Method for separating peak clusters in comprehensive two-dimensional chromatography based on a multi-modal model and related apparatus

Abstract

The application provides a full two-dimensional chromatographic peak family separation method and device based on a multi-modal model, a medium and a computer device, which comprises the following steps: constructing a mixed training data set for family classification training; performing supervised fine-tuning training on a pre-trained multi-modal large language model based on the mixed training data set; constructing a reinforcement learning fine-tuning data set, and training an initial family classification model in a reinforcement learning fine-tuning mode mainly based on a grouping relative strategy optimization; inputting preprocessed multi-modal input data into the trained family classification model for feature extraction to obtain a corresponding family classification result; and constructing a mixed training data set containing real samples and / or synthetic family samples, and enabling the pre-trained multi-modal large language model to learn a generation paradigm in the supervised fine-tuning stage, and introducing a comprehensive reward function in the reinforcement learning fine-tuning stage, so that the consistency of family division and the stability of quantitative statistics of the model under unknown samples and complex co-elution backgrounds can be improved.

Description

Technical Field

[0001] This invention relates to the field of mass spectrometry data analysis technology, and more specifically, to a method, apparatus, medium, and computer equipment for the separation of two-dimensional chromatographic peak families based on a multimodal model. Background Technology

[0002] In applications such as environmental monitoring, volatile organic compound (VOCs) apportionment, pollution source identification, and ozone formation potential assessment, two-dimensional chromatography-mass spectrometry (GC×GC-MS) data often exhibit characteristics such as numerous peaks, severe co-elution, significant baseline drift, and unstable signal-to-noise ratio. To obtain statistical results at the group level (e.g., volatile fractionation HVOCs / IVOCs / SVOCs, homologues, aromatic hydrocarbons, oxygen-containing / nitrogen-containing groups, etc.), traditional methods typically rely on: manual partitioning based on retention time thresholds, rule-based clustering, or a process of first matching spectral libraries and then grouping them according to manual experience. These methods, on the one hand, require extensive manual setting of thresholds and adjustment of clustering parameters, making them difficult to adapt to different instrument conditions, column combinations, and sample matrices; on the other hand, in the context of unknown substances, incomplete library coverage, and strong co-elution, they are prone to inconsistent group boundaries, cluster member drift, and non-reproducible results across batches, and the output often lacks a unified structured format and traceable attribution criteria. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method, apparatus, medium and computer equipment for the separation of two-dimensional chromatographic peaks based on a multimodal model, so as to overcome the above-mentioned disadvantages.

[0004] The above-mentioned technical objective of the present invention is achieved through the following technical solution: Firstly, a method for separating two-dimensional chromatographic peak families based on a multimodal model, comprising: S1. Construct a hybrid training dataset for family classification training, wherein the hybrid training dataset includes at least: family determination CoT, a real full two-dimensional chromatography-mass spectrometry dataset with sample labels, and / or a synthetic family sample dataset; wherein each training sample includes at least a multimodal input, a task instruction, and a structured output label corresponding to the task instruction; wherein the multimodal input includes at least a total ion flow map (TIC), at least one extracted ion flow map (EIC), and / or a fragment ion abundance matrix; S2. Based on the hybrid training dataset, supervised fine-tuning training is performed on the pre-trained multimodal large language model, enabling the model to learn to generate structured inference chains and structured family classification reports from the multimodal inputs and task instructions, thereby obtaining an initial family classification model; wherein, the structured family classification report includes at least: family label or cluster number, peak member list, and statistical results corresponding to the family division; S3. Construct a reinforcement learning fine-tuning dataset and train the initial family classification model using a reinforcement learning fine-tuning method that mainly optimizes the grouping relative strategy, so as to improve the consistency of family division and quantitative statistical stability of the model under unknown samples and complex co-outflow background, and obtain the family classification model. S4. Obtain the preprocessed multimodal input data, input the multimodal input data into the trained family classification model for feature extraction, and obtain the corresponding family classification result.

[0005] In one embodiment, the multimodal large language model includes: Visual encoder: used to extract peak morphology features and co-elution background features from TIC / EIC two-dimensional chromatographic tensors; Language decoder: Used to generate family labels and family classification criteria; Programmable Agent module: Used to call external knowledge bases to perform family prototype retrieval and consistency verification.

[0006] In one embodiment, the sample labels in the real full two-dimensional chromatography-mass spectrometry dataset include at least one or more of the following: The family or cluster label corresponding to the peak; Key ion set and / or fragment ion relative abundance labels for the peak; Peak retention time window and peak member relationship annotation; The sample labels are obtained through manual annotation and / or automatic annotation; the automatic annotation includes at least: family mapping annotation based on spectral library retrieval results, rule-based family grouping annotation based on molecular structure text description, or structural semantic annotation based on the output of a trained peak qualitative model.

[0007] In one embodiment, during the supervised fine-tuning training and / or reinforcement learning fine-tuning training, the model further outputs a peak embedding vector and constructs a contrastive learning or triplet learning objective based on positive pairs from the same family and negative pairs from different families, thereby reducing the embedding distance of samples from the same family and increasing the embedding distance of samples from different families to improve clustering stability and cross-batch transferability.

[0008] In one embodiment, the structured reasoning chain is a sequence of intermediate reasoning steps organized according to a preset template, and the sequence of intermediate reasoning steps includes at least one or more of the following step fields: Task understanding field: used to represent the meaning of input coordinates, task objectives, and output objectives; Task parsing fields: used to characterize the sub-objective decomposition, including at least feature recognition, family division strategy selection, region clustering, quantitative statistics and result organization; Step-by-step reasoning fields: including at least global perception and feature extraction, group recognition and spatial clustering, relationship and attribute association, and task-oriented quantization; Self-check field: Used to check the consistency of cluster discreteness, logical order of family division, and abnormal peak members; Confidence field: used to quantitatively characterize the uncertainty of family division boundaries or peak group affiliation; In this context, at least one field in the structured inference chain references the position interval parameter and / or threshold parameter of the peak in the RT1–RT2 space, so that the family partitioning rule can be reproduced.

[0009] In one embodiment, the structured output labels of the supervised fine-tuning training samples include at least: Inference chain label: A set of fields corresponding to the sequence of intermediate inference steps; Report tags: Structured table fields including volatility level or family number, family example, relative abundance percentage, and visualization suggestions; The relative abundance percentage is calculated based on the ratio of the integral value of the response to the corresponding peak group of each family to the total integral value of the response.

[0010] In one embodiment, the reinforcement learning fine-tuning phase employs a comprehensive reward function. The comprehensive reward function At least including: The format correctness score F is used to constrain the output to meet the structured protocol and be parsable; The molecular structure correctness score E evaluates the model's accuracy in molecular recognition and reasoning by calculating the distance between the predicted molecular structure and the actual molecular structure.

