Cross-omics sparse feature selection system and method based on hierarchical causal modeling

Abstract

The invention provides a cross-omics sparse feature selection system and method based on hierarchical causal modeling, and the system comprises a data input and preprocessing module which is used for receiving multi-omics original data of a multivariate sample; the hierarchical causal structure learning module is connected with the data input and adaptive preprocessing module and is used for constructing a cross-omics hierarchical causal topology; the causal-oriented sparse feature selection module is connected with the hierarchical causal structure learning module; and the model retraining and integration module is used for constructing a three-layer weighted integration discrimination model based on the screened markers, optimizing the fusion weight of each layer through a gradient descent algorithm, and outputting a final prediction result. According to the method, the protein-metabolism biological hierarchy relationship and serum-urine complementary information are fully utilized, and the method has good generalization ability and can be widely applied to marker mining and prediction modeling of cancers, metabolic diseases and the like, so that the accuracy and reliability of precise medical treatment are improved.

Description

Technical Field

[0001] This invention relates to the field of biomedical data processing and intelligent analysis technology, and in particular to a cross-omics sparse feature selection system and method based on hierarchical causal modeling. Background Technology

[0002] With the rapid development of precision medicine, multi-omics data (such as metabolomics and proteomics) from diverse samples (such as serum and urine) are playing an increasingly prominent role in disease diagnosis, efficacy monitoring, and personalized treatment.

[0003] Existing technologies have proposed various methods for omics feature selection and discriminative modeling, such as sparse regression based on LASSO (minimum absolute shrinkage and selection operator), feature importance scoring based on random forest (RF), and feature selection algorithms such as SFFS (forward floating feature selection). These methods aim to screen key features from high-dimensional omics data and build ensemble models for disease prediction and efficacy evaluation.

[0004] However, in multi-omics high-dimensional data scenarios, existing feature selection and modeling methods still generally suffer from the following drawbacks:

[0005] (1) Relevance / significance-driven, making it difficult to distinguish between causal contributions and spurious correlations. Univariate statistical screening, correlation screening, and most screening methods based on model weights / importance usually use the significance of differences, correlation strength, or predictive contribution as screening criteria, making it difficult to distinguish between features with real causal effects on efficacy or disease prediction results and those associated features caused by common driving factors, collinearity, confounding factors, or batch effects. This problem is more prominent when there are complex hierarchical dependencies between urine and serum, as well as between the proteome and metabolome, easily resulting in a set of candidate biomarkers that lack interpretability and reproducibility.

[0006] (2) Selection is unstable and reproducibility is poor under high-dimensional and small sample conditions. Body fluid proteomics and metabolomics data often have a structure in which the number of variables is much larger than the sample size, and measurement noise and missing values ​​are common. Whether it is a regularization method such as LASSO / Elastic Net, or a tree model importance and network screening method, the screening results may be highly sensitive to sample perturbation, training set partitioning, and parameter settings; under different resampling, different centers, or different batches, the selected feature set is prone to significant changes, resulting in insufficient reproducibility and difficulty in forming a stable feature panel for subsequent validation and implementation.

[0007] (3) Insufficient characterization of cross-fluid sources and cross-omics relationships limits generalization ability. Although multi-omics fusion modeling methods (such as PLS-DA, multi-omics factor analysis, stacking ensemble, etc.) can improve prediction performance, they often use splicing or simple ensemble methods to process data from different omics layers and different fluid sources, failing to explicitly characterize the directed dependence and hierarchical structure relationship between urine and serum, and between protein and metabolism. When there are cross-omics transmission effects or "upstream-driven-downstream-response" mechanisms, the lack of structural constraints will cause the model to have insufficient understanding of the relationship between features, thereby affecting the model's generalization performance on independent cohorts or real-world data.

[0008] (4) Lack of feature importance measurement oriented towards intervention semantics, resulting in insufficient interpretability and verifiability. Explanation methods such as the importance of tree models or ensemble learning, linear model weights, permutation importance, or SHAP are mostly still related to or predictive contributions, and their outputs are difficult to answer the question of "how much impact will virtual intervention or changes in the level of a certain molecular feature have on the discrimination results." The lack of importance assessment of intervention semantics will limit the biological interpretation of feature screening results, experimental validation design, and clinical usability justification.

[0009] (5) Batch effects, platform differences, and measurement noise can easily introduce biases. Humoromics data are easily affected by sample collection, processing procedures, mass spectrometry platforms, peak extraction, and quantification strategies, resulting in batch effects and systematic biases. Even with normalization or batch correction, existing technologies may still retain residual biases during the feature selection stage, causing the selected features to reflect more technical noise than real biological differences, which in turn leads to a decline in the performance of the discrimination model when applied across batches and centers.

