A multi-omics and phenotype association prediction method based on deep residual shrinkage network

Abstract

The present invention discloses a method for predicting the association between multi-omics and phenotypes based on a deep residual shrinkage network. The present invention first performs dimensionality reduction and sorting on each omics high-throughput data through differential analysis, uses a deep residual shrinkage neural network to learn the mapping relationship between omics and phenotypes, and on this basis fuses the above mapping relationship through confidence distance, combines the uncertain method of subjective logic to achieve prediction, and finally reversely derives the omics that is most closely related to the phenotype based on the classification confidence. The present invention integrates various omics to more accurately characterize complex human physiological activities, and more objectively analyzes and predicts phenotypes from multiple dimensions. At the same time, considering the high-dimensional redundancy characteristics of omics data, a deep residual shrinkage neural network with strong noise resistance is designed, combined with a subjective logic multi-omics fusion method based on confidence scores, so that the prediction performance of the present invention is significantly improved in terms of evaluation indicators such as accuracy, F1 score and AUC compared to existing methods.

Description

Technical Field

[0001] The present invention belongs to the field of computer application technology, and in particular relates to a multi-omics and phenotype association prediction method based on a deep residual shrinkage network. Background Art

[0002] With the rapid development of the new generation of biotechnology represented by high-throughput sequencing technology, biological scientists have mined massive amounts of human omics data from various cells, such as genomics, transcriptomics, and proteomics data. These bio-omics data contain a large amount of key information that can represent the essence of human life. However, these data have the characteristics of redundant heterogeneity, large noise interference, and complex association patterns. Traditional biostatistical analysis methods such as multivariate regression analysis have limitations such as narrow representation space and shallow feature depth, making it difficult to conduct complete and accurate association analysis and prediction of massive omics data.

[0003] The breakthroughs achieved in machine learning, especially deep learning, in recent years have solved the limitations of the above-mentioned traditional biostatistical analysis methods. Through the superposition of levels, it expands the shallow representation space into a deep level, and combines variable activation functions to achieve the fitting of nonlinear complex associations, which makes it good at processing large amounts of high-dimensional, noisy, redundant and complex nonlinear data. Therefore, deep learning has been widely used in the processing of various human omics data, and has been widely used in disease diagnosis, precision medicine, gene expression regulation and other fields.

[0004] On the other hand, the operation of the human physiological system is based on the close cooperation of various complex life activities. It is more in line with objective reality to study the association prediction problem of omics-phenotype from the perspective of multi-omics and multi-modality than single-omics and single-modality. Sappagh et al. proposed a multi-omics-phenotype prediction model based on stacked convolutional neural networks and bidirectional long short-term memory networks to dynamically analyze phenotypic changes by combining time series and background knowledge. Park et al. studied a multimodal method based on early fusion of methylation and transcription data, and combined differential analysis and multi-layer perceptron to improve the prediction accuracy of diseases that are difficult to diagnose in clinical practice. Wang et al. used graph convolutional neural networks to predict three omics, combined with the view association discovery network to construct the association relationship of omics and thereby improve the accuracy of phenotypic prediction. The above methods all combine deep learning and multi-omics fusion technology. However, in the deep learning part, only models with fewer layers and simpler structures are used. This is because omics data has the characteristics of high dimensionality, scarce samples, and significant noise compared to other data such as images. As a result, deep learning models with slightly deeper layers are difficult to train on omics data. In other words, sophisticated and complex models in other artificial intelligence application fields can hardly achieve improved association prediction performance in the biological field. Summary of the invention

[0005] In order to overcome the shortcomings of the prior art, the present invention provides a deep learning-driven omics-phenotype association prediction method for high-throughput multi-omics data, aiming at the characteristics of high dimensionality, small number of samples, and large noise interference in human omics data, which makes it difficult to train complex models. First, the differential analysis method is used to preprocess each omics, and the data is cleaned, screened and standardized. Secondly, a new neural network (deep residual shrinkage neural network) that supports multi-class omics data is designed to support the preliminary prediction of single omics-phenotype. Then, a confidence score evaluation model based on confidence distance is proposed, and the confidence score is embedded in the output features of the neural network to characterize the reliability of the preliminary prediction. Finally, a multi-omics fusion association prediction method based on uncertain subjective logic is designed to obtain the association prediction results and classification confidence of multi-omics fusion, which can be reversely deduced to the omics closely associated with the phenotype.

