Disease risk characterization prediction system and method

Abstract

The invention provides a disease risk characterization prediction system and method, and belongs to the technical field of biological information processing based on deep learning. Comprising an acquisition module used for acquiring a biomarker index of a to-be-detected target object; and the processing module is used for processing the obtained biomarker indexes of the target object to be detected by using a pre-trained disease risk assessment model to obtain a disease risk characterization result of the target object. According to the method, selection of biomarkers is incorporated into the construction process of the fusion model, and the fusion model uses REB as few as possible to represent most of feature information related to the disease; meanwhile, potential regulation effects among different biomarkers are obtained through feature derivation, so that the model obtains more effective characterization information; the fusion model balances the advantages and disadvantages of a plurality of basic learners, can mutually learn inconsistent information from the plurality of basic learners, and has better information extraction capability and stronger robustness.

Description

Technical Field

[0001] The present invention relates to the technical field of biological information processing based on deep learning, and in particular to a disease risk characterization and prediction system and method. Background Art

[0002] Genes do not act independently to generate or maintain brain function, but rather have co-regulated transcriptional profiles and play important roles in neurodevelopment and neuronal activity. Therefore, gene co-expression analysis based on transcriptome data is a promising approach for identifying genetic interactions and elucidating their impact on disease risk. Therefore, weighted gene co-expression network analysis (WGCNA) is used to analyze the potential signature network of disease co-expressed genes. For example, Radulescu et al. constructed a co-expression network based on WGCNA using transcriptome data from the brains of patients with schizophrenia (SCZ), and identified a signaling module containing many genes involved in synaptic signaling and neuroplasticity, which was also highly correlated with the diagnosis of SCZ and the polygenic risk score.

[0003] At present, artificial intelligence has been widely used in the study and diagnosis of diseases. Machine learning is an important branch of artificial intelligence, which mainly studies how computers can obtain useful predictive features from data through different algorithms or models. Machine learning (ML) models are used in clinical data collection to establish reliable disease risk or stratification models, but most of the existing research and analysis are based on functional or structural magnetic resonance imaging data. High-dimensional data reduction is an integral part of these image processing methods. ML methods may change the structure of the original data matrix and are insufficient to extract effective, reproducible and interpretable features for disease risk characterization and assessment; while research based on bioinformatics methods does not extract enough feature information from the data, and the analysis of disease biomarkers is inefficient and does not produce effective results. In addition, combining machine learning with RNA sequencing (RNA-seq) data can characterize disease risks, but the development of reliable machine learning models based on deep learning to characterize and assess disease risks remains an ongoing challenge. Summary of the invention

[0004] The object of the present invention is to provide a disease risk characterization and prediction method and system to solve at least one technical problem existing in the above-mentioned background technology.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a disease risk characterization prediction system, comprising: An acquisition module, used to acquire biomarker indicators of a target object to be detected; A processing module is used to process the acquired biomarker indicators of the target object to be detected using a pre-trained disease risk assessment model to obtain a disease risk characterization result of the target object; wherein, training the disease risk assessment model includes: the disease risk assessment model includes a basic learner and a meta-learner, the biomarker indicators of the target object are input into the basic learner, and the meta-learner outputs the disease risk characterization result; in training the disease risk assessment model, the basic learner is used to use the training data set as input features, and each basic learner outputs the prediction results and parameters of the biomarkers; the meta-learner is used to use the output of each basic learner as input, select the optimal basic learner combination through an optimization algorithm, integrate the results of the selected basic learners, and output the final probability prediction result of the disease risk.

[0006] As a further limitation of the first aspect, obtaining the training data set includes: based on a disease-related RNA-seq data set, obtaining differentially expressed protein-encoded biomarkers of potential pathways related to the disease to be characterized, defined as important biomarkers; using a mutual information method to evaluate the correlation between each important biomarker and the disease state to be characterized, and using a recursive feature elimination model to capture risk-essential biomarkers, that is, ranking each important biomarker according to its importance; applying polynomial feature derivation and mutual information filtering to the risk-essential biomarkers to identify other basic biomarkers related to the disease to be characterized, thereby obtaining the training data set.

[0007] As a further limitation of the first aspect, the mutual information method is used to evaluate the correlation between each important biomarker and the disease state to be characterized, and the recursive feature elimination model is used to capture the risk essential biomarkers, that is, the important biomarkers are ranked according to their importance, including: using the mutual information method to evaluate the correlation between each biomarker and the disease state, retaining biomarkers with mutual information values ​​between 0 and 0.052, using a pre-trained recursive feature elimination model to traverse all important biomarkers, and recording the importance of each important biomarker according to the ranking of the recursive feature elimination model. When the area under the curve value in the K-fold cross validation no longer increases with the decrease in the number of biomarkers, the number of biomarkers at this time is retained; wherein the pre-trained recursive feature elimination model is brought into the recursive feature elimination search process, and the recursive feature elimination parameters are set, that is, only one important biomarker is eliminated at a time, and the recursive feature elimination process will remove the least important important biomarker from the current data set each time.

[0008] As a further limitation of the first aspect, the training of the pre-trained recursive feature elimination model includes: using a Bayesian optimization algorithm based on the Tree-structured Parzen Estimator method, introducing 5-fold K-fold cross-validation data, optimizing and fine-tuning the hyperparameters of the gradient boosting tree algorithm model, and obtaining the recursive feature elimination model.

