A method and device for automatic preprocessing of mass cytometry data

Abstract

The application provides a mass spectrometry flow cytometry data full-automatic preprocessing method and device, and relates to the technical field of bioinformatics data processing. The preprocessing method comprises the following steps: acquiring original data, extracting signal characteristics, calculating single-cell probability scores of each event, separating single-cell target events from the original data based on the continuous single-cell probability scores, and outputting preprocessed single-cell data sets. The method realizes full-automatic unsupervised processing, eliminates artificial subjectivity and black box defects. By calibrating multi-dimensional original signals and fusing them according to semantic attributes, rough compression of multi-dimensional information is avoided, and independent semantics of features are retained to accurately distinguish complex boundaries. At the same time, by using a distribution quality index to dynamically and adaptively weight and perform nonlinear combination mapping, complex distribution data of different qualities are intelligently adapted, and finally, continuous probability scores with high accuracy are output, which significantly improves the reliability of single-cell extraction.

Description

Technical Field

[0001] This application relates to the field of bioinformatics data processing technology, and in particular to a fully automated preprocessing method and apparatus for mass spectrometry flow cytometry data. Background Technology

[0002] Mass cytometry is a single-cell analysis technique that combines the high-throughput capability of flow cytometry with the high resolution of mass spectrometry. It utilizes heavy metal isotopes for labeling and employs time-of-flight mass spectrometry for detection, enabling the simultaneous acquisition of a large amount of high-dimensional parameter information at the single-cell level. It is widely used in analytical scenarios such as the in-depth analysis of complex microenvironments. However, the raw data acquired during the detection process often contains a large amount of non-target event signals, such as abnormal artifacts generated during instrument detection, adhesion signals formed by the fusion of multiple targets in the detector, and low-quality impurities and impaired targets in the sample environment.

[0003] To extract high-quality single-cell targets from raw, high-dimensional, massive datasets, gating and screening preprocessing is typically required. Traditional data processing methods largely rely on manual delineation of regions on two-dimensional scatter plots to screen targets. With the rapid increase in data dimensionality, various automated data processing methods have gradually developed. These existing automated methods mainly include: parametric modeling and analysis of data distribution in high-dimensional space; thresholding by finding specific density distribution characteristics in one-dimensional or multi-dimensional space; and building pre-trained data screening models using neural networks to achieve automated separation of target events.

[0004] However, the aforementioned automated preprocessing methods have many limitations in practical applications. On the one hand, existing parametric hybrid modeling methods often directly compress and transform multi-dimensional feature information, ignoring the semantic independence between different detection dimensions and physical features, leading to ambiguous boundary determination. On the other hand, cutting methods based on conventional density features are usually only applicable to simple axis-aligned distributions and are difficult to handle complex, nonlinear fusion boundaries between adhering bodies and target bodies in multi-dimensional space. Furthermore, deep learning methods are highly dependent on large labeled sample sets, and the non-transparency of the model itself results in a lack of traceability in rigorous data analysis scenarios. At the same time, conventional normalization and clustering algorithms are prone to erroneous merging or over-segmentation of target groups when faced with data with a high proportion of adhering impurities due to unclear density boundaries.

[0005] In summary, existing data preprocessing methods struggle to balance interpretability and effective comprehensive analysis of multi-source heterogeneous features without annotation. When faced with diverse and complex data, current technologies often lack the ability to adaptively evaluate based on the data's inherent distribution quality. This can lead to the mechanical fragmentation or crude fusion of multidimensional features, resulting in insufficient accuracy and stability in high-quality single-cell target screening, failing to meet the demands for efficient and reliable processing of high-dimensional bioinformatics data. Summary of the Invention

[0006] In a first aspect, this application provides a fully automated preprocessing method for mass spectrometry flow cytometry data, including: Obtain raw mass cytometry data and extract signal features from at least two physical channel dimensions of each event in the raw data to construct a feature space; In the feature space, the single-cell probability score of each event is calculated using a hierarchical probability fusion model, including: First-level calibration step: Based on the signal features of each event in the feature space, calculate the original judgment signal of each event under multiple different evaluation dimensions, and calibrate and map the original judgment signal to a unified probability interval to obtain multiple independent probability signals corresponding to each event; The second-level semantic step is to group and fuse the multiple independent probability signals corresponding to each event according to preset semantic attributes to generate multiple intermediate features for characterizing the state of the event. The third-level decision-making step is to determine adaptive weights based on the distribution quality index of events in the feature space, and to use the adaptive weights to perform nonlinear combination mapping on the multiple intermediate features to output the continuous single-cell probability score of each event. Based on the continuous single-cell probability scores of each event, single-cell target events are separated from the original data, and a preprocessed single-cell dataset is output.

[0007] In an optional implementation, extracting signal features from at least two physical channel dimensions of each event in the original data to construct a feature space includes: Obtain the raw DNA content signal and event length signal; Based on a preset cofactor, an inverse hyperbolic sine transform is performed on the original DNA content signal and event length signal to maintain linear resolution in the low signal region and compress the dynamic range in the high signal region. The event length signal after inverse hyperbolic sine transformation is superimposed with a uniformly distributed random jitter over a preset interval to eliminate artifacts generated by discrete integer values ​​in density estimation, and a two-dimensional feature space is constructed based on the processed DNA content signal and the event length signal.

[0008] In an optional implementation, in the first-level calibration step, the original judgment signals under the multiple different evaluation dimensions include: clustering signal, orientation mixing signal, density signal, contour line signal, orientation signal, and quadrant signal; The step of calibrating and mapping the original determination signals to a unified probability interval specifically involves: By anchoring the Sigmoid function, each of the original decision signals is mapped to the [0,1] interval to obtain six independent probability signals; the six independent probability signals are: clustering probability signal, direction mixing probability signal, density probability signal, contour line probability signal, direction probability signal, and quadrant probability signal. In the second-level semantic step, the multiple intermediate features include: membership features, position features, and orientation safety features; the multiple independent probability signals corresponding to each event are grouped and fused according to preset semantic attributes, including: The clustering probability signal is fused with the directional mixing probability signal to generate the membership feature; The density probability signal is fused with the contour probability signal to generate the location feature; The direction probability signal is fused with the quadrant probability signal to generate the direction safety feature.

[0009] In an optional implementation, the specific calculation logic for grouping and fusing the multiple independent probability signals corresponding to each event according to preset semantic attributes includes: The membership feature is calculated by performing a geometric mean fusion of the clustering probability signal and the directional mixed probability signal based on adaptive weighting coefficients; and / or, The location feature is calculated by: calculating the geometric mean of the density probability signal and the contour probability signal to achieve soft logic AND gate decision; and / or, The directional safety feature is calculated as follows: the minimum value between the directional probability signal and the quadrant probability signal is taken, and an exponential modulation is applied to it based on the severity index that characterizes the degree of sample contamination, wherein the severity index is adaptively determined based on the shoulder-peak ratio and fragment ratio in the dimer direction and the fragment direction.

[0010] In an optional implementation, in the third-level decision-making step, adaptive weights are determined based on the distribution quality index of events in the feature space, and the adaptive weights are used to perform a nonlinear combination mapping on the plurality of intermediate features, including: Based on the statistical characteristics of event distribution in the feature space, the contamination degree, core quality, and confidence degree of the original data are calculated as the distribution quality indicators. The membership weight, location weight, and directional safety weight are calculated based on the distribution quality index; where the higher the core quality, the greater the location weight; and the higher the contamination degree, the greater the directional safety weight. After performing Logit transformation on the membership feature, the position feature, and the orientation safety feature respectively, the normalized membership weight, position weight, and orientation safety weight are linearly combined and an adaptive bias is added. The continuous single-cell probability score is then output through Sigmoid function mapping. When the event-by-event standard deviation of the multiple intermediate features exceeds the preset conflict threshold, the conflict detection mechanism is activated, and the continuous single-cell probability score of the event is pulled toward an intermediate uncertainty value through a smoothing gating function.

[0011] In an optional implementation, the process of acquiring the clustering signal relies on a direction-aware resampling clustering strategy, specifically including: Density clustering is performed in the feature space to identify single-cell principal clusters, and a two-component Gaussian mixture model is fitted within the single-cell principal clusters to obtain the corresponding two components; If the separation between the two components reaches a preset trigger threshold, structural fusion is determined to exist, triggering direction-aware power attenuation resampling, including: estimating the dimer direction vector, calculating the forward projection distance of each event to the single-cell density peak position in the dimer direction; based on the forward projection distance and adaptive attenuation scale parameters, calculating the retention probability according to the power attenuation function, and performing Bernoulli sampling on each event to selectively dilute the event density in the dimer direction; Density clustering is re-run on the resampled data cloud to obtain the clustering signals for each event.

[0012] In an optional implementation, the step of running density clustering in the feature space to identify single-cell principal clusters includes: Density clustering is performed based on the initial minimum cluster size and the initial minimum number of samples, and the proportion of the main cluster to all events is calculated. If the ratio does not reach the target threshold, then an adaptive parameter relaxation iteration is initiated: within the maximum number of attempts, the minimum cluster size and the minimum number of samples are reduced proportionally each time based on a preset decay coefficient, and density clustering is rerun under the constraint that the parameters are not lower than the preset absolute hard lower limit. The iteration terminates when the proportion reaches the target threshold, or after reaching the maximum number of attempts, the result with the largest proportion of principal clusters among all attempts is selected as the final single-cell principal cluster.

[0013] In an optional implementation, the power attenuation function is: S=1 / [1+(d hi / τ) p ]; Where S is the basic attenuation value; d hi τ is the forward projection distance; τ is the adaptive attenuation scale parameter; p is the preset attenuation index; Preferably, the adaptive decay scale parameter τ is determined by selecting a preset proportion of events as the core region centered on the single-cell density peak, and taking the 90th percentile of the forward projection subset in the core region as the adaptive decay scale parameter τ.

[0014] In an optional implementation, separating single-cell target events from the raw data based on the consecutive single-cell probability scores of each of the events includes: The continuous single-cell probability scores were spatially rescored using a score-weighted density gating method, and local averaging was performed in two-dimensional space based on spatial consistency using Nadaraya-Watson kernel regression. After stripping away the extremely high-scoring peak components, on the one-dimensional density distribution curve of the remaining scores, the first consecutive valley point where the slope of the density curve turns from negative to positive is searched from the high-scoring end to the low-scoring end as the separation threshold. By using the aforementioned separation threshold and combining it with connected component analysis, single-cell target events are separated.

[0015] In an optional implementation, before extracting signal features from at least two physical channel dimensions of each event in the original data to construct the feature space, a microsphere removal step is further included: For the cerium-140 microsphere calibration channel, an outlier shape metric is calculated based on the signal distribution of the cerium-140 microsphere calibration channel. When the outlier shape metric reaches a preset condition, the high percentile of the signal of the cerium-140 microsphere calibration channel is used as an adaptive threshold to filter events with signals higher than the adaptive threshold as microsphere candidate events. When the outlier shape metric does not reach the preset condition, it is determined that there is no significant microsphere population in the original data and microspheres are not removed. The candidate microsphere events are confirmed using DNA content signals containing Ir191 and Ir193. When the DNA content signal of a candidate microsphere event is lower than a preset low percentile threshold, the candidate microsphere event is determined to be a microsphere and deleted.