[0006] The present invention is roughly divided into two parts:

[0007] (1) Build a deep residual shrinkage neural network, make full use of the multi-omics features after high-throughput data preprocessing, and learn the association between omics features and their mapping relationship with phenotypes through supervised training. This network is mainly improved by combining the classic residual network in the field of computer vision with the attention mechanism and soft threshold processing. It can find unimportant features through the attention mechanism (specifically implemented as a sub-neural network), and then use soft threshold processing to eliminate these features and retain important features, thereby improving the noise resistance of the neural network. On the other hand, considering the characteristics of high dimensionality and small number of samples of omics data, the dense identity connection method is used to enable the neural network to obtain better prediction ability while retaining the easy-to-train characteristics of shallow neural networks while deepening the learning level. Therefore, this neural network can also be called a deep dense residual shrinkage neural network. This neural network and its corresponding preprocessing method can adapt to the preliminary learning of heterogeneous omics data such as sequences, arrays, and chips.

[0008] (2) Design an omics fusion method that combines confidence distance and subjective logic uncertainty method. This method is tightly coupled with the above-mentioned deep residual shrinkage neural network to jointly realize the prediction analysis of multi-omics-phenotype association. Calculate the confidence distance, and generate a confidence score based on the distance. Embed the probability output of the deep residual shrinkage neural network for each omics into the confidence score, thereby indicating the degree of trust each sample has in the prediction results of each omics. Combined with the uncertain evidence fusion theory based on subjective logic, the probability matrices of the multiple omics embedded in the confidence scores are fused, and the phenotype prediction results are analyzed to varying degrees from multiple perspectives. Finally, reliable multi-omics-phenotype association prediction results are obtained.

[0009] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0010] A multi-omics and phenotype association prediction method based on deep residual shrinkage network, the steps are as follows:

[0011] Step (1) Data preprocessing and deep residual shrinkage neural network construction

[0012] The main work of this step is to preprocess the high-dimensional multi-omics data separately (cleaning, screening and sorting, etc.). After the preprocessing, each omics data is preliminarily predicted by a new neural network proposed in the present invention, laying the foundation for the subsequent step (2) multi-omics fusion association prediction, as follows:

[0013] (1.1) Data preprocessing mainly realizes the conversion of the high-dimensional initial data representation space into a low-dimensional form for use by the neural network proposed in (1.2). In the present invention, the original data of each omics is in the form of a high-dimensional expression matrix. Due to the high depth of high-throughput sequencing, the huge amount of data, and the high noise and redundancy, the present invention uses a t-test-based differential analysis to perform preliminary cleaning, dimensionality reduction screening and sorting of each omics data during preprocessing. For a certain feature of all n samples of any omics, data samples are divided into k different phenotypes Represents the phenotypic label of the sample. The subscript j of i is a variable, indicating the specific category of the phenotypic label, ranging from 1 to k, such as Indicates the second sample of this feature, the phenotype is the data of the first category. The differential expression analysis is calculated based on formula (1):

[0014]

[0015] Among them, m and l represent the sample number, are the variances of the two samples respectively, n is the total number of samples, is the mean, and t is the statistic. Formula (1) describes the basic principle of feature difference analysis. Using a two-sided test, the confidence constant α is selected according to specific needs. The P value can be obtained by querying the t distribution table. The P value is a quantity that measures the statistical correlation between the feature and the label. The smaller the P value, the higher the correlation of the feature. According to the research conditions, the appropriate P value threshold is selected to screen all the features of each group.