[0010] (6) Encapsulation / search-based and metaheuristic methods are computationally expensive, prone to overfitting, and highly random. Encapsulation methods or metaheuristic feature selection such as forward / backward / stepwise selection, RFE (recursive feature elimination), PSO (particle swarm optimization), and GE (genetic algorithm) transform feature subset selection into an iterative search problem. However, they face the following shortcomings in high-dimensional omics scenarios: 1) Extremely large search space and high computational cost: The number of high-dimensional feature combinations grows exponentially, and iterative search requires repeated training and validation of the model, which is difficult to complete within reasonable resources and time; 2) Prone to overfitting and selection bias: Multiple rounds of search may "chase validation set metrics" on a limited number of samples, forming a feature set that depends on the training or validation partition; 3) Unstable results and high randomness: Metaheuristic algorithms usually involve random initialization and random mutation or perturbation processes. Different random seeds or parameter settings may lead to significantly different selected feature sets; 4) Still centered on prediction performance, lacking causal structure and intervention explanation: Such methods usually do not characterize the causal dependency structure across omics, making it difficult to ensure that the selected features have an interpretable mechanism path and intervention semantics.

[0011] In summary, current technologies lack a method that can simultaneously achieve feature importance measurement of causal semantics, characterization of cross-omics structural relationships, and stable and reproducible feature selection under multi-omics high-dimensional data conditions. Therefore, there is an urgent need to propose a new multi-omics sparse causal feature selection algorithm and its modeling process to overcome the above limitations and improve the accuracy, interpretability, and clinical applicability of biomarker mining. Summary of the Invention

[0012] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a cross-omics sparse feature selection system and method based on hierarchical causal modeling, which can solve the problems of low accuracy, insufficient interpretability and poor clinical applicability of biomarker mining in the prior art.

[0013] To achieve the above and other related objectives, this invention provides a cross-omics sparse feature selection system based on hierarchical causal modeling, comprising:

[0014] The data input and preprocessing module is used to receive multi-omics raw data of multivariate samples, and has an embedded tunable batch correction unit based on robust location statistics to remove systematic biases across batches and platforms through iterative weighting.

[0015] The hierarchical causal structure learning module, connected to the data input and adaptive preprocessing module, is used to construct a hierarchical causal topology across omics. It explicitly defines proteomics features as upstream potential regulatory variables and metabolomics features as downstream response variables using structural equation modeling, and introduces clinical confounding factors as covariates to characterize the directional dependencies between features.

[0016] The causal-guided sparse feature selection module is connected to the hierarchical causal structure learning module. It embeds a causal-gated sparse regularization mechanism and a counterfactual intervention importance calculation unit. It is used to automatically prune non-causal paths by optimizing the sparse loss function containing differentiable gated variables, and to screen out a set of sparse, stable biomarkers with causal semantics by evaluating the predicted changes under virtual intervention and the Bootstrap stability score.

[0017] The model retraining and ensemble module is used to build a three-layer weighted ensemble discriminant model based on the selected markers, and optimize the fusion weights of each layer through the gradient descent algorithm to output the final prediction results.

[0018] By adopting the above technical solutions, structural equation modeling (SEM) is introduced to explicitly characterize the "upstream-downstream" hierarchical dependency between the proteome and metabolome. Combined with causal gating sparse regularization and counterfactual intervention simulation, causal-oriented feature screening and biomarker mining are achieved for multi-omics data of multiple samples (serum and urine). Furthermore, by utilizing the principle of causal reasoning, the true causal contribution and spurious correlation are distinguished, improving the stability and biological interpretability of feature selection. Ultimately, a high-precision disease prediction and efficacy evaluation model is constructed.

[0019] In one embodiment of the present invention, the tunable batch correction unit performs the following operations:

[0020] a. Group the data according to potential sources of bias;

[0021] b. Calculate robust location statistics for each data set, including the median, weighted quantiles, or combinations thereof;

[0022] c. Weight the statistics based on preset tunable parameters to calculate the system deviation correction amount;

[0023] d. Update the data using the correction amount, wherein the tunable parameter is used to control the proportion of deviation removal in a single iteration;

[0024] e. Repeat the above process until the correction value converges.

[0025] 4. A cross-omics sparse feature selection method based on hierarchical causal modeling, characterized by the following steps:

[0026] S1. Data Acquisition and Adaptive Cleaning: Acquire multi-omics data of subjects, fill missing values ​​using KNN or correlation imputation, and remove outlier samples using the local anomaly factor algorithm.

[0027] S2. Robust System Bias Correction: Using an iterative algorithm based on robust location statistics and tunable parameters, batch effects and systematic biases in the data are eliminated.

[0028] S3. Hierarchical structural equation modeling: Construct a structural equation model that includes an upstream protein layer, a downstream metabolic layer, and a covariate layer, and set a unidirectional causal path constraint that only allows from upstream to downstream.

