Method and system for subsampling cells from single-cell genomics datasets

Abstract

The present disclosure relates to a computer-implemented method for extracting a subsample of cells from a plurality of cells in a single-cell genomics dataset, the method comprising: obtaining a single-cell genomics dataset represented in at least two dimensions, where information about each cell is represented in a first dimension and information about genetic features is represented in a second dimension; generating a cell-cell neighborhood graph from the single-cell genomics dataset, where the cell-cell neighborhood graph provides information about similarities in the genetic features of cells and where each cell is represented as a vertex in the cell-cell neighborhood graph; partitioning cells in the cell-cell neighborhood graph into seed cells and non-seed cells; assigning at least one first bounty to seed cells and at least one second bounty to non-seed cells in the cell-cell neighborhood graph; and traversing the cell-cell neighborhood graph using a prize-collecting Steiner tree algorithm to obtain a subsample of cells. The present disclosure further relates to a computer system for extracting a subsample of cells from a plurality of cells in a single-cell genomics dataset.

Description

Technical Field

[0001] The present disclosure relates to a computer-implemented method and system for extracting a subsample of cells from multiple cells in a single-cell genomics dataset. The method of the present disclosure shows an overview of a step-by-step process for subsampling cells from large-scale single-cell genomics datasets with high memory efficiency.

Background Art

[0002] Single-cell DNA genome sequencing involves separating single cells, amplifying the entire genome or regions of interest, constructing a sequencing library, and then applying next-generation DNA sequencing. In mammals, single-cell DNA sequencing has been widely applied to the study of normal physiology and diseases.

[0003] Single-cell genomics assays can be directed to determining the presence of base mutations in the genome, estimating the number of RNA molecules for each transcribed gene, identifying open regions of DNA in chromatin, identifying epigenetic marks on chromatin, and the like. Although there are numerous methods and protocols to achieve this, they all constantly generate sequenced fragments of multiple base sequences such that each sequence can be attributed to its original cell.

[0004] To obtain a count matrix, numerous computer-implemented methods have been developed to process and analyze such sequences from single-cell sequencing. A count matrix is typically a two-dimensional table where each row contains information about each cell and each column contains information about each genetic feature, or vice versa. An example of such a two-dimensional table is shown in FIG. 1. The quantified features can be, for example, the number of cleavage sites in each identified accessible region (in the case of single-cell ATAC sequencing), the number of complementary DNA (cDNA) fragments for each gene (in the case of single-cell RNA sequencing), and the like.

[0005] As is typical of single-cell datasets, they are highly redundant in the sense that cell clusters belong to the same cell type and are similar to one another. For common analytical purposes such as cell type identification and complex analysis, this redundancy in single-cell datasets leads to unnecessary data processing time and often requires large computing infrastructure. This issue becomes even more pronounced with large single-cell datasets, or when users attempt to analyze multiple single-cell datasets as a single unit.

[0006] As single-cell genomics moves toward generating data of previously unanalyzable sizes, reducing the size of single-cell datasets becomes necessary. A straightforward solution frequently applied to achieve this is to randomly sample cells from large, original single-cell datasets. However, this solution typically fails to represent the single-cell dataset manifold in an optimal way, resulting in suboptimal sampling results. Other existing methods for data sampling either do not select data in an optimal way and / or do not use computer memory efficiently in processing.

[0007] Therefore, there is a need for improved and efficient solutions to reduce the size of large single-cell genomics datasets while maintaining high computational efficiency and minimal information loss. [Overview of the project] [Problems that the invention aims to solve]

[0008] One object of this disclosure is to provide a technical implementation that enables downsampling of single-cell genomics datasets into smaller datasets by selecting subsamples of cells. Preferably, this solution can be fully integrated into readily available computing devices such as laptop computers. Another object of this disclosure is to perform data subsampling that preserves the original manifold of the original single-cell genomics dataset. [Means for solving the problem]