[0016] (1.2) Construct a deep residual shrinkage neural network, the input of which is the omics features screened in (1.1), and the output is a preliminary prediction vector of the phenotype. The deep residual shrinkage neural network is formed by stacking multiple residual shrinkage blocks with the same structure but different dimensions, and the residual shrinkage block is improved based on the classic bottleneck residual block. The structure of the classic bottleneck residual block consists of three convolution blocks (each convolution block consists of a convolution layer, a BatchNormalization (BN) layer and a RectifiedLinear Units (ReLU) layer stacked in sequence) and an identity connection. The residual shrinkage block proposed in the present invention makes two improvements on the basis of the classic bottleneck residual block. The first improvement is to add a soft threshold mechanism between the original third convolution block and the identity connection, which is specifically based on formula (2):

[0017]

[0018] In formula (2), the threshold is τ, x is the output of the third convolution block, and y is the output of x after soft threshold processing. The calculation of τ is the core of the soft threshold mechanism, which must meet three conditions at the same time: it should take different values ​​according to the noise conditions of different samples; it must be a positive number; it cannot be greater than the maximum absolute value of the input signal. In order to meet these three conditions, the present invention uses a sub-neural network to learn the threshold τ for each sample. The sub-neural network is composed of a global pooling layer, two fully connected layers, a ReLU layer, a BN layer and the final Sigmoid layer connected to each other. The operating mechanism of the sub-neural network is as follows: the input is x, then x is absolute valued, and then globally pooled to obtain a one-dimensional vector. The one-dimensional vector learns a scaling parameter z through two fully connected layers with activation functions and a BN layer. The scaling parameter is processed by the Sigmoid layer of formula (3) to obtain the scale parameter α:

[0019]

[0020] After obtaining α, the threshold can be calculated by combining formula (4):

[0021] τ=α·average ii,jj,cc |x ii,jj,cc | (4)

[0022] In formula (4), x ii,jj,cc is the feature map of the global pooling layer of the sub-neural network, and the three subscripts represent the width, height and channel of the feature map respectively. After the above process, the sub-neural network learns a threshold that satisfies the three conditions, so that the soft threshold mechanism can be further realized through formula (2), and finally the residual shrinkage block proposed in the present invention and even the entire deep residual shrinkage neural network have the anti-noise ability for each omics feature.

[0023] The second improvement is to improve the single identity connection of the classic bottleneck residual block into multiple identity connections. Specifically, the original identity connection is just a single connection from the input to the output of the bottleneck residual block. The present invention adds identity connections to the output of the residual block after the first and second convolution blocks respectively. This dense connection form realizes more comprehensive feature reuse, and ultimately makes the entire neural network easier to train when the depth increases.

[0024] The residual shrinkage block used in the present invention can be constructed by combining the classic bottleneck residual block with the above two improved methods. A specific number of residual shrinkage blocks of different dimensions are stacked as required, and specific processing layers such as full connection, pooling, convolution and softmax are added to completely construct the deep residual shrinkage neural network used in the present invention. The last layer of the neural network is a softmax layer, and its principle is based on the following formula (5):

[0025]

[0026] where f y is the input of the previous layer, C is the classification category, this processing satisfies the prediction probability of each classification is non-negative and the sum is 1, which means that the prediction probability of each predicted phenotype is positive and the sum is a complete set. After this processing, the prediction vector y can be formed pro =(c1,c2,…,c k ), c l (l∈1,2,...k) is the category of the predicted classification.

[0027] After processing in step (1), the high-dimensional noisy multi-omics data is input into the deep residual shrinkage neural network for prediction, which can obtain preliminary prediction vectors for each omics of each sample. These preliminary prediction vectors can achieve the goal of multi-omics and performance association prediction of the present invention through multi-omics fusion analysis in step (2).