[0029] S4. Causal Gated Sparse Screening: Learnable causal gating variables are introduced, and the model is trained using sparse regularization to automatically prune non-causal paths. Combined with counterfactual intervention scoring and multiple Bootstrap resampling stability assessments, the final stable causal biomarkers are determined.

[0030] S5. Multi-level integrated modeling: Each omics base model is trained based on causal biomarkers, and the final prediction result is output through a three-layer weighted integration strategy of omics layer, sample layer and body fluid layer and gradient descent weight optimization.

[0031] In one embodiment of the present invention, the causal gating sparse regularization mechanism is applied to the path coefficients of a structural equation model. Introducing continuously differentiable gated variables Define the effective path coefficient ,in A structure mask generated based on hierarchical constraints and prior biological knowledge;

[0032] The optimization objective function of the module includes parameters for the effective path coefficients. L1 norm sparse penalty term and gating variables Entropy regularization term to achieve automatic path selection and sparsification.

[0033] In one embodiment of the present invention, in step S4, the counterfactual intervention score is obtained by fixing other input parameters of the model, virtually assigning perturbations to the target features, and calculating the magnitude of the difference in the model's predicted output probability vector before and after the perturbation, which is used to quantify the causal contribution of the features to the prediction results.

[0034] In one embodiment of the present invention, step S4, the stability assessment specifically includes: performing multiple Bootstrap resampling operations on the original data; independently performing causal gating sparse filtering on each resampled data; counting the frequency of each feature being selected; and retaining features with a selection frequency higher than a preset threshold. Its characteristics.

[0035] In one embodiment of the present invention, the three-layer weighted integration strategy in step S5 has weights that automatically search for the optimal solution by minimizing the log loss function on the validation set and using the gradient descent algorithm, and constrains all weights to be non-negative and normalized to 1.

[0036] In one embodiment of the present invention, the method is applied to the efficacy assessment of diabetic nephropathy (DKD), wherein the causal markers include inflammatory factor receptors TNFR1 and TNFR2, complement components C3 and CFB, and lipid metabolites PC, LPC, SM, and Cer driven by them.

[0037] In one embodiment of the present invention, the method is applied to predict cancer immunotherapy response, wherein the causal biomarkers include immune cytokines in the tumor microenvironment and metabolic reprogramming products regulated by them.

[0038] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the cross-omics sparse feature selection method based on hierarchical causal modeling as described above.

[0039] As described above, the cross-omics sparse feature selection system and method based on hierarchical causal modeling of the present invention has the following beneficial effects:

[0040] 1. High causal reliability: Through explicit modeling and causal gating mechanism of SEM, spurious correlation features caused by confounding factors or reverse causality are effectively eliminated. The screened biomarkers are closer to the pathological nature of disease occurrence and have higher biological rationality.

[0041] 2. Stable screening results: Combining Bootstrap stability selection overcomes the model jitter problem caused by high-dimensional small sample data, significantly improving the reproducibility of biomarkers across centers and batches, which facilitates the subsequent development of clinical test kits.

[0042] 3. Highly interpretable and operable: It not only provides the predictive weights of features, but also provides an assessment of the "intervention value" of features through counterfactual intervention simulation. It can answer the question of "whether regulating this molecule can improve the efficacy", directly supporting translational medicine research and the discovery of new drug targets.

[0043] 4. Superior generalization ability: The three-layer integration strategy fully utilizes the complementary information of different body fluids (serum / urine) and different omics (protein / metabolism), combined with robust bias correction, so that the model performs significantly better than single-omics models or traditional clinical indicator models on independent test sets. Attached Figure Description

[0044] Figure 1 The diagram shown is a schematic representation of the system architecture of the hierarchical causal sparse feature selection framework disclosed in an embodiment of the present invention.

[0045] Figure 2 The diagram shows a hierarchical causal structure path disclosed in the embodiments of the present invention.

[0046] Figure 3 The diagram shows a heatmap of the causal biomarker screening results and a schematic diagram of the intervention effect as disclosed in the embodiments of the present invention.

[0047] Figure 4 The diagram shows a comparison of ROC curves for the single-mathematical model, the combined clinical indicator model, and the integrated model disclosed in this embodiment of the invention. Detailed Implementation

[0048] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.

[0049] Please see Figures 1 to 4 It should be understood that the structures, proportions, sizes, etc., illustrated in the accompanying drawings are merely for illustrative purposes to aid those skilled in the art and to facilitate understanding and reading. They are not intended to limit the scope of the invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, without affecting the effectiveness and purpose of the invention, should still fall within the scope of the technical content disclosed in this invention. Furthermore, the terms such as "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity and not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.