[0009] As a further limitation of the first aspect, polynomial feature derivation and mutual information filtering are applied to risk-essential biomarkers to identify other basic biomarkers associated with the disease to be characterized, including: subjecting the risk-essential biomarkers to multiple polynomial feature derivations, combining the derived biomarkers with the risk-essential biomarkers before polynomial derivation, and then filtering relatively useless information through mutual information analysis to obtain other basic biomarkers associated with the disease to be characterized.

[0010] In a second aspect, the present invention provides a disease risk characterization prediction method, comprising: Obtaining biomarker indicators of the target object to be detected; The biomarker indicators of the target object to be detected are processed using a pre-trained disease risk assessment model to obtain a disease risk characterization result of the target object; wherein, training the disease risk assessment model includes: the disease risk assessment model includes a basic learner and a meta-learner, the biomarker indicators of the target object are input into the basic learner, and the meta-learner outputs a disease risk characterization result; in training the disease risk assessment model, the basic learner is used to use the training data set as input features, and each basic learner outputs the prediction results and parameters of the biomarkers; the meta-learner is used to use the output of each basic learner as input, select the optimal basic learner combination through an optimization algorithm, integrate the results of the selected basic learners, and output the final probability prediction result of the disease risk.

[0011] As a further limitation of the second aspect, obtaining the training data set includes: based on the disease-related RNA-seq data set, obtaining differentially expressed protein-encoded biomarkers of potential pathways related to the disease to be characterized, defined as important biomarkers; using the mutual information method to evaluate the correlation between each important biomarker and the disease state to be characterized, and using a recursive feature elimination model to capture risk-essential biomarkers, that is, sorting each important biomarker according to its importance; applying polynomial feature derivation and mutual information filtering to the risk-essential biomarkers to identify other basic biomarkers related to the disease to be characterized, so as to obtain the training data set.

[0012] As a further limitation of the second aspect, the mutual information method is used to evaluate the correlation between each important biomarker and the disease state to be characterized, and the recursive feature elimination model is used to capture the risk essential biomarkers, that is, the important biomarkers are ranked according to their importance, including: using the mutual information method to evaluate the correlation between each biomarker and the disease state, retaining biomarkers with mutual information values ​​between 0 and 0.052, using a pre-trained recursive feature elimination model to traverse all important biomarkers, and recording the importance of each important biomarker according to the ranking of the recursive feature elimination model. When the area under the curve value in the K-fold cross validation no longer increases with the decrease in the number of biomarkers, the number of biomarkers at this time is retained; wherein the pre-trained recursive feature elimination model is brought into the recursive feature elimination search process, and the recursive feature elimination parameters are set, that is, only one important biomarker is eliminated at a time, and the recursive feature elimination process will remove the least important important biomarker from the current data set each time.

[0013] As a further limitation of the second aspect, the training of the pre-trained recursive feature elimination model includes: using a Bayesian optimization algorithm based on the Tree-structured Parzen Estimator method, introducing 5-fold K-fold cross-validation data, optimizing and fine-tuning the hyperparameters of the gradient boosting tree algorithm model, and obtaining the recursive feature elimination model.

[0014] As a further limitation of the second aspect, polynomial feature derivation and mutual information filtering are applied to risk-essential biomarkers to identify other basic biomarkers associated with the disease to be characterized, including: subjecting the risk-essential biomarkers to multiple polynomial feature derivations, combining the derived biomarkers with the risk-essential biomarkers before polynomial derivation, and then filtering relatively useless information through mutual information analysis to obtain other basic biomarkers associated with the disease to be characterized.

[0015] The beneficial effects of the present invention are as follows: the selection of biomarkers is incorporated into the construction process of the fusion model, and the fusion model uses as few REBs as possible to characterize most of the characteristic information related to the disease itself; the potential regulatory effects between different biomarkers are obtained through feature derivation, so that the model obtains more effective characterization information; the fusion model balances the advantages and disadvantages of multiple basic learners, and can learn inconsistent information from multiple basic learners, and has better information extraction capabilities and stronger robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0017] Figure 1 This is a functional principle block diagram of the disease risk characterization and prediction system described in an embodiment of the present invention.

[0018] Figure 2 This is a flow chart of the disease risk characterization and prediction method described in an embodiment of the present invention.

[0019] Figure 3 This is a flow chart of the disease risk assessment model training method described in an embodiment of the present invention.

[0020] Figure 4 This is a diagram of the enrichment analysis results described in an embodiment of the present invention.

[0021] Figure 5 It is a schematic diagram of the recursive feature elimination result described in an embodiment of the present invention.

[0022] Figure 6 It is a schematic diagram for comparing CV test results described in an embodiment of the present invention. DETAILED DESCRIPTION

[0023] To facilitate understanding of the present invention, the present invention is further explained below with reference to specific embodiments in conjunction with the accompanying drawings, and the specific embodiments do not constitute a limitation to the embodiments of the present invention.

[0024] Those skilled in the art should understand that the drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily necessary for implementing the present invention.

[0025] like Figure 1 As shown, in this embodiment, a disease risk characterization prediction system is first provided, which includes: an acquisition module, which is used to obtain biomarker indicators of the target object to be detected. A processing module is used to process the acquired biomarker indicators of the target object to be detected using a pre-trained disease risk assessment model to obtain a disease risk characterization result of the target object. In this embodiment, the above system is used to implement a disease risk characterization prediction method based on transcriptome analysis, such as Figure 2As shown, the method includes the following steps: using an acquisition module to acquire biomarkers of the target object to be detected. Then, using a processing module to process the acquired biomarkers of the target object to be detected using a pre-trained disease risk assessment model, and obtaining a result of characterizing the disease risk of the target object. A (susceptibility / risk) biomarker is a biomarker that is associated with an increased chance of a person suffering from a certain disease or condition, or in some cases, is associated with a reduced chance of the person suffering from a certain disease or condition. For example, in this embodiment, the above system is applied to the characterization and assessment of the risk of mental metabolic diseases. First, the acquisition module is used to acquire the biomarker index of the target object to be detected (e.g., manually inputting the relevant biomarker index using an input unit), and then, the processing module is used to process the acquired biomarkers, and finally the risk assessment result of mental metabolic diseases is output.