[0028] Step (2) Omics fusion analysis model combining confidence distance and subjective logic uncertainty algorithm

[0029] This step first calculates the confidence distance based on the preliminary prediction vector output by the deep residual shrinkage neural network in step (1), and then generates a confidence score based on the confidence distance. Then, multi-omics fusion based on the subjective logic uncertainty algorithm is performed according to the confidence score, and finally multi-omics phenotype-omics association prediction is achieved, as follows:

[0030] (2.1) Calculation of confidence distance and confidence score. For each omics sample, multiple probability vectors can be obtained by predicting the above deep residual shrinkage neural network. These vectors need to calculate the confidence distance and generate the confidence score. The calculation of the confidence distance is based on formula (6):

[0031]

[0032] y in formula (6) pre Represents y pro The vector after one-hot processing, and Represents the average value of the two vectors, Distance(y pro ,y pre ) is y pro With y pre The confidence distance between the two. The confidence score s can be obtained by simply weighting and normalizing the confidence distances of multiple omics, which directly represents the degree of trust in the prediction results of each omics. The confidence score s is used to expand the prediction probability vector y pre The formation of the mass function M can lay the foundation for the subsequent subjective logical uncertainty association prediction.

[0033] (2.2) Multi-omics association prediction based on subjective logic uncertainty evidence fusion theory, using the preliminary prediction vector y in step (2.1) above pro And confidence scores to achieve the final multi-omics-phenotype association prediction. The specific output is the final prediction result vector y res , the confidence score s of the vector end And the confidence distance vector rel between the final prediction result and each omics. In the hypothesis space, the mass function M is used to represent the probability of each situation in a specific sample space and its confidence. The mass function of each omics can be expressed as s i is the confidence score obtained in step (2.1), u is the number of omics species, is the initial prediction vector of the r-th omics. The total mass function M of the sample all It can be fused from the mass function of each group of omics as follows:

[0034]

[0035] To express the above formula (7) more specifically, for any After fusion, c q and q Calculate (q∈1,2,…u) by formula (8):

[0036]

[0037] Based on formula (7) and formula (8), the final multi-omics-phenotype prediction vector of each sample, which is derived from the fusion of u omics, can be calculated: and its confidence score s endIn addition, based on formula (6), the confidence distance between the fusion prediction vector and the initial prediction probability vector of each single group can be compared, that is, The rel can characterize the degree of association between each group of omics and the final prediction results of the phenotype, and the smaller the distance, the higher the degree of association is considered.

[0038] Beneficial effects of the present invention: The present invention is committed to improving the accuracy and reliability of omics-phenotype association prediction, integrating multiple omics to more accurately characterize complex human physiological activities, that is, more objectively analyzing and predicting phenotypes from multiple dimensions, and taking into account the high-dimensional redundancy characteristics of omics data, a deep residual shrinkage neural network with strong noise resistance is designed, combined with a subjective logic multi-omics fusion method based on confidence scores, so that the omics-phenotype association prediction performance of the entire model framework is significantly improved in terms of accuracy, F1 score, AUC (Area Under Curve) and other evaluation indicators compared with existing methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is the overall framework of the multi-phenotype-omics association prediction of the present invention.

[0040] Figure 2 It is a schematic diagram of the construction of the contraction residual block of the present invention.

[0041] Figure 3 Schematic diagram of the structure of the deep residual neural network of the present invention.

[0042] Figure 4 Schematic diagram of the multi-omics fusion method of the present invention.

[0043] Figure 5 This is a workflow diagram for multi-phenotype-omics association prediction of the present invention.

[0044] Figure 6 This is a performance comparison chart of the present invention and other similar models and algorithms. DETAILED DESCRIPTION

[0045] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0046] The present invention can be used for omics-phenotype association prediction, and the overall framework is as follows Figure 1 As shown, the general operation process is: the multi-omics data is pre-processed by performing differential analysis of the circular t test, and then the preliminary prediction of each omics is completed by the novel neural network proposed by the present invention, and then the preliminary prediction vector is evaluated by the confidence distance to obtain the confidence score, and finally based on the uncertain evidence fusion theory of subjective logic, the preliminary prediction vector is fused according to the confidence score of each omics to obtain the multi-omics-phenotype association prediction vector. The specific implementation of the present invention is described below with reference to an example.

