum visual feature.

[0010] In one possible implementation, based on the baseline fluctuation visual feature coding vector, the reaction mass spectrum visual feature coding vector is subjected to baseline fluctuation semantic correction to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector as the baseline fluctuation semantic correction reaction mass spectrum visual feature, including: performing univariate local unit feature coupling on the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors; based on the characteristic distribution characteristics of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors, determining the recursive inference attention weights of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursive inference attention weights; based on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursive inference attention weights, performing weighted modulation inference on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector.

[0011] In one possible implementation, the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector are subjected to univariate local unit feature coupling to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors, including: performing one-dimensional local implicit feature convolution coding on the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector respectively to obtain a set of baseline fluctuation visual local implicit feature vectors and a set of reaction mass spectrum visual local implicit coding feature vectors; performing monomer feature interaction on each corresponding group of baseline fluctuation visual local implicit feature vectors and reaction mass spectrum visual local implicit coding feature vectors in the set of baseline fluctuation visual local implicit feature vectors and the set of reaction mass spectrum visual local implicit coding feature vectors respectively to obtain the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors.

[0012] In one possible implementation, based on the set of attention weights recursively inferred from the baseline fluctuation-reaction mass spectrum visual local implicit interaction, weighted modulation inference is performed on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector, including: based on the set of attention weights recursively inferred from the baseline fluctuation-reaction mass spectrum visual local implicit interaction, weighted modulation is performed on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors to obtain the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors; the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors is input into a chain inference engine based on a forward LSTM model to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector.

[0013] In one possible implementation, based on the set of attention weights of the baseline fluctuation-reaction mass spectrum visual local implicit interaction recursively inferred attention, the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors is weighted modulated to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors, including: performing chain inference-driven nonlinear interference correction on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursively inferred attention weights to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction chain inference attention optimization weights; using the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction chain inference attention optimization weights, weighted modulating the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors to obtain the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors.

[0014] The present application provides a method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection, which divides the Mycobacterium tuberculosis culture samples into a control group and a drug-treated group, extracts samples at different time points and performs rapid mass spectrometry detection to obtain multi-stage original mass spectrometry data. It is converted into structured visual coding features through feature extraction to reflect the specific response pattern of the strain under the action of drugs. In order to eliminate the baseline fluctuation interference caused by non-drug factors, a baseline fluctuation semantic correction strategy is introduced to quantify the systematic fluctuations based on the time change of the control group, and the feature domain correction is performed on the drug-treated group data to remove the background noise and highlight the real drug response signal. Finally, the corrected features are used to distinguish the drug resistance category and output the confidence level, thereby improving the detection accuracy and reliability and providing a more solid technical support for the assessment of drug resistance. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings of the embodiments of the present application. Obviously, the drawings described below only relate to some embodiments of the present application and are not intended to limit the present application.

[0016] Figure 1 Schematic flow chart of the method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to an embodiment of the present application.

[0017] Figure 2 This is a schematic flow chart of step S4 in the method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection in an embodiment of the present application.

[0018] Figure 3 This is a schematic flow chart of step S5 in the method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to an embodiment of the present application.

[0019] Figure 4 This is a schematic flow chart of step S52 in the method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to an embodiment of the present application.

[0020] Figure 5 4 is a schematic flow chart of step S521 in the method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to an embodiment of the present application.

[0021] Figure 6 This is a schematic flow chart of step S523 in the method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] The following will clearly and completely describe the technical solutions in the embodiments of the present application with reference to the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present application, not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by ordinary technicians in this field without making creative efforts also fall within the scope of protection of this application.

[0023] To overcome the limitations of existing technologies, mass spectrometry (MS)-based detection methods are considered a promising new approach for detecting tuberculosis drug resistance due to their high throughput, high sensitivity, relatively rapid speed, and potential to directly detect changes in molecules (such as proteins, metabolites, and lipids) in biological samples. By comparing the mass spectra of Mtb samples before and after drug treatment, it is theoretically possible to capture the distinct physiological or metabolic responses of susceptible and resistant strains to drugs, thereby enabling the determination of drug resistance. However, mass spectrometry is inherently susceptible to multiple factors, including experimental conditions, sample handling, and instrument status, leading to baseline fluctuations and noise, which can severely impact the stability and accuracy of test results. In particular, when comparing samples at different time points or between different treatment groups, changes in the mass spectrometry signal caused by non-drug factors (i.e., baseline fluctuations) can mask or distort true biological differences resulting from drug action.

[0024] To address the above technical issues, the technical solution of this application proposes a mass spectrometry-based method for screening drug-resistant Mycobacterium tuberculosis (Mtb) bacteria. This method utilizes mass spectrometry to rapidly capture molecular changes in the strain's response to drugs, thereby enabling efficient and accurate drug resistance screening. The overall process is as follows: First, Mtb bacteria are cultured and divided into an untreated control group and a drug-treated group that receives the anti-tuberculosis drug to be tested. Before the experiment and after a period of drug exposure, samples are extracted from both groups for rapid processing and high-sensitivity mass spectrometry analysis to obtain raw mass spectrometry data at different time points and under different treatment conditions. Subsequently, feature extraction is performed on these raw mass spectrometry data, converting them into structured visual encoding features representing the initial baseline state, the baseline change of the control group over the same period of time, and the response state of the drug-treated group. These different states are then used to reflect the specific response pattern of Mtb to the drug, which is closely related to the strain's drug resistance. This facilitates the identification of drug resistance categories, improves the accuracy and reliability of screening results, and provides confidence in judgments, providing more solid technical support for drug resistance assessment.