[0026] The disease risk assessment model is composed of a basic learner and a meta-learner. The biomarker-related indicators of the target object are input into the basic learner, and the meta-learner outputs the prediction results of the disease risk. The basic learner uses the training data set to find the biomarker characteristics and disease risk prediction rules, construct a suitable objective function, iteratively update from the initial model and initial parameters, and feed back the trained model and parameters to the meta-learner; the meta-learner integrates the results of the basic learner training on multiple tasks to summarize the commonalities of biomarker characteristics and disease risk prediction rules, and performs fast and accurate reasoning on new tasks, and transmits the reasoning to the basic learner as the initial model and initial parameter values, or other parameters that can accelerate the training of the basic learner, while guiding the optimal behavior of the basic learner or exploring a specific new task, and extracting features related to the model and training on the task. Specifically, training the disease risk assessment model includes: the disease risk assessment model includes a basic learner and a meta-learner, the biomarker index of the target object is input into the basic learner, and the meta-learner outputs a disease risk characterization result; in training the disease risk assessment model, the basic learner is used to use the training data set as input features, and each basic learner outputs the prediction results and parameters of the biomarkers; the meta-learner is used to use the output of each basic learner as input, select the optimal basic learner combination through an optimization algorithm, integrate the results of the selected basic learners, and output the final probability prediction result of the disease risk.

[0027] In this embodiment, the training process of the basic learner and the meta learner includes: First, the input data is acquired and processed. In this embodiment, a total of 99 basic biomarkers are obtained. The steps include: first, differential expression analysis is performed based on RNA-seq data sets (including BrainGVEX, CommonMind Consortium CMC, and Lieber Institute for Brain Development LIBD data sets) to obtain 3228 differentially expressed biomarkers, which are then converted into 1800 effective protein coding biomarkers, 138 biomarkers are screened out through pathway enrichment analysis, 44 important biomarkers are obtained by applying quality control standards, and 6 risk essential biomarkers (REBs) are screened out using mutual information analysis and recursive feature elimination models. Polynomial features are derived for these 6 REBs, and a total of 111 biomarkers are obtained by combining the original REBs. Finally, 99 basic biomarkers for model input are finally determined through mutual information analysis and filtering. Among them, BrainGVEX is part of the BrainGVEX study of the Psychological Coding Consortium (Synaptic Connections doi:10.7303 / syn4590909).

[0028] Among them, the processing process for the basic learners is as follows: 2 extreme gradient boosting models (Xgboost), 2 gradient boosting decision trees (GBDT), 1 random forest (RF), 1 support vector machine (SVM), 1 KNN (K-nearest neighbor), 1 LR (logistic regression), a total of 8 candidate basic learners are used; the above 99 basic biomarker data are used as the input of each basic learner, and each basic learner is trained according to the objective function. Each basic learner uses 5-fold cross-validation to evaluate the performance, and finally each basic learner uses the prediction results and parameters of the biomarkers as output.

[0029] The processing process for the meta-learner is: use two candidate meta-learners (RF and LR), take the output of each base learner as input, select the optimal base learner combination and meta-learner through the TPE Bayesian optimization algorithm, perform ensemble learning on the selected base learner results, summarize the ensemble results, and form the final prediction model. The final model uses the probability prediction results of disease risk as output.

[0030] In this embodiment, the above-mentioned training data acquisition and model training process will be described in detail below.

[0031] In this embodiment, obtaining the training data set includes: based on the disease-related RNA-seq data set, obtaining differentially expressed protein-encoded biomarkers of potential pathways related to the disease to be characterized, which are defined as important biomarkers; using the mutual information method to evaluate the correlation between each important biomarker and the disease state to be characterized, and using a recursive feature elimination model to capture risk-essential biomarkers, that is, ranking each important biomarker according to its importance; applying polynomial feature derivation and mutual information filtering to the risk-essential biomarkers to identify other basic biomarkers related to the disease to be characterized, so as to obtain the training data set.

[0032] Among them, the mutual information method is used to evaluate the correlation between each important biomarker and the disease state to be characterized, and the recursive feature elimination model is used to capture the risk essential biomarkers, that is, to rank each important biomarker according to its importance, including: using the mutual information method to evaluate the correlation between each biomarker and the disease state, retaining biomarkers with mutual information values ​​between 0 and 0.052, using a pre-trained recursive feature elimination model to traverse all important biomarkers, and recording the importance of each important biomarker according to the ranking of the recursive feature elimination model. When the area under the curve value in the K-fold cross validation no longer increases with the decrease in the number of biomarkers, the number of biomarkers at this time is retained; wherein, the pre-trained recursive feature elimination model is brought into the recursive feature elimination search process, and the recursive feature elimination parameters are set, that is, only one important biomarker is eliminated at a time, and the recursive feature elimination process will remove the least important important biomarker from the current data set each time. The training of the pre-trained recursive feature elimination model includes: using a Bayesian optimization algorithm based on the Tree-structured Parzen Estimator method, introducing 5-fold K-fold cross-validation data, optimizing and fine-tuning the hyperparameters of the gradient boosting tree algorithm model, and obtaining the recursive feature elimination model.