[0011] Secondly, a two-dimensional chromatographic peak separation device based on a multimodal model includes: Dataset Construction Engine: Used to construct a hybrid training dataset for family classification training. The hybrid training dataset includes at least: Family Decision (CoT), a real full-two-dimensional chromatography-mass spectrometry dataset with family labels, and / or a synthetic family sample dataset. Each training sample includes at least a multimodal input, a task instruction, and a structured output label corresponding to the task instruction. The multimodal input includes at least a total ion flow map (TIC), at least one extracted ion flow map (EIC), and / or a fragment ion abundance matrix. Supervised fine-tuning training engine: used to perform supervised fine-tuning training on the pre-trained multimodal large language model based on the hybrid training dataset, enabling the model to learn to generate structured inference chains and structured family classification reports from the multimodal inputs and task instructions, thereby obtaining an initial family classification model; wherein, the structured family classification report includes at least: family label or cluster number, peak member list, and statistical results corresponding to the family division; Reinforcement learning fine-tuning engine: used to construct a reinforcement learning fine-tuning dataset, and to train the initial family classification model using a reinforcement learning fine-tuning method that mainly optimizes GRPO with a grouping relative strategy, so as to improve the consistency of family division and quantitative statistical stability of the model under unknown samples and complex co-outflow background, and obtain the family classification model; Family separation engine: used to acquire preprocessed multimodal input data, input the multimodal input data into the trained family classification model for feature extraction, and obtain the corresponding family classification results.

[0012] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0013] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.

[0014] In summary, the present invention has the following beneficial effects: by constructing a mixed training dataset containing real samples and / or synthetic family samples, and by enabling the pre-trained multimodal large language model to learn the generation paradigm of multimodal input + task instructions to generate a structured inference chain + structured family classification report during the supervised fine-tuning stage, and by introducing a comprehensive reward function containing chemical name family classification discrimination and format validity constraints during the reinforcement learning fine-tuning stage, the consistency of family classification and quantitative statistical stability of the model under unknown samples and complex co-flow backgrounds are improved. Attached Figure Description

[0015] Figure 1 This is a flowchart of the multimodal model-based full two-dimensional chromatographic peak family separation method of the present invention.

[0016] Figure 2 This is a structural diagram of the multimodal model-based two-dimensional chromatographic peak separation device in an embodiment of the present invention.

[0017] Figure 3 This is an internal structural diagram of a computer device in an embodiment of the present invention.

[0018] In the diagram: 1. Dataset construction engine; 2. Supervised fine-tuning training engine; 3. Reinforcement learning fine-tuning engine; 4. Family separation engine. Detailed Implementation

[0019] To make the objectives, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein.

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] Example 1 To address the aforementioned problems, this invention provides a method for the separation of two-dimensional chromatographic peak families based on a multimodal model, such as... Figure 1 As shown, it includes: S1. Construct a hybrid training dataset for family classification training, wherein the hybrid training dataset includes at least: family determination CoT, a real full two-dimensional chromatography-mass spectrometry dataset with sample labels, and / or a synthetic family sample dataset; wherein each training sample includes at least a multimodal input, a task instruction, and a structured output label corresponding to the task instruction; wherein the multimodal input includes at least a total ion flow map (TIC), at least one extracted ion flow map (EIC), and / or a fragment ion abundance matrix; Specifically, the multimodal large language model architecture is an open-source, free protocol pre-trained model, including but not limited to Qwen2.5VL and Qwen3VL. To process multimodal, full-two-dimensional mass spectrometry data, the multimodal large language model includes a visual encoder, a language decoder, and a programmable Agent module; and forms the following processing flow: the visual encoder receives the two-dimensional chromatographic tensor from TIC / EIC and outputs feature representations; after receiving the feature representations and task instructions, the language decoder generates a structured family classification report composed of family labels, cluster numbers, family division criteria, and statistical fields; when external verification or prototype retrieval is required, the programmable Agent module is triggered to call an external knowledge base, return the verification results, and feed them back to the language decoder for correcting or supplementing the family classification conclusions and supporting explanations.

[0022] To train the family classification model, construct a hybrid training dataset for family classification training. The hybrid training dataset shall include at least one or a combination of the following: 1. A real full two-dimensional chromatography-mass spectrometry dataset with sample labels, which is derived from full two-dimensional GC×GC (-TOF MS) data collected by actual instruments. It can reflect complex factors such as real baseline drift, noise level, peak tailing, co-elution and isomer interference, and is used to enable the model to learn the family division rules and output stability in real scenarios.

[0023] 2. Synthetic family sample datasets are a type of synthetic dataset generated using a low-rank synthesis engine. They aim to simulate and generate data that conforms to the characteristics of actual mass spectrometry instruments. Traditionally, generating high-quality labeled data requires significant manual intervention and experimentation. However, by using low-rank synthetic datasets, the synthesis engine can automatically generate a large amount of representative data, greatly reducing the cost and time of manual annotation. Low-rank synthetic datasets provide rich training samples for the model by simulating various complex situations, such as peak shape diversity, co-eluenting, and noise. By introducing ideal supervisory labels, the accuracy and reliability of the model are improved. Low-rank synthetic datasets and their generation process are an effective means to solve the difficulties of data annotation in chromatography-mass spectrometry data analysis. Synthetic data generated by the low-rank synthesis engine can simulate various real-world situations, such as peak shape changes and noise interference, and provide accurate supervisory labels for the model, thus significantly reducing the cost of manual annotation and improving the diversity and accuracy of training data. The mixed use of real and synthetic data, on the one hand, uses real data to match the instrument and actual sample distribution, and on the other hand, uses synthetic data to supplement family coverage and boundary samples, thereby reducing the cost of manual annotation and improving the robustness of family classification.

[0024] Multimodal inputs include: TIC / EIC / fragment ion abundance matrix. The total ion flux map (TIC) provides information such as global peak distribution, background baseline, overall response strength, and multi-peak clustering regions, supporting the model's global perception and regional prior establishment, such as identifying high-density peak regions, sparse peak regions, and background drift regions. The extracted ion flux map (EIC) provides local enhancement information related to specific ions / target compounds, making it easier for the model to locate peak clusters belonging to the same family in co-eluent or complex background conditions, or to identify the continuous distribution characteristics of homologue peaks within a family. In targeted family classification scenarios, EIC can select ion channels that are indicative of the target family; in non-targeted family classification scenarios, multiple EIC channels can be provided to enhance the multi-channel consistency information for family classification. The fragment ion abundance matrix characterizes the mass spectrometry fragmentation patterns corresponding to peaks, supporting the model to further distinguish families based on fragment ion characteristics when retention time and position alone are insufficient, such as distinguishing different chemical families like benzene series compounds, alkanes, and oxygen-containing VOCs. Task instructions are used to explicitly specify the analysis objectives and output specifications for family classification, triggering the model to execute different family classification task templates on the same input data. Task instructions can include at least one or more of the following: Family classification objectives primarily include: volatile partitioning, chemical family classification, or family aggregation statistics required for pollution pattern recognition. Discrimination criteria are suggested, such as emphasizing the RT1-RT2 position interval as the primary factor and fragmentation patterns as a secondary factor, or vice versa. Output structure constraints are used to specify output fields, field types, family enumeration sets, confidence / uncertainty expression methods, etc. Through task instructions, the model can learn the mapping relationship of the same input + different instructions = different structured outputs during the training phase, thereby improving the model's reusability and controllability in different application scenarios.