[0009] The method may be applied to a single-cell genomics dataset represented as a two-dimensional matrix containing quantified characteristics of each single cell assayed. As can be seen in the example in Figure 1, information about the cells (C1-Cn), such as identification such as a barcode, is given in the first dimension, while information about the genetic characteristics of each cell (F1-Fn) is represented in the second dimension. As will be recognized by those skilled in the art, genetic characteristics can be considered multidimensional in themselves, in the sense that each genetic characteristic is considered its own dimension. However, in the context of this disclosure, the terms “first” and “second” dimensions refer to the first and second dimensions of the matrix representing the single-cell genomics dataset.

[0010] A first aspect of this disclosure relates to a computer implementation method for extracting cell subsamples from multiple cells in a single-cell genomics dataset. The method is: - A step of obtaining a single-cell genomics dataset represented in at least two dimensions, wherein information about each cell is represented in the first dimension and information about genetic characteristics is represented in the second dimension, - A step of generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the aforementioned genetic features of cells, and cells are represented as vertices in the cell-cell neighborhood graph. -A step of dividing the cells in the cell-neighborhood graph into seed cells and non-seed cells, - In a cell-cell neighborhood graph, the steps include assigning at least one first prize to seed cells and at least one second prize to non-seed cells, - The method comprises the steps of traversing a cell-cell neighborhood graph using a prize-collecting Steiner tree algorithm to obtain a subsample of cells.

[0011] This single-cell genomics dataset may be obtained by a single-cell genomics sequencing process.

[0012] Implementing the steps described above has several benefits. It allows users with limited computing and memory resources to process the entire single-cell genomics dataset, then work with meaningful cell subsamples or advance the results for further applications. This dramatically reduces memory needs for large datasets while preserving the original diversity and data manifold. The resulting cell subsamples can acquire the full-spectrum information present in the original single-cell genomics dataset. This solution can be used and integrated into existing workflows or is available as a standalone module. "Manifold" is a term commonly understood by those skilled in the field of single-cell genomics and can be considered a high-dimensional surface. Numerous natural processes generate samples that lie on high-dimensional surfaces known as manifolds. While single-cell genomics datasets have hundreds to hundreds of thousands of features, the underlying biological processes of cellular states are simpler and can be represented in fewer dimensions. For this reason, manifold learning is popular in single-cell genomics, as cells can be conceptualized as existing on an unknown manifold that can be represented using fewer dimensions. The graphical representation of single-cell data is similar to a manifold, and sampling on this graph is similar to sampling on a data manifold.

[0013] Information about genetic features in the second of these dimensions may include quantified genetic features such as annotated genes, and / or gene locations in the coordinate form of the genome assembly, and / or the number of cleavage sites within the identified accessible region, and / or the number of cDNA fragments for each gene. Single-cell genomics datasets may be, but are not limited to, single-cell RNA sequencing datasets or single-cell ATAC sequencing datasets.

[0014] The steps of this method will be further explained and illustrated in the detailed description below. The step of generating a cell-cell neighborhood graph from a single-cell genomics dataset can be implemented using the k-nearest neighbor algorithm. The step of dividing the cells in the cell-cell neighborhood graph into seed cells and non-seed cells can be done by clustering and selection, but is not limited to these methods. The purpose of the step of dividing the cells in the cell-cell neighborhood graph into seed cells and non-seed cells is to prepare data for the steps of assigning prizes to the seed cells and non-seed cells in the cell-cell neighborhood graph and obtaining subsamples of cells across the cell-cell neighborhood graph using the Prize Collection Steiner Tree Algorithm (PCST).

