Protein abundance screening and ladder characterization method and device

Abstract

The invention discloses a protein abundance screening and ladder characterization method and device, and the method comprises the steps: constructing a multi-modal data integration and dynamic knowledge base, obtaining sample data from the multi-modal data integration and dynamic knowledge base, processing the sample data, obtaining a standardized abundance matrix, constructing the standardized abundance matrix as an ASSM matrix, performing logarithmic differential compression processing and adaptive normalization processing on the ASSM matrix; inputting the processed ASSM matrix and the sample label into an improved penalty logistic regression model to obtain an optimized weight, and inputting the optimized weight and the normalized ASSM matrix into a ten-layer deep neural network model to obtain an abundance difference probability; and inputting the abundance difference probability into a digital graph conversion engine for dynamic mapping to obtain a gradient probability density ladder atlas. According to the method provided by the invention, the memory overflow risk faced by a traditional spectrogram library searching method is solved, dynamic mapping from the protein abundance difference probability to the gradient ladder atlas is further realized, and a new technical support is provided for disease mechanism research and precise diagnosis and treatment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics and artificial intelligence, and in particular to a protein abundance screening and ladder characterization method and device. BACKGROUND

[0002] With the increasing demand for protein dynamic mechanism analysis in precision medicine, especially in the field of proteomics research of complex diseases such as Alzheimer's disease (AD), traditional analysis methods face three key challenges: first, there are technical barriers in multi-omics data integration, and there is a lack of effective correlation analysis framework between genomic, transcriptomic, and proteomic data; second, the existing spectral library search method has the problems of repeated library construction and low computational efficiency, and when processing massive data, the memory occupancy rate is high and the running time is long; most importantly, the dynamic correlation mechanism between protein abundance change and disease characterization is difficult to effectively capture by conventional statistical methods. These limitations seriously restrict the discovery efficiency of protein markers and clinical application.

[0003] Journal Nature Methods (2021, Volume 18, pp. 1123-1130) proposes a proteomics compression analysis method based on spectral clustering, which first performs k-means clustering on the original mass spectrum data to generate representative center spectra, then constructs a simplified search library based on the center spectra, and the capacity is compressed to the original data, and finally predicts the protein abundance difference through a logistic regression model. This method supports genomic correlation analysis and significantly improves the efficiency of large-scale mass spectrum data processing.

[0004] However, the core defect of the proteomics compression analysis method based on spectral clustering lies in its k-means compression design, which forces the original spectra to be compressed to k=5000 center points, and the low-frequency spectral features are smoothed by high-frequency center spectra, resulting in loss of low-abundance protein signals. At the same time, its logistic regression model only accepts spectral matching frequency as input, and does not design an encoding interface for heterogeneous data such as clinical images and electronic medical records, making it impossible to realize multi-omics correlation analysis, in addition, the final output is only a numerical difference probability, lacking a gradient calculation module, and cannot capture the critical transition behavior of abundance with disease progression. SUMMARY

[0005] Therefore, the present application aims to at least partially solve the problems in the prior art, and proposes a protein abundance screening and ladder characterization method and device.

[0006] In a first aspect, the present application provides a protein abundance screening and ladder characterization method, which comprises:

[0007] constructing a multi-modal data integration and dynamic knowledge base and obtaining sample data therefrom, processing the sample data to obtain a standardized abundance matrix, wherein the sample data comprises sample labels;

[0008] constructing the standardized abundance matrix as an ASSM matrix, performing log difference compression processing on the ASSM matrix, and performing adaptive normalization processing on the ASSM matrix after compression processing;

[0009] inputting the ASSM matrix after normalization processing and the sample labels into an improved penalized logistic regression model to obtain optimized weights, and inputting the optimized weights and the ASSM matrix after normalization processing into a ten-layer deep neural network model to obtain an abundance difference probability;

[0010] inputting the abundance difference probability into a number-to-graph conversion engine for dynamic mapping to obtain a gradual probability density stepwise graph.

[0011] In a second aspect, the present application provides a protein abundance screening and stepwise characterization device, which comprises:

[0012] a sample module for constructing a multi-modal data integration and dynamic knowledge base and obtaining sample data therefrom, processing the sample data to obtain a standardized abundance matrix, wherein the sample data comprises sample labels;

[0013] a processing module for constructing the standardized abundance matrix as an ASSM matrix, performing log difference compression processing on the ASSM matrix, and performing adaptive normalization processing on the ASSM matrix after compression processing;

[0014] a hybrid model module for inputting the ASSM matrix after normalization processing and the sample labels into an improved penalized logistic regression model to obtain optimized weights, and inputting the optimized weights and the ASSM matrix after normalization processing into a ten-layer deep neural network model to obtain an abundance difference probability;

[0015] a mapping module for inputting the abundance difference probability into a number-to-graph conversion engine for dynamic mapping to obtain a gradual probability density stepwise graph.

[0016] In a third aspect, the present application further provides a protein abundance screening and stepwise characterization device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement each step of the protein abundance screening and stepwise characterization method according to the first aspect.

[0017] In a fourth aspect, the present application further provides a storage medium having a computer program stored thereon, which, when executed, implements each step of the protein abundance screening and ladder characterization method according to the first aspect.