[0025] Sample labels include at least one or more of the following: A family label or cluster label corresponding to the peak. The family label indicates the chemical family or functional group to which the target peak belongs, such as HVOCs / IVOCs / SVOCs obtained by volatility classification, or alkane groups, aromatic hydrocarbon groups, oxygen-containing VOCs groups, etc., obtained by chemical category. The cluster label indicates the cluster number obtained by spatial clustering of the target peak on the RT1–RT2 plane, so that the model can learn the rule of grouping spatially similar and similar properties peaks into the same cluster. Key ion set and / or fragment ion relative abundance label for the peak. The key ion set helps the model learn the association between family / structure semantics and fragment patterns. The fragment ion relative abundance constrains the model to consider not only RT position but also fragment response consistency when classifying families, thereby reducing misclassification caused by relying solely on position partitioning. Peak retention time window and peak membership label. The retention time window supports peak localization, cluster input feature construction, and subsequent integral statistics. The peak membership label describes the peak-cluster-family membership relationship, thereby supporting the model to output a family member list and family-level statistics.

[0026] The sample labels are obtained through manual and / or automatic labeling to balance labeling quality and labeling cost. Manual labeling can be completed by analysts with experience in chromatography and mass spectrometry, and includes at least: verifying the RT1–RT2 range of the peak, peak member attribution, and cluster division boundaries; and determining family or cluster labels by combining experimental conditions with control blanks, standards, internal standards, and spikes. Automatic labeling includes at least the following: Family mapping labeling based on spectral library search results: spectral library search is performed on the mass spectra of the target peak to obtain candidate compounds or candidate categories. The names / categories of the retrieved compounds are mapped to a preset family system, thereby automatically generating family labels or family candidate sets. Samples with low search confidence or conflicts can be marked with a low-confidence / to-be-verified label for subsequent manual verification or as difficult samples in the reinforcement learning stage. Family grouping labeling based on molecular structure text description rules: when samples contain structural text information such as name, molecular formula, SMILES, and functional group descriptions, preset rules or dictionaries can be used for family grouping. Structural semantic labeling based on the output of a trained peak qualitative model: the trained peak qualitative model outputs candidate structural semantics for the peak, and the structural semantics are further mapped to family labels or cluster interpretation fields as weakly supervised labels. In the above automatic labeling process, a strategy of automatic labeling as the primary method and manual sampling correction as a supplement can be adopted. Automatic labeling first generates a large number of samples; manual verification of key samples, conflicting samples, and low-confidence samples forms a stable training loop.

[0027] The task instruction triggers the model to call the corresponding task template. The task instruction may include a task type field and optional parameter fields. The task type may include ozone formation potential assessment, volatile matter classification, pollution source apportionment, etc., and the parameter fields may include partition thresholds, output granularity, statistical caliber, etc. After receiving the task instruction, the model selects a matching task template, thereby constraining the fields, order, and calculation logic of the output structured report. The task template can be understood as a combination of steps for parsing input data, output field specifications, and calculation / validation rules, enabling the model to achieve multi-task output under the same set of model parameters.

[0028] The structured output labels of the supervised fine-tuning training samples include at least: inference chain labels: a set of fields corresponding to the sequence of intermediate inference steps; and report labels: structured table fields containing volatility level or family number, family example, relative abundance percentage, and visualization suggestions; wherein the relative abundance percentage is calculated based on the ratio of the response integral value of each family's corresponding peak group to the total response integral value.

[0029] Furthermore, the structured reasoning chain is a sequence of intermediate reasoning steps organized according to a preset template, and the sequence of intermediate reasoning steps includes at least one or more of the following step fields: Task understanding fields: These fields characterize the meaning of input coordinates, task objectives, and output objectives, clearly defining task boundaries and preventing the model from misinterpreting family classification as a single-peak qualitative analysis or merely drawing bounding boxes. For example, coordinate meanings include: RT1 and RT2 correspond to the first and second dimensions of retention time, respectively; intensity on the two-dimensional plane is represented by color / height / pixel value. Input channel meanings include: TIC characterizing the total response distribution, EIC characterizing the response distribution of target ions, and the fragment ion abundance matrix characterizing structure-related fragment patterns. Task objectives may include: volatility classification / family grouping / ozone generation potential assessment / pollution source apportionment, etc. Output objectives may include: family labels or cluster numbers, peak member lists, and corresponding statistical results such as integrals / proportions.

[0030] Task parsing fields: These are used to break down the task into executable sub-objective steps, mainly including: Feature identification, identifying peak regions, peak cluster distribution, baseline drift, and co-efferentiation background; Family division strategy selection, selecting division criteria based on task instructions, such as volatility, structural category, homologue sequence, source characteristics, etc.; Region clustering: aggregating similar / close peak clusters within the RT1-RT2 space, combining fragment pattern similarity when necessary; Quantitative statistics, calculating the integral intensity, proportion, RSD, and other statistics for each family (or cluster); Result organization, generating a structured analysis report, which can be output in HTML, XML, or JSON format, including entries such as family name, relative abundance, number of peaks, average response intensity, and contribution ratio to the target task.

[0031] The step-by-step inference field can be further refined into the following parts: global perception and feature extraction, used to outline signal-dense areas, low signal-to-noise ratio areas, and baseline drift areas, and to identify peak morphology (sharp, tailed, forward-extending, broad peak, double peak) and peak density; group identification and spatial clustering, which forms clusters based on RT1–RT2 positions, peak connectivity, and valley separation rules. For suspected homologous clusters, the following can be used as supplementary clustering criteria: sequential distribution along the RT direction, similar morphology, and regular spacing; relationship and attribute association, used to record the proximity relationships between record clusters and the attribute consistency within record clusters; task-oriented quantization, used to perform integration or statistics on each group / cluster to obtain the total intensity of each group. In summary, the step-by-step inference field is mainly used to leave traces of the inference process for verification and training constraints.

[0032] Self-check field: used to check the consistency of cluster discreteness, family division logical order, and abnormal peak members; to reduce the situation of seemingly reasonable but actually misclassified cases, and to provide optimizable signals for subsequent reinforcement learning fine-tuning. The self-check field at least includes: Cluster discreteness check: If the members of the same cluster are too discrete in the RT space, mark it as abnormal; Family division logical order check: For example, whether the interval relationship of volatility grading satisfies the order from high to low; Abnormal peak member check: If the fragment pattern of a certain peak is too different from the cluster prototype or the peak shape is significantly inconsistent, prompt suspected misclassification / require recheck.

[0033] Confidence field: used to quantitatively characterize the uncertainty of family division boundaries or peak group attribution; for example: give a lower confidence to the peaks near the boundary; give a risk prompt to the clusters with strong co-elution overlap, conflicting fragment patterns, and broken peak shapes; make the model uncertainty explicit, which is beneficial for subsequent manual recheck sorting and system risk control.