[0015] In other embodiments, this disclosure is a computer system, Memory and It comprises a processing unit, and the processing unit is - A step of loading a single-cell genomics dataset, represented in at least two dimensions, into memory, wherein information about each cell is represented in the first dimension and information about genetic characteristics is represented in the second dimension. - A step of generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of cells, and cells are represented as vertices in the cell-cell neighborhood graph. -A step of dividing the cells in the cell-neighborhood graph into seed cells and non-seed cells, - In a cell-cell neighborhood graph, the steps include assigning at least one first prize to seed cells and at least one second prize to non-seed cells, - Using a prize-collecting Steiner tree algorithm, the steps involve traversing the cell-cell neighborhood graph to obtain a subsample of cells, Further relating to computer systems configured to carry out this.

[0016] The present invention further relates to a computer program which, when executed by a computing device or computing system, has instructions for causing the computing device or computing system to carry out any embodiment of the method of the present disclosure for extracting subsamples of cells from multiple cells in a single-cell genomics dataset. The term "computer program" in this context should be interpreted broadly and must include, for example, a computer program running on a PC, a single-board computer or system on a chip, or on a mobile phone or tablet, or a computer program adapted to run as part of a software platform. [Brief explanation of the drawing]

[0017] The present invention will be described below with reference to the accompanying drawings, which are illustrative and not limiting the methods and systems of this disclosure for extracting cell subsamples from multiple cells in a single-cell genomics dataset. [Figure 1] Figure 1 shows an example of a single-cell genomics dataset represented in two dimensions within a matrix. [Figure 2] Figure 2 shows a flowchart illustrating one embodiment of the method of this disclosure for extracting cell subsamples from multiple cells in a single-cell genomics dataset. [Figure 3]Figure 3 shows a schematic diagram according to an embodiment of the system of the present disclosure for extracting cell sub-samples from multiple cells in a single-cell genomics dataset. [Figure 4] Figure 4 shows an example of a cell-cell neighborhood graph according to the method of the present disclosure for extracting cell sub-samples from multiple cells in single-cell genomics. [Figure 5] Figure 5 shows an example of a cell-cell neighborhood graph, in which the cells are divided into seed cells and non-seed cells. [Figure 6] Figure 6 shows an example of a cell-cell neighborhood graph, in which cell sub-samples are extracted from multiple cells in a single-cell genomics dataset. [Figure 7A] Figure 7A shows a visualization of a dendrogram including cell clusters. [Figure 7B] Figure 7B shows a visualization of a dendrogram including cell clusters. [Figure 8A] Figure 8A shows an example of clustering. [Figure 8B] Figure 8B shows an example of clustering. [Figure 9A] Figures 9A to 9C show a further example of clustering and cell sub-samples obtained as a result after traversing a cell-cell neighborhood graph using the prize-collecting Steiner tree algorithm (PCST). [Figure 9B] Figures 9A to 9C show a further example of clustering and cell sub-samples obtained as a result after traversing a cell-cell neighborhood graph using the prize-collecting Steiner tree algorithm (PCST). [Figure 9C] Figures 9A to 9C show a further example of clustering and cell sub-samples obtained as a result after traversing a cell-cell neighborhood graph using the prize-collecting Steiner tree algorithm (PCST). [Figure 10A]Figures 10A and 10B show examples of memory (RAM) usage for a subsampling method for a given data size and the time consumed to execute the subsampling method. [Figure 10B] Figures 10A and 10B show examples of memory (RAM) usage for a subsampling method for a given data size and the time consumed to execute the subsampling method. [Figure 11] Figure 11 shows a visualization of an example of multiple subsamplings. [Figure 12] Figure 12 shows a comparison of the degree of connectivity of subsampled cells between this method and random sampling. [Figure 13] Figure 13 shows the cluster sizes after subsampling. [Figure 14] Figure 14 shows a comparison between this implementation, random sampling, and other sampling tools regarding the degree of connectivity between subsampled cells and other subsampled cells in their original neighbors (non-zero degrees are shown). [Figure 15] Figure 15 shows a comparison between this implementation, random sampling, and other sampling tools regarding the degree of connectivity between subsampled cells and other subsampled cells in their original neighbors (non-zero degrees are shown). [Modes for carrying out the invention]

[0018] Recent advances in single-cell isolation techniques, along with methods for amplifying these genetic materials, have made it possible to explore genes, such as chromatin, RNA, and DNA, within a single cell. This field is known as single-cell genomics. In contrast to bulk sequencing, which uses gene expression levels—the average expression profile of individual cells—single-cell sequencing preserves information about the original cell for each molecule measured. Due to heterogeneity in cell populations, analyzing a single cell allows for the discovery of mechanisms not observed when studying bulk populations of cells. Single-cell sequencing can characterize individual cells at the molecular level.