[0047] The implementation scenario of this embodiment is the association prediction of low-grade glioma disease pathology classification. The specific implementation scheme discussed is only used to illustrate the implementation method of the present invention, and does not limit the scope of the present invention. More specifically, although the specific implementation scheme in this section is aimed at low-grade glioma, it is essentially an association prediction problem using four omics for binary classification, but the present invention can be used for any other phenotype (not just disease phenotype), any number of omics or more complex multi-classification problems. The following is a detailed description of the implementation scheme of the present invention for low-grade glioma disease association prediction, which is mainly divided into three parts: multi-omics data preprocessing, deep residual shrinkage neural network construction and prediction, confidence score calculation and multi-omics fusion association prediction, and combined with the overall framework Figure 1 and the overall process Figure 5 To elaborate:

[0048] (1) Multi-omics data preprocessing

[0049] The data used in this example comes from The Cancer Genome Atlas Program (TCGA), which includes four types of omics data: methylation data, mRNA data, miRNA data, and copy number variation data. The four types of omics raw data have been processed into matrix form, that is, the present invention is from the flow chart ( Figure 5 ) starts from the third box of the . The sample phenotypes are divided into two types - grade2 and grade3, with 210 and 212 numbers respectively. Taking a certain feature from mRNA as an example, two different types of samples are analyzed, and their feature means and sample variances are calculated respectively. The value of t is calculated based on formula (1). The P value can be obtained from the t distribution table by the t value and the confidence level α (taken as 0.05). The P value is used as the criterion for feature screening. The present invention takes the P value of 0.005 as the threshold. Each feature of each group is processed as described above, thereby each group 10 4 Reduce the dimension of data to 10 2 The data preprocessing part is completed, and the flowchart corresponds to the completion of "circular difference analysis". As for the phenotypic data, it can be simply processed into a one-hot vector according to the industry-recognized method.

[0050] (2) Construction and prediction of deep residual shrinkage neural network

[0051] like Figure 2 As shown in the figure, the residual shrinkage block is established. The left part of the figure represents the dense identity connection, the right part is the subnetwork for learning the threshold (the principle is based on the previous section, formula (3) and (4)), and the middle part represents the soft thresholding process (the principle is based on the previous section formula (2)). Figure 3Right connect multiple residual shrinkage blocks to form a residual shrinkage group, such as Figure 3 The four groups plus other processing modules can completely build a complete deep residual shrinkage neural network. The environment used in this embodiment is built based on python3.8, and the main package dependencies include numpy, tensorflow, pandas, scikit-learn, etc. The neural network is built using tensorflow. At the same time, the present invention is a general algorithm model with logical commonality, which can be built by most mainstream frameworks and environments, and is not limited to the tools or software used in this implementation case.

[0052] After the neural network is built, the preprocessed omics-phenotypes are input into the network for training. The early stopping mechanism is adopted for 2000 rounds of training. The ratio of training set to independent test set is 4:1, the dropout rate is 0.5, and the learning rate is 0.001. The multi-omics data of the independent test set is input into the trained neural network to generate preliminary prediction results of the four omics for each sample. The subsequent performance evaluation is based on the independent test set (5 independent test sets are randomly selected to ensure the reliability of the experimental part of this implementation plan). This part corresponds to "Preliminary prediction results calculated by neural network" in the flowchart.