[0050] like Figure 1 As shown, this invention provides a cross-omics sparse feature selection system based on hierarchical causal modeling, comprising:

[0051] 1. Data input and preprocessing module, used to receive multi-omics raw data of multivariate samples, and embeds a tunable batch correction unit based on robust location statistics, used to remove systematic biases across batches and platforms through iterative weighting.

[0052] The data input and preprocessing module includes a multi-source heterogeneous data interface unit, an adaptive preprocessing unit, a tunable batch correction unit based on robust position statistics, and a standardization unit.

[0053] In this embodiment, the multi-source heterogeneous data interface unit adopts a highly compatible data parsing interface, which can read and parse raw omics data in various formats. It can input multi-omics raw data of multiple samples, including serum proteomics, serum metabolomics, urine proteomics and urine metabolomics data. The data is usually organized in tabular form, with each row representing one sample and each column representing a molecular feature (such as protein or metabolite abundance).

[0054] The adaptive preprocessing unit automatically selects a processing strategy based on the sample type (serum or urine) and data distribution. In this embodiment, for missing values ​​and outliers commonly found in omics data, this unit incorporates adaptive decision-making logic.

[0055] Intelligent missing value imputation: Unlike traditional mean / median imputation, this system selects a strategy based on the feature missing mechanism. For non-random missing values ​​caused by low abundance, minimum value surrogate imputation is used; for random missing values, KNN (K-Nearest Neighbors) imputation or iterative imputation based on a weighted correlation network is used to preserve the local structure of the sample in the manifold space to the greatest extent.

[0056] Multidimensional outlier detection: Combining Mahalanobis distance and the Local Outlier Factor (LOF) algorithm. The LOF algorithm effectively identifies technical outliers (such as contaminated samples or failed preparations) that are not obvious in low-dimensional projections but deviate from the manifold in high-dimensional space by calculating the ratio of a sample point to its neighborhood density, and automatically removes them.

[0057] The tunable batch calibration unit is used to eliminate systematic biases caused by different batches, platforms, or bodily fluid sources, ensuring that downstream analysis is not affected by technical factors, and performs the following steps:

[0058] a. Group the data according to potential sources of bias;

[0059] b. Calculate the robust location statistic for each data set;

[0060] c. Weight the statistics based on preset tunable parameters to calculate the system deviation correction amount;

[0061] d. Update the data using the correction amount, wherein the tunable parameter is used to control the proportion of deviation removal in a single iteration;

[0062] e. Repeat the above process until the correction value converges;

[0063] Robust position statistics include, but are not limited to, median, weighted quantile or a combination thereof, and convergence conditions include, but are not limited to, the system bias correction being less than a threshold or the difference between two adjacent data updates being less than a threshold.

[0064] The standardized unit, due to the huge differences in matrix effects between serum and urine (such as protein concentrations differing by several orders of magnitude), uses Z-score normalization or probability integral transformation (PIT) to uniformly map data from different sources to a dimensionless distribution space, ensuring the comparability of path coefficients in subsequent SEM models.

[0065] The data input and preprocessing modules output clean, standardized multi-omics datasets, providing a reliable foundation for subsequent causal modeling.

[0066] 1. Hierarchical causal structure learning module, such as Figure 2 As shown, observed variables are represented by ellipses and circles (upstream proteins are in blue, downstream metabolism is in green), and standardized path coefficients are labeled with one-way arrows (thick arrows indicate significant paths, p<0.05).

[0067] The hierarchical causal structure learning module, connected to the data input and adaptive preprocessing module, is used to construct a hierarchical causal topology across omics. It uses structural equation modeling (SEM) to explicitly define proteomics features as upstream potential regulatory variables and metabolomics features as downstream response variables, and introduces clinical confounding factors as covariates to characterize the directional dependencies between features.

[0068] The hierarchical causal structure learning module includes an upstream-downstream hierarchical definition unit, a structural equation modeling unit, a parameter estimation and model fitting unit, and a unit incorporating prior knowledge.

[0069] The upstream-downstream hierarchical definition unit, based on the central dogma of molecular biology and metabolic regulation logic, has strict pre-defined variable definitions and hierarchical constraints:

[0070] 1) Variable definition

[0071] Upstream variables: proteomic features , ,··· ;

[0072] Downstream variables: metabolomics characteristics , ,··· ;

[0073] Confounding covariates include age, batch size, gender, and body mass index (BMI). These variables may simultaneously affect protein and metabolic levels and are included as covariates in the model to block spurious correlation pathways.

