TWI913719B - Feature creation method - Google Patents

Abstract

The invention provides a feature creation method, which includes the following steps. Diseases are tested separately by upstream somatics and downstream somatics to screen out somatic characteristics that are significantly different. Afterwards, the downstream somatics is used to test the upstream somatics, so as to select significant features. Next, the trans-omics features are created based on the types of upstream somatics and downstream somatics. Finally, feature selection and disease prediction model creation are performed.

Description

Feature creation methods

This invention relates to a feature building method, and more particularly to a feature building method for disease prediction.

Genetic information is transmitted sequentially through the genome, transcriptome, proteome, and metabolite, often referred to as upstream and downstream relationships. With advancements in molecular biology techniques, researchers use somatic data (genomes, transcriptomes, proteomes, and metabolites, etc.) to search for biomarkers between somatic studies and diseases, whether in a single or multiple somatic studies. Past methods of using combined multiple somatic studies to predict diseases extracted the correlation between somatic studies and diseases from a parallel perspective, ignoring the upstream and downstream relationships between somatic studies. However, the upstream and downstream relationships of somatic studies (e.g., the genome affects the downstream transcriptome, which in turn affects the downstream proteome) not only have the potential to assist models in prediction but also allow researchers to more comprehensively explain disease mechanisms from the bottom up.

This invention provides a feature establishment method that establishes trans-omics features through the upstream and downstream relationships of omics, enriches the feature pool, and aims to enhance the effect of disease prediction and further understand the upstream and downstream relationships of omics to explain the pathogenesis mechanism.

The feature establishment method of this invention includes the following steps: First, upstream and downstream voxels are tested against the disease to screen for voxel features showing significant differences. Then, downstream voxels are tested against upstream voxels to select features with significant differences. Next, voxel features are established based on the types of upstream and downstream voxels. Finally, feature selection and disease prediction model establishment are performed.

In one embodiment of the invention, the diseases include age-related macular degeneration, diabetic retinopathy, dementia, sarcopenia, or frailty.

In one embodiment of the present invention, upstream dynamics includes categorical variables and continuous variables.

In one embodiment of the present invention, categorical variables include single nucleotide polymorphisms (SNPs), and serial variables include ribonucleic acid (RNA).

In one embodiment of the present invention, the upstream physical tests for the disease include t-tests or chi-square tests, with t-tests used for continuous variables and chi-square tests used for categorical variables.

In one embodiment of the present invention, the downstream physical examination of the disease includes a t-test or a chi-square test, with the t-test used for continuous variables and the chi-square test used for categorical variables.

In one embodiment of the present invention, the downstream dynamics test on the upstream dynamics test includes the ANOVA test.

In one embodiment of the present invention, when the upstream dynamics is a continuous variable, the upstream dynamics value is standardized as the weight for adjusting the downstream dynamics weight, and the weight is multiplied by the downstream dynamics value to establish the serial dynamics features of the upstream and downstream dynamics.

In one embodiment of the present invention, feature selection includes testing serial physical feature values and diseases to screen for significant features.

In one embodiment of the present invention, feature selection includes selecting solid features and serial solid features using multi-machine learning methods, ranking the importance of features, and determining the number of features using slope plots of AUC and ACC and the elbow method.

In one embodiment of the present invention, the downstream organism includes proteins.

Based on the above, the feature establishment method of this invention approaches the issue from the perspective of upstream and downstream relationships in genetics, quantifying and establishing genetic information transmission features, namely, genetic features, to build a disease prediction model. In this way, the accuracy of the disease prediction model can be improved, and a more direct explanation of the disease mechanism can be provided.

Examples are provided below with accompanying figures for detailed explanation, but the examples provided are not intended to limit the scope of this disclosure. In addition, the terms "comprising," "including," "having," etc., used in this document are open-ended terms, meaning "including but not limited to."

Figure 1 is a flowchart of a method for establishing features according to an embodiment of the present invention.

