Biological target discovery method based on MOE strategy

Abstract

The present invention relates to a biological target discovery method based on MOE strategy, comprising: obtaining input data; assigning corresponding expert networks to different parts of the input data through a multi-expert allocator; all expert networks in an expert mixture model process the data parts assigned to them in parallel and output high-dimensional features; discovering potential drug targets based on the high-dimensional features; and perturbing the output of the expert network through a sensitivity analysis method to evaluate the contribution of each expert network to the final prediction result. The beneficial effect of the present invention is that the present invention dynamically assigns tasks to the most suitable expert network by introducing an expert mixture model, thereby improving the processing efficiency of the model and the accuracy of the results. In addition, the present invention uses the Gumbel operator to assign appropriate experts to different parts of the input data to ensure that each expert focuses on specific data characteristics, thereby improving the performance of the entire model.

Description

Technical Field

[0001] The present invention relates to the technical field of bioinformatics, and more specifically, to a biological target discovery method based on MOE strategy. Background Art

[0002] Accurate identification of biological targets is crucial in drug development and precision medicine. With the rapid accumulation of biological data, researchers have gradually relied on machine learning and deep learning techniques to process complex biological data. However, the diversity and complexity of biological data pose great challenges to existing methods, especially when processing large-scale, multimodal data. Existing methods are easily affected by noise and cannot fully explore the potential structure of the data, resulting in limited processing efficiency and result accuracy.

[0003] In addition, due to the "black box" characteristics of deep learning, the model has poor interpretability, making it difficult to provide a clear explanation for the biological target discovery process. In drug development, the interpretability of the model is particularly important because it not only determines the credibility of the results, but also affects the interpretation of biological significance. Therefore, existing technologies have significant deficiencies in processing complex data and providing interpretability. There is an urgent need for a biological target discovery method based on a mixture of experts (MOE) strategy that can balance accuracy and interpretability. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and propose a biological target discovery method based on the MOE strategy.

[0005] In the first aspect, a biological target discovery method based on the MOE strategy is provided, comprising:

[0006] Step 1: Get input data;

[0007] Step 2: Assign corresponding expert networks to different parts of the input data through a multi-expert allocator;

[0008] Step 3: All expert networks in the expert mixture model process their assigned data in parallel and output high-dimensional features.

[0009] Step 4: Discover potential drug targets based on the high-dimensional features;

[0010] Step 5: Perturb the output of the expert network through sensitivity analysis method to evaluate the contribution of each expert network to the final prediction result.

[0011] As a preference, it also includes:

[0012] Step 6: Generate a visualization chart, which is used to represent the role of each expert network in the entire biological target discovery process.

[0013] Preferably, step 2 comprises:

[0014] Step 2.1, extract the feature vector of the input data;

[0015] Step 2.2: According to the feature vector, use the Gumbel operator to assign corresponding expert networks to different parts of the input data.

[0016] Preferably, step 5 comprises:

[0017] Step 5.1: Make a small perturbation to the output of the expert network and calculate the sensitivity of the expert network to the final prediction result;

[0018] Step 5.2: Determine the contribution of each expert network to the final prediction result based on the calculated result of the sensitivity of the expert network to the final prediction result.

[0019] Preferably, in step 6, the expert network includes a feature contribution heat map and a task allocation map among experts.

[0020] In a second aspect, a biological target discovery system based on the MOE strategy is provided, which is used to execute any of the methods described in the first aspect, including:

[0021] An acquisition module is used to obtain input data;

[0022] An allocation module, used for allocating corresponding expert networks to different parts of the input data through a multi-expert allocator;

[0023] A processing module, which is used by all expert networks in the expert mixture model to process their assigned data parts in parallel and output high-dimensional features;

[0024] A discovery module, used to discover potential drug targets based on the high-dimensional features;

[0025] The evaluation module is used to perturb the output of the expert network through sensitivity analysis methods and evaluate the contribution of each expert network to the final prediction result.

[0026] According to a third aspect, a computer storage medium is provided, wherein a computer program is stored in the computer storage medium; when the computer program is executed on a computer, the computer executes any method described in the first aspect.

[0027] In a fourth aspect, an electronic device is provided, including:

[0028] Memory, used to store computer programs;

[0029] A processor is used to execute the computer program to implement any method as described in the first aspect.

[0030] The beneficial effects of the present invention are:

[0031] 1. The present invention introduces an expert mixture model to dynamically assign tasks to the most suitable expert network, thereby improving the processing efficiency of the model and the accuracy of the results.

[0032] 2. The present invention uses the Gumbel operator to assign appropriate experts to different parts of the input data to ensure that each expert focuses on specific data characteristics, thereby improving the performance of the entire model.