[0025] In particular, considering that mass spectrometry signals are susceptible to interference from non-biological factors and produce baseline fluctuations, such fluctuations will seriously affect the accuracy of result interpretation. Therefore, one of the core concepts of the present invention is to propose an accurate baseline correction strategy. Specifically, by comparing the mass spectrometry changes of the control group within the experimental time window, the systematic baseline drift or fluctuation is quantified, and this information is used to perform semantic correction on the reaction mass spectrometry data of the drug-treated group. In this way, the background noise can be stripped off and the specific signal truly caused by the interaction between the drug and the bacteria can be extracted. After this correction, the obtained baseline fluctuation semantically corrected reaction mass spectrum visual features can more purely and essentially reflect the specific response pattern of Mycobacterium tuberculosis to specific drugs, which is closely related to the drug-resistant phenotype of the strain.

[0026] Specifically, in the technical solution of this application, Figure 1 Schematic flow chart of the method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection in the embodiment of the present application. Figure 1 As shown, the method for screening tuberculosis drug-resistant substances based on mass spectrometry detection includes: S1, dividing the Mtb culture into a control group and a drug-treated group; S2, extracting a sample from the control group and processing and detecting it to obtain original mass spectrometry baseline sample data; S3, adding the anti-tuberculosis drug to be tested to the drug-treated group, and after a preset time, extracting a sample from the control group and the drug-treated group respectively and quickly processing and detecting them to obtain original mass spectrometry baseline transformation sample data and reaction mass spectrometry sample data; S4, determining the original mass spectrometry baseline fluctuation visual features based on the original mass spectrometry baseline sample data and the original mass spectrometry baseline transformation sample data; S5, based on the original mass spectrometry baseline fluctuation visual features, performing feature domain-based baseline fluctuation semantic correction on the reaction mass spectrometry sample data to obtain baseline fluctuation semantic correction reaction mass spectrometry visual features; S6, determining the drug resistance category and its confidence based on the baseline fluctuation semantic correction reaction mass spectrometry visual features.

[0027] Specifically, in step S1, Mtb cultures are divided into a control group and a drug-treated group. It should be understood that mass spectrometry signals are highly susceptible to non-drug-related factors (including but not limited to changes in the experimental environment, fluctuations in instrument status, minor deviations in sample processing, and the natural physiological evolution of the bacteria), which can manifest as baseline drift or noise. If not scientifically isolated and corrected for, these systemic or accidental background fluctuations can severely obscure or confound the true response induced by drug exposure, ultimately compromising the accurate identification of drug-resistant phenotypes. Therefore, dividing Mtb cultures into a control group and a drug-treated group is an advanced strategy for rigorous experimental control design, eliminating systemic interference, and accurately distinguishing drug-specific responses from basal changes. The control group reflects the baseline molecular expression and changes of the bacterial population under identical culture conditions, without drug stimulation, while the drug-treated group reflects the full range of physiological and molecular changes induced by drug exposure under defined conditions. By comparing mass spectrometry data from the two groups acquired at the same time point, signal variation caused by non-drug factors can be effectively quantified and isolated, thereby enabling higher confidence in drug resistance interpretation.

[0028] In a specific embodiment, a Mycobacterium tuberculosis strain obtained from a clinical or experimental setting is first cultured in a suitable culture medium for a period of time. After the cell density reaches a predetermined standard (e.g., OD600 = 0.5), the culture is divided equally into two groups. One group serves as a control group (i.e., a group without drug addition), and the other group serves as a drug-treated group (i.e., a group treated with a predetermined concentration of the anti-tuberculosis drug to be screened at the same volume or cell density).

[0029] Specifically, in step S2, a sample is extracted from the control group or the drug treatment group and processed and tested to obtain original mass spectrometry baseline sample data. It should be understood that mass spectrometry, as a highly sensitive molecular detection technology, not only indicates the molecular composition of the biological sample, but is also affected by various non-biological variables such as experimental environment, instrument status, culture conditions and even technical fluctuations in sample extraction and processing. Therefore, the mass spectrogram at any moment contains the background molecular composition information of the biological sample and the above-mentioned various complex background noises. If there is no authoritative and reliable reference baseline, that is, the original mass spectrometry baseline sample data, it is difficult to effectively identify the real difference between the subsequent drug response and non-specific background changes. By collecting the control group sample and obtaining its mass spectrometry data in the earliest stage of the experiment, before any drug intervention is applied, a standard baseline can be provided for all subsequent analyses, so that the mass spectrometry changes at different time points or under different treatment conditions can always be compared with the basic state, which is crucial for distinguishing drug-specific reactions from natural physiological fluctuations, system baseline drift, etc.

[0030] In one specific embodiment, at the start of the experiment (denoted as D0), a 1 mL sample was aseptically collected from the control group, and the sample processing process was immediately implemented. The processing process includes centrifugation to collect the bacterial cells, washing with pre-chilled PBS buffer to remove impurities from the culture medium, and then ultrasonic or chemical lysis using a standard lysis buffer to release intracellular proteins and metabolites. After protein determination and quantitative normalization, the processed sample undergoes enzymatic hydrolysis or relevant molecular derivatization steps (such as preparation of protein or metabolite adapters) and is directly analyzed by a high-resolution mass spectrometer (such as Q-TOF or Orbitrap LC-MS). The raw mass spectrometric signal is collected and saved as a standard format data file, i.e., the raw mass spectrometry baseline sample data. The raw mass spectrometry baseline sample data obtained through this operation serves as the molecular profile at time zero for all subsequent analyses, containing the biomolecular fingerprint of the strain in the absence of any drug or environmental interference. In particular, in other examples of the present application, a sample can also be extracted from the drug-treated group and processed and tested to obtain original mass spectrometry baseline sample data. It should be understood that at the start of the experiment, the drug-treated group has not yet been added with drugs. Therefore, the drug-treated group and the control group at this moment are still in the same state. Therefore, the data obtained by processing and monitoring them can also be used as original mass spectrometry baseline sample data.