[0034] In some embodiments, at least one field in the structured inference chain, such as the family identification and spatial clustering field or the family division strategy selection field, explicitly references the position interval parameters and / or threshold parameters of the peaks in the RT1-RT2 space, so that the family division rules are reproducible. That is, the model does not only describe what this area is, but writes out the executable rules such as HVOCs area = RT1 < P1 and RT2 < Q1, so that the same partition and family division results can be reproduced with the same threshold in subsequent experiments.

[0035] In some embodiments, for the report label, it at least includes: Volatility level or family number field: such as HVOCs / IVOCs / SVOCs or cluster number Cluster_k; Family example field: give example information such as the representative peak or representative compound name / homologous series / key ion set of this family; Relative abundance percentage field: used to characterize the relative contributions of each family; Visualization suggestion field: such as outputting suggestions such as a volatility distribution pie chart, a two-dimensional partition overlay chart, a family cluster heat map, or a family-abundance bar chart, etc., for subsequent result display and manual recheck.

[0036] Through the above settings, the supervised fine-tuning stage can simultaneously constrain the model to learn to output a reproducible inference chain according to the template, while outputting a structured family classification report and giving quantifiable relative abundance statistical results, thereby providing a basis for further improving the family division stability and statistical robustness in the subsequent reinforcement learning fine-tuning stage.

[0037] In some embodiments, the structured output labels correspond one-to-one with task instructions and are used to supervise the model in generating parsable, statistically valid, and database-ready family classification results. The structured output labels can be organized using tabular or JSON-style fields and include at least one or more of the following fields: 1. Family identifier field; 2. Peak set or region field corresponding to the family; 3. Family statistics field; 4. Confidence or consistency field. The structured output labels constrain the model's output format and support automatic statistics, avoiding the unparsable and unreproducible problems caused by only outputting natural language descriptions.

[0038] S2. Based on the hybrid training dataset, supervised fine-tuning training is performed on the pre-trained multimodal large language model, enabling the model to learn to generate structured inference chains and structured family classification reports from the multimodal inputs and task instructions, thereby obtaining an initial family classification model; wherein, the structured family classification report includes at least: family label or cluster number, peak member list, and statistical results corresponding to the family division; Specifically, the structured inference chain output by the model during the supervised fine-tuning phase refers to an explicitly parsable sequence of intermediate fields or intermediate result records. These records describe the key criteria and intermediate products used in the family classification process, such as the regional division results based on the RT1–RT2 space, a summary of clustering criteria, the correspondence between peak members and cluster numbers, consistency check results, and statistical definitions. The structured inference chain is a record of structured fields / steps, ensuring that the results are traceable, reproducible, and auditable. The structured family classification report output by the model includes at least the following fields: family label or cluster number; a list of peak members, such as a list of peak IDs, a list of peak boundary indices, or a set of RT position indices / peak numbers for peak members; and statistical results corresponding to the family division, such as total response / relative abundance / number of peaks / family percentage within the family.

[0039] In some embodiments, supervised fine-tuning training can be implemented as follows: For each training sample, the multimodal input and task instructions are used as model input, and the corresponding structured inference chain and structured family classification report are used as the supervision label output sequence to form a sample pair for training; the model output is constrained by a protocol, with the structured output using a fixed set of fields and a fixed syntax, such as JSON / key-value pairs / tabular fields, for subsequent automatic parsing, storage, and statistics; the loss of model prediction is optimized by minimizing the sequence loss between the predicted output and the supervision labels, enabling the model to learn the correctness of the output fields, the consistency between fields, and the correct caliber of the statistical results; after supervised fine-tuning is completed, the model can stably output structured family classification results when given task instructions, which can be used as the base model for subsequent reinforcement learning stages or online applications.

[0040] Furthermore, the structured reasoning chain is a sequence of intermediate reasoning steps organized according to a preset template, and the sequence of intermediate reasoning steps includes at least one or more of the following step fields: Task understanding fields: These fields characterize the meaning of input coordinates, task objectives, and output objectives, clearly defining task boundaries and preventing the model from misinterpreting family classification as a single-peak qualitative analysis or merely drawing bounding boxes. For example, coordinate meanings include: RT1 and RT2 correspond to the first and second dimensions of retention time, respectively; intensity on the two-dimensional plane is represented by color / height / pixel value. Input channel meanings include: TIC characterizing the total response distribution, EIC characterizing the response distribution of target ions, and the fragment ion abundance matrix characterizing structure-related fragment patterns. Task objectives may include: volatility classification / family grouping / ozone generation potential assessment / pollution source apportionment, etc. Output objectives may include: family labels or cluster numbers, peak member lists, and corresponding statistical results such as integrals / proportions.

[0041] Task parsing fields: These are used to break down the task into executable sub-objective steps, mainly including: Feature identification, identifying peak regions, peak cluster distribution, baseline drift, and co-efferentiation background; Family division strategy selection, selecting division criteria based on task instructions, such as volatility, structural category, homologue sequence, source characteristics, etc.; Region clustering: aggregating similar / close peak clusters within the RT1-RT2 space, combining fragment pattern similarity when necessary; Quantitative statistics, calculating the integral intensity, proportion, RSD, and other statistics for each family (or cluster); Result organization, generating a structured analysis report, which can be output in HTML, XML, or JSON format, including entries such as family name, relative abundance, number of peaks, average response intensity, and contribution ratio to the target task.

[0042] The step-by-step inference field can be further refined into the following parts: global perception and feature extraction, used to outline signal-dense areas, low signal-to-noise ratio areas, and baseline drift areas, and to identify peak morphology (sharp, tailed, forward-extending, broad peak, double peak) and peak density; group identification and spatial clustering, which forms clusters based on RT1–RT2 positions, peak connectivity, and valley separation rules. For suspected homologous clusters, the following can be used as supplementary clustering criteria: sequential distribution along the RT direction, similar morphology, and regular spacing; relationship and attribute association, used to record the proximity relationships between record clusters and the attribute consistency within record clusters; task-oriented quantization, used to perform integration or statistics on each group / cluster to obtain the total intensity of each group. In summary, the step-by-step inference field is mainly used to leave traces of the inference process for verification and training constraints.

[0043] Self-check field: used to perform consistency checks on cluster discreteness, family division logical order, and abnormal peak members; to reduce cases that seem reasonable but are actually misclassified, and to provide optimizable signals for subsequent reinforcement learning fine-tuning. The self-check field at least includes: Cluster discreteness check: If members of the same cluster are too discrete in the RT space, mark it as abnormal; Family division logical order check: For example, whether the interval relationship of volatility grading satisfies the order from high to low; Abnormal peak member check: If the fragmentation pattern of a certain peak is too different from the cluster prototype or the peak shape is significantly inconsistent, prompt suspected misclassification / require rechecking.

[0044] Confidence field: used to quantitatively characterize the uncertainty of family division boundaries or peak group attribution; for example: give a lower confidence to peaks near the boundary; give a risk prompt to clusters with strong co-elution overlap, conflicting fragmentation patterns, and broken peak shapes; make the model uncertainty explicit, which is conducive to subsequent manual rechecking and sorting and system risk control.

