A method, device and medium for analyzing the taste structure of compounds in food based on combined dimension reduction of UMAP and t-SNE

Abstract

The application discloses a food compound taste structure analysis method and device based on UMAP and t-SNE joint dimension reduction and a medium, and relates to the technical field of chemical data analysis. The method comprises the following steps: collecting compounds in food and corresponding taste descriptions, and constructing a compound-taste dataset; encoding the compounds to obtain high-dimensional molecular fingerprint data; first, using the UMAP algorithm to reduce the data to 10 dimensions to retain the core structure, and then using the t-SNE algorithm to further reduce to 2 dimensions to realize effective visualization and obtain a dimension reduction image; calculating the optimal cluster number, using the K-means clustering algorithm to perform clustering analysis on the dimension reduction image, and obtaining a clustering result; calculating the distribution and proportion probability of the compound taste and the molecular structure in each cluster, and constructing the correlation between the compound taste and the structure according to the distribution and proportion probability. The application can deeply analyze the taste structure relationship of the compounds in food, and provides a new prediction direction for food flavor research.

Description

Technical Field

[0001] This invention relates to the field of chemical data analysis technology, and in particular to a method, equipment, and medium for analyzing the taste structure of compounds in food based on UMAP and t-SNE combined dimensionality reduction. Background Technology

[0002] Food flavor is one of the key indicators for measuring food quality. Among these, taste is a core factor influencing consumer acceptance and product quality. Current research on food flavor largely focuses on volatile aroma components, while systematic analysis of taste compounds remains relatively weak. Taste perception involves various physicochemical properties, such as sweetness, bitterness, umami, astringency, and spiciness, which are typically caused by the interaction of compounds with oral receptors or mucous membranes. However, due to the complexity and multidimensionality of taste perception, identifying key taste substances from a large number of compounds and understanding their relationship with molecular structure still faces many challenges.

[0003] Existing technologies for analyzing taste structure mainly suffer from the following problems: First, taste-related compound data are usually high-dimensional and non-linear, making it difficult for traditional statistical methods to effectively capture their intrinsic structure; second, for certain flavor substances, especially compounds with low flavor thresholds and low content, it is difficult to fully identify them by instrument detection alone, often requiring expert sensory evaluation and experience-based inference, which lacks systematicity and predictability; in addition, the structure-activity relationship between taste and molecular structure is still unclear, making it difficult to achieve structure-based taste-oriented prediction and design.

[0004] To address these challenges, dimensionality reduction techniques have been introduced to process high-dimensional taste data and reveal its underlying chemical structure. Existing dimensionality reduction methods are mainly divided into two categories: linear and nonlinear. Linear dimensionality reduction methods, such as Principal Component Analysis (PCA), are computationally efficient but struggle to effectively capture the prevalent nonlinear relationships in taste data. Nonlinear dimensionality reduction methods, such as t-distributed random neighborhood embedding (t-SNE), can better preserve local data structure, but have limitations in global structure preservation and computational efficiency, especially for large-scale datasets, where computational costs are high and stable convergence is difficult. Unified Manifold Approximation and Projection (UMAP), as a relatively new nonlinear dimensionality reduction algorithm, not only effectively preserves the local and global structure of the data but also boasts higher computational efficiency, making it suitable for larger datasets.

[0005] Based on this, this invention proposes a joint dimensionality reduction approach combining UMAP and t-SNE. This leverages the advantages of UMAP in preserving overall structure and computational speed, while utilizing the fine-grained characterization of local structures by t-SNE. This allows for a more comprehensive and accurate reflection of the underlying structure of flavor compounds within the low-dimensional embeddings of their data. This method helps to systematically analyze the distribution patterns of flavor characteristics in food compounds, providing new technical means for predicting and inferring the structure of compounds in unknown flavor regions, and promoting the transformation of food flavor science from empirical description to computationally and predictably achievable goals. Summary of the Invention