[0031] Specifically, in step S3, the anti-TB drug to be tested is added to the drug-treated group. After a preset time, a sample is extracted from each of the control and drug-treated groups and rapidly processed and tested to obtain raw mass spectrometry baseline-shifted sample data and reaction mass spectrometry sample data. It should be understood that when mass spectrometry is used to detect anti-TB drug resistance, the molecular-level response of cells or bacteria to drug action truly reflects the essential differences between resistant and sensitive strains. For example, in Mycobacterium tuberculosis (Mtb), different strains exhibit significant differences in molecular metabolites, protein expression profiles, lipid composition, and other aspects under drug challenge. These differences can be intuitively captured by high-resolution mass spectrometry. To accurately capture these differences, it is necessary to eliminate variation caused by non-drug factors within the system, including experimental environment perturbations, instrument signal drift, sample batch effects, and natural bacterial growth or decay. This is why samples from drug-treated and corresponding control groups are collected simultaneously and analyzed at multiple time points before and after drug exposure. Only synchronous, controlled sampling and testing can effectively distinguish between true physiological or molecular model changes caused by drugs, rather than mistakenly confusing time-varying background fluctuations and system offsets as dominant characteristics of drug resistance. Therefore, by sampling the control group at the same time, establishing the original mass spectrometry baseline transformation sample data, and then synchronously measuring the reaction mass spectrometry sample data of the drug group, the difference mapping between the two can reasonably and rigorously represent the specific effect of the drug. This design effectively solves the false positives and false negatives caused by baseline fluctuations in existing mass spectrometry detection, and becomes the basis for screening for drug-resistant substances or distinguishing drug-resistant phenotypes of Mycobacterium tuberculosis.

[0032] In one specific embodiment, the control group is maintained under conventional culture conditions, while the drug-treated group receives a pre-set concentration of the anti-TB drug to be tested at a set starting time (T0). The drug concentration should refer to authoritative medical standards such as the WHO-recommended drug susceptibility cutoff point / minimum inhibitory concentration (MIC), ensuring that it can induce physiological responses in sensitive strains while effectively tracing the stable metabolic state of drug-resistant strains. Typical drugs include rifampin, isoniazid, and ethambutol, with a typical concentration of 1–10 μg / mL, and the duration of action can be set to 2 hours, 8 hours, 24 hours, etc.

[0033] The timing of sample collection is crucial. A short-term physiological stress response is typically considered 2 to 8 hours after drug addition, and samples are collected simultaneously with the corresponding time point for the control group. The principle of sampling is to rapidly terminate bacterial metabolic activity to prevent drug reaction withdrawal or prolongation, which could affect true signal capture. Subsequent reactions are typically inhibited by rapid ice baths, the addition of chemical terminators (e.g., methanol), or rapid freezing and centrifugation. The following steps can be employed: At the designated time point, equal volumes of bacterial culture (e.g., 1 mL) are collected from both the control and drug groups, immediately cooled in an ice bath, and then briefly centrifuged at 4°C to collect the cells. The pellet is washed with pre-chilled PBS to remove residual culture medium and exogenous interfering substances. Following washing, a lysis buffer containing denaturants, detergents, or mechanical sonication is used to disrupt the cell membrane and release intracellular molecules. Protein samples can then be further processed for protein quantification, reduction and alkylation, or enzymatic digestion (e.g., with trypsin) to generate peptide fragments. Metabolite samples can be extracted using a methanol-water system and clarified by centrifugation to obtain a metabolite mixture. All extracts were filtered through a 0.22 μm filter membrane and then directly analyzed on the machine.

[0034] Detection is performed using a liquid chromatography-tandem mass spectrometry (LC-MS / MS) system. Chromatographic conditions and gradient elution schemes should be optimized for the target molecule. Samples are added according to standard quantitative standards, and the mass spectrometer scan range is set (e.g., m / z 100-2000). Data acquisition is performed using full scan or target ion screening. To improve data timeliness and analytical efficiency, some analyses can utilize single-shot sampling, parallel automated injection, and automated data processing software to perform continuous testing of multiple samples through batch processing. After data acquisition, the raw mass spectrometry data is exported and converted to a unified format after system quality control and comparison with standards. This stage obtains raw mass spectrometry baseline shift sample data (corresponding to control group samples at the same time) and reaction mass spectrometry sample data (corresponding to drug group samples at the same time).

[0035] Specifically, in step S4, based on the original mass spectrum baseline sample data and the original mass spectrum baseline conversion sample data, the original mass spectrum baseline fluctuation visual characteristics are determined. It should be understood that, considering that the original mass spectrum data is usually a high-dimensional, complex signal sequence, which contains both the molecular fingerprint of the biological sample itself and a large amount of noise, baseline drift and systematic errors unrelated to the biological state change. Directly processing the original spectrogram for comparison and analysis is difficult and not robust. At the same time, the response of Mycobacterium tuberculosis to the drug may be reflected in a variety of subtle changes in the mass spectrum, such as the appearance / disappearance of a specific peak, intensity changes, peak shape changes or changes in the relative relationship between multiple peaks, and this information is often presented in the spectrogram in the form of a visual or pattern. Therefore, an effective method is needed to extract essential information that is biologically meaningful and useful for subsequent analysis from these raw data. Based on this, in the technical solution of the present application, based on the original mass spectrum baseline sample data and the original mass spectrum baseline conversion sample data, the original mass spectrum baseline fluctuation visual characteristics are determined.

[0036] In one embodiment, Figure 2 As shown, based on the original mass spectrum baseline sample data and the original mass spectrum baseline transformation sample data, the original mass spectrum baseline fluctuation visual feature is determined, including: S41, extracting sample mass spectrum visual features from the original mass spectrum baseline sample data and the original mass spectrum baseline transformation sample data respectively to obtain an original mass spectrum baseline sample visual feature coding vector and an original mass spectrum baseline transformation visual feature coding vector; S42, calculating the baseline fluctuation visual feature coding vector between the original mass spectrum baseline sample visual feature coding vector and the original mass spectrum baseline transformation visual feature coding vector as the original mass spectrum baseline fluctuation visual feature.