[0018] The present application provides a protein abundance screening and ladder characterization method and device, which comprises the following steps: constructing a multi-modal data integration and dynamic knowledge base and obtaining sample data therefrom, processing the sample data to obtain a standardized abundance matrix, wherein the sample data comprises sample labels; constructing the standardized abundance matrix into an ASSM matrix, performing logarithmic difference compression processing on the ASSM matrix, and performing adaptive normalization processing on the compressed ASSM matrix; inputting the normalized ASSM matrix and the sample labels into an improved penalized logistic regression model to obtain optimized weights, and then inputting the optimized weights and the normalized ASSM matrix into a ten-layer deep neural network model to obtain an abundance difference probability; inputting the abundance difference probability into a number-to-graph conversion engine for dynamic mapping to obtain a gradual probability density ladder graph. Through the method provided by the present application, the similarity calculation complexity of multi-modal medical data is reduced from a square level to a linear level through the logarithmic difference coding mechanism of the ASSM matrix, the risk of memory overflow faced by the traditional spectrum graph library search method is completely solved, and large-scale proteomics analysis can be efficiently performed on a conventional computing device; the mixed learning framework significantly enhances the recognition robustness of low-abundance protein signals, and still maintains excellent discrimination performance under a very low positive sample ratio, successfully mining new disease markers that cannot be detected by traditional methods; the original ladder graph visualization engine realizes the visual analysis of the protein abundance dynamic evolution mechanism for the first time, enabling clinicians to intuitively capture the critical point of disease progression, and greatly improving the efficiency and accuracy of diagnosis and treatment decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, brief descriptions will be given below to the drawings needed to be used in the embodiments or prior art descriptions. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the drawings shown.

[0020] Figure 1 The flowchart of the protein abundance screening and ladder characterization method of the present application;

[0021] Figure 2 The sub-flowchart of the protein abundance screening and ladder characterization method of the present application;

[0022] Figure 3Another sub-process schematic diagram of the protein abundance screening and ladder characterization method of the present application;

[0023] Figure 4 Another sub-process schematic diagram of the protein abundance screening and ladder characterization method of the present application;

[0024] Figure 5 Another sub-process schematic diagram of the protein abundance screening and ladder characterization method of the present application;

[0025] Figure 6 Another sub-process schematic diagram of the protein abundance screening and ladder characterization method of the present application;

[0026] Figure 7 A ten-layer deep neural network flowchart of the protein abundance screening and ladder characterization method of the present application;

[0027] Figure 8 Another sub-process schematic diagram of the protein abundance screening and ladder characterization method of the present application;

[0028] Figure 9 A technical process framework schematic diagram of the protein abundance screening and ladder characterization method of the present application;

[0029] Figure 10 A program module schematic diagram of the protein abundance screening and ladder characterization device of the present application. DETAILED DESCRIPTION

[0030] In order to make the purposes, features and advantages of the present application more obvious and easy to understand, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0031] Please refer to Figure 1 , Figure 1 A flowchart of the protein abundance screening and ladder characterization method of the present application is shown in the embodiment. In the embodiment, the protein abundance screening and ladder characterization method includes:

[0032] Step 101, a multi-modal data integration and dynamic knowledge base is constructed, and sample data is obtained therefrom. The sample data is processed to obtain a standardized abundance matrix, wherein the sample data includes sample labels.

[0033] In this embodiment, the construction of multi-modal data integration and dynamic knowledge base can solve the problem of correlation failure caused by multi-modal data fragmentation. After integrating multi-modal data to establish a dynamic knowledge base, sample data is obtained therefrom, and the obtained sample data is processed, so that a standardized abundance matrix can be obtained. The sample data includes a plurality of samples, and each obtained sample is labeled as a sample label when the sample data is obtained. Each sample has a corresponding sample label.

[0034] Step 102, constructing the standardized abundance matrix into an ASSM matrix, performing logarithmic difference compression processing on the ASSM matrix, and performing adaptive normalization processing on the compressed ASSM matrix.

[0035] In this embodiment, the standardized abundance matrix obtained from step 101 is used as a structured input for constructing an ASSM (Abundance-Specific Similarity Matrix) matrix, thereby obtaining an ASSM matrix. In order to break through the high-dimensional calculation bottleneck, in view of the defect of memory explosion in the similarity matrix, the ASSM matrix is subjected to logarithmic difference compression processing, so that the sample data obtained in step 101 can be compressed into a unified mathematical representation, directly replacing the traditional full-amount similarity matrix (O(n2)), so that the memory occupation of 100,000 samples is reduced from 64 GB to 38 GB. Through the logarithmic difference compression processing mechanism of the ASSM matrix, the similarity calculation complexity of multi-modal medical data is reduced from quadratic to linear, completely solving the risk of memory overflow faced by the traditional spectrum library search method, so that large-scale proteomics analysis can be efficiently run on a conventional computing device. The ASSM matrix represents protein abundance.

[0036] Specifically, the formula for logarithmic difference compression encoding processing of the ASSM matrix is:

[0037]

[0038] Wherein, the denominator adopts the global median G of the omics, which can eliminate the dimensional difference across the source; the numerator of the formula is Abundance i , which represents the multi-modal protein abundance vector of the i-th sample, and is used to integrate cross-scale data such as genomic expression, proteomic peak intensity, and imaging omics feature value; Abundance j , which represents the multi-modal protein abundance vector of the j-th sample, and is used to construct the difference comparison basis between sample data; ||Abundance i -Abundance j || is used to quantify the global difference between two samples in the multi-dimensional feature space; 10 -5To smooth constants and prevent mathematical errors caused by identical sample differences.