[0053] (3) Confidence score calculation and multi-omics fusion association prediction

[0054] Each sample corresponds to each omics. The initial prediction results of the neural network in the previous section are processed by formula (5) to generate a probability vector. The confidence distance between each probability vector and its prediction result is calculated based on formula (6). The confidence distances of multiple omics are simply weighted and normalized to obtain a confidence score, which is embedded in the probability vector. The above can correspond to Figure 4 The multi-omics fusion method knows the "j part of the probability vector with confidence score". Then, the subjective logic evidence fusion algorithm formed based on formula (7) and formula (8) fuses the probability vectors with confidence scores of the four omics and generates the final prediction result. Finally, based on formula (6), the similarity between the probability vector of the final prediction result and the probability vector of each omics is calculated (corresponding to Figure 4 Dissimilarity calculation) to obtain the correlation ranking between each group and phenotype, thus completing the multi-omics-phenotype association prediction process, corresponding to Figure 4 The second half. Take the association prediction of a sample with a specific phenotype of grade2 as an example. Figure 5After generating the confidence scores, the confidence scores of each group are compared. If the confidence scores are close and close to 1 at the same time, the weighted average of the two is taken as the final result; if one confidence score is close to 1 and the other is smaller (such as 0.7), the weight of the group with the smaller confidence score is reduced to generate the final result; if the confidence scores are both low, other groups are considered. In this way, the fusion of groups is performed two by two until all groups are fused to obtain the final multi-omics-phenotype result. The probability vector of the result is calculated with the probability vector of each group through formula (6) to calculate the degree of association. The smaller the value, the higher the degree of association between the group and the phenotype.

[0055] In addition, the performance of the model algorithm proposed in this invention has been experimentally proven to be significantly better than most existing methods by 4% to 8% in terms of accuracy, F1 score and AUC. Figure 6 Among them, MOGONET, MOMA and MDNNMD are models published by other scholars in top international journals in the past three years, and SVM is a traditional support vector machine model.