[0037] Specifically, in step S41, sample mass spectrometry visual features are extracted from the raw mass spectrometry baseline sample data and the raw mass spectrometry baseline transformed sample data, respectively, to obtain a raw mass spectrometry baseline sample visual feature encoding vector and a raw mass spectrometry baseline transformed visual feature encoding vector. This operation aims to convert the raw, noisy mass spectrometry signal into a low-dimensional, compact, and information-rich encoding representation. These visual feature encoding vectors represent a visual or pattern-based abstraction of the raw mass spectrometer. By extracting and encoding visual features, the goal is to filter out some noise and capture key patterns in the mass spectrometer that reflect the sample's molecular composition and state, such as important metabolite or protein marker signals, their relative abundance relationships, or other complex signal patterns. Specifically, feature extraction from the raw baseline data establishes a reference for the initial state; feature extraction from the baseline transformed data (data from the control group at the same time point) quantifies and captures nonspecific baseline fluctuations or biological background variation patterns that occur solely due to the passage of time or experimental manipulation in the absence of drug administration.

[0038] In one embodiment, sample mass spectrum visual features are extracted from the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data to obtain an original mass spectrum baseline sample visual feature encoding vector and an original mass spectrum baseline transformed visual feature encoding vector, respectively, including: passing the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data through a sample mass spectrum visual feature extractor based on a dilated convolutional neural network to obtain the original mass spectrum baseline sample visual feature encoding vector and the original mass spectrum baseline transformed visual feature encoding vector. Dilated convolutional neural network (Dilated CNN) is selected as a feature extraction tool because CNN is good at processing patterns in image data, and dilated convolution can effectively expand the receptive field and capture peak-to-peak correlations across a certain interval in the mass spectrum or more macroscopic spectral shape features, thereby more comprehensively and deeply understanding complex mass spectrum data.

[0039] In a specific embodiment, the network structure of the sample mass spectrum visual feature extractor based on the dilated convolutional neural network is as follows: the input layer corresponds to the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data, followed by 3-5 layers of dilated convolution, each layer of convolution kernel size is 64, and the dilation rate increases layer by layer (such as dilation rate = 1, 2, 4, 8) to capture the local details and global patterns of the signal. At the network output stage, a fully connected layer is used to downsample the convolution feature vector to a 128-dimensional compressed feature. The outputs are recorded as the original mass spectrum baseline sample visual feature encoding vector and the original mass spectrum baseline transformed visual feature encoding vector, respectively. Of course, this is only an example, and those skilled in the art can adjust the network architecture according to actual conditions.

[0040] Furthermore, because mass spectrometry is susceptible to non-drug factors (such as the experimental environment, instrument stability, natural changes in samples over time, and changes in culture medium components over time), its baseline or overall signal intensity can experience non-specific fluctuations or drift during the experiment. This fluctuation is not caused by the interaction between Mtb and the drug, but it is reflected in the mass spectrum. If the drug group is directly compared with the starting state or simply compared with the absolute difference between the drug group and the control group at a certain time point, the fluctuation caused by these non-drug factors will obscure the true drug effect signal, seriously affecting the accuracy of drug resistance judgment. To accurately assess the true impact of drugs on Mtb, these non-specific baseline fluctuations must be separated from the drug-induced response signal. Therefore, in order to quantify this time-dependent, drug-independent baseline fluctuation, the technical solution of this application further calculates a baseline fluctuation visual feature coding vector between the original mass spectrum baseline sample visual feature coding vector and the original mass spectrum baseline transformed visual feature coding vector. By comparing the mass spectral features (original mass spectrum baseline sample visual feature encoding vector) obtained at the start of the experiment in the untreated control group with the mass spectral features (original mass spectrum baseline transformed sample visual feature encoding vector) obtained at the same detection time point as the drug group, the mass spectral feature change pattern caused by all factors other than the drug effect during this period can be accurately captured. This difference or change, namely the calculated baseline fluctuation visual feature encoding vector, represents the projection of background noise, instrument drift, or natural physiological changes of the strain during normal growth in the feature space. In a specific embodiment, the feature-by-feature difference between the original mass spectrum baseline sample visual feature encoding vector and the original mass spectrum baseline transformed visual feature encoding vector is calculated to obtain the baseline fluctuation visual feature encoding vector.

[0041] Specifically, in step S5, based on the original mass spectrum baseline fluctuation visual features, the reaction mass spectrum sample data is subjected to a feature domain-based baseline fluctuation semantic correction to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature. It should be understood that since the original reaction mass spectrum visual feature coding vector contains the true biological response signal caused by the action of anti-tuberculosis drugs, it is also inevitably superimposed with the effects of baseline fluctuations caused by non-drug factors such as the experimental environment, instrument drift, sample processing, or natural changes in the bacteria themselves during the same time period. As quantified by the calculated baseline fluctuation visual feature coding vector, these non-specific fluctuations are objectively existing sources of interference. If not effectively processed, these baseline fluctuations will confuse the true drug response pattern, making it unreliable and inaccurate to judge drug resistance directly based on the uncorrected reaction feature vector, which may lead to false positive or false negative results. Therefore, in the technical solution of the present application, the reaction mass spectrum visual feature coding vector is further subjected to baseline fluctuation semantic correction based on the baseline fluctuation visual feature coding vector, thereby removing or suppressing the influence of the baseline fluctuation, so that the baseline fluctuation semantically corrected reaction mass spectrum visual feature coding vector obtained after correction can more purely and accurately reflect the specific molecular or physiological changes produced by the interaction between the drug and Mycobacterium tuberculosis.

[0042] In one embodiment, Figure 3 As shown, based on the original mass spectrum baseline fluctuation visual feature, the reaction mass spectrum sample data is subjected to a baseline fluctuation semantic correction based on a feature domain to obtain a baseline fluctuation semantically corrected reaction mass spectrum visual feature, including: S51, extracting sample mass spectrum visual features from the reaction mass spectrum sample data to obtain a reaction mass spectrum visual feature encoding vector; S52, based on the baseline fluctuation visual feature encoding vector, performing a baseline fluctuation semantic correction on the reaction mass spectrum visual feature encoding vector to obtain a baseline fluctuation semantically corrected reaction mass spectrum visual feature encoding vector as the baseline fluctuation semantically corrected reaction mass spectrum visual feature.