[0019] Single-cell genomics data can consist of a wide range of techniques aimed at profiling and quantifying gene expression (scRNA sequencing, SMART sequencing, sciRNA sequencing, Drop sequencing, InDrop), chromatin accessibility (scATAC sequencing, sciATAC sequencing), surface (CITE sequencing, DAB sequencing), and / or intracellular proteome (PEA), DNA methylation regions (scBS sequencing, sci-MET), DNA boundary histone (CutNTag, CutNRun) modification, or any combination of these modalities. Most of these laboratory protocols involve acquiring individual cells from the biological tissue of interest. The acquired cells or their extracted nuclei are tagged with unique barcodes to allow the origin of each molecule (DNA / cDNA fragment) to be traced back to the original cell. Many commercially available tools today use water droplets in oil in a microfluidic setup for cell / nucleus acquisition. Barcoded molecules are stored and sequenced using next-generation sequencing methods, typically based on one of the Illumina sequencing systems via a synthetic platform. Once the fragment is sequenced, or during the sequencing process, the identified sequence of DNA nucleotides is stored on digital media. This digital media ultimately stores the information of the sequenced fragment.

[0020] Single-cell genomics datasets are inherently large in size because they can contain millions of cells and a vast number of features. For example, a single-cell ATAC sequencing dataset can have up to 500,000 or even 1,000,000 features. Improved efficiency and cost-effectiveness of single-cell sequencing techniques are leading to the rapid generation of new datasets. These datasets offer numerous insights and are not only being generated in research fields but are increasingly being incorporated into medical diagnostics and pharmaceutical sciences. However, existing analytical platforms for these datasets are lagging behind the pace of data generation.

[0021] Large-scale single-cell genomics datasets, such as single-cell DNA sequencing, single-cell RNA sequencing, and single-cell ATAC sequencing, require significant computing infrastructure. This infrastructure requirement, along with the associated IT management costs, adds a considerable expense to derive actionable meaning from these datasets.

[0022] In the context of this disclosure, the term “cell” refers to a digital representation of a cell, not a physical cell. For example, a cell may be associated with a sequenced barcode. Cellular features may refer to the specific function, characteristics, or properties of that cell. In single-cell genomics sequencing, data analysis steps may include identifying valid cell barcodes and aligning the sequenced reads to the genome of the organism from which the cells originate. The aligned and sequenced reads are then processed by an analytical method, the choice of which depends on the type of single-cell genomics method used. For example, in single-cell RNA sequencing, the step immediately following alignment is to perform gene quantification. Regardless of the single-cell genomics method used, the objective of these initial stages of analysis is to generate a matrix, or matrix-like data structure, in which the cell and the obtained cell features form two axes of the matrix. In the case of single-cell RNA sequencing data, the cell-gene or cell-transcription matrix may be used. In the case of single-cell ATAC sequencing data, the cell-feature matrix can take the form of a cell-peak matrix.

[0023] This disclosure relates to a method for extracting cell subsamples from multiple cells in a single-cell genomics dataset in a computationally and memory-efficient manner, wherein the cell subsamples eliminate or reduce redundancy and provide cell subsamples that represent the single-cell genomics dataset.