[0074] 2) Hierarchical constraints (core: only allow proteins to target metabolism)

[0075] This embodiment applies hierarchical orientation constraints in the structural equation model (SEM):

[0076] The allowed directional relationship is: Pi→Mj

[0077] Forbidden direction: Mj→Pi

[0078] The structural equation modeling unit automatically constructs a SEM system containing two sets of simultaneous equations:

[0079] Protein layer equation (upstream layer):

[0080]

[0081] Metabolic layer equation (downstream layer, protein-driven):

[0082]

[0083] in, When the covariate takes the value 0, The baseline level;

[0084] When covariates and When the value is 0, The baseline level;

[0085] The coefficient of influence of covariates on protein;

[0086] : indicates protein metabolites The direct path coefficient;

[0087] The coefficient of influence of covariates on metabolites;

[0088] and : Indicates the error term;

[0089] Parameter estimation and model fitting unit: Path coefficients (representing causal strength) are calculated using maximum likelihood estimation, and the model's rationality is verified using fitting indices (such as RMSEA); at the same time, confounding factors (such as batch or age) are introduced as covariates to remove spurious correlations.

[0090] Incorporating prior knowledge units: Combining biological pathway databases (such as KEGG) as constraints enhances the biological reliability of causal relationships. In this embodiment, based on KEGG pathway / enzyme-substrate relationships, candidate protein-metabolism linkages are first generated, and then data is used for screening.

[0091] The hierarchical causal structure learning module outputs a hierarchical causal network diagram, which clearly shows which protein features may causally drive changes in metabolic features, as well as the transmission effect between serum and urine.

[0092] 3. The causal-guided sparse feature selection module is connected to the hierarchical causal structure learning module. Based on the above causal structure, it performs sparse screening to ensure that the selected features have causal semantics and stability.

[0093] The causal-guided sparse feature selection module includes a causal-gated sparse regularization mechanism and a counterfactual intervention importance calculation unit. It is used to automatically prune non-causal paths by optimizing a sparse loss function containing differentiable gated variables, and to screen out a set of sparse, stable biomarkers with causal semantics by evaluating the predicted changes under virtual intervention and the Bootstrap stability score.

[0094] Causal Gated Sparse Regularization Mechanism: Introduce sparse penalty terms (such as L1 norm) into the SEM framework to prioritize the retention of features with significant path coefficients. At the same time, design a "causal gating" mechanism to only allow upstream features to affect downstream features and automatically remove redundant features without causal paths.

[0095] Specifically, in the structural equation model, a gating variable is introduced for each candidate upstream-downstream path (using a differentiable continuous approximation binary gating function, restricting the gating variable to between 0 and 1, and discretizing it according to the threshold after training), so that the path coefficients participate in the model calculation in the form of "the product of the gating variable and the basic coefficients"; and the path coefficients that do not meet the hierarchical constraints are fixed to zero through a structural mask to ensure that only upstream features are allowed to affect downstream features.

[0096] Based on this, a sparse regularization penalty is applied to the effective path coefficients after gating, causing most path coefficients to shrink to zero, thereby automatically eliminating redundant features and paths without causal semantics, and outputting a sparse and stable set of causal markers.

[0097] The calculation method is as follows:

[0098] set up This indicates whether a candidate edge exists (determined by hierarchical rules and prior knowledge);

[0099] If Pi→Mj is allowed, then ;

[0100] If prohibited (e.g., metabolism → protein or not in the KEGG candidate), then ;

[0101] Formula for calculating coefficients:

[0102]

[0103] Ungated base path coefficients (can be positive or negative);

[0104] The gate variable (using a differentiable, continuous approximation binary gate function, restricting the gate variable to between 0 and 1, and discretizing it by a threshold after training) controls whether the path exists;

[0105] Counterfactual Intervention Importance Calculation Unit:

[0106] Virtual intervention: The system executes on the trained SEM model. This operation involves forcibly setting the value of a protein characteristic to a specific level (such as high or low expression), thereby severing its upstream dependencies.

[0107] Effect propagation: By using the pathway network of SEM, we observe how the intervention signal is transmitted to downstream metabolites and ultimately affects the model's response to the disease state. The predicted probability;

[0108] Intervention score calculation: The difference between the predicted probabilities before and after the intervention is calculated. The absolute value of the predicted probability after the intervention minus the original predicted probability is defined as the causal intervention effect score. This score reflects "how much the predicted outcome can be affected if this molecule is changed", providing interpretation for the intervention.

[0109] Stability evaluation unit: Multiple bootstrap resampling is used to repeat causal modeling on different subsets of data, calculate the consistency of feature selection (coefficient of variation or ranking stability), and remove unstable features.

[0110] The causal-guided sparse feature selection module outputs a sparse and stable subset of causal biomarkers, which typically contains only a few dozen high-contribution features, facilitating subsequent biological validation.

[0111] 4. Model retraining and ensemble module, used to build a three-layer weighted ensemble discriminant model based on the selected markers, and optimize the fusion weights of each layer through the gradient descent algorithm to output the final prediction results;

[0112] Basic model training unit: For each omics or fluid subset, the best algorithm is trained and selected using optimized features, employing hierarchical weighted ensemble and a three-layer voting mechanism.