[0043] Specifically, in step S51, the sample mass spectrum visual features are extracted from the reaction mass spectrum sample data to obtain a reaction mass spectrum visual feature encoding vector. It should be understood that the purpose of extracting features from the reaction mass spectrum data is to capture the physiological or molecular response pattern produced by the strain under the action of the drug. The reaction mass spectrum visual feature encoding vector is the cornerstone of subsequent baseline correction and final drug resistance judgment. In a specific embodiment, the reaction mass spectrum sample data is passed through the sample mass spectrum visual feature extractor based on the void convolutional neural network to obtain the reaction mass spectrum visual feature encoding vector.

[0044] Specifically, in step S52, based on the baseline fluctuation visual feature encoding vector, the reaction mass spectrum visual feature encoding vector is subjected to baseline fluctuation semantic correction to obtain a baseline fluctuation semantically corrected reaction mass spectrum visual feature encoding vector as the baseline fluctuation semantically corrected reaction mass spectrum visual feature. Specifically, the baseline fluctuation semantic correction process deconstructs the complex interaction between the baseline fluctuation global visual feature and the reaction mass spectrum global visual feature into a series of local interaction units, and aggregates this interaction influence information layer by layer through chain reasoning. This means that it does not simply subtract the baseline vector from the reaction mass spectrum visual feature encoding vector, but rather learns and models how the baseline fluctuation pattern semantically influences, superimposes, or alters the actual reaction pattern, thereby performing a more intelligent correction that is more adaptable to complex nonlinear relationships. More specifically, a one-dimensional convolution is first used to capture the local implicit features of the two input feature vectors (the baseline fluctuation visual feature encoding vector and the reaction mass spectrum visual feature encoding vector). Then, a single feature interaction engine is used to perform interaction modeling at the local level. This fine-grained interaction modeling can capture subtle interference patterns that baseline fluctuations may cause in specific local regions or peak shapes of the mass spectrum. Furthermore, by introducing chained reasoning attention and modifying the attention weight (based on the spatial interaction decomposition of interaction norms, etc.), the encoder can intelligently judge the importance of different local interaction information in the entire correction process, and weightedly integrate this information. Finally, serialized reasoning is performed through LSTM to generate a comprehensive baseline fluctuation semantic correction reaction mass spectrum visual feature encoding vector.

[0045] In one embodiment, Figure 4 As shown, in step S52, based on the baseline fluctuation visual feature coding vector, the reaction mass spectrum visual feature coding vector is subjected to baseline fluctuation semantic correction to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector as the baseline fluctuation semantic correction reaction mass spectrum visual feature, including: S521, performing single variable local unit feature coupling on the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors; S522, based on the baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors According to the characteristic distribution characteristics of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector in the set of vectors, the recursive inference attention weights of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector are determined to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursive inference attention weights; S523, based on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursive inference attention weights, the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors is weighted modulated inference to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector.

[0046] In one embodiment, Figure 5 As shown, in step S521, the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector are subjected to univariate local unit feature coupling to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors, including: S5211, the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector are subjected to one-dimensional local implicit feature convolution coding to obtain a set of baseline fluctuation visual local implicit feature vectors and a set of reaction mass spectrum visual local implicit coding feature vectors, which are expressed as:

[0047] ;

[0048] ;

[0049] in, is the baseline fluctuation visual feature encoding vector, is the reaction mass spectrum visual feature encoding vector, is a one-dimensional local latent feature convolutional coding, is the length of the one-dimensional convolution kernel, , , and They are the first, second, and third in the set of baseline fluctuation visual local implicit feature vectors. and Baseline fluctuation visual local latent feature vector, , , and They are the first, second, and third in the set of local implicit coding feature vectors of the reaction mass spectrum vision. and The local implicit encoding feature vector of the reaction mass spectrum visual yes and The number of vectors in and Same length.

[0050] S5212, performing monomer feature interaction on each corresponding group of baseline fluctuation visual local implicit feature vectors and reaction mass spectrum visual local implicit coding feature vectors in the set of baseline fluctuation visual local implicit feature vectors and the set of reaction mass spectrum visual local implicit coding feature vectors to obtain the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors, expressed as:

[0051] ;

[0052] in, is the first in the set of baseline fluctuation visual local latent feature vectors Baseline fluctuation visual local latent feature vector, is the first in the set of local implicit coding feature vectors of the reaction mass spectrum vision The local implicit encoding feature vector of the reaction mass spectrum visual It is the point product of position. It is added by position point. It is subtracted by position point, It is a cascade operation. is the set of implicit feature interaction response weight matrices implicit feature interaction response weight matrix, is the set of implicit feature interaction response bias vectors implicit feature interaction response bias vector, is the first in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vectors Baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vector.

[0053] Specifically, in the process of performing one-dimensional local implicit feature convolution encoding on the baseline fluctuation visual feature encoding vector and the reaction mass spectrum visual feature encoding vector respectively, the two feature vectors are first processed by a one-dimensional convolution operation to obtain a set of local implicit feature vectors representing the baseline fluctuation and the reaction mass spectrum respectively. This convolution encoding operation can effectively capture the structured local patterns within the feature vectors and extract the key local information in the mass spectrometry signal. The convolution kernel slides on the feature vector space, so that each local implicit feature vector can reflect the pattern and changes of the original signal within a certain receptive field, which greatly enhances the model's sensitivity to local signals in different spaces and lays a clear structural and hierarchical foundation for subsequent feature interaction links. Each local implicit feature vector itself is more expressive than the original input feature, and can focus on reflecting the subtle details of the peak shape, intensity changes and local correlation of a local segment in the mass spectrometry data.