[0024] The method may comprise a first step of loading a single-cell genomics dataset into computer memory in a partition that fits the memory. This may include converting data from text formats such as CSV (Comma Separated Values) and MTX (Matrix Market Exchange Format) into a format that can be efficiently loaded into memory within the partition. Disk-based data persistence formats may be used for this purpose. The purpose of these formats is to load any portion of a matrix or matrix-like dataset into memory. Such disk-based formats are implemented through various software libraries, such as HDF5, Zarr, and Apache Arrow.

[0025] As described above, the obtained single-cell genomics dataset may be represented in at least two dimensions, where information about each cell is represented in the first dimension and information about genetic features is represented in the second dimension. The information about genetic features in the second dimension may include quantified genetic features such as annotated genes and / or gene locations in the coordinate form of the genome assembly and / or the number of cleavage sites within the identified accessible region and / or the number of cDNA fragments for each gene.

[0026] The method may further comprise the step of generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of cells, and cells are represented as vertices in the cell-cell neighborhood graph. This step may include performing a k-nearest neighbor algorithm, where the k nearest neighbors of each cell are identified, cells are represented as vertices, and vertices are connected via edges if at least one of the vertices is the k nearest neighbor of another vertex. The similarity of the genetic features of cells can be quantified in various ways. For example, the similarity between two cells is inversely proportional to the distance between the cells in the graph. The distance between any pair of cells can be calculated by any suitable distance metric. For example, in the case of the Euclidean metric, the distance between cell C1 and another cell C2, each having 1 to N features, is...

[0027]

number

[0028] Angular metrics such as cosine distance are also frequently used because they exhibit scale invariance and can be useful when the range of values ​​varies considerably between cells (often due to sequencing depth or because cells were stored from two or more different experiments).

[0029] The method further comprises the step of dividing the cells in the cell-cell neighborhood graph into seed cells and non-seed cells. This step can be carried out, for example, by any preferred clustering method.

[0030] The inventors have found that, once cells are divided into seed cells and non-seed cells, using the Prize-Collecting Steiner Tree (PCST) algorithm to traverse the cell-cell neighborhood graph and obtain subsamples of cells is a particularly efficient way to obtain useful subsamples of cells. To successfully apply the PCST algorithm, different prizes are assigned to seed cells and non-seed cells. Different approaches exist for assigning prizes. Typically, seed cells are assigned high values, or several distinct high values, and non-seed cells are assigned low values, or several distinct low values. These values ​​may be between 0 and 1, high values ​​may be, for example, numbers greater than 0.5, and low values ​​may be, for example, numbers less than 0.5.

[0031] Figure 2 discloses an example of the method (1) of this disclosure for extracting a cell subsample from multiple cells in single-cell genomics. According to this example, method (100) is: - A step of obtaining a single-cell genomics dataset represented in at least two dimensions, wherein information about each cell is represented in the first dimension and information about genetic characteristics is represented in the second dimension (101), - A step of generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of cells, and cells are represented as vertices in the cell-cell neighborhood graph (102), -Cell-Cell Neighborhood Graph: Step (103) to divide the cells into seed cells and non-seed cells, - In a cell-cell neighborhood graph, the steps include assigning at least one first prize to seed cells and at least one second prize to non-seed cells, -The method comprises the step (104) of traversing a cell-cell neighborhood graph using a bounty-collecting Steiner tree algorithm (PCST) to obtain a subsample of cells.