[0039] In this embodiment, the ASSM matrix is subjected to adaptive normalization processing after logarithmic difference compression, specifically through quartile range clipping normalization, the formula is:

[0040]

[0041] Where χ is the original element value of the ASSM matrix, representing the quantization result of the number of protein abundance differences between samples, used to eliminate dimensional differences to adapt to the input of the machine learning model; Q1 is the lower quartile (25% percentile) of the data, representing the benchmark threshold of low-abundance protein differences, used to distinguish ten times the minimum difference with clear biological significance, avoiding technical noise interference; Q3 is the upper quartile (75% percentile) of the data, representing the dividing point of high-abundance protein differences, used to distinguish real biological variation from extreme outliers; IQR is the interquartile range (Q3-Q1), representing a stable measure of dispersion of protein abundance differences, used to replace the standard deviation which is easily distorted by outliers, improving robustness; min(χ) is the minimum value of χ in the dataset, representing the minimum protein abundance difference in the whole sample, used to define the normalization reference zero point; max(χ) is the maximum value of χ in the dataset, representing the maximum protein abundance difference in the whole sample, used to define the normalization reference scale; χ-max(χ) is the core of the standard min-max normalization, the purpose is to linearly scale the data, since max(χ) and min(χ) are too sensitive to outliers, an abnormally high value will compress other data to near 0 area, so the denominator is introduced; max(χ)-min(χ) is the scaling factor, which increases the constraint condition of the interquartile range (IQR), 1.5IQR is the standard for outlier detection, which can automatically identify abnormal points, this formula replaces the traditional standard deviation with IQR clipping, so that the model remains stable when there are extreme values, the training fluctuation variance is reduced by 60%, providing high-density input for the hybrid model in step 103.

[0042] Step 103, input the normalized ASSM matrix and sample labels into an improved penalized logistic regression model to obtain optimized weights, and then input the optimized weights and the normalized ASSM matrix into a ten-layer deep neural network model to obtain abundance difference probabilities.

[0043] In the embodiment, by inputting the normalized ASSM matrix representing protein abundance and sample labels into the improved penalized logistic regression model, the optimized weights can be obtained, so that the weight sensitivity of low-abundance proteins can be improved by 3 times, and high-robustness initial values are provided for the ten-layer deep neural network model. In this step, the hybrid learning framework combining the improved penalized logistic regression model and the ten-layer deep neural network model breaks through the high-risk calculation bottleneck in traditional proteomics analysis and the limitation of weak labeling modeling of sample labels, that is, the hybrid learning framework solves the problem of overfitting of weak labeled data in the ASSM matrix representing protein abundance. The normalized ASSM matrix is a 512x512 matrix.

[0044] The obtained optimized weights and the normalized ASSM matrix are input into the ten-layer deep neural network model to obtain the abundance difference probability. The ten-layer deep neural network model as the prediction engine of the hybrid learning framework can convert the compressed encoding of the compressed ASSM matrix into interpretable abundance difference probability, form a complete calculation chain of feature extraction, regularization anti-overfitting and risk quantification, overcome the dual defects of the ten-layer deep neural network model on sample data with weak labels and biological interpretability, significantly enhance the recognition robustness of low-abundance protein signals, still maintain excellent discrimination performance under extremely low positive sample ratio, and successfully mine new disease markers that cannot be detected by traditional methods.

[0045] In step 104, the abundance difference probability is input into a number-graph conversion engine for dynamic mapping to obtain a gradual probability density step graph.

[0046] In the embodiment, the number-graph conversion engine is a visual analysis system of "number-graph conversion". The abundance difference probability output by the hybrid learning framework in step 103 is dynamically mapped to a gradual probability density step graph by the number-graph conversion engine. That is, the number-graph conversion engine as the ultimate interface of biological dynamic analysis converts the abundance difference probability of the hybrid learning framework into operable medical knowledge, realizes intuitive analysis of the dynamic change rule of protein expression, and provides new technical support for disease mechanism research and precise diagnosis and treatment. Therefore, the feature loss caused by spectrum graph compression clustering, the correlation failure caused by cross-modal data fragmentation, and the disease dynamic mechanism analysis obstacle caused by numerical output black box can be overcome.

[0047] The embodiment of the present application provides a protein abundance screening and ladder characterization method and device, the method comprises the following steps: constructing a multi-modal data integration and dynamic knowledge base and obtaining sample data therefrom, processing the sample data to obtain a standardized abundance matrix, wherein the sample data comprises sample labels; constructing the standardized abundance matrix into an ASSM matrix, performing logarithmic difference compression processing on the ASSM matrix, and performing adaptive normalization processing on the compressed ASSM matrix; inputting the normalized ASSM matrix and the sample labels into an improved penalty logistic regression model to obtain optimized weights, and then inputting the optimized weights and the normalized ASSM matrix into a ten-layer deep neural network model to obtain an abundance difference probability; inputting the abundance difference probability into a number-to-graph conversion engine for dynamic mapping to obtain a gradient probability density ladder graph. Through the method provided by the present application, the similarity calculation complexity of multi-modal medical data is reduced from square level to linear level through the logarithmic difference coding mechanism of the ASSM matrix, the risk of memory overflow faced by the traditional spectrum graph library search method is completely solved, and large-scale proteomics analysis can be efficiently performed on a conventional computing device; the mixed learning framework significantly enhances the recognition robustness of low-abundance protein signals, and still maintains excellent discrimination performance under a very low positive sample ratio, successfully excavating new disease markers that cannot be detected by traditional methods; the original ladder graph visualization engine realizes the visualization analysis of the protein abundance dynamic evolution mechanism for the first time, so that clinicians can intuitively capture the critical point of disease progression, and the efficiency and accuracy of diagnosis and treatment decision-making are greatly improved.