[0054] Next, the corresponding vectors from the set of baseline fluctuation visual local implicit feature vectors and the set of reaction mass spectrum visual local implicit encoding feature vectors are paired and input into the monomer feature interaction link. A variety of orthogonal or synergistic interaction mechanisms are implemented through operators such as Hadamard products, point addition, point subtraction, or feature cascades. This monomer feature interaction is not just a superficial linear or nonlinear transformation, but also a systematic revelation of complex dependencies between local features, such as similarity, co-activation, and mutual inhibition. The effect of the interaction step is to accurately reveal the coupling relationship and response differences between the response induced by drug treatment and baseline fluctuations in each local region, effectively highlighting which scales of signal changes are driven by drug action and which are products of system background fluctuations. Each baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vector essentially condenses the dynamic association and information fusion between the baseline and reaction features in the same interval.

[0055] Specifically, in step S522, based on the characteristic distribution characteristics of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors, the recursive inference attention weight of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector is determined to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursive inference attention weights, which is expressed as:

[0056] ;

[0057] in, yes Middle eigenvalues, To calculate the square of the Euclidean norm of a vector, yes The number of eigenvalues ​​in , yes function, It is the first in the set of baseline fluctuation-response mass spectrometry visual local implicit interaction recursive inference attention weights Baseline fluctuation-response mass spectrometry visual local implicit interaction recursively infers attention weights.

[0058] Specifically, a recursive mechanism is used to generate recursive inference attention weights for each baseline fluctuation-response mass spectrometry visual local implicit interaction encoding vector. This step comprehensively considers the relative importance of each interaction response pair in the entire sequence and dynamically allocates attention based on its statistical distribution characteristics, covariation patterns, and local normative effects. The recursive inference mechanism can combine prior statistical differences with the sequential dependencies of the context. Through iterative optimization of attention weights, signal units that truly reflect drug-TB interactions maintain high weights in the global model, while background noise and non-specific components are gradually suppressed during the aggregation process.

[0059] In one embodiment, Figure 6 As shown, in step S523, based on the set of attention weights recursively inferred from the baseline fluctuation-reaction mass spectrum visual local implicit interaction, the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors is weighted modulated and inferred to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector, including: S5231, based on the set of attention weights recursively inferred from the baseline fluctuation-reaction mass spectrum visual local implicit interaction, the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors is weighted modulated to obtain the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors; S5232, the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors is input into a chain inference engine based on a forward LSTM model to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector.

[0060] In one embodiment, based on the set of attention weights recursively inferred from the baseline fluctuation-reaction mass spectrum visual local implicit interaction, weighted modulation is performed on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vectors to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation encoding vectors, including: performing chain inference-driven nonlinear interference correction on the set of attention weights recursively inferred from the baseline fluctuation-reaction mass spectrum visual local implicit interaction to obtain a set of chain inference-driven attention optimization weights for the baseline fluctuation-reaction mass spectrum visual local implicit interaction, expressed as:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] in, is the first in the set of baseline fluctuation visual local latent feature vectors Baseline fluctuation visual local latent feature vector, is the first in the set of local implicit coding feature vectors of the reaction mass spectrum vision The local implicit encoding feature vector of the reaction mass spectrum visual It is the first in the set of baseline fluctuation-response mass spectrum visual gradient covariation correlation factors. Baseline fluctuation-response mass spectrum visual gradient covariation correlation factor, It is the first in the set of baseline fluctuation-reaction mass spectrum visual feature deviation factors. Baseline fluctuations - response mass spectrum visual feature deviation factors, It is the first in the set of baseline fluctuation-reaction mass spectrum visual fluctuation energy level density. Baseline fluctuation-reaction mass spectrum visual fluctuation energy level density, The natural constant The logarithmic function value with base , is the first in the set of baseline fluctuation-response mass spectrometry visual cycle regularization compensation factors. Baseline fluctuation-response mass spectrometry visual period regularization compensation factor, is the first term in the set of baseline fluctuation-response mass spectra visual cycle induced phase evolution terms. Baseline fluctuation-response mass spectrum visual cycle induced phase evolution term, and They are and The corresponding weight coefficient is, It is the first in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction chain reasoning attention optimization weights Baseline fluctuation-response mass spectrometry for visual local implicit interaction chain reasoning and attention optimization weights.

[0066] The set of the baseline fluctuation-reaction mass spectrum visual local implicit interaction chain inference attention optimization weights is used to weighted modulate the set of the baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors to obtain the set of the baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors, which is expressed as:

[0067] ;

[0068] ;

[0069] in, It is the first in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction chain reasoning attention optimization weights Baseline fluctuation-response mass spectrum visual local implicit interaction chain reasoning attention optimization weight, is the first in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vectors Baseline fluctuation-response mass spectrum visual local implicit interaction encoding vector, , , and They are the first, second, and third in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors. and Baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation encoding vector, It is the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation encoding vectors.

[0070] Specifically, a chain-inference-driven nonlinear interference correction is performed on the set of recursively inferred attention weights for baseline fluctuation-response mass spectrometry visual local implicit interactions. The model first focuses on removing nonbiological noise and complex background perturbations from the mass spectral feature sequence. These noises, including batch variation, instrument baseline drift, and local deviations in sample preparation, often affect local attention allocation in a nonlinearly mixed manner. During the correction process, a chain-inference architecture is employed to perform nonlinear transformations and feedback recursion on the initially calculated recursively inferred attention weights, incorporating contextual dependencies and contextual relationships between local features in the sequence. By incorporating recursive mechanisms such as residual mapping, multi-layer perception, period regularization, and phase perturbation correction, the model automatically identifies attention jumps caused solely by noise, as well as anomalous changes under various canonical properties (such as gradient covariance, energy level density, feature deviation, period compensation, and phase transition), thereby smoothing, weakening, or normalizing these weights. This process integrates chain-information and dynamically adjusts each attention weight, resulting in a value that better represents the actual contribution and biological credibility of the corresponding local interaction unit in the global judgment. The final output weight set is the set of attention optimization weights for the baseline fluctuation-reaction mass spectrometry visual local implicit interaction chain reasoning. Its spatial distribution fully maps the true state of the signal after noise removal and information highlighting, and has a guiding optimization effect on the global feature flow.