[0045] In some embodiments, at least one field in the structured inference chain, such as the family identification and spatial clustering field or the family division strategy selection field, explicitly references the position interval parameters and / or threshold parameters of the peak in the RT1-RT2 space, so that the family division rules are reproducible. That is, the model does not only describe what this region is, but writes out executable rules such as HVOCs region = RT1 < P1 and RT2 < Q1, so that the same partition and family division results can be reproduced with the same threshold in subsequent experiments.

[0046] S3. Construct a reinforcement learning fine-tuning dataset, and use a reinforcement learning fine-tuning method mainly based on grouped relative strategy optimization to train the initial family classification model, so as to improve the family division consistency and quantitative statistical stability of the model under unknown samples and complex co-elution backgrounds, and obtain a family classification model; Specifically, based on the initial family classification model that has completed supervised fine-tuning, introducing reinforcement learning fine-tuning enables the model to still maintain stable output in some difficult scenarios, such as: a relatively high proportion of unknown samples or compounds not covered by the library, irregular peak shapes and unclear boundaries, or ambiguous family discrimination caused by inconsistent cross-channel responses and incomplete fragmentation information, so as to obtain a family classification model with stronger generalization ability. Compared with supervised fine-tuning training, reinforcement learning fine-tuning pays more attention to constraining the consistency, stability, interpretability, and usability of the model output through a reward function in real complex scenarios without a unique standard answer / existing multiple solutions, and is especially suitable for family division under unknown samples and co-elution backgrounds.

[0047] The reward function can be trained using various reinforcement learning algorithms, including GRPO, PPO, and CISPO. Based on task requirements, CISPO is the preferred training framework, especially when dealing with long inference chains and rare token problems, where its advantages are more pronounced. CISPO ensures high model stability during multi-step inference through a stable training process and effectively addresses the challenges posed by low-frequency samples. In some embodiments, reinforcement learning fine-tuning employs an algorithm primarily based on Group Relative Policy Optimization (GRPO) to train the initial family classification model. Its main steps include: for the same multimodal input and task instruction, the model samples K candidate structured outputs during the decoding phase, such as combinations of different family partition boundaries + peak member lists + statistical tables, denoted as... For each candidate output, first verify whether it meets the structured protocol, such as complete fields, parsability, valid values, and consistent references. Those that do not meet the requirements receive zero reward or a strong penalty to force the model to stably output results that can be stored in the database. For candidate outputs that pass format validation, calculate the comprehensive reward for each candidate based on the reward function. The relative advantage relationships are obtained by ranking the candidates within each group. A relative advantage estimate is constructed using the relative ranking within each group, and the model parameters are updated using the relative loss of GRPO, thereby increasing the probability of generating high-reward candidates and decreasing the probability of generating low-reward candidates.

[0048] In the aforementioned reinforcement learning fine-tuning process, generating multiple race division schemes for the same input and ranking them with rewards is equivalent to explicitly selecting more reasonable and stable race division strategies during training, reducing randomness and sampling bias. Race division often has multiple solutions, and GRPO utilizes relative advantage learning to avoid dependence on a single standard answer, making it suitable for unknown compounds and co-efferentiation scenarios.

[0049] In some embodiments, reinforcement learning fine-tuning can also employ a CISPO-based reinforcement learning algorithm. The loss function of CISPO is: ; in, Indicates the importance sampling ratio; This represents the maximum value of the clipping, which is a constant rather than a symmetrical range in CISPO; Represents the dominance function. Indicates the current policy in state Select action The probability of; This operation ensures that the importance sampling ratio after clipping is not involved in gradient calculation, but is treated as a constant to avoid affecting gradient updates. In CISPO, The operation ensures that the cropped ratio does not affect gradient updates, which allows for more stable training, especially when dealing with low-probability tokens.

[0050] Unlike the GRPO reinforcement learning method, the CISPO algorithm decouples the sampling weights of the dominance term from gradient calculation in its loss function. Specifically, GRPO updates the gradient by incorporating dominance terms, such as the difference between predicted and actual results, which can affect the matching accuracy of peaks across different categories, especially for rare or infrequent peaks. However, in CISPO, the dominance term is no longer associated with the gradient but exists as a constant, independent of gradient calculation. This allows all tokens to be included in the loss function, avoiding the problem of some peaks being pruned due to their infrequent occurrence. In this way, CISPO can improve the model's learning ability on unknown or low-frequency samples, enhancing its generalization ability.

[0051] Furthermore, the combination of CISPO and CoT methods offers significant advantages. In the CoT framework, stepwise inference provides explicit outputs by refining each inference step, while CISPO ensures that the optimization of these inference steps is unaffected by low-frequency samples through a stable training process. Within the CoT structure, CISPO helps stabilize the model's learning process when performing complex inference, especially in multi-round inference, effectively preventing excessive fluctuations and gradient explosion problems. Through this mechanism, CISPO reinforcement learning provides more stable and efficient training during CoT's stepwise inference process, ensuring that tokens in each inference step receive appropriate weights, thus avoiding the failure to learn the optimal solution due to low weights for rare tokens, as seen in GRPO.

[0052] CISPO's strategy ensures that even rare peaks are fully considered during training, thus improving the comprehensiveness of the training. Compared to traditional reinforcement learning algorithms, CISPO avoids the use of KL divergence and does not rely on controlling the update speed. Instead, it directly improves the stability and generalization ability of the training process by extracting the dominant terms. Therefore, CISPO not only improves the inference accuracy of CoT but also ensures more consistent and robust model performance when facing complex tasks.

[0053] Furthermore, to ensure consistency of family classification results across different batches, sample matrices, and instrument conditions, the model further outputs a peak embedding vector during supervised fine-tuning and / or reinforcement learning fine-tuning training. This peak embedding vector characterizes the family semantic features of the peaks and serves as a unified representation space for subsequent clustering and similarity retrieval. For multimodal inputs, the model first obtains a two-dimensional feature map via a visual encoder, then performs region convergence on the peak regions to obtain a visual feature representation of the peaks. Simultaneously, the model encodes the fragment ion abundance matrix or key ion set to obtain a spectral feature representation. The aforementioned visual and spectral features are then fused, such as through splicing, cross-attention, or gating fusion, and mapped to a fixed-dimensional space via a projection layer to obtain the peak embedding vector. In addition to the supervised loss for generating structured family classification reports, the above training process also introduces a metric learning objective for the embedding space, making peaks of the same family closer together and peaks of different families farther apart within the embedding space. Specifically, positive and negative pairs are first constructed, where positive pairs sample two different peaks from the same family label or the same cluster of mapped families. Sample peaks from different family labels to form negative examples The distance function is defined as follows: And set the interval parameters. The triplet loss can then be denoted as: In the fine-tuning stage of reinforcement learning, the same-family compactness / different-family separation of the embedding space can be used as a component of the reward function or as an auxiliary constraint for training, so as to reduce the embedding distance of same-family samples and increase the embedding distance of different-family samples, thereby improving clustering stability and cross-batch transferability.