[0033] 3. The present invention perturbs the expert output through sensitivity analysis method, evaluates the contribution of each expert to the final prediction result, and enhances the interpretability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A flowchart of a biological target discovery method based on MOE strategy;

[0035] Figure 2 A schematic diagram of a multi-expert allocator for a biological target discovery method based on the MOE strategy;

[0036] Figure 3 A schematic diagram of the MOE network backbone of a biological target discovery method based on the MOE strategy;

[0037] Figure 4 Schematic diagram of the structure of a biological target discovery system based on the MOE strategy. DETAILED DESCRIPTION

[0038] The present invention is further described below in conjunction with embodiments. The description of the following embodiments is only used to help understand the present invention. It should be noted that for ordinary persons in the art, without departing from the principle of the present invention, the present invention can also be modified in some ways, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

[0039] Embodiment 1:

[0040] In order to solve the problems of efficiency, accuracy and interpretability of existing methods in biological target discovery, such as Figure 1 As shown, the present invention provides a biological target discovery method based on the MOE strategy, which is performed by a multi-expert allocator, a MOE network and an interpretation module, including:

[0041] Step 1: Get input data.

[0042] In step 1, the input data includes RNA sequencing data of biological samples, single-cell expression data, single-cell transcriptome data, etc.

[0043] Step 2: Use a multi-expert allocator to assign corresponding expert networks to different parts of the input data.

[0044] like Figure 2 As shown, in step 2, the multi-expert allocator is used to use the Gumbel operator to allocate appropriate expert networks to different parts of the input data to ensure that each expert focuses on specific data features, thereby improving the performance of the entire model.

[0045] The multi-expert allocator includes a feature analysis unit and a Gumbel operator task allocation unit. The feature analysis unit is used to extract the feature vector of each sample, and the Gumbel operator task allocation unit is used to allocate appropriate expert networks to tasks of different feature parts of the feature vector.

[0046] Specifically, step 2 includes:

[0047] Step 2.1: Extract the feature vector of the input data through the feature analysis unit.

[0048] The input data is recorded as ,in, is the number of samples, is the feature dimension. For each sample, extract its feature vector , the feature analysis process can be expressed as:

[0049]

[0050] in, Indicates The feature vector of the samples.

[0051] Step 2.2: According to the feature vector, use the Gumbel operator to assign corresponding expert networks to different parts of the input data.

[0052] Specifically, for the feature vector , each expert network ,in is the number of expert networks, sampling independent noise values ​​from the Gumbel distribution , add these noise values ​​to the feature vector and calculate the temperature parameter Scaling is performed to obtain the adjusted probability, which is expressed as:

[0053]

[0054] in, For the feature vector , No. The probability that an expert selects the jth feature.

[0055] In order to determine the number of features assigned to each expert, this application introduces a hyperparameter , specifies the number of features each expert network will receive. The value of can be selected based on the properties of the data or optimized through cross-validation. , mask element is defined as:

[0056]

[0057] in, Represents a function that selects the indices of the top K features based on the adjusted probabilities. Mask matrix , where each Be an expert A binary mask vector of , indicating which vectors are selected. By entering data in and Hadamard product (element-wise multiplication) is applied between them to perform feature assignment. This operation is performed for each expert network Generate input data:

[0058]

[0059] in," ” denotes the Hadamard product.

[0060] Step 3: All expert networks in the expert mixture model process their assigned data in parallel and output high-dimensional features.

[0061] The MOE network backbone introduces an expert mixture model, which contains multiple expert networks. All expert networks process their assigned data in parallel, and the multi-expert allocator dynamically assigns tasks to the most suitable expert network, thereby improving the processing efficiency of the model and the accuracy of the results.

[0062] like Figure 3 As shown, the MOE network backbone is implemented in the following ways:

[0063] Expert Network The forward propagation process is:

[0064]

[0065] in, is the data allocated by the multi-expert allocator, It is the high-dimensional feature output by the expert network.

[0066] All expert networks process their assigned data in parallel, and the resulting feature matrix is:

[0067]

[0068] in, Represents a join operation.

[0069] Each expert network is implemented through multiple layers of nonlinear transformation functions:

[0070]

[0071] in, and are network parameters, is the activation function.

[0072] Through the loss function Fine-tune the expert network to ensure optimal performance under different task objectives. The process of updating parameters by gradient descent can be expressed as:

[0073]

[0074] in is the learning rate.

[0075] Step 4: Discover potential drug targets based on the high-dimensional features. Input to Classify in the classifier and get the classification result . Classification results Used to determine whether it is a drug target. If it is 0, it means it is not a drug target. If it is 1, it indicates a drug target.

[0076] Step 5: Perturb the output of the expert network through sensitivity analysis method to evaluate the contribution of each expert network to the final prediction result.