[0006] The purpose of this invention is to provide a method, device, and medium for analyzing the taste structure of compounds in food based on dimensionality reduction using a combination of UMAP and t-SNE, aiming to solve or improve at least one of the aforementioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following solution: A method for analyzing the taste structure of food compounds based on dimensionality reduction using a combination of UMAP and t-SNE includes: Collect compounds in food and their corresponding taste descriptions, and construct a compound-taste dataset; The compound is encoded to obtain molecular fingerprint data with a high-dimensional vector; the molecular fingerprint data is expressed in binary form. A staged dimensionality reduction approach was adopted using UMAP and t-SNE algorithms: First, the high-dimensional molecular fingerprint data was reduced to 10 dimensions using the UMAP algorithm. This step aims to efficiently capture and preserve the key features of the data in the global topological structure. Then, the obtained 10-dimensional data was used as input, and the t-SNE algorithm was used to further reduce it to 2 dimensions. This step aims to finely optimize the local proximity relationships between data points, thereby generating a two-dimensional image for visualization. The optimal number of clusters is calculated, and the K-means clustering algorithm is used to perform cluster visualization processing on the dimensionality-reduced image to obtain the clustering results; the clustering results include each type of cluster and its corresponding classification data; Based on the clustering results and the compound-taste dataset, the distribution and proportion of the taste and molecular structure of compounds in each cluster are calculated probabilistically, and the correlation between the taste and structure of compounds is constructed.

[0008] Optionally, the staged dimensionality reduction using UMAP and t-SNE algorithms includes the following specific steps: The UMAP dimensionality reduction algorithm was used to perform dimensionality reduction on each data point X in the molecular fingerprint data. i Perform nearest neighbor calculations and calculate the nearest neighbor based on the data point and each nearest neighbor point X. j The distance between them dEuclidean(x) ix j Determine the local neighborhood; Based on the topological clustering method, a high-dimensional topological structure is constructed using data points as nodes and the similarity between the local neighborhoods as edges; the similarity between the local neighborhoods is determined by Jaccard similarity calculation. Based on the high-dimensional topology, the similarity between each local neighborhood is converted into weights to construct a high-dimensional Fuzzy network. Based on this network, the high-dimensional structure of the data points is mapped to a 10-dimensional space to complete the first stage of dimensionality reduction. Based on the aforementioned 10-dimensional space, calculate X for every two data points. i With X j The conditional probability between them, to characterize the data point X in high-dimensional space. j As data point X i The conditional probability of the nearest neighbor; In the target 2D space, initialize the layout of data points and calculate the similarity between data points. Use the t-distribution to describe the joint probability distribution of data points in the low-dimensional space. By minimizing the Kullback-Leibler divergence between the probability distributions in the high-dimensional and low-dimensional spaces, and using the gradient descent algorithm to iteratively optimize the location of data points in the low-dimensional space, the data is finally mapped to a 2D space to obtain a dimensionality-reduced image for visualization.

[0009] Optionally, the nearest neighbor calculation uses Euclidean distance and Manhattan distance.

[0010] Optionally, the formula for calculating the Jaccard similarity is: J i，j = Among them, J i，j Indicates Jaccard similarity; N i and N j They represent data points x respectively i and x j The local neighborhood, N represents i and N j The intersection, N represents i and N j The union of .

[0011] Optionally, the process of calculating the optimal number of clusters and using the K-means clustering algorithm to perform cluster visualization processing on the dimensionality-reduced image to obtain the clustering results includes: Calculate the optimal number of clusters K, and randomly select K data points as initial centers; Each data point is assigned to the nearest cluster center; Calculate new cluster centers based on data points within the cluster; Repeatedly assign data points and update cluster centers until the cluster centers no longer change or the maximum number of iterations is reached, thus completing the iterative optimization and obtaining the cluster to which each data point belongs and the center position of each cluster.