[0054] In some embodiments, the multimodal large language model possesses generalization attribution capability: that is, for new compound peaks not appearing in the training set, it can still complete family inference based on multi-source evidence and output interpretable attribution basis. The model first uses the peak's position in the RT1–RT2 plane (or its interval parameter / threshold parameter) to perform coarse partitioning, for example, by volatility level or empirical partitioning rules to obtain a candidate family set; this step provides positional priors to narrow the candidate range and improve attribution efficiency; the model combines the peak morphology and its surrounding background to determine which family group's typical morphological distribution it is closer to, for example, homologous families often exhibit sequential distribution along the RT direction, with similar peak groups and regular spacing, and the model can use the spatial relationship of neighboring peak groups to enhance the judgment; the model can perform similarity matching between the peak's fragment ion pattern and a known family mass spectrometry feature library, which may contain representative fragment ion sets, relative abundance templates, or embedded prototypes of each family. Matching methods can include directly calculating similarity at the fragment abundance vector level, such as cosine similarity, or encoding fragment patterns into embedding vectors and performing nearest neighbor retrieval with family prototype embeddings. When the similarity exceeds a threshold, the peak can be assigned a corresponding family label using a high-confidence semi-supervised label. When the similarity is insufficient, a low-confidence attribution can be output and trigger manual review or further verification by the agent. For high-confidence attribution results, they can be used as weakly labeled samples for subsequent training or incremental learning queues. For low-confidence or conflicting samples, they can be reviewed through rule verification, spectral library supplementation, or by the agent calling an external knowledge base, and the review results can be written back as training data to form a closed loop of continuous iteration.

[0055] In some embodiments, the reinforcement learning fine-tuning phase employs a comprehensive reward function. Specifically, it includes: ; in, Let each represent the weight of its respective dimension, and satisfy the following conditions: . The score indicates the correctness of the format. The score represents the correctness of the molecular structure. The score represents the accuracy of task comprehension. Family separation correctness score.

[0056] The format correctness score indicates the correctness of the format. The key to the correctness of the format is to ensure that the content of the model output meets the predetermined structure requirements and that the data can be effectively extracted through regular expressions.

[0057] ; This indicates the number of fields that correctly conform to the predefined structure. This indicates the total number of all required fields. The output JSON must include necessary control fields, such as molecular formula, molecular structure, peak location, and region information. The JSON format must strictly adhere to the predetermined field order and hierarchy. This facilitates the extraction of necessary data in subsequent tasks using regular expressions or other parsing methods. Structured fields should be explicitly defined, such as molecular formula, molecular name, and peak region, ensuring that each field has a corresponding value. The output content must be consistent with the predetermined steps of the CoT model, without omissions or incorrect ordering. For example, molecular formula inference, peak detection, and region partitioning should be output sequentially, and each part requires structured processing.

[0058] The molecular structure correctness score is used to evaluate the model's accuracy in molecular recognition and reasoning by calculating the distance between the predicted and actual molecular structures. Since molecular formulas are not limited to InChI, InChIKey, SMILES, etc., molecular structures can be described using various other representations. To calculate the molecular structure correctness score, the structural similarity of different representations needs to be compared. Similarity can be calculated as follows: Cosine similarity: ;in, The vector representation of the predicted molecular formula. The vector representation of the true molecular formula; the closer the similarity value is to 1, the more similar the two molecular formulas are.

[0059] Tanimoto similarity: Where C is the intersection of two sets, A and B are the sizes of the two sets, and Tanimoto similarity measures the ratio of the intersection to the union of two sets, applicable to binary vectors. European distance: ; This represents the i-th dimension of the predicted molecular formula vector; This represents the i-th dimension of the true molecular formula vector.

[0060] Within the finely tuned multimodal qualitative model framework, a molecular formula embedding model is first trained using contrastive learning. This model maps molecular formulas to a vector space, generating a fixed-dimensional vector representation for each formula. These vector representations are used as part of a subsequent reward function to calculate the similarity between the predicted and actual molecular formulas. During this process, the model outputs a reward based on the similarity calculation, which is used for model optimization in reinforcement learning. In actual training, the embedding model effectively acts as the reward model, scoring the molecular formulas. The reward model calculates the reward based on the model's inference results; the embedding model trained through contrastive learning is responsible for calculating the similarity of molecular formulas and outputting the corresponding score. This score serves as the reward in reinforcement learning, helping to optimize the model's predictive ability.

[0061] The task comprehension accuracy score evaluates the model's understanding of the task during inference, particularly its handling of task-related key fields, inference steps, and information extraction. It ensures the model accurately parses and understands each inference step and correctly applies this information to guarantee a reasonable final output.

[0062] In this embodiment, the task understanding accuracy score includes the following components: Key field matching accuracy: By checking whether the fields extracted by the model during inference match the fields in the ground truth, we ensure that the key fields in the task are correctly identified and extracted during inference. ;in, Indicates the number of fields accurately identified by the model; This indicates the total number of all key fields required for the task. Consistency of inference steps: By comparing the degree of matching between the inference steps executed by the model and the actual inference process, it is ensured that the execution of each inference step conforms to the predetermined task objectives and logical relationships. ; This represents the number of incorrect inference steps performed by the model. This represents the total number of steps in the task reasoning process. The task comprehension accuracy score is obtained by weighted summation of key field matching accuracy and reasoning step consistency.

[0063] The family separation correctness score is used to evaluate the model's ability to correctly separate different families in a family separation task, ensuring the accuracy of separation for each cluster and checking whether the model can identify peaks of different categories and their relationships. The family separation correctness score includes the following categories: Cluster accuracy assesses whether the model correctly identifies and separates different clusters. Cluster accuracy focuses not only on the number of clusters but also on whether the number and category of molecules within each cluster match expectations. Bbox accuracy ensures that the bounding boxes (bboxes) surrounding the characteristic peaks within each cluster accurately select the characteristic peaks. Bbox accuracy can be comprehensively evaluated by the cross-union ratio (CUB) between the predicted and ground truth boxes and the center distance. Family name consistency verifies whether the names of the predicted clusters match the categories of the actual molecular clusters, especially for clusters with known chemical names or categories, ensuring that the cluster names or categories are consistent with the actual substance categories. For example, high volatile organic compounds (HVOCs) should be consistent with real HVOC clusters; molecular vector consistency is achieved by calculating the similarity of molecular vectors within a cluster to ensure that molecules within the cluster have similar chemical properties and structures, that is, the substances within the cluster have a high degree of similarity.

[0064] S4. Obtain the preprocessed multimodal input data, and input the multimodal input data into the trained family classification model for feature extraction to obtain the corresponding family classification result. Specifically, step S4 is used to input the multimodal chromatographic-mass spectrometry data of the real test sample into the trained family classification model after the completion of supervised fine-tuning and reinforcement learning fine-tuning, so that the model can automatically complete the following during the inference stage: extract peak groups and distribution features from the multimodal input, give family division / clustering results according to the task instructions; and output a structured family classification report for subsequent statistics, comparison and visualization.