[0113] The first layer weights a single algorithm model and weights different algorithm models from the same biological sample source.

[0114] The second layer weights a single sample type and integrates protein and metabolic models within the same body fluid.

[0115] The third layer is cross-fluid fusion, ultimately fusing the predicted results of serum and urine.

[0116] The weights are automatically calculated using gradient descent to ensure maximum prediction accuracy during training. The training objective is preferably log loss. The final weights are selected on the validation set using one of AUC, accuracy, or F1 score as the performance metric. All weights in each layer are non-negative, and the sum of the weights in the same layer is 1.

[0117] Performance evaluation: Calculate metrics such as AUC (area under the curve) and accuracy of the ROC (Receiving Receiver Operating Curve) on independent test sets, and verify the causal explanatory power of the model through counterfactual simulation.

[0118] 5. Output and Explanation Module: Visualized Causal Insights

[0119] Causal network visualization: Generate causal network diagrams, feature importance heatmaps, intervention effect simulation diagrams, etc., to intuitively display the hierarchical relationships between markers;

[0120] Biological explanation: Combining pathway enrichment analysis, we can explain the mechanistic significance of biomarker screening and support experimental validation design.

[0121] Clinical decision support: Outputs disease risk or efficacy response probability to aid in diagnosis and personalized treatment.

[0122] This embodiment overcomes the problems of spurious correlations, instability, and lack of intervention explanations in traditional methods by fully utilizing the biological hierarchy of protein-metabolism and complementary information from serum and urine through hierarchical causal modeling and sparse selection. The system is highly automated, interpretable, capable of processing high-dimensional, small-sample data, and has good generalization ability. It can be widely applied to biomarker mining and predictive modeling for cancer, metabolic diseases, etc., improving the accuracy and reliability of precision medicine.

[0123] Example 2:

[0124] This embodiment provides a selection method for a cross-omics sparse feature selection system based on hierarchical causal modeling as described in Embodiment 1, including the following steps;

[0125] S1. Data Acquisition and Adaptive Cleaning: Acquire multivariate paired samples (such as serum and urine) from subjects for proteomics and metabolomics testing; perform quality control on the raw data and fill missing values ​​using K-nearest neighbors (KNN) or correlation-based imputation methods based on sample correlation; and remove outlier samples by combining Mahalanobis distance and Local Outlier Factor (LOF) algorithms.

[0126] S2. Robust System Bias Correction: An iterative correction algorithm based on robust location statistics is used to eliminate batch effects. Specifically, the data is grouped according to potential sources of bias (such as batch or platform), and robust location statistics (such as median or weighted quantile) are calculated for each group of residual data. The statistics are weighted based on preset tunable parameters to obtain the correction amount and update the data. The iteration is repeated until the correction amount converges, ensuring that technical noise is removed while biological signals are preserved.

[0127] S3. Hierarchical structural equation modeling: Construct a structural equation model that includes an upstream protein layer, a downstream metabolic layer, and a confounding covariate layer; set hierarchical constraints, allowing only causal arrows from proteins to metabolites and prohibiting reverse connections; establish a linear equation system to explicitly model the direct regulatory effect of proteins on metabolism and the confounding effects of covariates.

[0128] S4. Causal Gated Sparse Screening: Introduce continuously differentiable gating variables and structural masks to the SEM path coefficients to define effective path coefficients; add a sparse penalty term (such as L1 regularization) for effective path coefficients to the loss function, and simultaneously estimate the model parameters and gating state through the optimization algorithm to force the coefficients of non-critical paths to shrink to zero, thereby screening out the feature set with significant direct or indirect transmission effects.

[0129] S5. Counterfactual Intervention Assessment and Stability Selection: Counterfactual inference is performed on the selected features. Other variables are fixed, and the values ​​of the target features are artificially changed. The change in the predicted probability output by the model is observed and defined as the "intervention importance score". At the same time, the original data is resampled multiple times using the Bootstrap method. Steps S3-S5 are repeated to count the frequency of each feature being selected (stability score). Only features with stability scores higher than the preset threshold and significant intervention scores are retained to form the final causal biomarker panel.

[0130] S6. Multi-layer ensemble modeling and output: Using the final biomarker panel, train the base models of each omics separately; construct a three-layer voting ensemble architecture, optimize each base model and the fusion weights of each layer through gradient descent guided by validation set performance; input the test samples into the ensemble model, and output the efficacy response probability or disease risk score.

[0131] Example 3:

[0132] This embodiment provides an application for efficacy discrimination of diabetic nephropathy (DKD) based on Embodiment 1 or Embodiment 2, and describes the entire process of constructing an efficacy discrimination model for patients with diabetic nephropathy (DKD) after receiving modified Huangfeng Decoction using the above system.