[0012] Optionally, the probability calculation results include: the probability of different functional groups and tastes appearing in each cluster, the probability of functional groups carrying different tastes, and the probability of tastes being generated by different functional groups.

[0013] The present invention provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described method for analyzing the taste structure of compounds in food based on the combined dimensionality reduction of UMAP and t-SNE.

[0014] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method described above for analyzing the taste structure of compounds in food based on joint dimensionality reduction using UMAP and t-SNE.

[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention discloses a method, device, and medium for analyzing the taste and structure of compounds in food based on dimensionality reduction using a combination of UMAP and t-SNE. The method includes collecting compounds in food and their corresponding taste descriptions, constructing a compound-taste dataset; encoding the compounds to obtain molecular fingerprint data with high-dimensional vectors; first, using the UMAP algorithm to reduce the data to 10 dimensions to preserve the core structure, then using the t-SNE algorithm to further reduce it to 2 dimensions for effective visualization, obtaining a dimensionality-reduced image; calculating the optimal number of clusters, and using the K-means clustering algorithm to perform cluster analysis on the dimensionality-reduced image to obtain clustering results; calculating the distribution and probability of compound taste and molecular structure in each cluster, and constructing the correlation between compound taste and structure. This invention can deeply analyze the taste and structure relationships of compounds in food, providing a new predictive direction for food flavor research. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1This is a flowchart illustrating the method for analyzing the taste structure of compounds in food in this embodiment; Figure 2 This is a data illustration of the calculation of the optimal number of clusters using the Kelly penalty function in this embodiment; Figure 3 This is a visualization of the data dimensionality reduction and clustering results in this embodiment; Figure 4 This is a texture-structure relationship diagram in the implementation of this embodiment. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The purpose of this invention is to provide a method, device, and medium for analyzing the taste structure of compounds in food based on UMAP and t-SNE combined dimensionality reduction, aiming to solve or improve at least one of the above-mentioned technical problems.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0021] like Figure 1 As shown, this invention provides a method for analyzing the taste structure of compounds in food based on joint dimensionality reduction using UMAP and t-SNE, including: Collect compounds in food and their corresponding taste descriptions, and construct a compound-taste dataset; The compound is encoded to obtain molecular fingerprint data with a high-dimensional vector; the molecular fingerprint data is expressed in binary form. A staged dimensionality reduction approach was adopted using UMAP and t-SNE algorithms: First, the high-dimensional molecular fingerprint data was reduced to 10 dimensions using the UMAP algorithm. This step aims to efficiently capture and preserve the key features of the data in the global topological structure. Then, the obtained 10-dimensional data was used as input, and the t-SNE algorithm was used to further reduce it to 2 dimensions. This step aims to finely optimize the local proximity relationships between data points, thereby generating a two-dimensional image for visualization. The optimal number of clusters is calculated, and the K-means clustering algorithm is used to perform cluster visualization processing on the dimensionality-reduced image to obtain the clustering results; the clustering results include each type of cluster and its corresponding classification data; Based on the clustering results and the compound-taste dataset, the distribution and proportion of the taste and molecular structure of compounds in each cluster are calculated probabilistically, and the correlation between the taste and structure of compounds is constructed.

[0022] As one specific implementation method, the following detailed implementation process is provided.

[0023] Step 1: Collect and filter compounds and their taste descriptions from food products to construct a "Taste-Compound" dataset, then proceed to Step 2. The "Taste-Compound" dataset in Step 1 mainly consists of compounds and their taste descriptors from existing databases and compounds and their taste descriptions reported in food-related literature. Furthermore, statistical analysis shows that a single compound in the dataset may have several taste characteristics, and the same taste characteristic may be generated by different compounds.

[0024] Step 2: Encode the compound into a computer-recognizable binary representation of a molecular fingerprint to obtain a high-dimensional vector, then proceed to Step 3. The computer-recognizable representation mentioned in Step 2 is obtained by encoding the compound into a binary representation of a molecular fingerprint using KNIME software. This invention uses the most commonly used molecular fingerprint: extended connection fingerprint (also called Morgan fingerprint), with a radius of 2 and a length of 1024 bits.