[0065] In one embodiment, for synthetic family sample data used for training, this application further provides a synthesis method, specifically including the following steps: A first-dimensional response vector is constructed on the time axis RT1, and a second-dimensional response vector is constructed on the time axis RT2. The first-dimensional response vector is used to describe the response distribution of the target peak along the RT1 direction, and the second-dimensional response vector is used to describe the response distribution of the target peak along the RT2 direction.

[0066] The first-dimensional response vector is: ; The second-dimensional response vector is: ; The first and second response vectors are parameterized randomly; a low-rank matrix is ​​constructed based on the first and second response vectors after parameterization.

[0067] For each component K, generate random parameters, including: ; in, This indicates the center position of the k-th characteristic peak on the X-axis; This indicates the center position of the k-th characteristic peak on the Y-axis; This represents the peak width of the k-th characteristic peak on the X-axis; This represents the peak width of the k-th characteristic peak on the Y-axis; This represents the tailing coefficient of the k-th characteristic peak. This represents the amplitude of the k-th characteristic peak.

[0068] The construction of a low-rank matrix can be achieved by outer product or low-rank superposition: for the case of a single peak, the randomized first-dimensional response vector and the second-dimensional response vector can be outer productd to obtain a two-dimensional peak matrix with a rank of 1; for the case of multiple peaks or overlapping peaks, multiple rank 1 matrices can be superimposed according to weights to form a low-rank matrix with a rank of K, so as to simulate the superposition effect of multiple compound peaks or multiple components of the same peak.

[0069] ; ; After adding noise to the low-rank matrix, preprocessing is performed to obtain the initial synthetic data, including: ; in, Indicates the noise variance. ; Noise types may include at least one: thermal noise / electronic noise, such as Gaussian noise, pink noise, baseline drift noise, chemical background noise such as solvent peaks / resident peaks, and random interference peaks. Preprocessing may include: intensity normalization, dynamic range compression, baseline removal or smoothing filtering, raster alignment, and mapping synthetic peaks to an RT1×RT2 grid resolution consistent with the real sample.

[0070] In some embodiments, preprocessing includes nonnegation and normalization to ensure that the signal conforms to the numerical range of the actual measurement. .

[0071] The initial synthetic data is ground truth adjusted to obtain synthetic data for training. In this embodiment, ground truth adjustment includes at least: generating matching ground truth labels for the synthetic data and performing consistency calibration. The ground truth labels may include: peak center coordinates RT1 and RT2, peak bounding boxes or peak profiles; quantitative correlation ground truth values ​​such as peak area / peak height; and compound entry information or structural characterization corresponding to the synthetic peaks. Furthermore, ground truth adjustment also includes operations such as drift perturbation and channel loss perturbation that are simultaneously reflected on the labels to ensure consistency between input and labels, thereby meeting the label consistency requirements of supervised training and contrastive learning training.

[0072] For each of the aforementioned low-rank synthetic data, pre-calculation is performed using analytical integration, including: ; in, This is the synthesized peak function generated by low-rank tensor decomposition.

[0073] In one embodiment, the adaptive area integral model synchronously outputs an area uncertainty index during inference. To mark high-risk quantitative results, the area uncertainty index The specific steps involved are as follows: By performing N forward inferences with random perturbations on the same input, the area set is obtained. ; Uncertainty index for calculating area ,include: .

[0074] Specifically, the adaptive area integral model can simultaneously output an area uncertainty index during inference. This is used to label high-risk quantitative results. The specific process for this step is as follows: Perform N forward inferences with random perturbations on the same input. During each inference, the input data undergoes random perturbations, causing the model to make multiple predictions for the same sample, thus generating a series of different predicted area values. Let these area values ​​be: ,in This represents the response area of ​​the target compound obtained in each inference. The area uncertainty index is calculated using the following formula. ,include: ;in: Standard deviation measures the dispersion of the predicted area and indicates the magnitude of variation in the predicted results. This represents the mean of the predicted area values, which serves as the center value of the area. The standardized coefficient of variation is expressed as the ratio of the standard deviation to the mean. If... A higher value indicates that the predicted area has greater uncertainty, suggesting that the result may have a higher risk.

[0075] Example 2 Please see Figure 2 A multimodal model-based two-dimensional chromatographic peak separation device, comprising: Dataset Construction Engine 1: Used to construct a hybrid training dataset for family classification training, wherein the hybrid training dataset includes at least: family determination CoT, a real full two-dimensional chromatography-mass spectrometry dataset with family labels, and / or a synthetic family sample dataset; wherein each training sample includes at least a multimodal input, a task instruction, and a structured output label corresponding to the task instruction; wherein the multimodal input includes at least a total ion flow map (TIC), at least one extracted ion flow map (EIC), and / or a fragment ion abundance matrix; Supervised fine-tuning training engine 2: used to perform supervised fine-tuning training on the pre-trained multimodal large language model based on the hybrid training dataset, enabling the model to learn to generate structured inference chains and structured family classification reports from the multimodal inputs and task instructions, thereby obtaining an initial family classification model; wherein, the structured family classification report includes at least: family label or cluster number, peak member list, and statistical results corresponding to the family division; Reinforcement learning fine-tuning engine 3: used to construct a reinforcement learning fine-tuning dataset, and to train the initial family classification model using a reinforcement learning fine-tuning method that mainly optimizes GRPO with a grouping relative strategy, so as to improve the consistency of family division and quantitative statistical stability of the model under unknown samples and complex co-outflow background, and obtain a family classification model. Family Separation Engine 4: Used to acquire preprocessed multimodal input data, input the multimodal input data into the trained family classification model for feature extraction, and obtain the corresponding family classification results.

[0076] Specific limitations regarding the multimodal model-based two-dimensional chromatographic peak family separation device can be found in the above-described limitations of the multimodal model-based two-dimensional chromatographic peak family separation method, and will not be repeated here. Each module in the aforementioned multimodal model-based two-dimensional chromatographic peak family separation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.

[0077] Those skilled in the art will understand that Figure 2The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the present application. The specific multimodal model-based two-dimensional chromatographic peak family separation device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0078] Example 3 A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the full two-dimensional chromatographic peak family separation method based on a multimodal model as described in Example 1.

[0079] Example 4 In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage medium. When the computer program is executed by the processor, it implements a two-dimensional chromatographic peak separation method based on a multimodal model.

[0080] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0081] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps: including: S1. Construct a hybrid training dataset for family classification training, wherein the hybrid training dataset includes at least: family determination CoT, a real full two-dimensional chromatography-mass spectrometry dataset with sample labels, and / or a synthetic family sample dataset; wherein each training sample includes at least a multimodal input, a task instruction, and a structured output label corresponding to the task instruction; wherein the multimodal input includes at least a total ion flow map (TIC), at least one extracted ion flow map (EIC), and / or a fragment ion abundance matrix; S2. Based on the hybrid training dataset, supervised fine-tuning training is performed on the pre-trained multimodal large language model, enabling the model to learn to generate structured inference chains and structured family classification reports from the multimodal inputs and task instructions, thereby obtaining an initial family classification model; wherein, the structured family classification report includes at least: family label or cluster number, peak member list, and statistical results corresponding to the family division; S3. Construct a reinforcement learning fine-tuning dataset and train the initial family classification model using a reinforcement learning fine-tuning method that mainly optimizes the grouping relative strategy, so as to improve the consistency of family division and quantitative statistical stability of the model under unknown samples and complex co-outflow background, and obtain the family classification model. S4. Obtain the preprocessed multimodal input data, input the multimodal input data into the trained family classification model for feature extraction, and obtain the corresponding family classification result.