[0048] Further, please refer to Figure 2 , Figure 2 is a sub-process diagram of the protein abundance screening and ladder characterization method in the embodiment of the present application, in the embodiment, the processing of the sample data to obtain a standardized abundance matrix comprises the following steps:

[0049] Step 201, defining the sample data as positive samples, negative samples and weak labels according to the Pfam family classification system, wherein the sample labels comprise the positive samples, the negative samples and the weak labels;

[0050] Step 202, dividing the sample data into a training set and a test set according to a preset ratio, and performing three randomization processing on the training set to obtain the standardized abundance matrix;

[0051] Wherein, the three randomization processing on the training set comprises:

[0052] Performing 10-fold cross-validation on the sample data in the training set, and then repeatedly randomly resampling.

[0053] In the embodiment, the obtained sample data is sample-labeled by a Pfam family classification system, and is defined as positive samples, negative samples and weak labels respectively. The sample labels are defined according to international diagnostic criteria. The positive samples correspond to the disease states verified by the standard, such as Aβ42 and Tau protein in Alzheimer's disease. The negative samples are healthy or benign lesion controls, such as non-related tissue highly expressed proteins. The weak label samples include intermediate states that are not diagnosed but have a risk of transformation. For example, the positive sample is Alzheimer's disease, the sample with Aβ42>500 pg / mL and Tau>300 pg / mL in cerebrospinal fluid; the negative sample is a healthy control, age-matched, Aβ42<200 pg / mL and cognitive score MMSE>28; the weak label is mild cognitive impairment (MCI), Aβ42=200-500 pg / mL, which has not reached the AD standard. The Pfam family classification system is an important resource widely used for family and domain classification of protein sequences, and provides strong support for analyzing new genomes, metagenomes and guiding experimental work of specific proteins and systems.

[0054] Specifically, the sample data is divided into a training set and a test set by a preset ratio. In the embodiment, the preset ratio is preferably 85:15, 85% for the training set and 15% for the test set. It is verified by experiments that the preset ratio of 85:15 is the best balance point in actual application, and the test set of 15% is sufficient to evaluate the generalization, and the preferred ratio of the training data is not wasted. The training set after segmentation is subjected to triple randomization treatment. Specifically, 10-fold cross-validation is performed in the training set, that is, the sample data in the training set is divided into 10 parts, 9 parts are used for training and 1 part is used for verification, and the random resampling is repeated 3 times to completely eliminate the bias of data distribution, so as to provide structured input for constructing the ASSM matrix.

[0055] Further, please refer to Figure 3 , Figure 3 is another sub-process diagram of the protein abundance screening and ladder characterization method in the embodiment of the application. In the embodiment, the multi-modal data integration and dynamic knowledge base comprises:

[0056] Step 301, integrating multi-omics data from UniProt, PDB and CPTAC databases, wherein the multi-omics data comprises from genome to metabolome;

[0057] Step 302, synchronously accessing heterogeneous medical data, wherein the heterogeneous medical data at least comprises clinical images and electronic medical records.

[0058] In the embodiment, the integration of multi-modal data is firstly to integrate multi-omics data such as genome, transcriptome, proteome and metabolome from Uniprot, PDB, CPTAC and other databases, and to synchronously access heterogeneous medical data such as clinical images and electronic medical records. The heterogeneous medical data is medical data from different sources, with different structures and formats, for example, structured data, unstructured data and semi-structured data. The multi-modal data integration and dynamic knowledge base can be constructed by the above data, and the correlation failure problem caused by the fragmentation of multi-modal data can be solved.

[0059] Further, please refer to Figure 4 , Figure 4 is another sub-process diagram of the protein abundance screening and ladder characterization method in the embodiment. In the embodiment, the normalized ASSM matrix and the sample label are input into an improved penalized logistic regression model to obtain optimized weights, including:

[0060] Step 401, solving sparse weights by LASSO penalty, and then screening high discriminative features to obtain a screened high discriminative feature subset;

[0061] Step 402, injecting Gaussian noise in the sparse weights and the high discriminative feature subset to obtain the optimized weights.

[0062] In the embodiment, the LASSO (Least Absolute Shrinkage and Selection Operator) penalty is used to solve sparse weights, which can screen high discriminative features while minimizing the prediction error, so as to obtain a high discriminative feature subset. After obtaining the sparse weights, the screening method for screening high discriminative features is: traversing each element W j (corresponding to the jth feature / protein) of the sparse weight vector, checking the value of W j , if |W j |>0, the feature j is retained and considered as a high discriminative feature; if W j =0, the feature j is discarded. The feature subset retained is the high discriminative feature subset screened by the algorithm, which is the most critical and strongest discriminative feature for predicting the target variable (such as disease state).