Claims

1. A multi-omics and phenotype association prediction method based on deep residual shrinkage network, characterized in that: Here are the steps: Step (1) Data preprocessing and deep residual shrinkage neural network construction (1.1) Data preprocessing: Use t-test-based differential analysis to perform preliminary cleaning, dimension reduction screening and sorting of each omics data; for a certain feature of all n samples of any omics, divide them into k data samples with different phenotypes Represents the phenotypic label of the sample. The subscript j of i is a variable, indicating the specific category of the phenotypic label, ranging from 1 to k. The differential expression analysis is calculated based on formula (1): Among them, m and l represent the sample number, are the variances of the two samples, n is the total number of samples, is the mean, t is the statistic; Formula (1) describes the basic principle of feature difference analysis. Using a two-sided test, the confidence constant α is selected according to specific needs. The P value can be obtained by querying the t distribution table. The P value is a quantity that measures the statistical correlation between the feature and the label. The smaller the P value, the higher the correlation of the feature. The P value threshold is selected according to the research conditions to screen all the features of each group; (1.2) Constructing a deep residual shrinkage neural network: The input of this neural network is the omics features screened in (1.1), and the output is the preliminary prediction vector of the phenotype; the deep residual shrinkage neural network is formed by stacking multiple residual shrinkage blocks with the same structure but different dimensions, and the residual shrinkage block is improved based on the classic bottleneck residual block; the structure of the classic bottleneck residual block consists of three convolution blocks and an identity connection, and each convolution block consists of a convolution layer, a BN layer, and a ReLU layer stacked in sequence; the residual shrinkage block makes two improvements on the basis of the classic bottleneck residual block. The first improvement is to add a soft threshold mechanism between the original third convolution block and the identity connection, which is specifically based on formula (2): In formula (2), the threshold is τ, x is the output of the third convolutional block, and y is the output of x after soft threshold processing; τ must meet three conditions at the same time: take different values ​​according to the noise conditions of different samples; must be a positive number; cannot be greater than the maximum absolute value of the input signal; to meet these three conditions, a sub-neural network is used to learn the threshold τ for each sample; the sub-neural network is composed of a global pooling layer, two fully connected layers, a ReLU layer, a BN layer and the last Sigmoid layer connected to each other; the operation mechanism of the sub-neural network is as follows: the input is x, then x is absolute valued, and then globally pooled to obtain a one-dimensional vector, the one-dimensional vector learns a scaling parameter z through two fully connected layers with activation functions and a BN layer, and the scaling parameter is processed by the Sigmoid layer of formula (3) to obtain the scale parameter α: After obtaining α, the threshold is calculated by combining formula (4): τ=α·average ii,jj,cc |x ii,jj,cc | (4) In formula (4), x ii,jj,cc is the feature map of the global pooling layer of the sub-neural network. The three subscripts represent the width, height and channel of the feature map respectively. The sub-neural network learns a threshold that satisfies the three conditions, and then further implements the soft threshold mechanism through formula (2), which ultimately enables the residual shrinkage block and the entire deep residual shrinkage neural network to have the anti-noise ability for each omics feature. The second improvement is to improve the single identity connection of the classic bottleneck residual block into multiple identity connections; specifically, the identity connection to the output of the residual block is added after the first and second convolution blocks respectively; By stacking a specific number of residual shrinkage blocks of different dimensions as required and adding specific processing layers such as full connection, pooling, convolution and softmax, a deep residual shrinkage neural network can be fully constructed. The last layer of the neural network is the softmax layer, and its principle is based on formula (5): where f y is the input of the previous layer, C is the classification category, and this processing satisfies that the prediction probability of each classification is non-negative and the sum is 1, that is, the prediction probability of each predicted phenotype is positive and the sum is a full set. After this processing, the prediction vector y is formed pro =(c1,c2,…,c k ), c l , l∈1,2,...k, is the category of the predicted classification; The multi-omics data are input into a deep residual shrinkage neural network for prediction, which can produce a preliminary prediction vector for each omics of each sample; Step (2) Omics fusion analysis model combining confidence distance and subjective logic uncertainty algorithm First, the confidence distance is calculated based on the preliminary prediction vector output by the deep residual shrinkage neural network in step (1). Then, the confidence score is generated from the confidence distance. Then, multi-omics fusion based on the subjective logic uncertainty algorithm is performed according to the confidence score, and finally multi-omics phenotype-omics association prediction is achieved, as follows: (2.1) Calculation of confidence distance and confidence score: For all samples of each omics, multiple probability vectors are predicted by deep residual shrinkage neural network. These vectors need to calculate the confidence distance and generate confidence scores. The calculation of confidence distance is based on formula (6): y in formula (6) pre Represents y pro The vector after one-hot processing, and Represents the average value of the two vectors, Distance(y pro ,y pre ) is y pro With y pre The confidence distance between the two; the confidence distances of multiple omics are weighted and standardized to obtain the confidence score s, which directly represents the degree of trust in the prediction results of each omics; the confidence score s is used to expand the prediction probability vector y pre The mass function M is formed, which lays the foundation for the subsequent subjective logic uncertainty association prediction; (2.2) Multi-omics association prediction based on subjective logic uncertainty evidence fusion theory: Using the preliminary prediction vector y in step (2.1) pro and confidence scores to achieve the final multi-omics-phenotype association prediction; the specific output is the final prediction result vector y res , the confidence score s of the vector end And the confidence distance vector rel between the final prediction result and each omics; in the hypothesis space, the mass function M is used to represent the probability of each situation in a specific sample space and its confidence. The mass function of each omics is expressed as s i is the confidence score obtained in step (2.1), u is the number of omics species, is the initial prediction vector of the r-th omics; the total mass function M of the sample all It is fused from the mass function of each group, as follows: A more specific statement of formula (7): For any After fusion, c q and q Calculated by formula (8), where q∈1,2,…u: Based on formula (7) and formula (8), the final multi-omics-phenotype prediction vector of each sample, which is derived from the fusion of u omics, is calculated: and its confidence score s end In addition, based on formula (6), the confidence distance between the fusion prediction vector and the initial prediction probability vector of each single group is compared, that is, The rel can characterize the degree of association between each group of omics and the final prediction results of the phenotype, and the smaller the distance, the higher the degree of association is considered.

Images