[0033] Furthermore, polynomial feature derivation and mutual information filtering are applied to the risk essential biomarkers to identify other basic biomarkers related to the disease to be characterized, including: subjecting the risk essential biomarkers to multiple polynomial feature derivations, combining the derived biomarkers with the risk essential biomarkers before polynomial derivation, and then filtering relatively useless information through mutual information analysis to obtain other basic biomarkers related to the disease to be characterized. Among them, polynomial derivation refers to a process of fusing features to form new features. From the perspective of mathematical space, it is to map points in low-dimensional space to high-dimensional space, which is a kind of feature engineering operation.

[0034] Objective Machine learning (ML) is an important branch of artificial intelligence (AI). Its application to RNA sequencing (RNA-seq) data is a promising method for characterizing disease risks. In this example, taking the characterization and evaluation of psychometabolic diseases as an example, a reliable ML model is established based on transcriptome analysis to characterize the risk of schizophrenia (SCZ). The above-mentioned disease risk characterization prediction system, method and training of the disease risk assessment model are described in detail as follows.

[0035] In order to explore reliable risk signature biomarkers, extract effective replicable and interpretable features (this feature will appear in different patients with a certain disease, is universal and the appearance of this feature can be reasonably explained in the occurrence and development of the disease) for the risk characterization and assessment of mental metabolic diseases, and establish a reliable machine learning model based on transcriptome analysis, in this embodiment, traditional bioinformatics analysis is combined with machine learning models, and a new machine learning model (dynamic secondary learning risk fusion model, i.e. DYSLER fusion model) is trained as a disease risk assessment model, and traditional bioinformatics analysis is combined with machine learning models to characterize the risk of mental metabolic diseases, so as to supplement the understanding of the risk characterization and assessment of mental metabolic diseases based on identified basic risk biomarkers.

[0036] like Figure 3 As shown, in this embodiment, the construction and training of the disease risk assessment model includes the following processes: identifying important biomarkers REB in disease-related pathways, capturing and amplifying REB (i.e., capturing risk-essential biomarkers and ranking them according to the importance of biomarkers), and constructing a new DYSLER fusion model for training. Specifically: First, important biomarkers in potential pathways related to mental metabolic diseases are obtained using a variety of traditional bioinformatics tools. Secondly, a new method for identifying REB is designed, and REBs are identified by recursive feature elimination combined with multiple machine learning models and feature engineering techniques, that is, they are ranked according to the importance of risk biomarkers, and then polynomial feature derivation and mutual information filtering are applied to further identify other biomarkers related to mental metabolic diseases, and the capture and amplification of REB are completed. Finally, a new DYSLER fusion model is constructed based on Bayesian optimization to improve the reliability and stability of the risk characterization of mental metabolic diseases. The new DYSLER fusion model includes a basic learner and a meta-learner based on Bayesian optimization to enhance the reliability and stability of characterizing disease risks.

[0037] Specifically, the process of establishing the DYSLER fusion model mainly includes: using the Tree-structured Parzen Estimator (TPE) method to perform Bayesian optimization (a Bayesian optimization algorithm based on a tree structure) on the hyperparameters of the basic learner; using 5-fold cross validation to evaluate the model performance; using a designed random process to fuse the selected basic learners and meta-learners; and finally, packaging the entire process into a complete DYSLER fusion model based on the TPE algorithm. In general, the DYSLER fusion model is an integrated learning model that improves the accuracy and stability of disease risk assessment through the combination of multiple basic learners and a meta-learner. In the specific application process of the DYSLER fusion model, the biomarker indicators of the target object are input, and the disease risk probability value is finally output through the meta-learner. Furthermore, the characterization and classification results of high risk / low risk of disease can be given according to the disease risk probability value. For example, a probability threshold is set, and the disease risk probability value is at or above the set threshold, and it is rated as high risk, and the disease risk probability value is below the threshold, and it is rated as low risk.

[0038] In a specific embodiment, the process of obtaining important biomarkers includes: combining traditional bioinformatics with RNA-seq data sets to identify important biomarkers in potential pathways related to psycho-metabolic diseases. For example, using the R packages of limma, Deseq2 and EdgeR for differential analysis, the three packages are used to preliminarily screen the biomarkers that are differentially expressed between the disease group and the healthy control group. Using the org.Hs.eg.db package, the ENSEMBL IDs of the differentially expressed biomarkers are converted to ENTREZID IDs. Using the clusterProfiler R package, based on the gene ontology (GO) biological process (BP) and the Kyoto Encyclopedia of Genes and Genomes (KEGG), effective protein-coding biomarkers are combined for pathway enrichment analysis, and quality control standards are applied to obtain a list of differentially expressed protein-coding biomarkers enriched in potential pathways related to psycho-metabolic diseases, while reducing confounding effects, and obtaining important biomarkers for subsequent analysis. The quality control standards are prior art, such as the same quality control standards described in the original PsychENcode paper by Gandal et al., and will not be repeated here.