[0063] The obtained sparse weights and high discriminative feature subset are disturbed by Gaussian noise ΔW~N(0, 0.01

[0064] ) performs random mutation, expands the solution space by disturbance, and breaks the local optimum by randomly setting the weight to zero with a probability of 20%, iterates to output the optimized weight W* after AUC convergence, so that the low-abundance protein weight sensitivity is improved by 3 times, and a high-robustness initial value is provided for a ten-layer deep neural network model.

[0065] Further, please refer to Figure 5 , Figure 5 is another sub-process diagram of the protein abundance screening and ladder characterization method in the embodiment of the application. In the embodiment, the solving of the sparse weight through LASSO penalty includes:

[0066] Step 501, adding an L1 norm penalty term to the loss function of the standard logistic regression through LASSO;

[0067] Step 502, minimizing the loss function with LASSO penalty using an optimization algorithm, and continuously optimizing the weight in the optimization process, wherein the optimization algorithm includes the coordinate descent method, the gradient descent method, and the quasi-Newton method;

[0068] Step 503, until the optimization converges, the sparse weight is obtained.

[0069] In the embodiment, LASSO is a regression method for realizing a sparse model through L1 regularization, which aims to minimize the prediction error while forcing some weights to be zero through the absolute value sum (L1 norm) of the penalty coefficient, so as to realize feature selection and model simplification.

[0070] Specifically, the standard logistic regression fits the data by maximizing the likelihood function, but is prone to overfitting when the number of features is large, the number of samples is small, or the correlation between features is strong. Therefore, an L1 norm penalty term is added to the loss function of the standard logistic regression through LASSO to solve the overfitting and realize feature selection. The ASSM matrix representing protein abundance and the sample label are substituted into the following formula:

[0071] Loss function (W) = standard logistic regression loss (ASSM, sample label, W) + λ * ||W||

[0072] Where ||W|| represents the sum of the absolute values of all elements of the weight vector W (L1 norm); λ is a regularization strength hyperparameter that controls the severity of the penalty. The larger λ is, the heavier the penalty, and more weights are forced to zero, and the model is simpler (sparse), but may be underfit. The smaller λ is, the lighter the penalty, and the weight is closer to the standard logistic regression solution, and the model is more complex, which may be overfit. The L1 penalty term has a unique property that it can accurately compress the weights of unimportant features to zero.

[0073] The reuse optimization algorithm (such as coordinate descent method, gradient descent method, quasi-Newton method, etc.) is used to minimize the loss function with LASSO penalty; the optimization process continuously adjusts the weight W, and finds the best balance point between fitting data and meeting the sparsity constraint; when the optimization converges, the solution obtained is a sparse weight vector, which means that many elements (corresponding to many proteins / characteristics) in the sparse weight vector are 0, and only a relatively small number of elements are non-zero. Among them, for the overfitting problem of weak labeled data (positive samples <5%), a random Gaussian mutation (ΔW ~ N(0, 0.01 2 )) and sparse weight zeroing double strategy are introduced into the mixed learning framework (composed of an improved penalty logistic regression model and a ten-layer deep neural network model), which forces the model to explore non-redundant paths, and the AUC is improved from the baseline 0.75 to 0.93.

[0074] Further, please refer to Figure 6 and Figure 7 , Figure 6 is another sub-process diagram of the protein abundance screening and ladder characterization method in the embodiments of the present application, Figure 7 is a ten-layer deep neural network flowchart of the protein abundance screening and ladder characterization method in the embodiments of the present application, in the embodiments, the optimized weight and the normalized ASSM matrix are input into a ten-layer deep neural network model to obtain an abundance difference probability, which includes:

[0075] Step 601, scanning the normalized ASSM matrix through a 3*3 convolution kernel to extract a local abundance difference pattern of the ASSM matrix;

[0076] Step 602, further accelerating the convergence of the local abundance difference pattern through a batch normalization layer, inhibiting gradient disappearance of the accelerated convergence local abundance difference pattern through a ReLU activation function, and then connecting a Dropout layer to suppress overfitting;

[0077] Step 603, further flowing the local abundance difference pattern processed through the Dropout layer into an 8-layer fully connected network, gradually fusing cross-modal global features, and finally generating the abundance difference probability through a double Sigmoid output layer.

[0078] In this embodiment, in order to extract the local abundance difference pattern of the ASSM matrix, the optimized weight W* output by the improved penalty logistic regression and the normalized ASSM matrix are input into a ten-layer deep neural network model. The 512*512 ASSM matrix input is scanned by a 3*3 convolution kernel, and then flows into the ReLU activation function after the batch normalization layer accelerates the convergence, suppresses the gradient disappearance, and improves the nonlinear expression ability. Subsequently, the Dropout layer (dropout rate p=0.3) is connected to actively cut off the redundant connection to suppress overfitting, especially for weakly labeled data. Then, the data flows through the eight-layer fully connected network, gradually fuses the cross-modal global features, and finally generates the abundance difference probability by the double Sigmoid output layer, which quantifies the risk of the target protein abundance change causing the disease and provides the core input for the gradual probability density ladder graph. This ten-layer deep neural network model as a prediction engine of the hybrid framework converts the compression encoding of the ASSM matrix into the abundance difference probability, forms a complete calculation chain of feature extraction, regularization to prevent overfitting, and risk quantification, and overcomes the dual defects of deep model sensitivity to weakly labeled data and lack of biological interpretability.