[0071] The optimized set of chain inference attention weights is used to weighted modulate the baseline fluctuation-response mass spectrometry visual local implicit interaction encoding vectors. This is essentially a feature weighting process aimed at information aggregation and signal-to-noise ratio improvement. Each local implicit interaction feature vector is linearly or nonlinearly scaled according to its corresponding optimized weight. Important feature segments that truly reflect the response of drug stimulation or drug-resistant molecules are given higher weights, thereby gaining opportunities for amplification and priority expression in subsequent chain processing or global expression, while regions with high noise and low importance are significantly suppressed or compressed. This weighted modulation not only ensures that the model output focuses on drug resistance-related signals, but also shields noise and potential biases, improving the discriminative ability and robustness of downstream global feature inference.

[0072] Specifically, in step S5232, the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors is input into the chain inference engine based on the forward LSTM model to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector, which is expressed as:

[0073] ;

[0074] in, is the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation encoding vectors, is the forward LSTM encoding, is the baseline fluctuation semantic correction response mass spectrum visual feature encoding vector.

[0075] Specifically, after preliminary feature extraction, interaction modeling, and meticulous weighted modulation, the corresponding mass spectrometry segment already contains information about the biological response and background noise, representing a unit that deeply expresses local features. Because mass spectrometry signals exhibit temporal correlations in both structural and physiological terms, these local feature vectors are not only independent of each other but also possess complex internal dependencies and dynamic evolution.

[0076] The LSTM network, with its recursive structure and gating mechanism, fully leverages this environment, performing forward recursive processing on the entire sequence of encoded vectors. The model's input, forget, and output gates precisely control feature absorption and filtering at each step according to the temporal nature of the mass spectrometry data, effectively preventing the accumulation of noise and irrelevant information while meticulously preserving key biological interaction traces. The hidden state at each moment integrates current input with historical information to model long-term dependencies and dynamic trends between different mass spectrometry segments, ensuring global coordination and temporal consistency within the feature sequence. Through this structure, the chained inference process achieves high-level abstraction from local to global features, effectively sorting out response signals at different scales and gradually removing residual effects of non-drug interference and baseline drift. The final output hidden state, serving as the visual feature encoding vector for the baseline fluctuation semantically corrected reaction mass spectra, deeply integrates all temporal linkages and nonlinear interactions inherent in the input sequence, resulting in a highly expressive, comprehensive feature that reflects the true biological response of the drug to Mycobacterium tuberculosis.

[0077] Specifically, in step S6, the drug resistance category and its confidence level are determined based on the baseline fluctuation semantically corrected mass spectrum visual features. It should be understood that the baseline fluctuation semantically corrected mass spectrum visual feature encoding vector obtained through the critical baseline fluctuation semantic correction process has largely eliminated complex baseline fluctuations and noise interference caused by non-drug factors such as the experimental environment, instrument drift, and sample processing. This corrected baseline fluctuation semantically corrected mass spectrum visual feature encoding vector is no longer a raw, noisy representation of mass spectrometry data, but rather a highly refined, pure, and direct representation of the true molecular response pattern of Mycobacterium tuberculosis to a specific anti-tuberculosis drug. It is the most reliable and representative information carrier for distinguishing between sensitive and resistant strains. Therefore, the drug resistance category and its confidence level are further determined based on the baseline fluctuation semantically corrected mass spectrum visual feature encoding vector. Specifically, an appropriate discriminant algorithm (e.g., a machine learning-based classifier that uses the corrected baseline fluctuation semantically corrected mass spectrum visual feature encoding vector as input) can be applied to identify and distinguish the drug resistance category of Mycobacterium tuberculosis. In addition, the confidence level is calculated when the drug resistance category is determined. The purpose is to evaluate the reliability of the discrimination results and provide a reference for decision-making for clinicians.

[0078] In one specific embodiment, a softmax multi-classifier is used to process the baseline fluctuation semantically corrected response mass spectral visual feature encoding vector to determine the drug resistance category and its confidence level. Specifically, the baseline fluctuation semantically corrected response mass spectral visual feature encoding vector is fed through the input layer into a fully connected layer. After undergoing linear transformation (weighting and biasing) in the fully connected layer, it is fed into the output layer, which employs a softmax multi-classification architecture to output a posterior probability distribution for each category. The model's final classification result prioritizes the category corresponding to the highest probability, and the probability value for that category is output as the judgment confidence level. The confidence level reflects the "distance" or degree of similarity between the sample's corrected features and the inherent category patterns. A higher confidence level indicates greater confidence in the model's judgment. If all confidence levels are low, the sample's response pattern does not match any known category in the training set. This can be considered suspicious / unclassified, aiding clinical decision-making.

[0079] During model training, a large number of clinical or standard tuberculosis samples with known phenotypes are collected and processed through the same process to obtain confidence scores and corresponding labels (such as "susceptible," "isoniazid-resistant," "rifampicin-resistant," and "multidrug-resistant"). Training uses the cross-entropy loss function, combined with mini-batch gradient descent to optimize model parameters. K-fold cross-validation and early stopping are also employed to prevent overfitting. To address class imbalance, loss function weights can be adjusted or data augmentation techniques can be employed to enhance the model's predictive power for rare drug-resistant phenotypes.

[0080] In summary, the present application provides a method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection. The method divides the Mycobacterium tuberculosis culture samples into a control group and a drug-treated group, extracts samples at different time points and performs rapid mass spectrometry detection to obtain multi-stage original mass spectrometry data. It is converted into structured visual coding features through feature extraction to reflect the specific response pattern of the strain under the action of drugs. In order to eliminate the baseline fluctuation interference caused by non-drug factors, a baseline fluctuation semantic correction strategy is introduced to quantify the systematic fluctuations based on the time change of the control group, and the feature domain correction is performed on the drug-treated group data to remove the background noise and highlight the real drug response signal. Finally, the corrected features are used to distinguish the drug resistance category and output the confidence level, thereby improving the detection accuracy and reliability and providing a more solid technical support for the assessment of drug resistance.