[0039] Specifically, in this example, a systematic search of schizophrenia (SCZ) RNA-seq datasets was performed on the PsychENCODE Synapse portal, and four cohorts were included, including 631 SCZ patients and 1091 healthy controls: BrainGVEX (95 SCZ patients and 259 controls in dataset 1), CMC (264 SCZ patients and 294 controls in dataset 2), LIBD (175 SCZ and 318 controls in dataset 3), and Human Brain Collection Core HBCC (97 SCZ and 220 controls in dataset 4). First, the three RNA-seq datasets (BrainGVEX, CMC, and LIBD) were combined to identify important biomarkers in potential SCZ-related pathways, and after combining the results of various differential analyses, a total of 3228 differentially expressed biomarkers (including 2312 upregulated and 916 downregulated biomarkers) were identified. After converting ENSEMBL IDs to ENTREZID IDs, a total of 1,800 valid protein-coding biomarkers were obtained (including 1,147 upregulated and 653 downregulated biomarkers). In the pathway enrichment analysis, the significance thresholds were set at p-value < 0.05 and q-value < 0.05. Under this condition, 167 upregulated biomarkers and 171 downregulated biomarkers were enriched in 73 upregulated pathways and 92 downregulated pathways, respectively. Among the enriched significant pathways, 39 pathways related to the central nervous system and immunity (27 upregulated pathways and 12 downregulated pathways) were selected as potential SCZ-related pathways, which contained 138 differentially expressed biomarkers. In order to minimize potential confounding effects. The same quality control standards as described in the original PsychENcode paper by Gandal et al. were further applied in this example, and finally 44 important biomarkers enriched in 32 potential SCZ-related pathways were obtained. As shown in Figure 4 As shown, a subset of 32 potential SCZ-related pathways was highlighted in this example, including 28 significant biomarkers in only 10 relevant pathways. Figure 4 In the figure, different colors of the right semicircle represent different SCZ-related pathways, and different colors of the left semicircle represent the expression abundance of different genes in the SCZ group.

[0040] In this embodiment, the capture of risk-essential biomarkers is ranked according to the importance of the biomarkers, including: first, selecting a basic model: using a more effective gradient boosting tree algorithm model. A tree-based Bayesian optimization algorithm is used to introduce 5-fold CV (K-fold cross validation) data to optimize and fine-tune the model's hyperparameters. Secondly, the trained gradient boosting tree algorithm model is brought into the recursive feature elimination search process, and the recursive feature elimination parameter (n_features_to_select = 1) is set, that is, only one biomarker is eliminated at a time. The recursive feature elimination process will remove the least important biomarker from the current data set each time, and then repeat the model training step based on the hyperparameters of the input model to traverse all biomarkers. Finally, the importance of each biomarker is recorded according to the ranking of the recursive feature elimination evaluator. When the area under the curve (AUC) value in CV no longer increases as the number of biomarkers decreases, the number of biomarkers at this time will be retained. In this embodiment, in order to obtain more biomarkers based on REB, polynomial feature derivation (up to 3 times) was performed on the captured risk-essential REB, and then the derived biomarkers were combined with the risk-essential REB captured before the derivation, and finally the relatively useless feature information was filtered out by mutual information analysis (the threshold value was 10% of its MIV average value). The mutual information value MIV is used to reflect the correlation between the biomarker and the disease. In this embodiment, mutual information analysis was used multiple times to screen biomarkers for characterizing disease risk. The count matrices of all data sets were standardized and mutual information analysis was performed to filter out relatively useless feature information in the entire data set. The mutual information method was used to evaluate the correlation between each biomarker and the disease state.

[0041] In this embodiment, the count matrix of all data sets is standardized (is the expression value of the biomarker in each sample, is the average expression value of the biomarker in all samples, and is the standard deviation corresponding to the average), and then mutual information analysis is performed to filter out relatively useless feature information in the entire data set. First, the mutual information method is used to evaluate the correlation between each biomarker and the disease state. Considering that the mutual information value is the difference between the product of the marginal distribution of the joint distribution (as shown in the calculation formula of the KL divergence), in this embodiment, the mutual information value MIV is calculated using the calculation formula of the KL divergence to evaluate the correlation between each biomarker and the disease state. When the mutual information value is between 0-0.052, the corresponding biomarker is retained as a candidate marker. The calculation formula of the KL divergence is: ; in, Represents KL divergence, which is used to measure the difference between two probability distributions; Biomarkers x and disease statesy The joint probability distribution of p(x) Biomarkers x The marginal probability distribution of ; p(y) Indicates disease status y The marginal probability distribution of ; x∈X ,express x From the feature space X Samples in y∈Y ,express y From the label space Y The samples in .

[0042] Specifically, in this embodiment, redundant feature information is filtered out through mutual information analysis, and finally 38 important biomarkers are left, such as Figure 5 Shown is the recursive feature elimination curve, the X-axis represents the number of biomarkers that have fallen off, the Y-axis represents the model AUC value after 5 cross-validations, the shaded area is the 95% confidence interval, the dots represent the results of each experiment, and the dotted line represents the final number of biomarkers that have fallen off. Figure 5 The results of 5 recursive feature eliminations are shown, and the change in the area under the curve (AUC) relative to the number of features is plotted. The optimal number of biomarkers to be retained is determined based on the time point at which the AUC value no longer increases with the decrease in the number of biomarkers. The remaining 6 important biomarkers, including C3AR1, CELSR1, GFAP, HTR2B, NEUROD6 and TREM2. In order to make full use of the known features of the above 6 REBs, polynomial feature derivation was used for expansion, and the biomarkers obtained after expansion (a total of 105) were merged with the above 6 original important biomarkers. These 111 biomarkers were filtered again using mutual information analysis, and finally 99 biomarkers remained. These 99 biomarkers were the input of the base learner.