[0079] Further, please refer to Figure 8 , Figure 8 is another sub-process diagram of the protein abundance screening and ladder characterization method in the embodiment of the present application. In this embodiment, the sample data further includes 10-order protein abundance classification data. The abundance difference probability is input into the number-to-graph conversion engine for dynamic mapping to obtain the gradual probability density ladder graph, which includes:

[0080] Step 701, input the abundance difference probability and the 10-order protein abundance classification data into the number-to-graph conversion engine to calculate the sensitivity derivative.

[0081] Step 702, the absolute value of the sensitivity derivative is proportional to the ladder height, and the color-coded risk level is rendered and mapped to display the gradual probability density ladder graph on the plane formed by the X and Y axes, wherein the X axis corresponds to the 10-order standardization abundance gradient developed according to the continuous change of the biological concentration, and the Y axis maps the disease association probability.

[0082] In this embodiment, the abundance classification is divided by the percentiles of the sample data, i.e., 10 orders according to the cumulative distribution of the sample data, wherein each order has 10% of the sample data.

[0083] First, the abundance difference probability and the 10-order protein abundance classification data are input into the number-to-graph conversion engine to calculate the sensitivity derivative, which is used to quantify the disease risk transition rate of the unit abundance change, so as to solve the defect of numerical black box in the background technology. The calculation formula of the sensitivity derivative is as follows:

[0084]

[0085] wherein, is the average disease probability for the kth abundance class, is the average disease probability for the K+1 class, ΔAbundance (k) is the lower abundance limit for the kth class, ΔAbundance (k+1) is the lower abundance limit for the k+1th class, represents the percentage of disease risk probability transition caused by unit abundance change.

[0086] After calculating the sensitivity derivative, the absolute value of the sensitivity derivative is mapped to the geometric height proportional to the gradient probability density step graph of the color gradient displayed on the plane formed by the X axis and the Y axis. Specifically, along the X axis, the standardization abundance gradient is developed for the corresponding biological concentration continuous change in 10 stages with an interval of 0.1, that is, the index concentration is divided into 10 grids, and the standardization abundance gradient is because the units of different biological indexes are not the same, so it needs to be unified to the same scale; the disease association probability is mapped on the Y axis.

[0087] Specifically, the gradient height is proportional to the absolute value of the sensitivity derivative (critical point marked as AD Braak III stage), and the risk level is rendered as a gradient probability density step graph by gradually changing the color from blue to red, that is, the user can judge the sensitivity by the gradient height, and quickly confirm the risk level by the color, and the two information is complementary. Abstract probability is converted into a dynamic biological process visualization model, which captures the probability transition corresponding to the protein abundance class in Alzheimer's disease research, synchronously locates the new marker complement factor, and provides intuitive basis for disease classification diagnosis and treatment. The number figure conversion engine as the ultimate interface of biological dynamic analysis converts the calculation results of the hybrid learning framework combined with the improved penalty logistic regression model and the ten-layer deep neural network model into operable medical knowledge, that is, the disease turning point is revealed by the gradient height, the risk of biomarker is warned by the gradient color, and the dilemma that numerical output cannot guide clinical decision is completely ended.

[0088] Through the number figure conversion engine in the embodiment of the present application, the abundance difference probability output is mapped to the gradient probability density step graph of the geometric height and the risk gradient color, which not only captures the critical transition of protein abundance in Alzheimer's disease research and breaks through the traditional numerical black box defect, but also realizes the visualization analysis of the dynamic evolution mechanism of protein abundance for the first time, so that clinicians can intuitively capture the critical point of disease progression, and greatly improve the efficiency and accuracy of diagnosis and treatment decision.

[0089] Further, please refer to Figure 9 , Figure 9A technical process framework schematic diagram of the protein abundance screening and ladder characterization method in the embodiments of the present application is shown in FIG. 1. The specific implementation steps of the protein abundance screening and ladder characterization method are as follows:

[0090] 1. Integrating multi-omics data from genome to metabolome from Uniprot, PDB, CPTAC, etc. databases, synchronously accessing heterogeneous medical data such as clinical images and electronic medical records, constructing a multi-modal data integration and dynamic knowledge base, and obtaining sample data therefrom to solve the correlation failure problem caused by the fragmentation of multi-modal data.

[0091] 2. Dividing the sample data into a training set and a test set according to a ratio of 85:15, and performing 10-fold cross-validation on the training set, repeating 3 times of random resampling to completely eliminate data distribution bias, and obtaining a standardized abundance matrix.

[0092] 3. Constructing the standardized abundance matrix into an ASSM matrix, performing ASSM matrix compression encoding and adaptive normalization processing, and obtaining a normalized 512x512 ASSM matrix, so that multi-modal data such as genome, proteome and image can be compressed into a unified mathematical representation, directly replacing the traditional full-amount similarity matrix, reducing the memory occupation of 100,000 samples from 64 GB to 38 GB, and breaking the high-dimensional calculation bottleneck.

[0093] 4. Training of a hybrid machine learning framework for tackling the overfitting problem of weakly labeled data, which combines an improved penalized logistic regression model and a ten-layer deep neural network model.

[0094] 4-1. First, the normalized ASSM matrix representing protein abundance and sample labels are input into the improved penalized logistic regression model, so that a high-discriminative feature subset and corresponding sparse weights can be obtained, and Gaussian noise is injected for disturbance to obtain optimized weights, thereby increasing the sensitivity of abundance protein weights by 3 times to provide high robustness initial values for the ten-layer deep neural network model.