Referring to Figure 1, in steps S10 and S12, upstream and downstream immunological characteristics are used to test for diseases to screen for immunological features with significant differences. In this embodiment, diseases may include metabolic syndromes, eye-related diseases (such as age-related macular degeneration or diabetic retinopathy), or aging-related diseases (such as dementia, sarcopenia, or frailty), but this invention is not limited to these. Upstream immunological characteristics may include categorical and continuous variables. Categorical variables may be, for example, single nucleotide polymorphisms (SNPs), and continuous variables may be, for example, RNA (transcriptional organisms). Downstream immunological characteristics may be, for example, proteins. The tests performed on diseases using upstream immunological characteristics include t-tests or chi-square tests. T-tests are used for continuous variables, and chi-square tests are used for categorical variables. Downstream immunological tests for the disease include t-tests or chi-square tests. The t-test is used for continuous variables, and the chi-square test is used for categorical variables. A process for screening immunological features with significant differences is, for example, using a significance level of 0.05. Features with a significance level less than 0.05 are considered significant and can be added to the next stage of the process.

Next, referring to Figure 1, in step S20, downstream genotyping is used to test upstream genotyping to select features with significant significance. When upstream genotyping is a categorical variable, such as a single nucleotide polymorphism (SNP), the genotype phenotype can include AA, AB, BB, and O0. When upstream genotyping is a continuous variable, such as RNA (transcriptional molecules), it is a continuous variable describing the abundance of the molecule in the body. Taking upstream genotyping as a categorical variable, tandem genotyping features are established based on the upstream-downstream relationship of genotyping in biology (i.e., genes produce RNA through transcription, RNA then produces peptides through translation, and further combines to form proteins). Utilizing the idea that different SNP phenotypes lead to different protein expression levels, tandem genotyping features of protein expression levels are established using SNP phenotypes. Before establishing features, in order to establish upstream and downstream features with significant differences, downstream atomic features are tested against upstream atomic features to select significant features, such as SNPs and proteins, using ANOVA. The process for selecting significant features is, for example, by setting the significance level to 0.05. If it is less than 0.05, it is considered a significant atomic feature and can be added to the next stage of the process.

Table 1 shows an example of patient data. Upstream dynamics (Category Variable, SNP) Upstream dynamics (continuous variables, RNA) Downstream biology (proteins) Patient 01 AA 76.82 791911 Patient 02 AB 86.74 2424990 Patient 03 AA 168.31 878780 Patient 04 00 58.88 336298

Next, referring to Figure 1, in step S30, establish serial volumetric features based on the types of upstream and downstream volumetrics.

When the upstream protein phenotype is a categorical variable, in the patient data example in Table 1, there are three SNP phenotypes (i.e., AA, AB, and 00). Three SNP-protein phenotypes (i.e., SNP_AA_ProteinX, SNP_AB_ProteinX, and SNP_00_ProteinX) are established to describe the relationship between the SNP phenotype and the downstream protein expression level. For example, in the SNP_AA_ProteinX feature, patient 01 has an SNP phenotype of AA and a protein expression level of 791911. Therefore, patient 01's phenotype value in SNP_AA_ProteinX is 791911. Similarly, Patient 02 has an SNP phenotype of AB and a protein expression level of 2,424,990. However, because its SNP phenotype is not AA, Patient 02 has a string eigenvalue of 0 in SNP_AA_ProteinX. Based on this rule, string eigenvalues corresponding to SNP phenotypes and protein expression levels are established (as shown in Table 2). Table 2 SNP_AA_ProteinX SNP_AB_ProteinX SNP_00_ProteinX Patient 01 791911 0 0 Patient 02 0 2424990 0 Patient 03 878780 0 0 Patient 04 0 0 336298

When the upstream variance features are continuous variables, the upstream variance values are standardized and used as weights for adjusting the downstream variance weights (affecting the concept of multiplier). The aforementioned weights are then multiplied by the downstream variance values to establish the serial variance features of the upstream and downstream variances (as shown in Table 3). Table 3 Weight RNA_ProteinX Patient 01 W ₁ = 76.82 / ( 76.82+86.74+168.31+58.88 ) W ₁ *791911 Patient 02 W ₂ = 86.74 / (76.82+86.74+168.31+58.88) W ₂ *2424990 Patient 03 W ₁₃ = 168.31 / (76.82+86.74+168.31+58.88) W ₃ * 878780 Patient 04 W ₄ = 58.88 / (76.82+86.74+168.31+58.88) W ₄ *336298

Finally, referring to Figure 1, in steps S40 and S50, feature selection and disease prediction model building are performed. Feature selection can be carried out in various ways, such as testing the values of hematologic features against the disease to screen out significant features, or by using multi-machine learning methods to select features from hematologic and hematologic features separately, ranking the importance of features, and using the steep slope plot of AUC and ACC and the elbow method to determine the number of features, so as to find the features that have an impact on the disease among the features of each type, and finally building a disease prediction model.