[0082] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for separating peak families in full two-dimensional chromatography based on a multimodal model, characterized in that, Includes the following steps: S1. Construct a hybrid training dataset for family classification training, wherein the hybrid training dataset includes at least: family determination CoT, a real full two-dimensional chromatography-mass spectrometry dataset with sample labels, and / or a synthetic family sample dataset; wherein each training sample includes at least a multimodal input, a task instruction, and a structured output label corresponding to the task instruction; wherein the multimodal input includes at least a total ion flow map (TIC), at least one extracted ion flow map (EIC), and / or a fragment ion abundance matrix; S2. Based on the hybrid training dataset, supervised fine-tuning training is performed on the pre-trained multimodal large language model, enabling the model to learn to generate structured inference chains and structured family classification reports from the multimodal inputs and task instructions, thereby obtaining an initial family classification model; wherein, the structured family classification report includes at least: family label or cluster number, peak member list, and statistical results corresponding to the family division; S3. Construct a reinforcement learning fine-tuning dataset and train the initial family classification model using a reinforcement learning fine-tuning method that mainly optimizes the grouping relative strategy, so as to improve the consistency of family division and quantitative statistical stability of the model under unknown samples and complex co-outflow background, and obtain the family classification model. S4. Obtain the preprocessed multimodal input data, input the multimodal input data into the trained family classification model for feature extraction, and obtain the corresponding family classification result.

2. The method for separating peak families in full two-dimensional chromatography based on a multimodal model according to claim 1, characterized in that, The multimodal large language model includes: Visual encoder: used to extract peak morphology features and co-elution background features from TIC / EIC two-dimensional chromatographic tensors; Language decoder: Used to generate family labels and family classification criteria; Programmable Agent module: Used to call external knowledge bases to perform family prototype retrieval and consistency verification.

3. The method for separating peak families in full two-dimensional chromatography based on a multimodal model according to claim 1, characterized in that, The sample labels in the real full two-dimensional chromatography-mass spectrometry dataset include at least one or more of the following: The family or cluster label corresponding to the peak; Key ion set and / or fragment ion relative abundance labels for the peak; Peak retention time window and peak member relationship annotation; The sample labels are obtained through manual annotation and / or automatic annotation; the automatic annotation includes at least: family mapping annotation based on spectral library retrieval results, rule-based family grouping annotation based on molecular structure text description, or structural semantic annotation based on the output of a trained peak qualitative model.

4. The method for separating peak families in full two-dimensional chromatography based on a multimodal model according to claim 1, characterized in that, In the supervised fine-tuning training and / or reinforcement learning fine-tuning training, the model further outputs peak embedding vectors and constructs contrastive learning or triplet learning objectives based on positive pairs of the same family and negative pairs of different families, so as to reduce the embedding distance of samples of the same family and increase the embedding distance of samples of different families, thereby improving clustering stability and cross-batch transferability.

5. The method for separating peak families in full two-dimensional chromatography based on a multimodal model according to claim 1, characterized in that, The structured reasoning chain is a sequence of intermediate reasoning steps organized according to a preset template, and the sequence of intermediate reasoning steps includes at least one or more of the following step fields: Task understanding field: used to represent the meaning of input coordinates, task objectives, and output objectives; Task parsing fields: used to characterize the sub-objective decomposition, including at least feature recognition, family division strategy selection, region clustering, quantitative statistics and result organization; Step-by-step reasoning fields: including at least global perception and feature extraction, group recognition and spatial clustering, relationship and attribute association, and task-oriented quantization; Self-check field: Used to check the consistency of cluster discreteness, logical order of family division, and abnormal peak members; Confidence field: used to quantitatively characterize the uncertainty of family division boundaries or peak group affiliation; In this context, at least one field in the structured inference chain references the position interval parameter and / or threshold parameter of the peak in the RT1–RT2 space, so that the family partitioning rule can be reproduced.

6. The method for separating peak families in full two-dimensional chromatography based on a multimodal model according to claim 1, characterized in that, The structured output labels of the supervised fine-tuning training samples include at least: Inference chain label: A set of fields corresponding to the sequence of intermediate inference steps; Report tags: Structured table fields including volatility level or family number, family example, relative abundance percentage, and visualization suggestions; The relative abundance percentage is calculated based on the ratio of the integral value of the response to the corresponding peak group of each family to the total integral value of the response.

7. The method for separating peak families in full two-dimensional chromatography based on a multimodal model according to claim 1, characterized in that, The fine-tuning phase of reinforcement learning employs a comprehensive reward function. The comprehensive reward function At least including: The format correctness score F is used to constrain the output to meet the structured protocol and be parsable; The molecular structure correctness score E evaluates the model's accuracy in molecular recognition and reasoning by calculating the distance between the predicted molecular structure and the actual molecular structure.

8. A two-dimensional chromatographic peak separation device based on a multimodal model, comprising: Dataset Construction Engine: Used to construct a hybrid training dataset for family classification training. The hybrid training dataset includes at least: Family Decision (CoT), a real full-two-dimensional chromatography-mass spectrometry dataset with family labels, and / or a synthetic family sample dataset. Each training sample includes at least a multimodal input, a task instruction, and a structured output label corresponding to the task instruction. The multimodal input includes at least a total ion flow map (TIC), at least one extracted ion flow map (EIC), and / or a fragment ion abundance matrix. Supervised fine-tuning training engine: used to perform supervised fine-tuning training on the pre-trained multimodal large language model based on the hybrid training dataset, enabling the model to learn to generate structured inference chains and structured family classification reports from the multimodal inputs and task instructions, thereby obtaining an initial family classification model; wherein, the structured family classification report includes at least: family label or cluster number, peak member list, and statistical results corresponding to the family division; Reinforcement learning fine-tuning engine: used to construct a reinforcement learning fine-tuning dataset, and to train the initial family classification model using a reinforcement learning fine-tuning method that mainly optimizes GRPO with a grouping relative strategy, so as to improve the consistency of family division and quantitative statistical stability of the model under unknown samples and complex co-outflow background, and obtain the family classification model; Family separation engine: used to acquire preprocessed multimodal input data, input the multimodal input data into the trained family classification model for feature extraction, and obtain the corresponding family classification results.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the full two-dimensional chromatographic peak family separation method based on a multimodal model as described in any one of claims 1-7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the full two-dimensional chromatographic peak family separation method based on a multimodal model as described in any one of claims 1-7.

Images