A multi-omics data analysis method

Abstract

The present invention proposes a multi-omics data analysis method, comprising the following steps: S1, collecting multi-omics data of biological information, wherein the multi-omics data includes any two or more data from genomics, proteomics, metabolomics, proteomics, epigenomics, transcriptomics, and lipidomics; S2, inputting the collected multi-omics data into a MUMA model to train the model; S3, inputting the multi-omics data to be tested into the trained MUMA model to obtain classification results and related explanations. The present invention can utilize meta-learning technology to extract generalizable and interpretable knowledge from different heterogeneous data sources and achieve high performance. The proposed method also includes several explicit and concise meta-learners to help automatically reweight samples from multiple sets of data.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a multi-omics data analysis method. Background Art

[0002] In molecular biology, omics primarily encompasses genomics, proteomics, metabolomics, transcriptomics, lipidomics, immunoomics, glycomics, and RNAomics. Multi-omics, on the other hand, involves using two or more omics research methods to jointly explain scientific questions at the species, gene, and metabolite levels, thereby better understanding disease processes and the metabolic pathways of substances within the body.

[0003] In recent years, researchers have used a variety of high-throughput molecular analysis technologies to generate multi-omics data that measure different layers of cellular regulation. Analyzing these data in an integrated manner is ideal for providing new insights into the molecular interactions that underlie life and disease, or for developing better classifiers for biological samples, such as cancer subtypes.

[0004] However, biological data is subject to varying degrees of noise during both data generation and acquisition. Current multi-omics analysis methods often ignore the high noise characteristics of biological data during modeling, thereby reducing the ability to identify biologically relevant patterns and increasing the risk of overfitting or underfitting model training. Some multi-omics analysis techniques take high noise into account, such as using weighted sample loss functions to reduce the impact of noise on predictive models. However, these techniques require manual pre-set sample weight functions and require additional parameter tuning, resulting in limited practicality. Summary of the Invention

[0005] The present invention aims to at least address the challenges and shortcomings of existing multi-biomedical data analysis methods, reduce the negative impact of high noise in biological multi-omics data samples on prediction models, and particularly innovatively proposes a multi-omics data analysis method.

[0006] In order to achieve the above-mentioned object of the present invention, the present invention provides a multi-omics data analysis method, comprising the following steps:

[0007] S1, collecting multi-omics data of biological information, wherein the multi-omics data includes any two or more data of genomics, proteomics, metabolomics, proteomics, epigenomics, transcriptomics, and lipidomics;

[0008] S2, input the collected multi-omics data into the MUMA model to train the model;

[0009] The MUMA model is:

[0010]

[0011] represents the weight of the i-th sample in the j-th omics;

[0012] m represents the total number of omics;

[0013] n represents the total number of samples;

[0014] represents the loss function of the i-th sample in the j-th omics on the training set;

[0015] β (j) represents the model parameters of the j-th classifier;

[0016] λ (j) represents the hyperparameter that controls the sparsity of model j;

[0017] ||·||1 represents a norm;

[0018] Q(v (j) ,v (k) ) means v (j) ,v (k) Perform Q function operation;

[0019] Is a distance metric penalty function;

[0020] S3: Input the multi-omics data to be tested into the trained MUMA model to obtain the classification results and related explanations.

[0021] Furthermore, the objective function of the MUMA model is:

[0022]

[0023] Represents β (j) It's about Θ (j) function;

[0024] Represents finding a set of β (j) Minimize the E function;

[0025] β (j) represents the model parameters of the j-th classifier;

[0026] Θ (j) represents the hyperparameters of the j-th classifier;

[0027] m represents the total number of omics, i.e., the total number of classifiers;

[0028] Indicates that Θ (j) Input meta-learning learner function

[0029] represents the loss function of the i-th sample in the j-th omics on the training set;

[0030] λ (j) represents the hyperparameter that controls the sparsity of the j-th classifier;

[0031] ||β (j) ||1 means β (j) The one-norm of .