[0016] In an optional implementation, the fully automated preprocessing method for mass cytometry data further includes a three-layer progressive artifact removal step: The first layer: extreme value detection is performed based on the event length signal, and robust Z-score is calculated based on the ratio of DNA content signals containing Ir191 and Ir193 in the transform domain to detect ratio anomalies, thereby screening abnormal events. The second layer: the events are segmented according to the collection time sequence. Based on the statistics of the event length signal and DNA content signal of each segment, the change point detection algorithm is used to identify the mutation segments that deviate from the overall distribution and the events falling into them are marked as anomalies. The third layer: For the events marked in the first and second layers, a robust Z-score is calculated based on multiple quality control channel signals such as event length, DNA content signal, viability, and microsphere channel. The number of channels exceeding the threshold in each event is counted, and those exceeding the preset number are marked as high-confidence artifacts confirmed by multiple channels. The final events removed in the artifact removal step are the union of the events marked in the first and second layers.

[0017] In an optional implementation, after outputting the preprocessed single-cell dataset, the following steps are also included: The extracted signal features also include the viability label channel signal; the distribution pattern of the viability label channel signal is determined by peak detection using kernel density estimation. When it exhibits a bimodal distribution, the intersection point of the two components in the two-component Gaussian mixture model is used as an adaptive valley threshold to remove dead cells; when it exhibits a unimodal distribution, percentile, inflection point, and peak value plus multiple times the absolute median difference are tried sequentially as backoff threshold strategies. Based on a tree-like hierarchical classification structure with four preset levels, a unique final classification label is assigned to each event in the original data; Preferably, the final classification label includes at least one of single cell, dual cell, fragment, microsphere, artifact, and dead cell.

[0018] Secondly, this application provides a fully automated preprocessing device for mass spectrometry flow cytometry data, comprising: The extraction module is used to acquire raw data from mass spectrometry flow cytometry and extract signal features from at least two physical channel dimensions of each event in the raw data to construct a feature space. The calculation module is used to calculate the single-cell probability score of each event in the feature space through a hierarchical probability fusion model, including: First-level calibration step: Based on the signal features of each event in the feature space, calculate the original judgment signal of each event under multiple different evaluation dimensions, and calibrate and map the original judgment signal to a unified probability interval to obtain multiple independent probability signals corresponding to each event; The second-level semantic step is to group and fuse the multiple independent probability signals corresponding to each event according to preset semantic attributes to generate multiple intermediate features for characterizing the state of the event. The third-level decision-making step is to determine adaptive weights based on the distribution quality index of events in the feature space, and to use the adaptive weights to perform nonlinear combination mapping on the multiple intermediate features to output the continuous single-cell probability score of each event. The output module is used to separate single-cell target events from the original data based on the continuous single-cell probability scores of each event, and output the preprocessed single-cell dataset.

[0019] Thirdly, this application provides a computer device, the computer device including a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the fully automated preprocessing method for mass spectrometry flow cytometry data as described in any of the foregoing embodiments.

[0020] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed on a processor, implements the fully automated preprocessing method for mass spectrometry flow cytometry data according to any one of the foregoing embodiments.

[0021] The fully automated processing method provided in this application enables fully automated, unsupervised processing of mass spectrometry flow cytometry data, effectively avoiding the subjectivity, non-reproducibility, and time-consuming inefficiency caused by manual intervention. By establishing a hierarchical probabilistic fusion architecture, the entire processing flow has high transparency and interpretability, effectively overcoming the black-box defects of existing deep learning algorithms, and eliminating the need to rely on large-scale manually labeled datasets, thus significantly improving the standardization level and automation efficiency of data processing.

[0022] In the first-level calibration and second-level semantic processing, the original judgment signals from multiple different evaluation dimensions are mapped to a unified probability interval and strictly grouped and fused according to preset semantic attributes. This significantly improves the boundary ambiguity problem caused by the compression of multidimensional information into a single probability in traditional parametric hybrid models. This processing method fully preserves the independent semantic information of different physical and evaluation dimensions, enabling the system to accurately distinguish whether the boundary ambiguity of an event in multidimensional space is caused by distance or directional deviation, thus representing the true physical state of complex target events more meticulously and accurately.

[0023] In the third-level decision processing, the distribution quality index of events in the feature space is used to dynamically determine adaptive weights, and intermediate features are nonlinearly combined and mapped accordingly, giving the method a strong dynamic adaptability. Faced with data containing a high proportion of heterogeneous impurities or complex non-axis-aligned boundaries, the dynamic weighting mechanism can intelligently adapt to the data quality fluctuations of different samples, overcoming the limitation of traditional one-dimensional density threshold cutting in handling nonlinear boundaries. The final output continuous single-cell probability score has extremely high reliability, significantly improving the accuracy and robustness of separating high-quality single-cell target events from massive high-dimensional heterogeneous noise. Attached Figure Description

[0024] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and therefore should not be considered as a limitation on the scope of protection of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a schematic diagram of the hardware operating environment involved in the embodiment of the fully automated preprocessing method for mass spectrometry flow cytometry data of this application; Figure 2 This is a flowchart illustrating Example 1 of the fully automated preprocessing method for mass spectrometry flow cytometry data in this application; Figure 3 This is a detailed flowchart of step S100 in Example 2 of the fully automated preprocessing method for mass spectrometry flow cytometry data in this application. Figure 4 This is a detailed flowchart of step S200 in Example 3 of the fully automated preprocessing method for mass spectrometry flow cytometry data in this application. Figure 5 This is a detailed flowchart of step S300 in Example 5 of the fully automated preprocessing method for mass spectrometry flow cytometry data in this application. Figure 6 An overview diagram of the overall process and hierarchical probabilistic fusion architecture (HPF) of the method provided for the experimental examples in this application; Figure 7 This is a diagram illustrating the principle of microsphere and artifact removal and the distribution of data preprocessing in the experimental examples of this application; Figure 8 This is a heatmap and median statistics of the single-cell calibration probability signal based on the zero layer of HPF in the experimental examples of this application; Figure 9 This is a single-cell probability heatmap and median statistics plot based on the first-layer semantic fusion features of HPF in the experimental examples of this application. Figure 10This is a single-cell score distribution map based on weighted density gating spatial rescoring and valley threshold detection in the experimental examples of this application; Figure 11 This is a schematic diagram of the one-dimensional distribution and two-dimensional scatter plot of dead cells detected based on GMM two-component fitting in the experimental examples of this application. Figure 12 This is a diagram showing the multi-stage progressive gating and the final classification results of the GatingTree in the experimental examples of this application; Figure 13 This is a diagram showing the results of fully automated single-cell separation on tangential, intersecting, and fused samples in the experimental examples of this application. Figure 14 This is a schematic diagram of the functional modules and connection methods of the fully automated preprocessing device for mass spectrometry flow cytometry data in this application. Detailed Implementation

[0026] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0027] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0028] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0029] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0030] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0031] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0032] like Figure 1 The diagram shown is a structural schematic of the hardware operating environment of the terminal involved in the embodiments of this application.

[0033] This application's embodiment of a fully automated preprocessing system (device) for mass spectrometry flow cytometry data can be a PC, or a mobile terminal device such as a smartphone, tablet, or laptop. The system may include: a processor 1001, such as a CPU; a network interface 1004; a user interface 1003; a memory 1005; and a communication bus 1002. The communication bus 1002 is used to establish communication between these components. The user interface 1003 may include a display screen, an input unit such as a keyboard, or a remote control; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a stable memory, such as a disk storage device. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001. Optionally, the fully automated preprocessing system for mass spectrometry flow cytometry data may also include RF (Radio Frequency) circuitry, audio circuitry, a Wi-Fi module, etc. In addition, the system can also be equipped with other sensors such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, which will not be described in detail here.

[0034] Those skilled in the art will understand that Figure 1 The system shown is not intended to limit it and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a data interface control program, a network connection program, and a fully automated preprocessing program for mass spectrometry flow cytometry data.

[0035] In summary, the method provided in this application achieves efficient, accurate and highly automated carbon monoxide concentration detection through automatic sample gas identification, real-time baseline updating and dynamic closed-loop cleaning based on real-time residual state.

[0036] Example 1 Reference Figure 2 This embodiment provides a fully automated preprocessing method for mass cytometry data. During mass cytometry, the detection process generates raw data containing tens of thousands or even millions of events. These events not only include real single-cell targets but also contain a large number of artifacts, cell debris, and duplexes formed by multiple cell adhesions. To ensure the accuracy of downstream data analysis, these events need to be cleaned and preprocessed. Specifically, this includes: Step S100: Obtain raw mass cytometry data and extract signal features from at least two physical channel dimensions of each event in the raw data to construct a feature space.

[0037] The aforementioned “event” refers to a set of independent particle signals detected by a single flow cytometer reading; the “physical channel dimension” reflects the properties of the event under specific physical or chemical labeling.

[0038] As an optional implementation, the physical channel dimension may include, but is not limited to, channels reflecting the cellular nucleic acid content and length channels reflecting the duration of an event in the detector. By extracting these features, the messy raw data is mapped into a mathematical feature space with spatial geometric relationships (such as distance and density), laying the foundation for subsequent algorithm analysis.

[0039] After constructing the feature space, this method uses a hierarchical probability fusion model to calculate the single-cell probability score of each event. Traditional single algorithms are often prone to getting stuck in local optima or making misjudgments when dealing with complex data, while the hierarchical probability fusion model adopted in this application solves this problem through three progressive levels.

[0040] Specifically, step S200 involves calculating the single-cell probability score of each event in the feature space using a hierarchical probability fusion model, including: Step S210, first-level calibration step: Based on the signal features of each event in the feature space, calculate the original judgment signal of each event under multiple different evaluation dimensions, and calibrate and map the original judgment signal to a unified probability interval to obtain multiple independent probability signals corresponding to each event.

[0041] The model calculates the original decision signals for each event under multiple different evaluation dimensions based on the signal characteristics of each event in the feature space. The evaluation dimension refers to the quantitative evaluation of the event from different algorithmic or geometric perspectives, such as spatial clustering, local spatial density, or feature vector direction. Since these original decision signals of different dimensions often have completely different physical dimensions and value ranges (e.g., density may be hundreds or thousands of counts, while distance is a small floating-point number), direct numerical calculations would cause large numerical signals to mask small numerical signals. Therefore, this step further calibrates and maps these original decision signals to a unified probability interval (usually [0,1]), thereby obtaining multiple independent probability signals corresponding to each event. This mapping can be achieved through a sigmoid growth function (such as the sigmoid function) or other nonlinear normalization methods, ensuring fair comparison of heterogeneous signals on the same mathematical scale.

[0042] Step S220, second-level semantic step: group and fuse the multiple independent probability signals corresponding to each event according to preset semantic attributes to generate multiple intermediate features for characterizing the state of the event.