[0077] In step 5, the contribution of each expert to the final prediction result is evaluated through the explanation module to enhance the interpretability of the model; the explanation module includes a sensitivity analysis unit and an expert evaluation unit.

[0078] Step 5 includes:

[0079] Step 5.1: The sensitivity analysis unit makes a small perturbation to the output of the expert network and calculates the sensitivity of the expert network to the final prediction result.

[0080] Specifically, for On samples, the expert network The output is , making a small perturbation , and observe the effect of the disturbance on the final prediction result The sensitivity calculation formula is:

[0081]

[0082] in, Indicates expert Sensitivity to prediction results.

[0083] Step 5.2: According to the calculation result of the sensitivity of the expert network to the final prediction result, the contribution of each expert network to the final prediction result is determined by the expert evaluation unit.

[0084] Specifically, the calculation formula of step 5.2 is:

[0085]

[0086] in, Indicates Expert The average contribution to the result, Indicates Samples in Expert The sensitivity value on .

[0087] Embodiment 2:

[0088] Based on Example 1, Example 2 of the present application provides a more specific biological target discovery method based on the MOE strategy, including:

[0089] Step 1: Get input data;

[0090] Step 2: Use a multi-expert allocator to assign corresponding expert networks to different parts of the input data.

[0091] Step 3: All expert networks in the expert mixture model process their assigned data in parallel and output high-dimensional features.

[0092] Step 4: Discover potential drug targets based on the high-dimensional features.

[0093] Step 5: Perturb the output of the expert network through sensitivity analysis method to evaluate the contribution of each expert network to the final prediction result.

[0094] Step 6: Generate a visualization chart, which is used to represent the role of each expert network in the entire biological target discovery process.

[0095] The interpretation module may also include a visualization generation unit. The visualization generation unit is used to generate a series of visualization charts, such as feature contribution heat maps and task allocation charts among experts, to help users understand the role of each expert in the entire biological target discovery process.

[0096] Through the above steps, the present invention realizes efficient discovery and accurate analysis of biological targets, and provides a reliable solution suitable for drug development and biomarker research. For example, the input protein data set can be feature extracted through the MOE model. First, the input data is divided into several parts and dynamically assigned to different expert networks through a multi-expert allocator. Each expert network focuses on different data characteristics, such as amino acid sequence features, spatial structure features, or functional domain information. Each expert network uses a deep neural network to extract features from the data portion assigned to it and outputs a high-dimensional feature representation.

[0097] Next, the output of each expert network is interpreted through sensitivity analysis methods to analyze the contribution of each expert to the final prediction results. Specifically, a small perturbation is made to the output of each expert network to observe the impact on the final prediction results and calculate the sensitivity index. The results of these sensitivity analyses can be presented in graphical form, such as heat maps, to intuitively evaluate the role of each expert in the entire target discovery process.

[0098] When identifying potential drug targets, the output of the expert network is combined to discover key biological targets by analyzing the binding sites or functional regions of specific genes. For example, a specific functional region of some proteins may be marked as a high-contribution region in multiple expert networks, indicating that the functional region may be a potential drug target. The rationality is verified by sensitivity analysis charts, and these regions are further experimentally verified to determine their potential as drug targets. Ultimately, this MOE-based approach can help effectively identify and verify biological targets and improve the efficiency and accuracy of drug development.

[0099] In addition, this tool can also be used to infer cell differentiation trajectories. For example, single-cell RNA sequencing data can be used as input to perform feature analysis on different cell types and differentiation states through a multi-expert allocator. Each expert network can focus on different gene expression patterns or cell features to extract high-dimensional feature representations of specific differentiation stages. Then, the contribution of different experts to the inference of differentiation trajectories is evaluated by sensitivity analysis methods to identify key genes or pathways in the differentiation process. Combined with the output of the expert network, a cell differentiation tree can be constructed to clarify the transition path of cells from progenitor cells to each terminal cell type, helping to reveal the underlying mechanism of cell differentiation. In this way, not only can the accuracy of cell differentiation trajectory inference be improved, but also the interpretability of the results can be enhanced, thereby providing new tools and perspectives for cell biology research.

[0100] Based on the above embodiments, the present application has been experimentally verified as follows:

[0101] Dataset:

[0102] (1) PBMC 10k (10x Genomics): A dataset of peripheral blood mononuclear cells (PBMCs) from healthy donors, containing RNA sequencing data of approximately 10,000 single cells.

[0103] (2) Tabula Muris (Mouse Cells): Single-cell transcriptomics data of mouse cells, covering a variety of different organs and tissues.

[0104] (3) Human Pancreas (Single-Cell Expression): Single-cell expression data of human pancreas, commonly used to study insulin secretion and pancreatic function.