[0032] Furthermore, the MUMA model is implemented based on multiple MLPs as classifiers, with each omics corresponding to an MLP. The input of each MLP is the loss value of the sample, and the output is the weight of the sample. The Sigmoid activation function is used on the output sample weight to ensure that the output is in [0, 1].

[0033] Traditional research typically uses specific weight functions to model continuous phenomena. However, MLPs can dynamically approximate a variety of weight functions and adapt to various noise scenarios. This allows for sample-adaptive weight updates through a mapping mechanism from model loss to an explicit weight function based on meta-learning theory.

[0034] Furthermore, the MLPs are regularized by the interaction term l interactive (β; Θ) interacting;

[0035]

[0036] ||||2 represents the two-norm;

[0037] Indicates that Θ (j) Input meta-learning learner function

[0038] represents the loss function of the j-th omics on the training set;

[0039] Θ (j) represents the hyperparameters of the j-th classifier;

[0040] Indicates that Θ (k) Input meta-learning learner function Ψ (k) ;

[0041] represents the loss function of the k-th omics on the training set;

[0042] m represents the total number of omics, i.e., the total number of classifiers;

[0043] Different omics share common knowledge through their interaction-based regularization terms, helping the model learn across different omics data.

[0044] Furthermore, the meta-learning process of MLP includes:

[0045] Automatically learn the hyperparameter Θ in a meta-learning manner using unbiased data that is much smaller than the number of training samples. (j) ,Θ (j) represents the hyperparameters of the j-th MLP;

[0046]

[0047] M(j) represents the meta-knowledge of unbiased samples;

[0048] To define symbols;

[0049]

[0050] M represents the number of omics in the meta-learning process;

[0051] represents the loss function of the i-th sample in the j-th omics of unbiased data;

[0052] represents unbiased data D meta(j) The label of the i-th sample under the j-th omics;

[0053] represents the jth classifier model of the meta-learning process;

[0054] represents unbiased data D meta(j) The value of the i-th sample under the j-th omics;

[0055] λ (j) represents the hyperparameter that controls the sparsity of the j-th classifier;

[0056] ||β (j) ||1 means β (j) The one-norm of ;

[0057] β (j) are the model parameters of the j-th classifier.

[0058] Furthermore, it also includes: using an alternative optimization search algorithm to optimize the MUMA model:

[0059] (1) Calculation of β (j)* (Θ (j) ) is an approximate solution of Use the objective loss function in equation (2) to apply the gradient descent method to the small batch of sample data to obtain the optimal value β(j)* (Θ (j) ) is an approximate solution of

[0060]

[0061] represents the model parameters of the j-th classifier at time t;

[0062] Where α represents the step size;

[0063] n represents the total number of training samples;

[0064] (2) Fixed β (j) and update Θ (j) :According to equation (6), the classifier parameter update formula Depends on Θ (j) ,therefore, is Θ (j) function, for Θ (j) Then, we use equation (7) and metadata to update the meta-parameter Θ (j) :

[0065]

[0066] where Θ (j)t+1 Represents the meta-parameters of the j-th classifier at time t+1;

[0067] Θ (j)t represents the meta-parameters of the j-th classifier at time t;

[0068] M (j) Representing meta-knowledge of a very small number of unbiased samples;

[0069] δ represents the step size;

[0070] (3) Fixed Θ (j) and update β (j) :When Θ (j)t+1 When fixed, the parameter β of the j-th classifier (j) The solution is:

[0071]

[0072] in Represents the model parameters of the j-th classifier at time t+1;

[0073] represents the model parameters of the j-th classifier at time t;

[0074] α represents the step size;

[0075] n represents the total number of training samples.

[0076] In summary, due to the adoption of the above technical solutions, the present invention is able to extract generalizable and interpretable knowledge from different heterogeneous data sources using meta-learning techniques and achieve high performance. The proposed method includes several explicit and concise meta-learners that help automatically reweight samples from multiple sets of data. Specifically, the method of the present invention utilizes a multi-layer perceptron (MLP) as a universal approximator of almost all continuous functions to parameterize the weighting function. Each specific MLP reweights each omics data according to its uniqueness. In addition, MUMA includes an interaction-based regularization term to help the model learn interactive knowledge across different omics data.