[0043] After obtaining multiple independent probability signals, the model groups and fuses the probability signals corresponding to each event according to preset semantic attributes. Here, "semantic attributes" refer to the logical meaning represented by the probability signals when evaluating the event state. For example, some signals collectively reflect the spatial safety of the event's location, while others collectively reflect the degree of membership of the event in a certain community. By grouping and fusing probability signals with relevant semantics, multiple intermediate features are generated to characterize the state of the event.

[0044] It should be noted that the 'event state (or event status)' mentioned in this embodiment does not refer to the conventional physical state such as the biological activity of cells, but specifically refers to the topological and geometric distribution state of each event in a high-dimensional feature space. Specifically, this state synthesis can characterize the certainty of an event belonging to the main population of single cells (corresponding to membership feature), the core safety level in the spatial density distribution (corresponding to position feature), and the degree of safety of its deviation from the contamination direction of common impurities (such as duplexes or fragments) (corresponding to directional safety feature).

[0045] This semantic-based grouping and fusion mechanism effectively avoids the "one-vote veto" phenomenon caused by directly multiplying multidimensional features, thus enhancing the robustness of feature representation.

[0046] Step S230, third-level decision-making step: Determine adaptive weights based on the distribution quality index of events in the feature space, and use the adaptive weights to perform nonlinear combination mapping on the multiple intermediate features to output the continuous single-cell probability score of each event.

[0047] To address the significant data quality variations among different detection samples, this step introduces an adaptive mechanism. The system determines adaptive weights based on a global distribution quality index of events in the feature space (e.g., a statistical indicator representing the separation of high-density regions in the feature space or the overall noise ratio). Subsequently, these adaptive weights are used to perform a nonlinear combination mapping on the previously generated intermediate features. In this process, the distribution quality index acts as a "feedback regulator," for example, adaptively increasing the weights of certain intermediate features sensitive to adhesion when samples are severely clustered. Finally, this level outputs a continuous single-cell probability score for each event, which precisely quantifies the likelihood that each event belongs to a real, high-quality single cell.

[0048] Step S300: Based on the continuous single-cell probability scores of each event, separate the single-cell target events from the original data and output the preprocessed single-cell dataset.

[0049] Since a smooth and high-confidence continuous score has been obtained, the system can separate single-cell target events from the original data, for example, by defining the boundary by setting a preset threshold or adaptively finding the valley of the score distribution curve, and finally outputting a cleaned and preprocessed single-cell dataset for downstream high-dimensional bioinformatics analysis.

[0050] In summary, this method achieves fully automated unsupervised processing of mass spectrometry flow cytometry data, eliminating the subjectivity and inefficiency of manual operation, and overcoming the shortcomings of deep learning models, such as lack of interpretability and reliance on labeled data. By calibrating and mapping multidimensional raw decision signals and grouping and fusing them according to semantic attributes to generate intermediate features, it effectively avoids the crude compression of multidimensional information by traditional hybrid models, preserves the independent semantics of different physical properties, and accurately distinguishes complex boundary states. Furthermore, by dynamically determining adaptive weights using data distribution quality indicators and performing nonlinear combination mapping on intermediate features, the processing can intelligently adapt to data with different qualities and complex nonlinear distributions, breaking through the limitations of traditional fixed-dimensional threshold cutting. Finally, it outputs highly reliable continuous single-cell probability scores, significantly improving the accuracy and robustness of single-cell target extraction.

[0051] Example 2 Reference Figure 3Based on the aforementioned embodiments, in order to construct a feature space that can accurately reflect the true physical state of an event, this embodiment extracts signal features from at least two physical channel dimensions of each event in the original data. Specifically, this includes the extraction and standardization transformation of DNA content signals and event length signals. The specific process is as follows: Step S100 involves extracting signal features from at least two physical channel dimensions of each event in the original data to construct a feature space, including: Step S110: Obtain the original DNA content signal and event length signal.

[0052] In mass cytometry, the DNA content signal (usually labeled with specific metal isotopes, such as Ir191 or Ir193 channels) directly reflects the abundance of nucleic acid within the detected event and is the core basis for determining whether it is an intact cell. The event length signal records the time span that the ion cloud of the event takes to pass through the time-of-flight detector, mainly used to help determine the target volume and whether adhesion has occurred. These two constitute the basic physical dimensions for identifying single cells.

[0053] Step S120: Based on a preset cofactor, perform an inverse hyperbolic sine transform on the original DNA content signal and event length signal to maintain linear resolution in the low signal region and compress the dynamic range in the high signal region.

[0054] The raw signal output from mass spectrometry flow cytometry instruments has an extremely large dynamic range, and is often accompanied by background noise interference in the very low-value region. The introduction of the inverse hyperbolic sine transform utilizes its unique mathematical properties: in the low-signal region, this transform exhibits an approximately linear characteristic, thus stably preserving the distribution characteristics of small signals and avoiding numerical divergence or distortion in low-signal or zero-value regions caused by conventional logarithmic transforms; while in the high-signal region, the transform smoothly transitions to a logarithmic characteristic, effectively compressing the dynamic range of extremely high signals. A preset cofactor is used to adjust the scaling ratio of the linear to logarithmic interval transition (for example, the cofactor is often set to 5 in mass spectrometry flow cytometry data processing). Through this transform, channel signals of different orders of magnitude are normalized to a similar numerical scale.

[0055] In a preferred embodiment, the preset cofactor can be dynamically and adaptively adjusted based on the signal-to-noise ratio (SNR) of the current detection channel. When the standard deviation of the detected signal channel's noise floor is σ... noise At that time, the cofactor is calculated as 5.0 × (σ noise / σ ref ), where σ refThis represents the reference noise level under standard instrument conditions. This signal-to-noise ratio-based adaptive mechanism enables the feature space construction to be compatible with detection instruments from different generations and with different background noise levels.

[0056] Step S130: The event length signal after inverse hyperbolic sine transformation is superimposed with a uniformly distributed random jitter of a preset interval to eliminate artifacts generated by discrete integer values ​​in density estimation, and a two-dimensional feature space is constructed based on the processed DNA content signal and the event length signal.

[0057] Furthermore, a uniformly distributed random jitter within a preset interval is superimposed on the event length signal after the inverse hyperbolic sine transform. Due to the sampling mechanism of the instrument, the event length is usually recorded as a discrete integer value. When performing high-dimensional spatial distance calculations or local density estimation, these numerous overlapping discrete integers cause the data to exhibit discontinuous strip-like clusters in the feature space, i.e., forming density estimation artifacts. Such artifacts can easily interfere with subsequent density-sensitive algorithms (such as density clustering), causing erroneous cluster truncation. To eliminate this phenomenon, this embodiment introduces a uniformly distributed random jitter within a preset interval (e.g., adding random noise distributed in the range [-0.5, 0.5]) onto the event length signal, causing events that originally overlapped on the same integer scale to produce small, continuous shifts locally.

[0058] Finally, a two-dimensional feature space is constructed based on the DNA content signal and the event length signal after inverse hyperbolic sine transform and superimposed random jitter processing. This feature space eliminates local extremum interference and dimensional differences in discrete data, making the data distribution continuous and smooth, and providing a reliable underlying coordinate system for subsequent accurate spatial state assessment and probability fusion calculation.

[0059] In some embodiments, due to the extremely complex data distribution in the multidimensional feature space, a single evaluation index is insufficient to comprehensively and accurately define single-cell targets. Therefore, this embodiment introduces multiple raw judgment signals with different evaluation dimensions in the first-level calibration step.

[0060] Specifically, in the first-level calibration step, the original judgment signals under the multiple different evaluation dimensions include: clustering signal, direction mixing signal, density signal, contour line signal, direction signal, and quadrant signal.

[0061] It should be noted that the above six signals represent six independent physical and geometric evaluation perspectives: among them, the clustering signal is used to characterize whether an event belongs to the main group in the unsupervised topology; the directional mixed signal characterizes the probability of the event occurring based on the statistical distribution model; the density signal reflects the degree of clustering in the local feature space where the event is located; the contour line signal reflects the hierarchical height of the event in the overall density distribution topology; the directional signal is used to characterize the angular characteristics of the event deviating from the core of the group; and the quadrant signal is used to mark the absolute region coordinates of the event in the feature space.

[0062] Because the calculation principles of the six original decision signals are different, and their numerical ranges and physical dimensions are completely different (e.g., including real distances, integer counts, and angle vectors), they cannot be directly weighted. Therefore, in step S210, the original decision signals are respectively calibrated and mapped to a unified probability interval, specifically as follows: Step S211: By anchoring the Sigmoid function, each of the original decision signals is mapped to the [0,1] interval to obtain six independent probability signals. The six independent probability signals are: clustering probability signal, direction mixing probability signal, density probability signal, contour probability signal, direction probability signal, and quadrant probability signal.

[0063] In this process, the anchored sigmoid function smoothly transforms the deviation into a standard probability value non-linearly by setting anchor points (such as specific density thresholds or distance thresholds) corresponding to the physical meaning of each signal. After mapping, the dimensional differences are eliminated, resulting in the six independent probability signals mentioned above.

[0064] After obtaining the standardized probability signals, directly performing full fusion can easily lead to extreme weak signals in one dimension masking high-confidence signals in other dimensions. Therefore, in the second-level semantic step, a semantic attribute-based grouping fusion mechanism is employed to generate multiple intermediate features representing the state of the event. These intermediate features specifically include: membership features, location features, and orientation safety features.

[0065] In the second-level semantic step, the multiple intermediate features include: membership features, positional features, and orientation safety features. Step S220 involves grouping and fusing the multiple independent probability signals corresponding to each event according to preset semantic attributes, including: Step S221: The clustering probability signal and the directional mixing probability signal are fused to generate the membership feature. Since both the clustering and directional mixing signals are logically used to answer the question of whether an event "belongs to the target group," their fusion can provide a high-confidence determination of group membership.

[0066] Step S222: The density probability signal and the contour probability signal are fused to generate the location feature; both density and contour signals focus on the spatial aggregation state of events, and their combination can accurately characterize whether an event is located in a high-density security center area.

[0067] Step S223: The direction probability signal and the quadrant probability signal are fused to generate the direction safety feature. Since duplex data and cell debris are usually located in specific spatial directions and quadrants, fusing these two signals reflecting spatial orientation and regional coordinates can effectively construct a safety barrier targeting the direction of structural contamination.

[0068] Through the aforementioned grouping mechanism, the six heterogeneous signals are decoupled and recombined into three intermediate features with clear physical representations. This preserves the independence of multi-perspective evaluation and provides structured data support for subsequent dynamic integrated decision-making.

[0069] In some specific embodiments, transforming independent probability signals into intermediate features that can accurately characterize the event state requires matching corresponding mathematical calculation logic for different semantic attributes. Specifically, for the grouping and fusion of membership features, location features, and orientation safety features, step S200, which groups and fuses the multiple independent probability signals corresponding to each event according to preset semantic attributes, includes the following specific calculation logic: (1) The membership feature is calculated by performing geometric averaging fusion of the clustering probability signal and the directional mixed probability signal based on adaptive weighting coefficients.