[0105] (4) Mouse Brain (Smart-seq2): Single-cell transcriptome data of mouse brain generated using Smart-seq2 sequencing technology, suitable for neuroscience research.

[0106] (5) Human Liver (Single-Cell RNA-seq): Single-cell RNA sequencing data of human liver, used to study liver disease and liver function

[0107] As shown in Table 1, the results of the method of the present invention on the above four data sets are compared with common methods such as LogisticRegression, Random Forest, SVM, and XGBoost to show the performance differences on different data sets. All experiments were run on a server with an NVIDIA A100 GPU to ensure the computational efficiency of large-scale data. Each model was run 5 independent experiments on each data set, the accuracy of each experiment was recorded, and the average and standard deviation were calculated.

[0108] Table 1

[0109]

[0110] As shown in the above results, the present invention has significant advantages over these common methods. In terms of performance, the accuracy of the present invention is improved by 6.6% on the PBMC 10k (10x Genomics) data set; 6.3% on the Tabula Muris (MouseCells) data set; 6.3% on the HumanPancreas (Single-Cell Expression) data set; 6.8% on the Mouse Brain (Smart-seq2) data set; and 6.3% on the Human Liver (Single-Cell RNA-seq) data set. In terms of stability, the method of the present invention shows lower variance on all test data sets compared with the comparative method. This shows that the method of the present invention is more robust.

[0111] It should be noted that the parts in this embodiment that are the same or similar to those in Embodiment 1 can be referenced to each other and will not be described in detail in this application.

[0112] Embodiment 3:

[0113] Based on Example 1, Example 3 of the present application provides a biological target discovery system based on the MOE strategy, such as Figure 4 As shown, including:

[0114] An acquisition module is used to obtain input data;

[0115] An allocation module, used for allocating corresponding expert networks to different parts of the input data through a multi-expert allocator;

[0116] A processing module, which is used by all expert networks in the expert mixture model to process their assigned data parts in parallel and output high-dimensional features;

[0117] A discovery module, used to discover potential drug targets based on the high-dimensional features;

[0118] The evaluation module is used to perturb the output of the expert network through sensitivity analysis methods and evaluate the contribution of each expert network to the final prediction result.

[0119] Specifically, the system provided in this embodiment is a system corresponding to the method provided in Embodiment 1. Therefore, the parts in this embodiment that are the same or similar to those in Embodiment 1 can be referenced to each other and will not be repeated in this application.

Claims

1. A biological target discovery method based on MOE strategy, characterized in that: include: Step 1: Get input data; Step 2: Assign corresponding expert networks to different parts of the input data through a multi-expert allocator; The multi-expert allocator includes a feature analysis unit and a Gumbel operator task allocation unit. The feature analysis unit is used to extract the feature vector of each sample, and the Gumbel operator task allocation unit is used to allocate appropriate expert networks to tasks of different feature parts of the feature vector. Step 2 includes: Step 2.1, extract the feature vector of the input data; Step 2.2, according to the feature vector, use the Gumbel operator to assign corresponding expert networks to different parts of the input data; Step 3: All expert networks in the expert mixture model process their assigned data in parallel and output high-dimensional features. Step 4: Discover potential drug targets based on the high-dimensional features; Step 5: Perturb the output of the expert network through sensitivity analysis method to evaluate the contribution of each expert network to the final prediction result.

2. The biological target discovery method based on MOE strategy according to claim 1, characterized in that: Also includes: Step 6: Generate a visualization chart, which is used to represent the role of each expert network in the entire biological target discovery process.

3. The biological target discovery method based on MOE strategy according to claim 2, characterized in that: Step 5 includes: Step 5.1: Make a small perturbation to the output of the expert network and calculate the sensitivity of the expert network to the final prediction result; Step 5.2: Determine the contribution of each expert network to the final prediction result based on the calculated result of the sensitivity of the expert network to the final prediction result.

4. The biological target discovery method based on MOE strategy according to claim 3, characterized in that: In step 6, the visualization chart includes a feature contribution heat map and a task allocation chart among experts.

5. A biological target discovery system based on MOE strategy, characterized in that: Used to perform the method according to any one of claims 1 to 4, comprising: An acquisition module is used to obtain input data; An allocation module, used for allocating corresponding expert networks to different parts of the input data through a multi-expert allocator; A processing module, which is used by all expert networks in the expert mixture model to process their assigned data parts in parallel and output high-dimensional features; A discovery module, used to discover potential drug targets based on the high-dimensional features; The evaluation module is used to perturb the output of the expert network through sensitivity analysis methods and evaluate the contribution of each expert network to the final prediction result.

6. A computer storage medium, characterized in that: The computer storage medium stores a computer program; when the computer program is executed on a computer, the computer executes any one of the methods described in claims 1 to 4.

7. An electronic device, characterized in that: include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 4.

Images