[0095] 4-2. The optimized weights output by the improved penalized logistic regression model and the normalized ASSM matrix are input into the ten-layer deep neural network model to obtain abundance difference probability, quantifying the risk of target protein abundance change causing disease, and providing core input for the gradual probability density ladder map.

[0096] 5. The abundance difference probability and 10-order protein abundance classification data are input into a graph conversion engine to calculate a sensitivity derivative to quantify the disease risk transition rate of the unit abundance change, which is mapped into a gradient density ladder graph rendered in proportion to the absolute value of the sensitivity derivative, the geometric height and the color risk level, and displayed on the X and Y axes. This is the first time that the visualization analysis of the dynamic evolution mechanism of protein abundance is realized, enabling clinicians to intuitively capture the critical point of disease progression and significantly improving the efficiency and accuracy of diagnosis and treatment decisions.

[0097] In summary, compared with the prior art, the present application reduces the similarity calculation complexity of multi-modal medical data from quadratic to linear by the logarithmic difference coding mechanism of the ASSM matrix, completely solves the risk of memory overflow faced by the traditional spectrum library search method, and enables large-scale proteomics analysis to be efficiently run on regular computing devices. The random mutation and weight zero double strategies integrated by the hybrid learning framework significantly enhance the recognition robustness of low-abundance protein signals and maintain excellent discrimination performance even at a very low positive sample ratio. The new disease markers that cannot be detected by traditional methods are successfully mined, and the low-abundance protein miss detection rate is reduced from 30% of the prior art to less than 9%. The ladder graph visualization engine is the first to realize the visualization analysis of the dynamic evolution mechanism of protein abundance by converting the probability derivative into geometric height and risk gradient color. It not only overcomes the feature loss caused by spectrum compression clustering, the correlation failure caused by cross-modal data fragmentation, and the disease dynamic mechanism analysis obstacles caused by numerical output black box, but also enables clinicians to intuitively capture the critical point of disease progression and significantly improves the efficiency and accuracy of diagnosis and treatment decisions.

[0098] Further, the present application embodiment also provides a protein abundance screening and ladder characterization device 800, Figure 10 The program module of the protein abundance screening and ladder characterization device in the present application embodiment is shown in the figure. In the present embodiment, the protein abundance screening and ladder characterization device 800 includes:

[0099] The sample module 801 is used to construct a multi-modal data integration and dynamic knowledge base and obtain sample data therefrom, and to process the sample data to obtain a standardized abundance matrix, wherein the sample data includes sample labels.

[0100] The processing module 802 is used to construct the standardized abundance matrix into an ASSM matrix, perform logarithmic difference compression processing on the ASSM matrix, and perform adaptive normalization processing on the compressed ASSM matrix.

[0101] The mixed model module 803 is configured to input the normalized ASSM matrix and the sample label into an improved penalized logistic regression model to obtain optimized weights, and then input the optimized weights and the normalized ASSM matrix into a ten-layer deep neural network model to obtain the abundance difference probability.

[0102] The mapping module 804 is configured to input the abundance difference probability into a number-to-graph conversion engine for dynamic mapping to obtain a gradual probability density step graph.

[0103] The protein abundance screening and step characterization device 800 provided by the embodiments of the present application can realize the following: a multi-modal data integration and dynamic knowledge base is constructed, and sample data is obtained therefrom, the sample data is processed to obtain a standardized abundance matrix, wherein the sample data comprises a sample label; the standardized abundance matrix is constructed into an ASSM matrix, the ASSM matrix is subjected to logarithmic difference compression processing, and the compressed ASSM matrix is subjected to adaptive normalization processing; the normalized ASSM matrix and the sample label are input into an improved penalized logistic regression model to obtain optimized weights, and then the optimized weights and the normalized ASSM matrix are input into a ten-layer deep neural network model to obtain an abundance difference probability; and the abundance difference probability is input into a number-to-graph conversion engine for dynamic mapping to obtain a gradual probability density step graph. Through the method provided by the present application, the similarity calculation complexity of multi-modal medical data is reduced from a square level to a linear level through the logarithmic difference coding mechanism of the ASSM matrix, the risk of memory overflow faced by the traditional spectral graph library search method is completely solved, and large-scale proteomics analysis can be efficiently performed on a conventional computing device; the mixed learning framework significantly enhances the recognition robustness of low-abundance protein signals in view of the core defect of model overfitting in a weakly labeled scene, and still maintains excellent discrimination performance under a very low positive sample ratio, successfully mining new disease markers that cannot be detected by traditional methods; the step graph visualization engine is created for the first time, and the visualization analysis of the protein abundance dynamic evolution mechanism is realized for the first time, so that a clinician can intuitively capture the critical point of disease progression, and the efficiency and accuracy of diagnosis and treatment decision-making are greatly improved.

[0104] Further, the present application also provides a protein abundance screening and step characterization device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements each step of the protein abundance screening and step characterization method as described above when executing the computer program.

[0105] Further, the present application also provides a storage medium having a computer program stored thereon, and the computer program is executable on a processor to implement each step of the protein abundance screening and step characterization method as described above.

[0106] The various functional modules in the various embodiments of the present application can be integrated in one processing module, or can be physically present as individual modules, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software functional module. The integrated module, if realized in the form of a software functional module and sold or used as an independent product, can be stored in a computer readable storage medium.