[0070] Clustering probability signals focus on unsupervised topological partitioning, while directional mixed probability signals focus on parameter fitting of statistical models. Using geometric mean fusion establishes a "soft logic AND gate" relationship between the two; that is, only when an event exhibits a high single-cell membership probability at both the clustering topology and statistical model levels can a high membership feature value be obtained. Any low probability in a single dimension will significantly drag down the final membership value through the geometric mean mechanism. Simultaneously, the introduction of adaptive weighting coefficients allows the system to adaptively adjust the dependence weights on clustering results or model fitting results based on the current data distribution of the samples, thereby improving adaptability to heterogeneous data.

[0071] (2) The location feature is calculated by calculating the geometric average of the density probability signal and the contour probability signal to realize the soft logic AND gate decision.

[0072] Density probability signals reflect the degree of clustering in the local feature space of an event, while contour probability signals reflect the topological height of the event within the overall distribution. In actual detection data, some impurities may form high-density local clusters in edge regions. By calculating the geometric mean of these two signals, a soft logic AND gate is implemented to ensure that only events located in high-density regions and situated in the core highlands (i.e., non-edge regions) of the overall topology are assigned high-position feature values, thereby effectively filtering out interference from locally clustered impurities.

[0073] (3) The calculation method of the directional safety feature is as follows: take the minimum value between the directional probability signal and the quadrant probability signal, and apply exponential modulation to it based on the severity index that characterizes the degree of sample contamination, wherein the severity index is adaptively determined based on the shoulder peak ratio and fragment ratio in the dimer direction and fragment direction.

[0074] Specifically, the quantitative calculation of severity is as follows: the severity of pollution is estimated from two dimensions, namely the dimer direction and the fragment direction, and the maximum value of the two is taken as the comprehensive severity.

[0075] Dimer severity: Based on the projection of events within the main cluster onto the dimer direction, calculate the robust median med and robust absolute median difference MAD, defining a 2σ threshold = med + 2 × MAD. Calculate the proportion of events exceeding this threshold frac_main in the main cluster and the proportion of all events exceeding this threshold frac_all, obtaining the shoulder ratio = frac_all / frac_main. Dimer severity is the value that restricts (shoulder ratio - 1) / 4 to the interval [0, 1].

[0076] Fragmentation severity: Q, the 10th percentile of the principal cluster density distribution 10 As a low-density threshold, calculate the fragmentation ratio = (of all events below Q) 10 (proportion) / (prior clusters with values ​​below Q) 10 The fragmentation severity is a value that limits (fragmentation ratio - 1) / 4 to the range [0, 1].

[0077] The overall severity is the maximum of the dimer severity and the fragmentation severity, limited to the range [0.20, 1.0]. The lower limit of 0.20 ensures minimal directional safety constraints even in clean samples. Finally, a severity exponential modulator is applied to the minimum value, achieving adaptive adjustment of the penalty intensity according to the degree of data contamination.

[0078] Example 3 Reference Figure 4Based on the aforementioned embodiments, when fusing multi-dimensional intermediate features, a uniform fixed weight allocation model is often insufficient to adapt to data pattern fluctuations caused by different batches or quality test samples. Therefore, this embodiment introduces an adaptive dynamic weighting and nonlinear combination mechanism based on the global statistical distribution characteristics of the samples in the third-level decision-making step to output high-confidence continuous single-cell probability scores.

[0079] Specifically, in the third-level decision-making step, step S230, determining adaptive weights based on the distribution quality index of events in the feature space, and using the adaptive weights to perform a nonlinear combination mapping on the multiple intermediate features, includes: Step S231: Based on the statistical characteristics of event distribution in the feature space, calculate the contamination degree, core quality, and confidence degree of the original data as the distribution quality index.

[0080] Specifically, contamination level characterizes the proportion of potential cell debris or impurities formed by the adhesion of multiple targets in the sample; core quality characterizes the cohesion and structural clarity of the main target population (i.e., high-density single-cell cloud) in the feature space; and confidence level reflects the statistical reliability of the spatial distribution pattern of the sample. By extracting these three macroscopic indicators, the system can "perceive" the difficulty and distribution pattern of the current sample being processed.

[0081] Step S232: Calculate the corresponding membership weight, position weight, and directional safety weight based on the distribution quality index; wherein, the higher the core quality, the greater the position weight; and the higher the contamination degree, the greater the directional safety weight.

[0082] Building upon the preceding steps, the system calculates the corresponding membership weights, location weights, and directional safety weights based on the distribution quality index, achieving a dynamic balance across different evaluation perspectives. The dynamic adjustment logic is as follows: when high core quality is detected, indicating a clear sample structure and regular boundaries, the system adaptively increases the location weight, making the final decision more dependent on the spatial physical coordinates and density topology of the event. Conversely, when high contamination is detected, indicating significant overlap and adhesion within the sample, relying solely on spatial location is prone to misjudgment. In this case, the system adaptively increases the directional safety weight, strengthening the penalty and removal of events in high-risk contamination directions. This dynamic game-theoretic mechanism allows the feature fusion process to automatically adapt to the physical condition of the samples themselves.

[0083] Step S233: After performing Logit transformation on the membership feature, the position feature, and the orientation safety feature respectively, the normalized membership weight, position weight, and orientation safety weight are linearly combined and an adaptive bias is added. Then, the continuous single-cell probability score is output through Sigmoid function mapping.

[0084] After determining the normalized weights for each direction, the system performs a nonlinear combination mapping to calculate the final score. Considering that the membership features, positional features, and orientation safety features are all probability values ​​within the [0,1] interval, directly performing linear weighting in the probability space would lead to severe edge distortion and loss of nonlinear relationships. Therefore, this embodiment performs a Logit transformation (i.e., log-odds transformation) on the three intermediate features, smoothly stretching the probability values ​​originally limited to the [0,1] interval to the real number axis space. Subsequently, in the real number domain, the aforementioned normalized membership weights, positional weights, and orientation safety weights are used to linearly combine the transformed features, and an adaptive bias derived based on the overall confidence level is added. Finally, the weighted summation of the real numbers is remapped back to the [0,1] interval via the Sigmoid function, outputting the continuous single-cell probability scores for each event. This process of nonlinear transformation-linear combination-nonlinear mapping mathematically constructs a rigorous probability fusion channel, capable of keenly capturing and amplifying the subtle advantages and disadvantages between features.

[0085] Specifically, the quantitative calculation of the distribution quality index is as follows: Contamination level: Calculate the robust 2σ threshold of the principal cluster projection values ​​in the projection space along the dimer direction. The ratio of the proportion of all events actually exceeding this threshold to the expected proportion (0.0228) is defined as the shoulder ratio. The contamination level is the value that restricts (shoulder ratio - 1) / 8 to the interval [0, 1].

[0086] Core quality: Fisher kurtosis and skewness are calculated for the log density distribution of events within the main cluster, and then normalized to core sharpness and core symmetry, respectively. The weighted combination is core quality = 0.6 × core sharpness + 0.4 × core symmetry.

[0087] Confidence level: Based on the consistency between core quality and (1-contamination level), it is estimated by the standard deviation of the two and restricted to the [0,1] interval.

[0088] Based on the above indicators, dynamic adjustment and nonlinear mapping are performed: Logit transformation is applied to the three intermediate features respectively. Linear mapping is used to calculate the weights: membership baseline weight w_1 = 1.0; position weight w_2 = 1.0 × (0.5 + core quality); orientation safety weight w_3 = 0.6 × (0.7 + 0.6 × contamination); bias b = 0.0 - 0.3 × contamination. When the confidence level < 0.5, the weights are back to the baseline value through linear interpolation. To prevent distortion, safety clamping is applied to the weights: membership and position weights are clamped in [0.3, 1.5], orientation weight is clamped in [0.2, 1.2], and bias is clamped in [-0.5, 0.5]. Finally, the weights are normalized, and a weighted linear combination of the Logit-transformed features is performed in the real number domain with the bias added. Then, the result is remapped back to the [0, 1] interval using the Sigmoid function, outputting a continuous single-cell probability score.

[0089] Step S234: When the event-by-event standard deviation of the multiple intermediate features exceeds the preset conflict threshold, the conflict detection mechanism is activated, and the continuous single-cell probability score of the event is pulled toward the intermediate uncertainty value through a smoothing gating function.

[0090] Furthermore, to prevent misjudgments by the algorithm in extreme edge cases, this embodiment also designs a low-level conflict detection mechanism. During the calculation process, the system synchronously calculates the event-by-event standard deviation of the multiple intermediate features corresponding to each event. When some events exhibit extreme contradictions on different intermediate features (e.g., extremely high membership but extremely low directional safety, causing the standard deviation to exceed the preset conflict threshold), the system determines that the event is in the blind spot of the evaluation perspective or a highly overlapping fuzzy boundary. At this time, the conflict detection mechanism is activated. By introducing a smoothing gating function, a damping effect is applied to the fusion score, forcibly pulling the continuous single-cell probability score of the event towards an intermediate uncertainty value near 0.5. This mechanism effectively avoids the system's "overconfidence" in controversial data, ensuring that the score distribution curve forms a clear trough at the boundary between single cells and impurities, thereby providing a safe and clear cutting basis for subsequent separation steps.

[0091] In some implementations, the clustering signal serves as an important original decision signal characterizing the topological structure of the event space, and the accuracy of its acquisition directly affects the reliability of subsequent semantic fusion.

[0092] When conventional density clustering methods process samples with a high proportion of duplexes (i.e., detection signals generated by the adhesion of two or more cells), the edge density of the duplex community may be similar to that of the main single-cell community, causing the two to form a high-density "bridge" in the feature space, which is then incorrectly merged into the same cluster by the algorithm.

[0093] To overcome this limitation, in this embodiment, the process of acquiring the clustering signal relies on a direction-aware resampling clustering strategy, and the specific implementation logic is as follows: (1) Density clustering is performed in the feature space to identify the single-cell principal cluster, and a two-component Gaussian mixture model is fitted inside the single-cell principal cluster to obtain the corresponding two components. This operation constitutes an internal structure verification mechanism, which aims to detect population heterogeneity under high-density masking.

[0094] (2) If the separation between the two components reaches a preset trigger threshold, it is determined that there is structural fusion and triggers direction-sensing power attenuation resampling, including: estimating the dimer direction vector, calculating the forward projection distance of each event to the single cell density peak position in the dimer direction; based on the forward projection distance and the adaptive attenuation scale parameter, calculating the retention probability according to the power attenuation function, and performing Bernoulli sampling on each event to selectively dilute the event density in the dimer direction.

[0095] Subsequently, the system evaluates the spatial separation between the two components. If the separation between the two components reaches a preset trigger threshold, it indicates that there are actually two statistically significantly different groups within the main cluster, and the current clustering result is determined to have a structural fusion of single cells and duplexes. After detecting the structural fusion, the system triggers a direction-sensing power attenuation resampling process.

[0096] Specifically, the process first estimates the dimer orientation vector. In the feature space of mass cytometry data, dimers typically have higher nucleic acid content and longer event lengths compared to single cells. This dimer orientation vector points to a specific angle in the feature space where these physical quantities simultaneously increase. Next, the system calculates the forward projection distance from each event in the main cluster to the single-cell density peak position along the dimer orientation. This projection distance quantifies the degree to which each event deviates from the single-cell core and penetrates deeper into the dimer contamination region.