[0025] Step 3: Staged dimensionality reduction using UMAP and t-SNE algorithms: First, the high-dimensional molecular fingerprint data is reduced to 10 dimensions using the UMAP algorithm. This step aims to efficiently capture and preserve key features of the data in the global topological structure. Then, using the resulting 10-dimensional data as input, the t-SNE algorithm is used to further reduce it to 2 dimensions. This step aims to finely optimize the local proximity relationships between data points, thereby generating a 2D image for visualization, before proceeding to Step 4. In this embodiment, the dimensionality reduction algorithm described in Step 3 combines the currently fastest and best-preserving global structure-based Unified Manifold Approximation and Projection Dimensionality Reduction method, UMAP, with the best local structure characterization method, t-SNE.

[0026] The calculation steps of the combined UMAP and t-SNE algorithm in this embodiment mainly include the following parts: ① Local Neighborhood Construction. UMAP first constructs a local neighborhood for each data point X based on the given dataset. i Calculate its relationship with its nearest neighbor X. j The distance between them dEuclidean(x) i ,x jTypically, Euclidean distance or Manhattan distance is used to determine the k nearest neighbors of each data point based on a preset parameter k. More preferably, this invention uses Euclidean distance to calculate the distance between data points, and the calculation formula is as follows: dEuclidean(x i ,x j )= Where, x i and x j These represent two data points; n is the dimension of the data point; x i,k and x j,k They represent data points x respectively i and x j The coordinate value in the k-th dimension. More preferably, in this invention, the preset parameter k=15.

[0027] ② Topology Construction. UMAP uses a method called "topology clustering," which constructs the topology between data points based on the similarity in local neighborhoods. This process can be understood as building a connected graph in the data space, where data points represent nodes and the similarity between local neighborhoods represents edges. More preferably, this embodiment uses Jaccard similarity to measure the similarity between local neighborhoods, with the following formula: J i，j = J i，j Indicates Jaccard similarity; N i and N j They represent data points x respectively i and x j The local neighborhood, N represents i and N j The size of the intersection, N represents i and N j The size of the union of the values. The Jaccard similarity calculated in this way characterizes the data point x. i and x j The degree of overlap between neighborhoods. In UMAP, Jaccard similarity is usually used as a weight to construct the weight matrix W, and then to construct the topology of the local neighborhood.

[0028] ③ Fuzzy Network Embedding. Using information from the topology, a representation called a "fuzzy network" (fuzzysimplicial set) is constructed, where the similarity between each data point and its neighbors is converted into weights. Based on the fuzzy network, the high-dimensional structure of the data points is mapped to a low-dimensional space. This process aims to preserve the local relationships between data points in the high-dimensional space and retain these local relationships as much as possible in the low-dimensional space. This process involves three key hyperparameters: n_neighbors (the number of neighbors for each data point), min_dist (the minimum distance between points in the embedding space), and n_components (the dimension of the embedding space, i.e., the dimension after dimensionality reduction). In this invention, n_neighbors = 15, min_dist = 0.1, and n_components = 10.

[0029] ④ Probability Distribution Construction. The 10-dimensional data output from UMAP is used as input to t-SNE. In the 10-dimensional space, the similarity probability between data points is calculated: Where σ i The bandwidth parameter is adaptively determined through a preset perplexity parameter. More preferably, in this invention, the preset parameter perplexity is 30.

[0030] ⑤ Two-dimensional space optimization embedding. The 10-dimensional data is further reduced to 2-dimensionality using the t-SNE algorithm. In the 2-dimensional embedding space, the t-distribution is used to define the similarity between data points: KL divergence is used to measure the difference between high-dimensional and low-dimensional spatial distributions: P and Q are two probability distributions used to represent the distribution of data points in high-dimensional and low-dimensional spaces.