[0077] Specifically, it has the following advantages:

[0078] 1) It can dynamically approximate various weight functions and adapt to various noise conditions, that is, the loss of input samples and the weight of output samples. Traditional methods require manual preset fixed weight functions and cannot adapt to data. (Related to noise removal)

[0079] 1) MUMA has a universal approximation capability that can dynamically approximate a variety of weight functions. This capability is crucial for multi-omics data analysis, as these data are expected to be complex and diverse. To the best of our knowledge, this is the first work to treat multi-omics weight functions in an explicit and concise form.

[0080] 2) MUMA encourages different omics to share common knowledge through its interaction-based regularization term. The confidence level of a sample in a given omics may be determined by leveraging the recommendations of the majority of other omics, similar to a voting strategy. Thus, the robustness of MUMA is improved.

[0081] 3) During the integration process, MUMA can simultaneously predict phenotypes and select significant multi-omics features, providing an efficient classification and gene selection tool, especially under high-noise conditions.

[0082] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. DETAILED DESCRIPTION

[0083] The following describes embodiments of the present invention in detail, wherein the same or similar reference numerals throughout the embodiments represent the same or similar elements or elements having the same or similar functions. The embodiments described below are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0084] The method of the present invention is implemented based on the MUMA model, which includes the MUMA framework and the MUMA objective function. The MUMA model is implemented using multiple MLPs, one for each omics. Each MLP takes the sample loss as input and outputs the sample weight. The MUMA also includes an interaction-based regularization term to help the model learn across different omics data. The meta-learning process refers to how the parameters within the MLP should be optimized.

[0085] MUMA framework

[0086] Consider a typical classification problem of a multi-omics dataset in Represents the value of the i-th sample under the j-th omics, whose dimension is p j ,y j represents the label of the i-th sample, and n represents the total number of training samples. Define f(x (j) ,β (j) ) is the classifier model, where β (j) is the model parameter to be estimated under the j-th omics. In this paper, it is always set to a logistic regression model and is referred to as a classifier for convenience below. Definition is the loss function of the i-th sample in the j-th omics in the training set. For the convenience of representation, it is expressed as:

[0087] We propose a weighted multi-omics analysis framework as follows:

[0088]

[0089] represents the importance of the i-th sample in the j-th omics, i.e., the weight; these weights will attempt to reduce the negative impact of high levels of noise and distribution inconsistency in multi-omics data on the prediction model.

[0090] m represents the total number of omics;

[0091] λ (j) represents the hyperparameter that controls the sparsity of model j;

[0092] ||·||1 represents a one-norm; this forces the model to be sparse, leading to strong biological interpretation.

[0093] Q(v (j) ,v (k) ) means v (j) ,v (k) Perform Q function operation;

[0094] is a distance metric penalty function used to describe the relationship between sample importances of different omics. We use the Euclidean norm as the Q function. It promotes Interactions between omics and samples. This prior structure is designed with the assumption that different omics will share knowledge about sample confidence. This assumption is reasonable and fundamental because the recommended information from other omics is often the main factor in determining the sample confidence level in a specific omics.

[0095] Objective function of MUMA

[0096] In equation (1), we can predetermine the weight update function to update the sample weight Like the strategies of self-paced learning (SPL) or focal loss, those strategies that specify the update function require certain assumptions about the training data. In addition, they usually involve a series of additional hyperparameters, which further limits their application. To address these issues, based on Equation (1), we further specify the proposed MUMA, which includes a mapping mechanism from model loss to an explicit weight function based on meta-learning theory to achieve adaptive sample re-weighing or selection. The goal of the MUMA model is as follows:

[0097]

[0098] Represents β (j) It's about Θ (j) function;

[0099] β (j) represents the model parameters to be estimated under the j-th omics;

[0100] Θ (j) represents the hyperparameters of the j-th omics, i.e., the parameters of the j-th learning machine (MLP);

[0101] Represents finding a set of β (j) Minimize the E function; since there are m prediction models, the value of j is 1 to m, and these different prediction models rely on the interactive regularization term l interactive (β; Θ) interact.