[0133] 1. Data preparation:

[0134] Two hundred patients diagnosed with DKD were recruited and divided into a "good efficacy group" (n=100) and a "poor efficacy group" (n=100) based on clinical outcomes (eGFR reduction rate and improvement in proteinuria) after 6-12 months of treatment.

[0135] Pre-treatment baseline serum and urine samples were collected and analyzed using TMT-labeled quantitative proteomics and untargeted metabolomics. Approximately 4,000 protein features and 1,500 metabolite features were obtained.

[0136] The data underwent adaptive preprocessing to remove two outlier samples, and KNN was used to fill in approximately 5% of the missing values.

[0137] 2. Robust bias correction:

[0138] Because the samples were collected in three batches, the original data showed a significant batch clustering effect when subjected to PCA. The robust correction algorithm of this invention (setting...) is then applied. =0.5), after 5 iterations, PCA showed that the samples were clustered by disease group, the systematic differences between batches were eliminated, and the biological differences between groups were preserved.

[0139] 3. Causal Feature Screening: Proteins were set as upstream variables, metabolites as downstream variables, and age, sex, BMI, and baseline eGFR as covariates. Causal gated sparsity screening and 100 Bootstrap stability selections were performed.

[0140] like Figure 2 , Figure 3 As shown, 12 core causal biomarkers were ultimately selected:

[0141] Upstream kinesin: TNFR1, TNFR2 (inflammatory cytokine receptors, known to be closely related to the progression of DKD), CFB, C3 (complement system activation marker), A2M.

[0142] Downstream response metabolites: PC 36:2, LPC 16:0, LPC 18:1 (glycerophospholipids), SM d18:1 / 16:0, Cer d18:1 / 16:0 (sphingolipids), TG 52:2, DG 36:2 (glycerides).

[0143] These biomarkers were significantly upregulated in the group with poor efficacy, with a coefficient of variation <0.2, indicating high stability. During the model retraining and ensemble phases, the optimized causal biomarkers were used to train the base algorithms separately. After selecting the models with the best performance in each omics group, a three-layer weighted voting process was implemented: the first layer was single-omics weighting, the second layer was single-sample type weighting (serum / urine), and the third layer was cross-fluid fusion. The weights were optimized using gradient descent to ensure maximum accuracy in efficacy discrimination.

[0144] Finally, following the above process, this implementation case compared the performance of the single model and the ensemble model on the test set. ROC AUC evaluation showed that the ensemble model achieved an AUC of 0.944, significantly higher than the single model (AUC 0.743-0.875), and superior to traditional clinical indicators (such as demographic characteristics like age, sex, and BMI, and the combined clinical biochemical indicators like proteinuria, eGFR, serum creatinine, and serum urea nitrogen, with an AUC of 0.667). Figure 4 (As shown). The ensemble model fully utilizes the hierarchical relationships of causal biomarkers, significantly improving the accuracy and generalization ability of efficacy prediction.

[0145] Overall, this case study demonstrates how to construct an efficient and accurate DKD efficacy discrimination model through cross-sample multi-omics data preprocessing, hierarchical causal modeling, sparse screening and integration, significantly improving predictive performance and uncovering a panel of biomarkers with causal semantics. This provides an effective tool and method for clinical research in the fields of serum and urine omics.

[0146] Example 4:

[0147] The method of this invention is also applicable to the study of other complex diseases. This embodiment provides a method for predicting the response to immune checkpoint inhibitor (ICI) therapy in non-small cell lung cancer (NSCLC):

[0148] Upstream variables: immune cytokines and immune checkpoint proteins in tumor tissue or plasma.

[0149] Downstream variables: plasma metabolites (such as lactate, kynurenine, and tryptophan).

[0150] Process: This system constructs a causal network of "cytokines → metabolites" and screens out pathway-related metabolic features.

[0151] Results: Compared to using tumor mutational burden (TMB) or PD-L1 expression alone, the multi-omics causal model of this system can more accurately identify patients with "immune cold" tumors, with an AUC increase of more than 15%, and suggests the potential value of combining metabolic regulators (such as IDO1 inhibitors) with immunotherapy.

[0152] Example 4:

[0153] The present invention also provides a computer-readable storage medium (such as a hard disk, optical disk, flash memory, etc.) storing a computer program thereon. When the computer program is executed by a processor, it can implement the various steps of the cross-omics sparse feature selection method based on hierarchical causal modeling described in the above embodiments, including data cleaning, bias correction, SEM modeling, causal gating screening, stability assessment, and ensemble prediction.