[0097] After acquiring the projection distance, the system calculates the retention probability of each event based on the forward projection distance and the adaptive attenuation scale parameter, according to the power attenuation function. The power attenuation function provides a non-linear probability attenuation mechanism: events with small projection distances (close to the single-cell core) are assigned a retention probability close to 1; while for events with large projection distances (deep into the duplex and connecting bridge regions), the retention probability decreases power-law with increasing distance. The adaptive attenuation scale parameter is adaptively determined based on the statistical distribution characteristics of the current data projection distance, controlling the starting position and intensity of probability attenuation. Based on this, the system performs Bernoulli sampling on each event. Through this random discarding mechanism based on retention probability, the system selectively dilutes the event density in the dimer direction, artificially weakening the "connecting bridge" between single cells and duplexes, while completely preserving the true density topology of the single-cell core region.

[0098] (3) Re-run density clustering on the resampled data cloud to obtain the clustering signals for each event.

[0099] Finally, on the data cloud after resampling and breaking density bridges, the system re-runs density clustering. This time, because the density of the contiguous regions has been effectively diluted, the density clustering algorithm can accurately find the natural breakpoints in density, thus successfully separating the single-cell principal clusters from the duplexes, ultimately outputting the clustering signals that accurately reflect the true topological affiliation of each event. This orientation-aware resampling clustering strategy effectively enhances the robustness and boundary characterization ability of unsupervised clustering algorithms when facing highly heterogeneous and contaminated samples.

[0100] Example 4 Based on the aforementioned embodiments, extracting clustering signals typically relies on density-based clustering algorithms (e.g., density-based spatial clustering algorithms). However, due to the extremely high heterogeneity of biological samples detected by mass spectrometry flow cytometry, the density distribution and scattering of data from different batches vary greatly. If fixed clustering parameters are used, for sparse samples, it is very easy for the true single-cell population to be over-segmented or misclassified as background noise, thus making it impossible to extract complete single-cell principal clusters.

[0101] To address this issue, in this embodiment, the process of running density clustering in the feature space to identify single-cell principal clusters specifically includes an adaptive parameter relaxation iteration mechanism, implemented as follows: (1) Perform density clustering based on the initial minimum cluster size and the initial minimum number of samples, and calculate the proportion of the main cluster to all events.

[0102] These two initial parameters are typically set to relatively strict empirical values. The minimum cluster size defines the minimum number of events required to form an independent cluster, while the minimum number of samples controls the density threshold of the core point's neighborhood to adjust the algorithm's conservatism. After the initial clustering, the system counts the number of events in each cluster, selects the cluster with the most events as the initial single-cell master cluster, and calculates the proportion of events contained in this master cluster to the total number of events. In typical high-quality samples, single-cell events should constitute the absolute majority; therefore, this proportion can serve as a reliable indicator for evaluating the completeness of the clustering results.

[0103] (2) If the proportion does not reach the target threshold, then start the adaptive parameter relaxation iteration: within the maximum number of attempts, the minimum cluster size and the minimum number of samples are reduced proportionally based on the preset decay coefficient, and density clustering is rerun under the constraint that the parameters are not lower than the preset absolute hard lower limit.

[0104] In this step, the system compares the ratio with a preset target threshold. If the ratio does not reach the target threshold (e.g., indicating that a large number of real single-cell events are excluded from the main cluster due to overly strict parameters), an adaptive parameter relaxation iteration is initiated. In each iteration, the system progressively reduces the minimum cluster size and the minimum number of samples based on a preset decay coefficient (e.g., multiplying the current parameters by a decay coefficient less than 1), thereby relaxing the clustering restrictions and allowing connections to relatively sparse data regions. Considering that excessive parameter reduction may cause the algorithm to mistake small fragments for independent communities or lead to excessive consumption of computing resources, the system introduces parameter constraints during the iteration process, ensuring that the adjusted parameters are not lower than a preset absolute hard lower limit. Under the condition of satisfying this constraint, the system reruns density clustering using the updated parameters and recalculates the proportion of the main cluster.

[0105] Specifically, the iterative control logic for adaptive parameter relaxation includes a maximum number of attempts K (e.g., K=4) and a decay coefficient γ (e.g., γ=0.70). The parameters for the k-th attempt (k=0,1,...,K-1) are set as follows: Minimum cluster size (k) = max(30, [initial minimum cluster size × γ)] k ]); Minimum number of samples (k) = max(5, [initial minimum number of samples × γ)] k ).

[0106] Taking a total number of events N=100,000 participating in clustering as an example, the adaptive parameter reduction process is shown in Table 1: Table 1. Example of parameter adaptive relaxation iteration process

[0107] (3) When the proportion reaches the target threshold, the iteration is terminated, or after the maximum number of attempts is reached, the result with the largest proportion of the main cluster among all attempts is selected as the final single-cell main cluster.

[0108] This iterative mechanism is controlled by a dual exit condition: when the calculated proportion of the main cluster reaches the target threshold in a certain attempt, indicating that the single-cell entity has been completely identified, the system immediately terminates the iteration and uses the current clustering result as the final output; if, due to the extreme distribution of the samples themselves, the system still fails to meet the target threshold requirement after reaching the maximum number of attempts, a bottom-line guarantee mechanism is triggered. The system will traverse and compare the historical records of all iterative attempts and select the result with the largest proportion of the main cluster as the final single-cell main cluster. This mechanism, which combines adaptive parameter relaxation with bottom-line guarantee, allows the clustering process to dynamically adapt to heterogeneous samples with different densities, effectively avoiding the failure of fixed parameters under complex distributions and ensuring the stability and completeness of the clustering signal extraction.

[0109] Example 5 Based on the aforementioned embodiments, in order to effectively dilute the event density of the adhesion region while preserving the data structure of the real single-cell core region to the greatest extent possible without loss when breaking the "density connection bridge" between single cells and duplexes, the system employs a probabilistic calculation mechanism based on a nonlinear power attenuation function during the orientation-aware resampling process. Specifically, the power attenuation function is: S=1 / [1+(d hi / τ) p ]; Where S is the calculated base decay value, i.e., the probability that each event will be retained during subsequent Bernoulli sampling; d hi τ is the forward projection distance from the event to the single-cell density peak position in the dimer direction, which quantifies the physical degree of the event's shift towards the contaminated area; τ is an adaptive decay scale parameter used to define the spatial boundary at which the decay effect begins to have a significant effect; p is a preset decay index used to control the steepness of the probability decay curve when it crosses the boundary.

[0110] Through this nonlinear function, when the forward projection distance d of the event... hi When smaller (i.e., within the single-cell safety region), due to d hi The value of / τ is less than 1, and after high-order power p operations, it tends to 0, making the retention probability S extremely high, close to complete retention; while when the projected distance d of the event is less than 1, it tends to 0, making the retention probability S extremely high, close to complete retention; hi When the scale parameter τ is exceeded and penetrates into the adhesion bridge region, (d hi / τ) pThe value of increases sharply, causing the retention probability S to drop precipitously. This mathematical mechanism ensures gentle protection of the single-cell body and strong stripping of the boundary adhesion region during resampling.

[0111] Furthermore, since the absolute physical scale (e.g., the dispersion of cell communities) of mass spectrometry flow cytometry samples from different batches varies significantly in the feature space, using a fixed distance value as the attenuation boundary will lead to a decrease in system robustness. Therefore, the adaptive attenuation scale parameter τ is dynamically determined based on the data distribution pattern of the sample itself. The specific implementation process is as follows: A preset proportion of events are selected as the core region centered on the single-cell density peak, and the 90th percentile of the forward projection subset in the core region is taken as the adaptive attenuation scale parameter τ.

[0112] First, the system locates the density peak (i.e., the core point with the highest local density) of the main cluster of single cells in the feature space. Using this as the center, it selects a set of events with a predetermined proportion of high local density (e.g., the top half), defining this as the core region of the single cell. This core region represents the most representative subset of pure single cells in the sample, least affected by dimer contamination. Subsequently, the system extracts the forward projection distances of all events within this core region along the dimer direction, forming a forward projection subset. To find a robust statistical boundary that covers the vast majority of normal single cells while excluding extreme outliers, the system calculates the 90th percentile of this forward projection subset and directly assigns this percentile value to the adaptive decay scale parameter τ. Using this dynamic determination method, the decay scale τ can be strictly decoupled from and adaptively matched to the actual physical span of the currently processed sample. If the current sample's single-cell community distribution is relatively dispersed, the calculated τ value will increase accordingly, thus delaying the triggering position of decay and avoiding accidental damage to normal edge cells. If the sample community is highly compact, the τ value will decrease accordingly, ensuring timely and accurate density dilution of the adjacent adhering duplexes. This data-driven parameter adaptation mechanism greatly enhances the generalization ability and processing accuracy of this automatic preprocessing method across samples and instrument scenarios.

[0113] In some embodiments, after calculating the continuous single-cell probability scores for each event, these continuous probability values ​​need to be converted into hard classification boundaries to physically separate high-quality single-cell target events from the original data. Considering the microscopic fluctuations of high-dimensional biological data, directly using a fixed score threshold (such as an absolute 0.5) for segmentation often leads to misjudgment of boundary regions. Therefore, this embodiment constructs an adaptive score-weighted density gating and separation method, the specific implementation process of which is as follows: refer to Figure 5Step S300, based on the consecutive single-cell probability scores of each event, separates single-cell target events from the original data, including: Step S310: Spatial rescoring of the continuous single-cell probability scores is performed using a score-weighted density gating method, and local averaging is performed in two-dimensional space based on spatial consistency using Nadaraya-Watson kernel regression.

[0114] In a two-dimensional feature space, based on the spatial consistency assumption, events that are physically close are likely to have the same true class attributes. To eliminate microscopic isolation noise that may be introduced by independent computation, the system uses Nadaraya-Watson kernel regression to perform local averaging in the two-dimensional space. Specifically, this kernel regression method introduces a kernel function (e.g., a Gaussian kernel) to assign weights to the spatial distance between events as independent variables, with closer events receiving larger weights. Subsequently, the weighted average of the scores of all neighbors within a local neighborhood of an event is calculated to update the event's original probability score. After spatial re-scoring, the scores become highly smooth in spatial distribution, eliminating discrete extrema from anomalous changes and providing stable data support for subsequently depicting clear boundary distribution curves.

[0115] Step S320: After stripping the extremely high-scoring peak components, on the one-dimensional density distribution curve of the remaining scores, search for the first consecutive valley point where the slope of the density curve changes from negative to positive from the high-scoring end to the low-scoring end as the separation threshold.

[0116] In typical mass cytometry samples, there are numerous high-quality single cells with extremely well-defined characteristics. These events form a massive density peak near the maximum score (e.g., close to 1.0). Directly searching for the valley on this curve often masks the subtle fluctuations in the intermediate score range, causing the valley detection algorithm to fail. Therefore, the system first isolates the extremely high-score peak component in its computational logic (e.g., ignoring the extremely high density range with scores above 0.95), amplifying and focusing on the one-dimensional density distribution curve of the remaining ambiguous score range. On this curve, the system follows a directional search logic from the high-score end to the low-score end, calculating the first derivative (i.e., the slope) of the density curve. The system locates the first consecutive valley point where the slope of the density curve turns from negative to positive. This valley point statistically represents the optimal separation saddle between the tail of the high-score single-cell population distribution and the head of the low-score impurity population distribution. The score corresponding to this consecutive valley point is determined as the adaptive separation threshold for the current sample.