[0043] In this embodiment, the constructed novel DYSLER fusion model consists of a basic learner and a meta-learner, and the types, quantities, and corresponding hyperparameters of the basic learners and meta-learners need to be optimized. When modeling DYSLER, these selected basic learners and meta-learners are model fused through a designed random process, and the results of disease risk characterization are output from the meta-learner as the final result. The above process is packaged into a complete DYSLER fusion model based on the tree-structured Parsons estimator (TPE) algorithm. The results of disease risk characterization are output from the meta-learner as a reliable final result.

[0044] Specifically, in this embodiment, the model fusion of these selected basic learners and meta-learners through the designed random process includes: a) initializing an empty model pool; b) for each optimized basic learner: using 5-fold cross validation to evaluate the performance, if AUC > 0.65, the model is added to the model pool; c) randomly selecting 3-6 models in the model pool; d) assigning random weights (the sum is 1) to the selected models; e) using the weighted average method to merge the prediction results of the selected models; f) inputting the merged results into the meta-learner; g) using 5-fold cross validation to evaluate the overall performance; h) repeating steps c-g 100 times to select the combination with the best performance.

[0045] In the specific training, 99 biomarkers after polynomial feature derivation and mutual information filtering are used as input features, and the input features need to be standardized and preprocessed ( x-μ ) / σ ,in, x is the original value, μ is the average value, σ is the standard deviation.

[0046] For the hyperparameters of the base learner, the TPE algorithm is used for Bayesian optimization. The objective function is to maximize the average AUC value of 5-fold cross validation, and the number of optimization iterations is 100. Among them, the early stopping condition is: if the AUC does not improve within 20 iterations, stop, and use 4 CPU cores for parallel search. The specific optimization process is: a) For each base learner, use the TPE algorithm to sample within the given parameter range; b) Use the sampled parameters to configure the training model; c) Perform 5-fold cross validation and calculate the average AUC; d) Update the prior distribution of the TPE algorithm; e) Repeat steps ad until the number of iterations or the early stopping condition is reached.

[0047] Finally, the TPE algorithm was used to globally optimize the entire DYSLER fusion model. The optimization parameters were: the selection and number of base learners (3-6), the selection of meta learners (RF or LR), and the initial range of fusion weights for the fusion of base learners and meta learners; the optimization goal was to maximize the AUC of 5-fold cross-validation; the number of optimization iterations was 50; and the early stopping strategy was to stop if there was no improvement in the AUC within 10 iterations.

[0048] Model training and evaluation: Use the optimized DYSLER fusion model configuration to train the model on the entire training set, evaluate the model performance on the validation set, calculate the AUC, accuracy, precision, recall, and F1 score, and perform the final evaluation on the external test set (HBCC dataset).

[0049] Specifically, in this embodiment, a Bayesian optimization method based on TPE (a tree-structured Bayesian optimization algorithm) based on 5-fold CV is used to perform Bayesian optimization on the hyperparameters of the basic learner.

[0050] The objective functions of each basic learner are introduced as follows: The objective function of Xgboost consists of three parts: the first term is the loss function term, which measures the difference between the model prediction value and the true value, the second term γT is the regularization term that controls the complexity of the tree, and the last two terms are L1 and L2 regularization terms, which are used to prevent the model from overfitting. The objective function of Xgboost is: ; in, m represents the number of samples, Indicates t In the iteration i The predicted value of samples; is the first-order derivative, which represents the gradient of the loss function to the current predicted value. is the second-order derivative, which represents the second-order gradient of the loss function to the current prediction value; γT is the regularization term, γ is the regularization coefficient, and T is the total number of leaf nodes; α represents the regularization coefficient of L1, and λ represents the regularization coefficient of L2; Indicates j The prediction score of each leaf node.

[0051] In the objective function of Xgboost, T is defined as the number of leaf nodes, and L1 regularization and L2 regularization are used to control the complexity of the model. and are all predictors of the model output. Numerically, for leaf nodes j Any sample on i , = .

[0052] The objective function of GBDT is: ; ; in, Indicates t At the iteration, the sample i The prediction result of the integrated machine learning algorithm is t The sample number at the iteration i The prediction results; Indicates t- At 1 iteration, the sample i The prediction results; η Represents the learning rate, which is used to control the step size of each iteration; Representation sample i In the t The pseudo residual at the iteration, the label that needs to be fitted during training is the sample The pseudo residual ; The formula calculates the negative gradient, that is, the negative value of the gradient of the loss function to the previous round of prediction results. This value is used as the pseudo residual of the current round to train the new weak learner. Representation sample i The true label of Indicates t- Prediction results at 1 iteration; ∂l represents the partial derivative of the loss function, ∂ is the symbol of partial derivative. is the result of weak estimator fitting. Among them, the loss in the GBDT model is the cross entropy loss, and the evaluation criterion of the weak estimator is the square error.

[0053] The objective function of RF is: ; ; in, It is a dataset The samples in Indicates the random forest t The function of a decision tree, is an indicator function that returns 1 when the condition is met and 0 otherwise; π represents the decision tree generation function in random forest; represents the complete dataset, Represents a randomly sampled data subset; Represents random forest for samples The final prediction result of Y Represents a label space.

[0054] The objective function of SVM is: ; in, ω and b are the hyperplane parameters, ω is the weight vector, b is the bias term; is the Lagrange multiplier; Represents a sample, N Indicates the total number of samples.

[0055] The objective function of LR is: ; is the log-likelihood function, where urepresents model parameters; * represents vector product.

[0056] The loss function of the KNN algorithm is: ; in, k Indicates the number of neighbors; Indicates j categories; Representation sample x of k A set of nearest neighbor samples.