[0031] Step 4: Calculate the optimal number of clusters. Apply the K-means clustering algorithm to cluster the dimensionality-reduced data and visualize it to obtain the corresponding clusters, which serve as the classification of all compounds. Then proceed to Step 5. The optimal number of clusters mentioned in Step 4 is calculated using the Kelly penalty function. The cluster number that minimizes the Kelly penalty function is selected as the optimal number of clusters. The calculation formula is as follows: K(k) = - Where K(k) is the Kelly penalty function, representing the penalty value when the number of clusters is k; λ is the probability of observing data given the number of clusters k; λ is the penalty coefficient, used to balance the penalty term and the model fit. Figure 2 As shown in the diagram, the optimal number of clusters in this embodiment is 5.

[0032] Furthermore, in the K-means clustering mentioned in step 4, K data points are selected as initial cluster centers based on the calculation results, and then each data point is assigned to the nearest cluster center. Next, the center point of each cluster is recalculated, and the assignment of data points within each cluster is updated. This process iterates repeatedly until the cluster centers no longer change significantly or the predetermined maximum number of iterations is reached, finally outputting the cluster and cluster center to which each data point belongs. In this embodiment, Euclidean distance is used to calculate the distance between two given points A(x1,y1) and B(x2,y2), and the calculation formula is as follows: Among them, (x) i ,y i () are the coordinates of points A and B in two-dimensional space.

[0033] Step 5: Analyze the clustering results, such as... Figure 3 As shown. The distribution and proportion of compound taste and structure in each cluster are calculated separately, and the two are correlated before proceeding to step 6. Specifically, the distribution and proportion of compound taste in each cluster mentioned in step 5 are calculated using the following formula: A (Distribution of texture in each cluster) = Frequency of this texture in this cluster / Total frequency of this texture in the dataset B (the proportion of each texture in each cluster) = frequency of occurrence of that texture in that cluster / total number of compounds in that cluster In this embodiment, the structure of the compound was calculated using KNIME software to obtain the functional groups contained in the compound. Furthermore, the distribution ratio of these functional groups in each cluster was calculated using the following formula: C (distribution of functional groups in each cluster) = frequency of the functional group in that cluster / total frequency of the functional group in the dataset D (the proportion of functional groups in each cluster) = frequency of the functional group in the cluster / total number of compounds in the cluster Step 6: Analyze the corresponding relationship between compound taste and the distribution and proportion of functional groups in each cluster. Combine this with the probability of each functional group producing a particular taste, and the probability of each functional group producing that taste, to establish a correlation between compound taste and structure. The correlation results are as follows: Figure 4 As shown, the thickness of the lines represents the degree of correlation; the thicker the lines, the greater the correlation, and the thinner the lines, the smaller the correlation. This aims to provide predictive direction for the study of the taste structure of baijiu.