[0102] Indicates that Θ (j) Input meta-learning learner function

[0103] represents the loss function of the i-th sample in the j-th omics on the training set;

[0104] Compared to equation (1), in equation (2), we transform is the meta-learning learner function The input is the loss value of the classifier model on sample i in omics j, and the classifier model is a logistic regression classification model with L1 norm. The loss value is passed Get; the output is the weight of the sample. We denote l interactive (β; Θ) is We model each (j) The MLP uses a sigmoid activation function to ensure that the output lies in the range [0, 1]. Because the weight of a sample ranges from [0, 1], a sample value of 0 would be useless for training the predictive model. Traditional research typically uses specific weight functions to simulate continuous phenomena. However, the MLP can approximate any continuous function with arbitrary precision, making it a more versatile and flexible tool for scientific analysis.

[0105] MUMA's meta-learning process

[0106] Our goal is to automatically learn the hyperparameter Θ in a meta-learning manner. (j) (j=1, 2, ..., m). Specifically, assume that we have a small amount (much smaller than n, which is an empirical value, where n represents the total number of training samples) of unbiased data (i.e., the data is clean and the data distribution is balanced). M (j) Represents the meta-knowledge of a very small number of unbiased samples (much less than n, an empirical value, where n represents the total number of training samples), where the number of meta-samples in the j-omics group is Θ. (j)* It can be updated as follows.

[0107]

[0108] To define symbols;

[0109] β (j) (Θ (j) ) is about Θ (j) function;

[0110] at the same time In the unbiased data D meta(j) Calculated in D. That is, the parameters of the j-th MLP are meta(j) On optimization.

[0111] M represents the number of omics in the meta-learning process;

[0112] represents unbiased data D meta(j)The label of the i-th sample under the j-th omics;

[0113] represents the jth classifier model of the meta-learning process;

[0114] represents unbiased data D meta(j) The value of the i-th sample under the j-th omics;

[0115] MUMA's algorithm

[0116] We use an alternative optimization search algorithm to optimize the MUMA model. The detailed optimization process is as follows:

[0117] Initialization. During this process, some optimization variables and parameters are initialized.

[0118] For each j (j=1, 2, ..., m) execute:

[0119] (1) Calculation of β (j)* (Θ (j) ) is an approximate solution of Use the objective loss function in equation (2) to apply the gradient descent method to the small batch of sample data to obtain the optimal value β (j)* (Θ (j) ) is an approximate solution of

[0120]

[0121] Where α represents the step size;

[0122] (2) Fixed β (j) and update Θ (j) According to equation (4), the classifier parameter update formula is Depends on Θ (j) , note that this is a variable rather than a quantity. Therefore, is Θ (j) function, which can be used for Θ (j) Then we can use equation (5) and metadata to update the meta-parameter Θ (j) :

[0123]

[0124] where Θ (j)t+1 Represents the meta-parameters of the j-th classifier at time t+1;

[0125] Θ (j)t represents the meta-parameters of the j-th classifier at time t;

[0126] M(j) Representing meta-knowledge of a very small number of unbiased samples;

[0127] δ represents the step size. (The step size α is different for different models)

[0128] (3) Fixed Θ (j) and update β (j) When Θ (j)t+1 When fixed, the parameter β of the j-th classifier (j) The solution is:

[0129]

[0130] in Represents the model parameters of the j-th classifier at time t+1;

[0131] represents the model parameters of the j-th classifier at time t;

[0132] α represents the step size;

[0133] n represents the total number of training samples;

[0134] It is represented by Equation 2, where the parameters are β at the tth moment in the jth omics and Θ at the t+1th moment in the jth omics.

[0135] The above steps can be performed by any popular deep learning framework. The parameters of both the classifier and the meta-learner are updated according to the values ​​they calculated in the previous step, resulting in a stable learning process.