[0117] Step S330: Using the separation threshold and combined with connected component analysis, single-cell target events are separated.

[0118] In practice, to prevent isolated impurity points at the spatial edge from having rescores that happen to exceed the separation threshold due to local signal anomalies (i.e., forming false positive spatial "islands"), the system performs spatial connectivity analysis after initial screening using the separation threshold. Only events initially screened that maintain physical connectivity with the main single-cell community in two-dimensional spatial topology are determined as the final single-cell target events. This strategy, which combines numerical thresholding with spatial connectivity analysis, significantly filters out free false positive noise, effectively improving the purity and reliability of the final preprocessed single-cell dataset.

[0119] Example 6 Based on the aforementioned embodiments, in the mass spectrometry flow cytometry detection process, to standardize the data and calibrate instrument fluctuations, calibration microspheres containing specific metal isotope labels (e.g., microspheres containing Ce, Eu, Ho, etc.) are typically added to the original sample. If the signals from these microspheres are mixed into subsequent single-cell clustering analysis, they will severely interfere with the topology of the feature space. Therefore, before extracting signal features from at least two physical channel dimensions of each event in the original data to construct the feature space, the system performs an independent microsphere removal step. This step, based on the distribution of the calibration channel signals, uses an adaptive percentile threshold for microsphere detection, specifically including the following: In step S100, before extracting signal features from at least two physical channel dimensions of each event in the original data to construct the feature space, step S400, the microsphere removal step, is also included: Step S410: For the cerium-140 microsphere calibration channel, calculate the outlier shape metric based on the signal distribution of the cerium-140 microsphere calibration channel; Step S420: When the outlier shape measurement reaches the preset condition, the high percentile of the signal of the cerium-140 microsphere calibration channel is used as the adaptive threshold, and events with signals higher than the adaptive threshold are selected as microsphere candidate events. Step S430: When the outlier shape measurement does not meet the preset conditions, it is determined that there is no significant microsphere population in the original data and the microspheres are not removed. The candidate microsphere events are confirmed using DNA content signals containing Ir191 and Ir193. When the DNA content signal of a candidate microsphere event is lower than a preset low percentile threshold, the candidate microsphere event is determined to be a microsphere and deleted.

[0120] An outlier shape metric is calculated based on the signal distribution of the cerium-140 microsphere calibration channel to characterize the prominence of the high-value tails of the signal relative to the main distribution. For example, the difference between the 99th percentile and the median of the channel signal is taken and then divided by its interquartile range (the difference between the 75th percentile and the 25th percentile): sep = (P99 - median) / (P75 - P25). When sep is greater than or equal to a preset threshold (e.g., 3), it is determined that a significant microsphere population exists in the original data and the process proceeds to the cutting decision; otherwise, it is determined that no significant microsphere population exists and no events are removed. When proceeding to the cutting decision, the system adaptively selects the cutting percentile according to the degree of outlier: when sep is greater than a preset strong separation threshold (e.g., 5), the 90th percentile is used; otherwise, the 95th percentile is used. The Ce140 signal value corresponding to this percentile is used as the adaptive threshold, and events with Ce140 signals higher than this threshold are selected as microsphere candidate events.

[0121] The aforementioned outlier shape metric is used to determine whether there is a clear distinguishing structure between the microsphere signal peaks and the background in the original data: when the outlier degree is strong (large sep), it indicates the existence of isolated high-signal microsphere peaks, and the 90th percentile can be used to fully include the microspheres; when the outlier degree is relatively weak, a more stringent 95th percentile is used to avoid mis-inclusion. Since the cut-off threshold is taken from the percentile of the signal distribution of the current sample itself, rather than a fixed absolute signal value, this screening can adapt to the differences in signal intensity of different samples, batches, and instrument conditions.

[0122] The mechanism for confirming microsphere candidate events using DNA content signals containing Ir191 and Ir193 is as follows: In actual detection, some real cells may have high microsphere channel signals due to non-specific adsorption of calibration isotopes or spatial adhesion to the microspheres, thus being included in the microsphere candidate list during the initial screening. Since the calibration microspheres themselves do not contain nucleic acids (their DNA content signal is extremely low), while real cells have significant DNA content signals, the system takes the low percentile (e.g., the 10th percentile) of the DNA content (the sum of Ir191 and Ir193) signal of all events as the DNA reference threshold; only candidate events with DNA content signals below this reference threshold are confirmed as microspheres and deleted, while other candidate events with significant DNA content signals are retained. Since the outlier metric and cutoff threshold are adaptively determined based on the current data distribution, this mechanism can automatically adapt to signal differences between different batches and instrument conditions, effectively avoiding the false deletion of real cells.

[0123] During data acquisition in mass cytometry, instrument artifacts of non-biological origin are often mixed into the raw data due to minor blockages in the sample inlet tubing, gas flow fluctuations, or transient instability in the detector voltage. If these artifact events are included in subsequent cluster analysis, they will generate spurious data subclusters. Conventional cleaning methods often use simple single-channel threshold truncation, which can easily lead to the false rejection of normal events.

[0124] To address this issue, this embodiment constructs a three-layer progressive artifact removal process, which effectively reduces the probability of false positives and accidental deletions while ensuring artifact removal efficiency. The specific implementation process is as follows: The fully automated preprocessing method for mass cytometry data also includes a three-layer progressive artifact removal step: The first layer: extreme value detection based on event length signal, and robust Z-score calculation based on the ratio of DNA content signals containing Ir191 and Ir193 in the transform domain to detect ratio anomalies, thereby screening anomalous events.

[0125] This layer adopts a single-point statistical perspective, using two types of signals for judgment: First, the event length signal, reflecting the time span of the ion cloud passing through the detector. The system uses the low and high percentiles of its distribution as boundaries, classifying events that are too short or too long as suspected anomalies. Second, the signals from the Ir191 and Ir193 DNA isotope channels. Their abundance ratio is constant in nature, constituting a physical invariant. The system performs an inverse hyperbolic sine transform on these two signals and takes the difference (i.e., the logarithm of the ratio). This is calculated only for events where both channel signals are above the noise floor, and a robust Z-score (i.e., a standard score calibrated by the difference between the median and the absolute median) is applied to this ratio. When the absolute value exceeds a preset threshold, it is judged as a ratio anomaly. The union of these two types of anomalous events is labeled as the first type of candidate artifact.

[0126] The second layer: uses a change point detection algorithm to identify abrupt segments in a time-series signal sequence that deviate from the standard deviation.

[0127] Artifacts caused by instrument flow path instability often exhibit temporal clustering, manifesting as signal drift over a continuous period. The system sorts events according to acquisition time sequence and segments them by a fixed number of events. For each segment, the median of the transformed event length signal and DNA content signal (the sum of Ir191 and Ir193) is calculated. A change point detection algorithm (such as the PELT change point analysis algorithm) is applied to the median sequence of each segment to identify abnormal segments that deviate from the overall distribution by a preset multiple of the absolute median difference. Continuous events falling into these abnormal segments are marked as second-class candidate artifacts.

[0128] The third layer: Multi-channel consensus confirmation is performed only on the union of candidate events output from the first and second layers to determine whether to perform final removal. The system calculates robust Z-scores for multiple available quality control channel signals (event length, DNA content signal, viability marker, and microsphere Ce140 channel), and counts the number of channels whose absolute Z-score exceeds a preset threshold for each event. When the number of channels exceeding the threshold for a labeled event reaches the preset number, it is marked as a high-confidence artifact confirmed by multi-channel analysis and finally removed; otherwise, it is retained as a normal or special biological signal.

[0129] This multi-layer design achieves cross-validation from single-point statistics and temporal distribution to multi-channel physical logic through the synergistic effect of "first and second layer high-sensitivity candidate delineation" and "third layer multi-channel confidence verification". This ensures the artifact removal rate while improving the reliability of the judgment and the robustness of data cleaning.

[0130] In some embodiments, after separating single-cell target events from the raw data, to ensure the data purity of downstream high-dimensional analysis and provide traceable statistical classification for all detected events, this fully automated data preprocessing method, after outputting the preprocessed single-cell dataset, also includes steps for processing viability signals and assigning global event labels. The specific implementation is as follows: After outputting the preprocessed single-cell dataset in step S300, the following steps are also included: Step S500: The extracted signal features also include the viability label channel signal; the distribution pattern of the viability label channel signal is determined by peak detection through kernel density estimation. When it is bimodal, the intersection of the two components in the binary Gaussian mixture model is used as an adaptive valley threshold to remove dead cells; when it is unimodal, percentile, inflection point and peak value plus multiple absolute median difference are tried in sequence as backoff threshold strategies.

[0131] In mass cytometry, the damaged cell membrane of dead cells allows viability-labeling dyes (such as cisplatin-labeled Pt195 channels) to enter the cell interior in large quantities, resulting in significantly higher signal intensity in these channels compared to live cells. However, the absolute signal intensity of live and dead cells fluctuates due to factors such as sample batch, staining conditions, and instrument background, making it easy to misjudge using a fixed threshold. Therefore, this embodiment employs an adaptive gating strategy. The system extracts the viability-labeled channel signal and smooths it using kernel density estimation (KDE) to generate a one-dimensional signal density distribution curve. Subsequently, a peak detection algorithm is used to determine the distribution pattern of this curve.

[0132] If the peak detection results show a bimodal distribution, it indicates the simultaneous presence of significant live and dead cell populations in the current sample. The system fits a two-component Gaussian mixture model onto this distribution, estimating the normal parameters for both the low-signal (live cells) and high-signal (dead cells) populations, and extracting the intersection point of these two Gaussian component curves. This intersection point statistically represents the boundary with the minimum classification error rate. The system uses this as an adaptive valley threshold, classifying events above this threshold as dead cells and removing them.

[0133] If the peak detection results show a single-peak distribution, it indicates that the current sample activity is extremely high, and the number of dead cells is insufficient to form an independent statistical peak. To prevent forced fitting of a mixture model from causing live cells to be incorrectly segmented, the system initiates a backoff threshold strategy. Under this strategy, the system tries different single-peak truncation algorithms in order of priority: first, it tries a percentile threshold based on a preset empirical proportion; second, it tries to find an inflection point threshold where the curvature of the main peak tail changes abruptly; finally, it tries to use a safety boundary threshold based on the main peak value plus multiple times the absolute median difference (MAD, a robust statistical measure of dispersion). Through this multiple backoff mechanism, the algorithm is ensured to run stably under extremely pure samples.

[0134] Step S600: Based on a tree-like hierarchical classification structure containing four preset levels, assign a unique final classification label to each event in the original data.

[0135] Furthermore, the final classification label includes at least one of single cell, duodenome, fragment, microsphere, artifact, and dead cell.