[0034] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0035] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for analyzing the taste structure of compounds in food based on combined dimensionality reduction using UMAP and t-SNE, characterized in that, include: Collect compounds in food and their corresponding taste descriptions, and construct a compound-taste dataset; The compound is encoded to obtain molecular fingerprint data with a high-dimensional vector; The molecular fingerprint data is expressed in binary form; A staged dimensionality reduction approach was adopted using UMAP and t-SNE algorithms: First, the high-dimensional molecular fingerprint data was reduced to 10 dimensions using the UMAP algorithm, capturing and preserving the key features of the data in the global topological structure; then, using the obtained 10-dimensional data as input, the t-SNE algorithm was used to reduce the 10-dimensional data to 2 dimensions, optimizing the local proximity relationship between data points, thereby generating a dimensionality-reduced image for visualization. Calculate the optimal number of clusters, and use the K-means clustering algorithm to perform cluster visualization processing on the dimensionality-reduced image to obtain the clustering results; The clustering results include various clusters and their corresponding classification data; Based on the clustering results and the compound-taste dataset, the distribution and proportion of compound taste and molecular structure in each cluster are calculated probabilistically, and the correlation between compound taste and structure is constructed. The process of generating the dimensionality-reduced image includes: The first phase involves reducing high-dimensional data to 10 dimensions based on UMAP: The UMAP dimensionality reduction algorithm was used to perform dimensionality reduction on each data point X in the molecular fingerprint data. i Perform nearest neighbor calculations and calculate the nearest neighbor based on the data point and each nearest neighbor point X. j The distance between them dEuclidean(x) i x j Determine the local neighborhood; distance dEuclidean(x) i x j The formula for calculating ) is: dEuclidean(x i ,x j )= Where, x i and x j These represent two data points respectively; It is the dimension of the data points; and They represent data points x respectively i and x j In the Dimensional coordinates; Based on the topological clustering method, a high-dimensional topological structure is constructed using data points as nodes and the similarity between the local neighborhoods as edges. The similarity between the local neighborhoods is determined by Jaccard similarity calculation. The formula for calculating Jaccard similarity is: J i，j = Among them, J i，j Indicates Jaccard similarity; and They represent data points x respectively i and x j The local neighborhood, express and The intersection, express and The union of; Based on the high-dimensional topology, the similarity between each local neighborhood is converted into weights to construct a high-dimensional Fuzzy network. Based on the high-dimensional Fuzzy network, the high-dimensional structure of the data points is mapped to a 10-dimensional space to complete the first stage of dimensionality reduction. In the second stage, the 10-dimensional data is further reduced to 2 dimensions based on t-SNE: Based on 10-dimensional space, calculate X for every two data points. i With X j The conditional probability between them, to characterize the data point X in high-dimensional space. j As data point X i The conditional probability of the nearest neighbor; In the target 2D space, initialize the layout of data points and calculate the similarity between data points. Use the t-distribution to describe the joint probability distribution of data points in the low-dimensional space. By minimizing the Kullback-Leibler divergence between the probability distributions in the high-dimensional space and the low-dimensional space, and using the gradient descent algorithm to iteratively optimize the location of data points in the low-dimensional space, the data is finally mapped to a 2-dimensional space to obtain a dimensionality-reduced image for visualization. In the first stage of UMAP dimensionality reduction, nearest neighbor calculation uses Euclidean distance and Manhattan distance; in the second stage of t-SNE dimensionality reduction, the conditional probability is calculated based on Euclidean distance and Manhattan distance or a standardized cosine distance. In the second stage, the low-dimensional mapping is optimized by minimizing the Kullback-Leibler divergence between the probability distributions of the high-dimensional space and the low-dimensional space; The process of calculating the optimal number of clusters, using the K-means clustering algorithm to perform cluster visualization processing on the dimensionality-reduced image, and obtaining the clustering results includes: Calculate the optimal number of clusters K, and randomly select K data points as initial centers; Each data point is assigned to the nearest cluster center; Calculate new cluster centers based on data points within the cluster; Repeatedly assign data points and update cluster centers until the cluster centers no longer change or the maximum number of iterations is reached, thus completing the iterative optimization and obtaining the cluster to which each data point belongs and the center position of each cluster.

2. The method for analyzing the taste structure of compounds in food based on UMAP and t-SNE combined dimensionality reduction according to claim 1, characterized in that, The probability calculation results include: the probability of different functional groups and textures appearing in each cluster, the probability of functional groups carrying different textures, and the probability of textures being generated by different functional groups.

3. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program and the processor runs the computer program to enable the electronic device to perform the method for analyzing the taste structure of compounds in food based on the combined dimensionality reduction of UMAP and t-SNE according to any one of claims 1-2.

4. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the method for analyzing the taste structure of compounds in food based on joint dimensionality reduction of UMAP and t-SNE as described in any one of claims 1-2.

Images