Once a disease prediction model is established, it can be applied to assist in diagnosis or early warning. After obtaining a sample from the subject (e.g., blood sample), current molecular biology techniques can be used to obtain the subject's SNP, RNA, and other somatic data, and to establish the somatic features required for the model. Then, applying the established prediction model to the subject's data allows for analysis of the subject's probability of developing the disease, leading to further clinical diagnosis or treatment. For example, in the case of sarcopenia, in addition to current diagnostic methods based on slow walking speed, poor grip strength, weight loss, low energy, or low activity levels, this invention's somatic disease prediction model provides a somatic perspective (i.e., biochemical data within the body) to assist in diagnosis. In other words, this invention also provides a method for assessing health and disease risks and a method for predicting diseases. After establishing a disease prediction model using the feature-based method of this invention, the established prediction model is applied to the subject's data to analyze the subject's probability of developing the disease. The analysis results can be used to assist in diagnosis, and then relevant clinical diagnosis or treatment can be carried out.

In summary, this invention provides a feature establishment method that approaches the problem from the perspective of upstream and downstream relationships in genetics (e.g., the genome influences downstream transcriptosomals, which in turn influence downstream proteomics). It quantifies and establishes genetic information transmission features, also known as string DNA features, enriching the feature pool to build disease prediction models. This aims to enhance disease prediction effectiveness and further elucidate upstream and downstream relationships in genetics to explain pathogenesis. In this way, the accuracy of disease prediction models can be improved, and more direct explanations of disease mechanisms can be provided. It also enables researchers to explain disease mechanisms more comprehensively from the bottom up.

S10, S12, S20, S30, S40, S50: Steps

Figure 1 is a flowchart of a method for establishing features according to an embodiment of the present invention.

S10, S12, S20, S30, S40, S50: Steps

Claims (11)

A feature establishment method includes: testing upstream and downstream voxels against a disease to screen voxel features with significant differences; testing the upstream voxels against the downstream voxels to select features with significant differences; establishing serial voxel features based on the types of the upstream and downstream voxels; and performing feature selection and disease prediction model establishment. The feature creation method as described in claim 1, wherein the disease includes age-related macular degeneration, diabetic retinopathy, dementia, sarcopenia, or frailty. The feature creation method as described in claim 1, wherein the upstream variance includes categorical variables and continuous variables. The feature establishment method as described in claim 3, wherein the categorical variable includes single nucleotide polymorphisms (SNPs) and the serial variable includes ribonucleic acid (RNA). The feature establishment method as described in claim 3, wherein the upstream immunology assay for the disease includes a t-test or a chi-square test, wherein the t-test is used for the continuous variable and the chi-square test is used for the categorical variable. The feature establishment method as described in claim 3, wherein the downstream immunology assay for the disease includes a t-test or a chi-square test, wherein the t-test is used for the continuous variable and the chi-square test is used for the categorical variable. The feature building method as described in claim 1, wherein the test of the upstream voxel with the downstream voxel includes an ANOVA test. The feature establishment method as described in claim 3, wherein when the upstream variance is the continuous variable, the upstream variance value is standardized as the weight for adjusting the downstream variance weight, and the weight is multiplied by the downstream variance value to establish the serial variance features of the upstream variance and the downstream variance. The feature establishment method as described in claim 1, wherein the feature selection includes testing serial feature values and the disease to screen for significant features. The feature creation method as described in claim 1, wherein the feature selection includes selecting stereo features and serial stereo features using multi-machine learning methods, ranking the importance of features, and determining the number of features using the slope plot of AUC and ACC and the elbow method. The feature creation method as described in claim 1, wherein the downstream genome includes proteins.