[0154] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A cross-omics sparse feature selection system based on hierarchical causal modeling, characterized in that, include: The data input and preprocessing module is used to receive multi-omics raw data of multivariate samples, and has an embedded tunable batch correction unit based on robust location statistics to remove systematic biases across batches and platforms through iterative weighting. The hierarchical causal structure learning module, connected to the data input and adaptive preprocessing module, is used to construct a hierarchical causal topology across omics. It explicitly defines proteomics features as upstream potential regulatory variables and metabolomics features as downstream response variables using structural equation modeling, and introduces clinical confounding factors as covariates to characterize the directional dependencies between features. The causal-guided sparse feature selection module is connected to the hierarchical causal structure learning module. It embeds a causal-gated sparse regularization mechanism and a counterfactual intervention importance calculation unit. It is used to automatically prune non-causal paths by optimizing the sparse loss function containing differentiable gated variables, and to screen out a set of sparse, stable biomarkers with causal semantics by evaluating the predicted changes under virtual intervention and the Bootstrap stability score. The model retraining and ensemble module is used to build a three-layer weighted ensemble discriminant model based on the selected markers, and optimize the fusion weights of each layer through the gradient descent algorithm to output the final prediction results.

2. The cross-omics sparse feature selection system based on hierarchical causal modeling according to claim 1, characterized in that: The tunable batch correction unit performs the following operations: a. Group the data according to potential sources of bias; b. Calculate the robust location statistic for each data set; c. Weight the statistics based on preset tunable parameters to calculate the system deviation correction amount; d. Update the data using the correction amount, wherein the tunable parameter is used to control the proportion of deviation removal in a single iteration; e. Repeat the above process until the correction value converges.

3. The cross-omics sparse feature selection system based on hierarchical causal modeling according to claim 1, characterized in that: The causal gated sparse regularization mechanism in the path coefficients of structural equation models Introducing continuously differentiable gated variables Define the effective path coefficient ,in A structure mask generated based on hierarchical constraints and prior biological knowledge; The optimization objective function of the module includes parameters for the effective path coefficients. L1 norm sparse penalty term and gating variables Entropy regularization term to achieve automatic path selection and sparsification.

4. A cross-omics sparse feature selection method based on hierarchical causal modeling, characterized in that, Includes the following steps: S1. Data Acquisition and Adaptive Cleaning: Acquire multi-omics data of subjects, fill missing values ​​using KNN or correlation imputation, and remove outlier samples using the local anomaly factor algorithm. S2. Robust System Bias Correction: Using an iterative algorithm based on robust location statistics and tunable parameters, batch effects and systematic biases in the data are eliminated. S3. Hierarchical structural equation modeling: Construct a structural equation model that includes an upstream protein layer, a downstream metabolic layer, and a covariate layer, and set a unidirectional causal path constraint that only allows from upstream to downstream. S4. Causal Gated Sparse Screening: Learnable causal gating variables are introduced, and the model is trained using sparse regularization to automatically prune non-causal paths. Combined with counterfactual intervention scoring and multiple Bootstrap resampling stability assessments, the final stable causal biomarkers are determined. S5. Multi-level integrated modeling: Each omics base model is trained based on causal biomarkers, and the final prediction result is output through a three-layer weighted integration strategy of omics layer, sample layer and body fluid layer and gradient descent weight optimization.

5. The cross-omics sparse feature selection system based on hierarchical causal modeling according to claim 4, characterized in that: In step S4, the counterfactual intervention score is obtained by fixing other input parameters of the model, virtually assigning perturbations to the target features, and calculating the magnitude of the difference in the model's predicted output probability vector before and after the perturbation. This score is used to quantify the causal contribution of the features to the prediction results.

6. The cross-omics sparse feature selection system based on hierarchical causal modeling according to claim 4, characterized in that: In step S4, the stability assessment specifically includes: performing multiple Bootstrap resampling operations on the original data; independently performing causal gating sparse filtering on each resampled data; counting the frequency of selection for each feature; and retaining features with a selection frequency higher than a preset threshold. Its characteristics.

7. The cross-omics sparse feature selection system based on hierarchical causal modeling according to claim 4, characterized in that: The three-layer weighted ensemble strategy in step S5 uses the gradient descent algorithm to automatically search for the optimal solution by minimizing the log loss function on the validation set, and constrains all weights to be non-negative and normalized to 1.

8. The cross-omics sparse feature selection system based on hierarchical causal modeling according to claim 4, characterized in that: The method is applied to the efficacy assessment of diabetic nephropathy (DKD), and the causal markers include inflammatory factor receptors TNFR1 and TNFR2, complement components C3 and CFB, and lipid metabolites PC, LPC, SM, and Cer driven by them.

9. The cross-omics sparse feature selection system based on hierarchical causal modeling according to claim 4, characterized in that: The method is applied to predict cancer immunotherapy response, and the causal biomarkers include immune cytokines in the tumor microenvironment and metabolic reprogramming products regulated by them.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 4 to 9.

Images