[0136] Since the preprocessing steps involve multiple modules such as artifact removal, microsphere verification, single-cell scoring, and dead cell identification, directly assigning labels to events may lead to logical conflicts (for example, an event may simultaneously possess the attributes of dead cells and microspheres at the underlying feature level). Therefore, the system performs global label assignment based on a tree-like hierarchical classification structure with four preset levels. This tree structure sets a strict logical order of judgment according to the priority of biology and physics (for example, instrument artifacts and calibration microspheres, as non-biological signals, are placed at the top level for priority removal, followed by fragment removal at the biological dimension, then classification of single cells and duplexes at the morphological dimension, and finally viability assessment under the single-cell branch).

[0137] Specifically, the four-level tree-like hierarchical classification structure includes: the first level, which distinguishes between non-biological signals and biological signals, and divides events into microsphere / artifact sets and cellular event sets; the second level, which performs morphological division on the cellular event sets, distinguishing between single-cell events and non-single-cell events; and the third level, which includes two parallel sub-levels: the third level A, which further divides the single-cell events into live cells and dead cells based on vitality channel signals; and the third level B, which classifies the non-single-cell events into duplexes and cell fragments by jointly determining them based on directional projection from the single-cell center.

[0138] Based on this tightly structured tree-like path, each event in the original data is guided to a unique leaf node, thus being assigned a unique final classification label. This classification result not only outputs a clean target dataset but also provides a detailed, conflict-free global event archive for the overall quality assessment of the detected samples.

[0139] The following specific experimental examples further illustrate this application. However, it should be understood that these experimental examples are merely for the purpose of further explanation and should not be construed as limiting this application in any way.

[0140] Experiment Example 1: Full workflow and intermediate result verification of fully automated, unsupervised mass spectrometry flow cytometry data preprocessing This experimental example aims to demonstrate the working mechanism of each core module of the system in this application through a complete data processing flow. (Refer to...) Figure 6 (a) shows the complete data flow from raw FCS data through preprocessing, microsphere removal, artifact removal, density modeling, HPF single-cell scoring, dead cell removal, and final classification. Figure 6 Figure (b) shows the complete mathematical architecture of the zeroth to second layers of HPF hierarchical probabilistic fusion.

[0141] exist Figure 7 Figures (a) to (d) illustrate the preprocessing characteristics of microspheres and artifact removal. The selected microsphere candidate events were validated using physical constants independent of the statistical model by calculating the scatter plot and histogram of the Ir191 / Ir193 isotope ratio.

[0142] Reference Figure 8 Figures (a) to (g) show the heatmap distribution and median statistics of the six heterogeneous original decision signals of the zero layer of the HPF after calibration and mapping to the [0,1] probability interval, confirming that the heterogeneous signals successfully eliminated the dimensional differences.

[0143] Reference Figure 9 Images (a) to (d) show the distribution of the first-layer feature (membership, location, orientation security) heatmaps generated after semantic grouping and fusion.

[0144] Reference Figure 10 Figures (a) to (d) show the results of score-weighted density-gated (SWD) spatial rescoring using Nadaraya-Watson kernel regression. The smoothed scores successfully searched for the trough point where the slope changes from negative to positive at the high-score end of the distribution, and this adaptive threshold accurately distinguished between the single-cell core population and the impurity population.

[0145] Reference Figure 11 (a) and (b) illustrate the process of fitting the bimodal viability marker signal with a GMM during the dead cell removal step, and using the intersection of the two components as an adaptive threshold.

[0146] Reference Figure 12 In (a) to (i), during the validation of human peripheral PBMC cell samples, based on the tree-like hierarchical classification structure (GatingTree), the system assigns a unique final classification label to each event, realizing progressive gating classification from the original confounding events to the gradual stripping away of various biological / physical impurities.

[0147] Experiment Example 2: Robustness Verification of Fully Automated Single-Cell Isolation on Multimorphological Samples This experimental example aims to examine the single-cell separation effect of the fully automated preprocessing method provided in this application on mass spectrometry flow cytometry samples with different density distributions and varying difficulty levels.

[0148] Experimental methods: Human peripheral blood mononuclear cell (PBMC) samples covering three morphological difficulty levels were selected for validation. The sample classification criteria were based on the topological relationship between single cell and dimer outlines, categorized into three types: tangential (S-type), intersecting (I-type), and fused (M-type). The detection equipment employed was a Zhejiang Pukang Medical Mass Spectrometry Flow Cytometry (MSFLO) system. Samples were input into the fully automated preprocessing system described in this application, enabling end-to-end processing without manual intervention.

[0149] Experimental Results and Analysis: refer to Figure 13 (Fully automated single-cell isolation results) All samples, including those of S, I, and M types, were successfully processed using the fully automated system. The overall single-cell yield ranged from 36.5% to 73.4%, consistent with results obtained through manual gating.

[0150] Analysis shows that even in the most challenging fusion-type (M-type) samples, where the single-cell main cluster and the biplex cloud are almost completely merged, the algorithm in this application does not crash, robustly characterizes the nonlinear boundary, and accurately outputs high-quality scatter plots of single-cell candidate events. This verifies the adaptability and high robustness of the proposed process on samples with different levels of challenge.

[0151] Experiment Example 3: Direction-Aware Power Attenuation Resampling (DAPD) Strategy Ablation Experiment This experimental example aims to examine the actual improvement effect of the Direction-Aware Power Attenuation Resampling (DAPD) strategy introduced in the density clustering stage of this application on solving the "doublet erroneous merging" problem compared with traditional density clustering algorithms.

[0152] Experimental methods: Six representative highly contaminated cell samples (containing different degrees of separation and contamination) were selected, and the results of FlowJo manual gating were used as the reference standard.

[0153] Experimental example group: Single cell identification was performed using the complete two-stage strategy of this application, which includes separation detection and DAPD resampling.

[0154] Comparative group: Single cell identification was performed using the naive HDBSCAN clustering algorithm without resampling.

[0155] Experimental Results and Analysis: Six highly contaminated cell samples with varying degrees of separation and contamination were selected to compare the complete two-stage strategy with the traditional HDBSCAN algorithm. Experimental data are shown in Table 2. Table 2. Comparison of DAPD ablation experimental performance

[0156] Ablation experiments showed that: (1) The complete two-stage strategy of this application did not lead to performance degradation (ΔF1≥0) in all 6 samples. (2) The average F1 score improved by +0.050, with the maximum improvement reaching +0.136; (3) Through comparative data analysis, it was found that the typical failure modes of traditional HDBSCAN are: incorrectly merging dimer clusters with single-cell main clusters, resulting in a significant decrease in precision (e.g., sample B, down to 0.740), or misclassifying real single cells as non-single cells when dimer clouds are incorrectly split, resulting in a decrease in recall. The DAPD strategy of this application effectively overcomes the inherent defects of existing isotropic clustering methods.

[0157] Experiment Example 4 (including comparative examples): Fully automated processing of large-scale, extremely contaminated samples and comparative evaluation of mainstream algorithms This experimental example aims to examine the comprehensive performance of the preprocessing method (especially the HPF architecture) in the face of large sample sizes and extremely high pollution and poor data distribution, and to verify its advantages over existing mainstream automated gating algorithms.

[0158] Experimental methods: Forty-nine samples were selected for mass spectrometry flow cytometry analysis (33 samples from POWCLIN instrument and 16 samples from Fluidigm instrument). These included extremely high-contamination samples with a contamination index reaching a maximum of 1.0 and a shoulder-to-peak ratio as high as 26.79 (the dimer cloud of this sample was highly fused with the single-cell main cluster, approaching the challenge level of enzymatically digested tissue suspensions of solid tumors).

[0159] A comprehensive performance evaluation was conducted using the results of manual gate control as a reference standard.

[0160] Experimental case group (M1): The HPF hierarchical probabilistic fusion fully automated preprocessing method of this application was adopted.

[0161] Comparative groups (M2-M6): The existing Cleanet algorithm, cytofQC algorithm, flowClust multivariate mixture model, 2DGMM algorithm and conventional percentile cut-off method were used respectively.

[0162] Experimental Results and Analysis: The results of a comprehensive comparative evaluation of 49 samples are shown in the table below: Table 3. Evaluation Results of Mainstream Algorithms

[0163] Experimental analysis shows that: (1) In the treatment of extremely high contamination samples (shoulder peak ratio 26.79), the HPF architecture of this application did not collapse, the F1 score still reached 0.929, and the single cell yield remained at 55.7%, which confirmed the extremely high data robustness.

[0164] (2) In the comparison of all 49 samples, this application (M1) is the only method with an F1 score ≥ 0.90 in all samples (accounting for 100%), and its performance indicators (F1 score, precision, recall, and yield bias) are significantly better than all other comparative algorithms (Wilcoxon test p < 6.2 × 10⁻⁶). -12 ).

[0165] (3) The system exhibits high stability across instruments, with an F1 difference of only 0.002 and a Lin consistency correlation coefficient (CCC) as high as 0.991, which fully demonstrates that the method has high generalization ability and industrial-grade stability.

[0166] refer to Figure 14 This application also provides a fully automated preprocessing device for mass spectrometry flow cytometry data, comprising: Extraction module 10 is used to acquire raw data from mass spectrometry flow cytometry and extract signal features of at least two physical channel dimensions of each event in the raw data to construct a feature space. Calculation module 20 is used to calculate the single-cell probability score of each event in the feature space through a hierarchical probability fusion model, including: First-level calibration step: Based on the signal features of each event in the feature space, calculate the original judgment signal of each event under multiple different evaluation dimensions, and calibrate and map the original judgment signal to a unified probability interval to obtain multiple independent probability signals corresponding to each event; The second-level semantic step is to group and fuse the multiple independent probability signals corresponding to each event according to preset semantic attributes to generate multiple intermediate features for characterizing the state of the event. The third-level decision-making step is to determine adaptive weights based on the distribution quality index of events in the feature space, and to use the adaptive weights to perform nonlinear combination mapping on the multiple intermediate features to output the continuous single-cell probability score of each event. Output module 30 is used to separate single-cell target events from the original data based on the continuous single-cell probability scores of each event, and output the preprocessed single-cell dataset.

[0167] In specific industrial applications, the above modules can be integrated and deployed on a local workstation or a cloud server. The extraction module is configured with a standard data interface for communicating with the flow cytometer hardware (e.g., a data parsing engine for parsing standard FCS format files); the output module can further link with a cloud database, synchronizing the preprocessed single-cell dataset and event-level continuous probability scores to the cloud via a REST API to support downstream weighted differential abundance statistics and uncertainty propagation analysis.

[0168] It is understood that the apparatus in this embodiment corresponds to the fully automated preprocessing method for mass spectrometry flow cytometry data in the above embodiments, and the options in the above embodiments are also applicable to this embodiment, so they will not be described again here.

[0169] This application also provides a computer device, which includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the fully automated preprocessing method for mass spectrometry flow cytometry data as described in any of the foregoing embodiments.

[0170] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0171] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0172] This application also provides a computer storage medium storing a computer program, which, when executed on a processor, implements the fully automated preprocessing method for mass spectrometry flow cytometry data according to any one of the foregoing embodiments.