[0081] The basic principles of the present invention have been described above in conjunction with specific embodiments. However, it should be noted that the advantages, strengths, and effects mentioned in the present invention are merely illustrative and non-limiting, and should not be construed as necessarily possessed by each embodiment of the present invention. Furthermore, the specific details of the above embodiments are provided for illustrative purposes and to facilitate understanding, and are not intended to be limiting. These details do not necessarily limit the present invention to being implemented using these specific details.

[0082] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be encompassed therein. Any reference to a figure in a claim should not be construed as limiting the claim to which it relates.

[0083] In addition, it is obvious that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units recited in a device claim can also be implemented by one unit through software or hardware.

[0084] Finally, it should be noted that the above description has been provided for purposes of illustration and description. Furthermore, the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to be limiting. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art will appreciate that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection, characterized in that: include: Mtb cultures were divided into control and drug-treated groups; Extract a sample from the control group or the drug-treated group, process and detect the sample to obtain raw mass spectrometry baseline sample data; Adding the anti-tuberculosis drug to be tested to the drug-treated group, and after a preset time, extracting a sample from each of the control group and the drug-treated group, and rapidly processing and testing the samples to obtain raw mass spectrum baseline shift sample data and reaction mass spectrum sample data; Determining original mass spectrum baseline fluctuation visual features based on the original mass spectrum baseline sample data and the original mass spectrum baseline transformation sample data; Extracting sample mass spectrum visual features from the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data respectively to obtain an original mass spectrum baseline sample visual feature coding vector and an original mass spectrum baseline transformed visual feature coding vector; Calculating a baseline fluctuation visual feature coding vector between the original mass spectrum baseline sample visual feature coding vector and the original mass spectrum baseline transformation visual feature coding vector as the original mass spectrum baseline fluctuation visual feature; Extracting sample mass spectrum visual features from the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data to obtain an original mass spectrum baseline sample visual feature coding vector and an original mass spectrum baseline transformed visual feature coding vector, respectively, including: passing the original mass spectrum baseline sample data and the original mass spectrum baseline transformed sample data through a sample mass spectrum visual feature extractor based on a dilated convolutional neural network to obtain the original mass spectrum baseline sample visual feature coding vector and the original mass spectrum baseline transformed visual feature coding vector; Based on the original mass spectrum baseline fluctuation visual feature, the reaction mass spectrum sample data is subjected to a baseline fluctuation semantic correction based on a feature domain to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature, including: extracting a sample mass spectrum visual feature from the reaction mass spectrum sample data to obtain a reaction mass spectrum visual feature encoding vector; based on the baseline fluctuation visual feature encoding vector, the reaction mass spectrum visual feature encoding vector is subjected to a baseline fluctuation semantic correction to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature encoding vector as the baseline fluctuation semantic correction reaction mass spectrum visual feature; The drug resistance category and its confidence level are determined based on the baseline fluctuation semantic correction reaction mass spectrum visual features.

2. The method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to claim 1, characterized in that: Based on the baseline fluctuation visual feature coding vector, performing baseline fluctuation semantic correction on the reaction mass spectrum visual feature coding vector to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector as the baseline fluctuation semantic correction reaction mass spectrum visual feature, including: Performing univariate local unit feature coupling on the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors; Determining the recursive inference attention weights of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector based on the characteristic distribution characteristics of each baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vector in the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursive inference attention weights; Based on the baseline fluctuation-reaction mass spectrum visual local implicit interaction recursively inferring a set of attention weights, the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vectors is subjected to weighted modulation inference to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature encoding vector.

3. The method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to claim 2, characterized in that: The baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector are subjected to univariate local unit feature coupling to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors, including: Performing one-dimensional local implicit feature convolution coding on the baseline fluctuation visual feature coding vector and the reaction mass spectrum visual feature coding vector respectively to obtain a set of baseline fluctuation visual local implicit feature vectors and a set of reaction mass spectrum visual local implicit coding feature vectors; Each corresponding group of baseline fluctuation visual local implicit feature vectors and reaction mass spectrum visual local implicit coding feature vectors in the set of baseline fluctuation visual local implicit feature vectors and the set of reaction mass spectrum visual local implicit coding feature vectors are subjected to monomer feature interaction to obtain the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors.

4. The method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to claim 3, characterized in that: Based on the baseline fluctuation-reaction mass spectrum visual local implicit interaction recursively inferring a set of attention weights, weighted modulation reasoning is performed on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vectors to obtain a baseline fluctuation semantic correction reaction mass spectrum visual feature encoding vector, including: Based on the baseline fluctuation-reaction mass spectrum visual local implicit interaction recursively inferring a set of attention weights, weighted modulating the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vectors to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation encoding vectors; The set of the baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors is input into a chain inference engine based on a forward LSTM model to obtain the baseline fluctuation semantic correction reaction mass spectrum visual feature coding vector.

5. The method for screening drug-resistant substances of Mycobacterium tuberculosis based on mass spectrometry detection according to claim 4, characterized in that: Based on the baseline fluctuation-reaction mass spectrum visual local implicit interaction recursively inferring a set of attention weights, weighted modulating the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction encoding vectors to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation encoding vectors, including: Performing a chain inference-driven nonlinear interference correction on the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction recursive inference attention weights to obtain a set of baseline fluctuation-reaction mass spectrum visual local implicit interaction chain inference attention optimization weights; The set of baseline fluctuation-reaction mass spectrum visual local implicit interaction chain inference attention optimization weights is used to weighted modulate the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction coding vectors to obtain the set of baseline fluctuation-reaction mass spectrum visual local implicit interaction modulation coding vectors.

Images