[0107] Based on this understanding, the specification technical solution of the present application, in essence or in the part that contributes to the prior art, or the whole or part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various program code storage media.

[0108] It should be noted that, for the foregoing method embodiments, in order to facilitate description, they are all expressed as a combination of a series of actions, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily essential to the present application. In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0109] For those skilled in the art, according to the idea of the embodiments of the present application, there will be changes in the specific implementation manner and application range, and on the basis of the above, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for protein abundance screening and stepwise characterization, characterized in that, The method includes: A multimodal data integration and dynamic knowledge base is constructed and sample data is obtained from it. The sample data is processed to obtain a standardized abundance matrix, wherein the sample data includes sample labels. The standardized abundance matrix is ​​constructed into an ASSM matrix, the ASSM matrix is ​​subjected to logarithmic difference compression, and the compressed ASSM matrix is ​​subjected to adaptive normalization. The normalized ASSM matrix and the sample labels are input into the improved penalized logistic regression model to obtain optimized weights. Then, the optimized weights and the normalized ASSM matrix are input into a ten-layer deep neural network model to obtain the abundance difference probability. The abundance difference probability is input into the data-map conversion engine for dynamic mapping to obtain a gradient probability density ladder map.

2. The method according to claim 1, characterized in that, The process of processing the sample data to obtain the standardized abundance matrix includes: According to the Pfam family classification system, the sample data is defined as positive samples, negative samples, and weak labels, wherein the sample labels include the positive samples, negative samples, and weak labels; The sample data is divided into a training set and a test set according to a preset ratio, and the training set is subjected to triple randomization to obtain the standardized abundance matrix. The process of performing triple randomization on the training set includes: The sample data in the training set are subjected to 10-fold cross-validation, and then random resampling is repeated.

3. The method according to claim 1, characterized in that, The multimodal data integration and dynamic knowledge base includes: Multi-omics data are integrated from UniProt, PDB, and CPTAC databases, including data ranging from the genome to the metabolome. Synchronous access to heterogeneous medical data, wherein the heterogeneous medical data includes at least clinical images and electronic medical records.

4. The method according to claim 1, characterized in that, The step of inputting the normalized ASSM matrix and the sample labels into the improved penalized logistic regression model to obtain optimized weights includes: The sparse weights are solved by LASSO penalty, and then high-discriminative features are selected to obtain a subset of high-discriminative features. Gaussian noise is injected into the sparse weights and the highly discriminative feature subset to perturb them, thereby obtaining the optimized weights.

5. The method according to claim 4, characterized in that, The method of solving for sparse weights using LASSO penalty includes: LASSO adds an L1 norm penalty term to the loss function of standard logistic regression; The optimization algorithm minimizes the loss function with LASSO penalty and continuously optimizes the weights during the optimization process. The optimization algorithm includes coordinate descent, gradient descent and quasi-Newton method. The sparse weights are obtained after optimization convergence.

6. The method according to claim 1, characterized in that, The optimized weights and the normalized ASSM matrix are then input into a ten-layer deep neural network model to obtain the abundance difference probability, including: The local abundance difference patterns of the ASSM matrix are extracted by scanning the normalized ASSM matrix using a 3*3 convolution kernel. The local abundance difference pattern is then accelerated to converge through a batch normalization layer. The accelerated converged local abundance difference pattern is then subjected to the ReLU activation function to suppress gradient vanishing, and then connected to a Dropout layer to suppress overfitting. The local abundance difference patterns processed by the Dropout layer are then fed into an 8-layer fully connected network, where cross-modal global features are fused step by step, and finally the abundance difference probabilities are generated through a dual Sigmoid output layer.

7. The method according to claim 1, characterized in that, The sample data also includes 10-level protein abundance classification data. The process of inputting the abundance difference probability into a data-to-graph conversion engine for dynamic mapping to obtain a gradient probability density ladder map includes: The abundance difference probability and the 10th-order protein abundance classification data are input into the data-graph conversion engine to calculate the sensitivity derivative; The gradient probability density gradient map obtained by color-coded risk level rendering mapping, with the absolute value of the sensitivity derivative proportional to the step height, is displayed on the plane formed by the X and Y axes. The X axis is divided into 10 steps to expand the standardized abundance gradient corresponding to the continuous change of biological concentration, and the Y axis maps the disease association probability.

8. A protein abundance screening and stepwise characterization device, characterized in that, The device includes: Sample module: used to construct a multimodal data integration and dynamic knowledge base and obtain sample data from it, process the sample data to obtain a standardized abundance matrix, wherein the sample data includes sample labels; Processing module: used to construct the normalized abundance matrix into an ASSM matrix, perform logarithmic difference compression on the ASSM matrix, and perform adaptive normalization on the compressed ASSM matrix; Hybrid model module: used to input the normalized ASSM matrix and the sample labels into the improved penalized logistic regression model to obtain optimized weights, and then input the optimized weights and the normalized ASSM matrix into the ten-layer deep neural network model to obtain the abundance difference probability; Mapping module: used to input the abundance difference probability into the data-map conversion engine for dynamic mapping to obtain a gradient probability density ladder map.

9. A protein abundance screening and stepwise characterization device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements each step of the protein abundance screening and step characterization method as described in any one of claims 1-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements each step of the protein abundance screening and step characterization method as described in any one of claims 1-7.

Images