[0173] The computer storage medium can be a readable storage medium, a non-volatile storage medium, or a volatile storage medium. For example, the computer storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0174] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0175] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0176] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0177] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A fully automated preprocessing method for mass spectrometry flow cytometry data, characterized in that, include: Obtain raw mass cytometry data and extract signal features from at least two physical channel dimensions of each event in the raw data to construct a feature space; In the feature space, the single-cell probability score of each event is calculated using a hierarchical probability fusion model, including: First-level calibration step: Based on the signal features of each event in the feature space, calculate the original judgment signal of each event under multiple different evaluation dimensions, and calibrate and map the original judgment signal to a unified probability interval to obtain multiple independent probability signals corresponding to each event; The second-level semantic step is to group and fuse the multiple independent probability signals corresponding to each event according to preset semantic attributes to generate multiple intermediate features for characterizing the state of the event. The third-level decision-making step is to determine adaptive weights based on the distribution quality index of events in the feature space, and to use the adaptive weights to perform nonlinear combination mapping on the multiple intermediate features to output the continuous single-cell probability score of each event. Based on the continuous single-cell probability scores of each event, single-cell target events are separated from the original data, and a preprocessed single-cell dataset is output.

2. The fully automated preprocessing method for mass spectrometry flow cytometry data as described in claim 1, characterized in that, The step of extracting signal features from at least two physical channel dimensions of each event in the original data to construct a feature space includes: acquiring the original DNA content signal and event length signal; performing an inverse hyperbolic sine transform on the original DNA content signal and event length signal based on a preset cofactor to maintain linear resolution in the low-signal region and compress the dynamic range in the high-signal region; superimposing a uniformly distributed random jitter of a preset interval on the event length signal after the inverse hyperbolic sine transform to eliminate artifacts generated by discrete integer values ​​in density estimation; and constructing a two-dimensional feature space based on the processed DNA content signal and event length signal; and / or, In the first-level calibration step, the original judgment signals under multiple different evaluation dimensions include: clustering signal, orientation mixing signal, density signal, contour line signal, orientation signal, and quadrant signal; The step of calibrating and mapping the original determination signals to a unified probability interval specifically involves: By anchoring the Sigmoid function, each of the original decision signals is mapped to the [0,1] interval to obtain six independent probability signals; the six independent probability signals are: clustering probability signal, direction mixing probability signal, density probability signal, contour line probability signal, direction probability signal, and quadrant probability signal. In the second-level semantic step, the multiple intermediate features include: membership features, position features, and orientation safety features; the multiple independent probability signals corresponding to each event are grouped and fused according to preset semantic attributes, including: The clustering probability signal is fused with the directional mixing probability signal to generate the membership feature; The density probability signal is fused with the contour probability signal to generate the location feature; The direction probability signal is fused with the quadrant probability signal to generate the direction safety feature.

3. The fully automated preprocessing method for mass spectrometry flow cytometry data as described in claim 2, characterized in that, The specific calculation logic for grouping and fusing the multiple independent probability signals corresponding to each event according to preset semantic attributes includes: The membership feature is calculated by performing a geometric mean fusion of the clustering probability signal and the directional mixed probability signal based on adaptive weighting coefficients; and / or, The location feature is calculated by: calculating the geometric mean of the density probability signal and the contour probability signal to achieve soft logic AND gate decision; and / or, The directional safety feature is calculated as follows: the minimum value between the directional probability signal and the quadrant probability signal is taken, and an exponential modulation is applied to it based on a severity index characterizing the degree of sample contamination, wherein the severity index is adaptively determined based on the shoulder-to-peak ratio and fragment ratio in the dimer and fragment directions; and / or, In the third-level decision-making step, adaptive weights are determined based on the distribution quality index of events in the feature space, and the adaptive weights are used to perform a nonlinear combination mapping on the multiple intermediate features, including: Based on the statistical characteristics of event distribution in the feature space, the contamination degree, core quality, and confidence degree of the original data are calculated as the distribution quality indicators. The membership weight, location weight, and directional safety weight are calculated based on the distribution quality index; where the higher the core quality, the greater the location weight; and the higher the contamination degree, the greater the directional safety weight. After performing Logit transformation on the membership feature, the position feature, and the orientation safety feature respectively, the normalized membership weight, position weight, and orientation safety weight are linearly combined and an adaptive bias is added. The continuous single-cell probability score is then output through Sigmoid function mapping. When the event-by-event standard deviation of the multiple intermediate features exceeds a preset conflict threshold, a conflict detection mechanism is activated, using a smoothing gating function to pull the continuous single-cell probability score of the event toward an intermediate uncertainty value; and / or, The process of acquiring the clustering signal relies on a direction-aware resampling clustering strategy, specifically including: Density clustering is performed in the feature space to identify single-cell principal clusters, and a two-component Gaussian mixture model is fitted within the single-cell principal clusters to obtain the corresponding two components; If the separation between the two components reaches a preset trigger threshold, structural fusion is determined to exist, triggering direction-aware power attenuation resampling, including: estimating the dimer direction vector, calculating the forward projection distance of each event to the single-cell density peak position in the dimer direction; based on the forward projection distance and adaptive attenuation scale parameters, calculating the retention probability according to the power attenuation function, and performing Bernoulli sampling on each event to selectively dilute the event density in the dimer direction; Density clustering is re-run on the resampled data cloud to obtain the clustering signals for each event.

4. The fully automated preprocessing method for mass spectrometry flow cytometry data as described in claim 3, characterized in that, The process of performing density clustering in the feature space to identify single-cell principal clusters includes: Density clustering is performed based on the initial minimum cluster size and the initial minimum number of samples, and the proportion of the main cluster to all events is calculated. If the ratio does not reach the target threshold, then an adaptive parameter relaxation iteration is initiated: within the maximum number of attempts, the minimum cluster size and the minimum number of samples are reduced proportionally each time based on a preset decay coefficient, and density clustering is rerun under the constraint that the parameters are not lower than the preset absolute hard lower limit. The iteration terminates when the proportion reaches the target threshold, or after reaching the maximum number of attempts, the result with the largest proportion of principal clusters among all attempts is selected as the final single-cell principal cluster; and / or, The power attenuation function is: S=1 / [1+(d hi / t) p ]; Where S is the basic attenuation value; d hi τ is the forward projection distance; τ is the adaptive attenuation scale parameter; p is the preset attenuation index.

5. The fully automated preprocessing method for mass spectrometry flow cytometry data as described in claim 4, characterized in that, The method for determining the adaptive decay scale parameter τ is as follows: select a preset proportion of events as the core region centered on the single cell density peak, and take the 90th percentile of the forward projection subset in the core region as the adaptive decay scale parameter τ.

6. The fully automated preprocessing method for mass spectrometry flow cytometry data as described in claim 2, characterized in that, The process of separating single-cell target events from the original data based on the continuous single-cell probability scores of each of the aforementioned events includes: spatially rescoring the continuous single-cell probability scores using a score-weighted density gating method; performing local averaging in two-dimensional space based on spatial consistency using Nadaraya-Watson kernel regression; after removing extremely high-score peak components, searching the first continuous valley point where the slope of the density curve changes from negative to positive on the one-dimensional density distribution curve of the remaining scores, from the high-score end to the low-score end, as a separation threshold; using the separation threshold in conjunction with connected component analysis to separate single-cell target events; and / or, Before extracting signal features from at least two physical channel dimensions of each event in the original data to construct a feature space, a microsphere removal step is included: for the cerium-140 microsphere calibration channel, an outlier shape metric is calculated based on the signal distribution of the cerium-140 microsphere calibration channel; when the outlier shape metric reaches a preset condition, events with signals higher than the adaptive threshold are selected as microsphere candidate events using the high percentile of the signal in the cerium-140 microsphere calibration channel as an adaptive threshold; when the outlier shape metric does not reach the preset condition, it is determined that there is no significant microsphere population in the original data and microspheres are not removed; and the microsphere candidate events are confirmed using DNA content signals containing Ir191 and Ir193; when the DNA content signal of a microsphere candidate event is lower than a preset low percentile threshold, the microsphere candidate event is determined to be a microsphere and deleted; and / or, After outputting the preprocessed single-cell dataset, the process further includes: extracting signal features that also include viability label channel signals; determining the distribution pattern of the viability label channel signals through peak detection using kernel density estimation; when the distribution is bimodal, using the intersection of the two components in the binary Gaussian mixture model as an adaptive valley threshold to remove dead cells; when the distribution is unimodal, sequentially trying percentile, inflection point, and peak value plus multiple times the absolute median difference as backoff threshold strategies; and assigning a unique final classification label to each event in the original data based on a tree-like hierarchical classification structure containing four preset levels.

7. The fully automated preprocessing method for mass spectrometry flow cytometry data as described in claim 6, characterized in that, The fully automated preprocessing method for mass spectrometry flow cytometry data also includes a three-layer progressive artifact removal step: The first layer: extreme value detection is performed based on the event length signal, and a robust Z-score is calculated based on the ratio of DNA content signals containing Ir191 and Ir193 in the transform domain to detect ratio anomalies, thereby screening abnormal events. The second layer: the events are segmented according to the collection time sequence. Based on the statistics of the event length signal and DNA content signal of each segment, the change point detection algorithm is used to identify the mutation segments that deviate from the overall distribution and the events falling into them are marked as anomalies. The third layer: For the events marked in the first and second layers, a robust Z-score is calculated based on multiple quality control channel signals such as event length, DNA content signal, viability, and microsphere channels. The number of channels exceeding a threshold in each event is counted, and those exceeding a preset number are marked as high-confidence artifacts confirmed by multiple channels. The artifact removal step ultimately removes the union of the events marked in the first and second layers; and / or, The final classification label includes at least one of single cell, duodenome, fragment, microsphere, artifact, and dead cell.

8. A fully automated preprocessing device for mass spectrometry flow cytometry data, characterized in that, include: The extraction module is used to acquire raw data from mass spectrometry flow cytometry and extract signal features from at least two physical channel dimensions of each event in the raw data to construct a feature space. The calculation module is used to calculate the single-cell probability score of each event in the feature space through a hierarchical probability fusion model, including: First-level calibration step: Based on the signal features of each event in the feature space, calculate the original judgment signal of each event under multiple different evaluation dimensions, and calibrate and map the original judgment signal to a unified probability interval to obtain multiple independent probability signals corresponding to each event; The second-level semantic step is to group and fuse the multiple independent probability signals corresponding to each event according to preset semantic attributes to generate multiple intermediate features for characterizing the state of the event. The third-level decision-making step is to determine adaptive weights based on the distribution quality index of events in the feature space, and to use the adaptive weights to perform nonlinear combination mapping on the multiple intermediate features to output the continuous single-cell probability score of each event. The output module is used to separate single-cell target events from the original data based on the continuous single-cell probability scores of each event, and output the preprocessed single-cell dataset.

9. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the fully automated preprocessing method for mass spectrometry flow cytometry data according to any one of claims 1-7.

10. A computer storage medium, characterized in that, It stores a computer program that, when executed on a processor, implements a fully automated preprocessing method for mass spectrometry flow cytometry data according to any one of claims 1-7.

Images