[0057] Finally, the training optimization fusion process of the above basic learners and meta-learners is encapsulated into a complete DYSLER fusion model (i.e., the final disease risk assessment model), whose Bayesian optimization is based on the TPE algorithm, in which the hyperparameters of the basic learners are frozen. When modeling DYSLER, when the hyperparameters of these models are determined by the TPE-based optimization method, the type and number of basic learners and meta-learners can be selected, and then the model fusion of the selected basic learners and meta-learners is performed through the designed random process. The results of disease risk characterization prediction are output from the meta-learner as a reliable final result.

[0058] Finally, in this example, the performance of the DYSLER fusion model constructed above was compared with other ML models to identify risk features based on 5-fold CV of the entire dataset (including BrainGVEX, CMC, and LIBD). The performance of the DYSLER fusion model was compared with other models (including common methods such as Catboost, LightGBM, Xgboost, and common models such as RF, SVM, and decision trees). In CV detection, the DYSLER fusion model performed best with an average AUC value of 77%, as shown in Figure 2. Figure 6 shown.

[0059] To better understand the identified SCZ REBs, in this example, three large-scale genetic datasets (the Psychiatric Genomics Consortium PGC, CMC, and UK Biobank LIBD-Liber Institute for Brain Development UKB) were used to annotate REBs through PRS, as shown in Table 1 .

[0060] Table 1

[0061] For each of the 6 REBs identified, a summary of the different levels of annotation is provided in Table 1. When the PGC3-SCZ-derived PRS was used to predict the UKB-SCZ phenotype (shown in Table 2), the genome-wide The best PRS for PGC3-SCZ and 6-REB levels were all significantly associated with SCZ status (permutation P < 0.00001 or 0.01). When the CMC-SCZ data with a smaller sample size were used, only the genome-wide PRS was significant (permutation P < 0.00001). However, the direction of effect (all odds ratios [ORs] > 1.1) was consistent across all different settings. Although significant predictive power was expected when using PRS constructed from genome-wide SNPs, the results suggest that the polygenic risk for SCZ due to 6-REB SNPs remains above random chance. To leverage the raw data and GWAS summary statistics, PRS were calculated using PRSice2 to assess how the polygenic risk for SCZ in PGC3-SCZ (training data) predicted the phenotype in CMC-SCZ or UKB-SCZ (target data). There were no overlapping samples between the training and target data. To reconcile genetic variation between each pair of training and target data, the analysis was restricted to common SNPs present in both datasets (minor allele frequency > 0.05). SNPs with imputation scores below 0.5 in the training data and palindromic SNPs with ambiguous alleles (A / T and G / C) on both strands were removed. Two PRSs (polygenic risk scores) were constructed, one capturing SNPs with genome-wide significance and the other including only SNPs corresponding to REBs. Standardized PRSs were generated and then associated with binary phenotypes by logistic regression, adjusted for sex and principal components determined by eigenvalues ​​(3 for CMC-SCZ and 4 for UKB-SCZ). PRSice2 was also used to determine the best PRS scores at different P thresholds (5 × 10-8, 10-7, 10-6, 10-5, 104.001.01.05, 0.1, 0.2, 0.3, 0.4, 0.5, and 1). Permutations (100,000 times) were performed to correct for multiple testing at different thresholds and to avoid overfitting. These annotations provide additional supporting evidence linking these REBs to ScZ disease or its related brain functions.

[0062] Table 2

[0063] In this embodiment, the polygenic scoring data of psychometabolic diseases in the UK Biobank database was used for analysis. The polygenic risk of diseases caused by the screened 6-REB SNPs was higher than that of random combinations, and the identified REBs were highly expressed in the central nervous system, which was closely related to neuroinflammation and neurodevelopment, and was consistent with the general theory of etiology of psychometabolic diseases. The significant changes of the above REB in patients with psychometabolic diseases were confirmed by RT-qPCR or western blot, and the behavioral characteristics of psychometabolic diseases were observed in some of the above-screened REB-deficient mice, further indicating that these REBs may have an important effect on neuroinflammation, neurodevelopment, and the metabolism of neural factors. In this embodiment, a variety of metrics and screening criteria were used to generate data sets for downstream analysis, and the screened disease-related risk biomarkers can be verified by animal models. Among the six REBs analyzed in this example, the mRNA levels of five REBs (C3AR1, CELSR1, GFAP, NEORUD6, and TREM2) showed significant changes in the treatment of model mice with MK-801-induced psychometabolic diseases, and the changes in these genes showed changes consistent with the results of human samples in the RNA-seq datasets (BrainGVEX, CMC, LIBD, and HBCC).

[0064] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative work on the basis of the technical solution disclosed in the present invention should be included in the scope of protection of the present invention.

Claims

1. A disease risk characterization prediction system, characterized in that: include: An acquisition module, used to acquire biomarker indicators of a target object to be detected; A processing module is used to process the acquired biomarker indicators of the target object to be detected using a pre-trained disease risk assessment model to obtain a disease risk characterization result of the target object; wherein, training the disease risk assessment model includes: the disease risk assessment model includes a basic learner and a meta-learner, the biomarker indicators of the target object are input into the basic learner, and the meta-learner outputs the disease risk characterization result; in training the disease risk assessment model, the basic learner is used to use the training data set as input features, and each basic learner outputs the prediction results and parameters of the biomarkers; the meta-learner is used to use the output of each basic learner as input, select the optimal basic learner combination through an optimization algorithm, integrate the results of the selected basic learners, and output the final probability prediction result of the disease risk.