[0136] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. A multi-omics data analysis method, characterized in that: The following steps are involved: S1, collecting multi-omics data of biological information, wherein the multi-omics data includes any two or more data of genomics, proteomics, metabolomics, proteomics, epigenomics, transcriptomics, and lipidomics; S2, input the collected multi-omics data into the MUMA model to train the model; The MUMA model is: represents the weight of the i-th sample in the j-th omics; m represents the total number of omics; n represents the total number of samples; represents the loss function of the i-th sample in the j-th omics on the training set; β (j) represents the model parameters of the j-th classifier; λ (j) represents the hyperparameter that controls the sparsity of model j; ||·||1 represents a norm; Q(v (j) ,v (k) ) means v (j) ,v (k) Perform Q function operation; Is a distance metric penalty function; The MUMA model is implemented based on multiple MLPs as classifiers. Each omics corresponds to an MLP. The input of each MLP is the loss value of the sample, and the output is the weight of the sample. The Sigmoid activation function is used on the output sample weight to ensure that the output is in [0, 1]. The interaction regularization term l between MLPs interactive (β; Θ) interacting; || ||2 represents the two-norm; Indicates that Θ (j) Input meta-learning learner function Ψ i (j) ; represents the loss function of the j-th omics on the training set; Θ (j) represents the hyperparameters of the j-th classifier; Indicates that Θ (k) Input meta-learning learner function Ψ (k) ; represents the loss function of the k-th omics on the training set; m represents the total number of omics, i.e., the total number of classifiers; The meta-learning process of MLP includes: Automatically learn the hyperparameter Θ in a meta-learning manner using unbiased data that is much smaller than the number of training samples. ( j ) ,Θ (j) represents the hyperparameters of the j-th MLP; M(j) represents the meta-knowledge of unbiased samples; To define symbols; M represents the number of omics in the meta-learning process; represents the loss function of the i-th sample in the j-th omics of unbiased data; represents unbiased data D meta(j) The label of the i-th sample under the j-th omics; represents the jth classifier model of the meta-learning process; represents unbiased data D meta(j) The value of the i-th sample under the j-th omics; λ (j) represents the hyperparameter that controls the sparsity of the j-th classifier; ||β (j) ||1 means β (j) The one-norm of ; β (j) is the model parameter of the j-th classifier; S3: Input the multi-omics data to be tested into the trained MUMA model to obtain the classification results and related explanations.

2. A multi-omics data analysis method according to claim 1, characterized in that: The objective function of the MUMA model is: Represents β (j) It's about Θ (j) function; Represents finding a set of β (j) Minimize the E function; β (j) represents the model parameters of the j-th classifier; Θ (j) represents the hyperparameters of the j-th classifier; m represents the total number of omics, i.e., the total number of classifiers; Indicates that Θ (j) Input meta-learning learner function represents the loss function of the i-th sample in the j-th omics on the training set; λ (j) represents the hyperparameter that controls the sparsity of the j-th classifier; ||β (j) ||1 means β (j) The one-norm of .

3. A multi-omics data analysis method according to claim 1, characterized in that: Also includes: Optimize the MUMA model using an alternative optimization search algorithm: (1) Calculation Approximate solution of Use the objective loss function in equation (2) to apply the gradient descent method to the small batch of sample data to obtain the optimal value Approximate solution of represents the model parameters of the j-th classifier at time t; Where α represents the step size; n represents the total number of training samples; (2) Fixed β (j) and update Θ (j) :According to equation (6), the classifier parameter update formula Depends on Θ (j) ,therefore, is Θ (j) function, for Θ (j) Then, we use equation (7) and metadata to update the meta-parameter Θ (j) : where Θ (j)t+1 Represents the meta-parameters of the j-th classifier at time t+1; Θ (j)t represents the meta-parameters of the j-th classifier at time t; M (j) Representing meta-knowledge of a very small number of unbiased samples; δ represents the step size; (3) Fixed Θ (j) and update β (j) :When Θ (j)t+1 When fixed, the parameter β of the j-th classifier (j) The solution is: in Represents the model parameters of the j-th classifier at time t+1; represents the model parameters of the j-th classifier at time t; α represents the step size; n represents the total number of training samples.

Images