2. The disease risk characterization prediction system according to claim 1, characterized in that: Acquiring the training data set includes: based on the disease-related RNA-seq data set, obtaining differentially expressed protein-encoded biomarkers of potential pathways related to the disease to be characterized, which are defined as important biomarkers; using the mutual information method to evaluate the correlation between each important biomarker and the disease state to be characterized, and using a recursive feature elimination model to capture risk-essential biomarkers, that is, ranking each important biomarker according to its importance; applying polynomial feature derivation and mutual information filtering to the risk-essential biomarkers to identify other basic biomarkers related to the disease to be characterized, so as to obtain the training data set.

3. The disease risk characterization prediction system according to claim 2, characterized in that: The mutual information method is used to evaluate the correlation between each important biomarker and the disease state to be characterized, and the recursive feature elimination model is used to capture the essential risk biomarkers, that is, the important biomarkers are ranked according to their importance, including: using the mutual information method to evaluate the correlation between each biomarker and the disease state, retaining biomarkers with mutual information values ​​between 0 and 0.052, using a pre-trained recursive feature elimination model to traverse all important biomarkers, recording the importance of each important biomarker according to the ranking of the recursive feature elimination model, and retaining the number of biomarkers at this time when the area under the curve value in the K-fold cross validation no longer increases with the decrease in the number of biomarkers; wherein, the pre-trained recursive feature elimination model is brought into the recursive feature elimination search process, and the recursive feature elimination parameters are set, that is, only one important biomarker is eliminated at a time, and the recursive feature elimination process will remove the least important important biomarker from the current data set each time.

4. The disease risk characterization prediction system according to claim 3, characterized in that: The training of the pre-trained recursive feature elimination model includes: using a Bayesian optimization algorithm based on a Tree-structured Parzen Estimator method, introducing 5-fold K-fold cross-validation data, optimizing and fine-tuning the hyperparameters of the gradient boosting tree algorithm model, and obtaining the recursive feature elimination model.

5. The disease risk characterization prediction system according to claim 1, characterized in that: Polynomial feature derivation and mutual information filtering are applied to risk essential biomarkers to identify other basic biomarkers related to the disease to be characterized, including: subjecting the risk essential biomarkers to multiple polynomial feature derivations, combining the derived biomarkers with the risk essential biomarkers before polynomial derivation, and then filtering relatively useless information through mutual information analysis to obtain other basic biomarkers related to the disease to be characterized.

6. A disease risk characterization prediction method, characterized in that: include: Obtaining biomarker indicators of the target object to be detected; The biomarker indicators of the target object to be detected are processed using a pre-trained disease risk assessment model to obtain a disease risk characterization result of the target object; wherein, training the disease risk assessment model includes: the disease risk assessment model includes a basic learner and a meta-learner, the biomarker indicators of the target object are input into the basic learner, and the meta-learner outputs a disease risk characterization result; in training the disease risk assessment model, the basic learner is used to use the training data set as input features, and each basic learner outputs the prediction results and parameters of the biomarkers; the meta-learner is used to use the output of each basic learner as input, select the optimal basic learner combination through an optimization algorithm, integrate the results of the selected basic learners, and output the final probability prediction result of the disease risk.

7. The disease risk characterization prediction method according to claim 6, characterized in that: Acquiring the training data set includes: based on the disease-related RNA-seq data set, obtaining differentially expressed protein-encoded biomarkers of potential pathways related to the disease to be characterized, which are defined as important biomarkers; using the mutual information method to evaluate the correlation between each important biomarker and the disease state to be characterized, and using a recursive feature elimination model to capture risk-essential biomarkers, that is, ranking each important biomarker according to its importance; applying polynomial feature derivation and mutual information filtering to the risk-essential biomarkers to identify other basic biomarkers related to the disease to be characterized, so as to obtain the training data set.

8. The disease risk characterization prediction method according to claim 7, characterized in that: The mutual information method is used to evaluate the correlation between each important biomarker and the disease state to be characterized, and the recursive feature elimination model is used to capture the essential risk biomarkers, that is, the important biomarkers are ranked according to their importance, including: using the mutual information method to evaluate the correlation between each biomarker and the disease state, retaining biomarkers with mutual information values ​​between 0 and 0.052, using a pre-trained recursive feature elimination model to traverse all important biomarkers, recording the importance of each important biomarker according to the ranking of the recursive feature elimination model, and retaining the number of biomarkers at this time when the area under the curve value in the K-fold cross validation no longer increases with the decrease in the number of biomarkers; wherein, the pre-trained recursive feature elimination model is brought into the recursive feature elimination search process, and the recursive feature elimination parameters are set, that is, only one important biomarker is eliminated at a time, and the recursive feature elimination process will remove the least important important biomarker from the current data set each time.

9. The disease risk characterization prediction method according to claim 8, characterized in that: The training of the pre-trained recursive feature elimination model includes: using a Bayesian optimization algorithm based on the Tree-structured Parzen Estimator method, introducing 5-fold K-fold cross-validation data, optimizing and fine-tuning the hyperparameters of the gradient boosting tree algorithm model, and obtaining the recursive feature elimination model.

10. The disease risk characterization prediction method according to claim 6, characterized in that: Polynomial feature derivation and mutual information filtering are applied to risk essential biomarkers to identify other basic biomarkers related to the disease to be characterized, including: subjecting the risk essential biomarkers to multiple polynomial feature derivations, combining the derived biomarkers with the risk essential biomarkers before polynomial derivation, and then filtering relatively useless information through mutual information analysis to obtain other basic biomarkers related to the disease to be characterized.

Images