[0032] The steps of the method of this disclosure for extracting cell subsamples from multiple cells in a single-cell genomics dataset are described in further detail below. As will be understood by those skilled in the art, not all steps are necessarily required to carry out the invention. Some of these steps are optional. Data processing The analysis process for sequenced fragments begins by converting the data into the FASTQ data format, which allows humans to decode the sequences. The sequenced fragments may be subjected to sequencing quality control to remove low-quality nucleotides or to remove all low-quality fragments together. This quality control is optional depending on the setting. These quality-controlled sequenced fragments (and any supplementary information present for each fragment, such as cell barcodes or sequencing) can then be aligned to a genome / transcriptome, or an arbitrarily selected portion of the genome, using genomic alignment tools. This particular step can also be performed in an alternative way, for example, by using a pseudo-alignment strategy to identify the genomic portion from which the fragment originates. Once the genomic location of each fragment (either in terms of precise coordinates or genomic entities) is determined, intermediate steps may exist, depending on the modality being accessed, such as peak recall in the case of single-nucleus ATAC sequencing and single-cell cut&tag. Barcode information for each fragment, or genomic information derived from them (e.g., peaks), can be collated to create a cell-feature matrix. This cell-feature matrix represents single-cell genomics data in a shareable form that allows for viable analysis. The actual storage format of this cell-feature can be HDF5, CSV, LOOM, ZARR, or any other columnar data storage format. Assay identification This step can be performed once the single-cell genomics dataset is stored in memory. In this method, in the first step, a number of different assays present in the dataset can be identified. Advances in single-cell genomics methods now allow sequencing of different genomic features from the same cell. One technique available for this purpose is cell-indicated sequencing of transcriptomes and epitopes (CITE sequencing), which allows for the quantification of both gene expression (transcript amount and diversity) and cell surface markers from the same partner. Thus, one embodiment of the method of this disclosure for extracting cell subsamples from multiple cells in a single-cell genomics dataset further comprises the step of identifying the genetic features of assays and sequences in the single-cell genomics dataset. Filtering and excluding details In one embodiment of the method of the present disclosure for extracting a subsample of cells from multiple cells in a single-cell genomics dataset, the cells are filtered and excluded in a further step. This step may include filtering and excluding cells having a number of features below a predetermined lower limit and / or above a predetermined upper limit, or filtering and excluding cells based on other properties.

[0033] In such cases, where cells having a number of features below a predetermined lower limit and / or above a predetermined upper limit are filtered and excluded, or where a threshold is specified for classifying cells as having too many or too few features, the method may construct a normal distribution based on the same mean and sample variance and exclude cells below a certain percentage subtracted from the distribution. Other preferred methods for performing feature filtering may be conceivable. Normalization The methods of this disclosure may further include a step of normalizing information about genetic features before generating a cell-cell neighborhood graph. This information normalization step may be used to provide comparable quantified genetic features. This step may include specifying the normalization to be applied. To speed up this process, normalization may be applied when the single-cell genomics dataset is loaded into memory. In the case of single-cell RNA sequencing (scRNA sequencing), a library size normalization method may be used. In this normalization step, the feature values ​​of each cell are divided by the sum of the values ​​from each cell, and the resulting value is multiplied by a constant scalar. This normalization step ensures that the effect of sequencing depth of individual cells in downstream analysis is minimized. In the case of single-cell assays for transposase-accessible chromatin using sequencing (scATAC sequencing), the method may apply TI-IDF normalization (word frequency, document frequency reciprocal). The TF value is calculated by dividing each cell by the sum of its values, and the IDF value is calculated by dividing the total number of cells by a vector of the number of cells in which each gene exists. The scalar product of TF-IDF gives a normalized value. For cell surface markers, this method allows for CLR (centered log-ratio) normalization. Feature Selection The method of this disclosure may further include a step of performing feature selection by selecting a subset of smaller features from information about genetic features before generating a cell-cell neighborhood graph. This step may rely on the assumption that only a subset of smaller genetic features can acquire cellular diversity and that many features are redundant to one another. As a result of feature selection, the subset of smaller genetic features is used for further analysis.

[0034] The choice of feature selection method may depend on the type of single-cell genomics dataset being used. For scRNA sequencing datasets, feature selection may be performed based on gene variance. Since gene variance and mean expression are frequently matched, mean-variance trends can first be eliminated by dividing the genes into a desired number of bins based on their mean expression values. The gene with the minimum variance in each bin is selected, and stepwise regression is performed between the mean and variance to fit a curve through the selected genes from each bin. Using the resulting curve, the corrected variance is obtained by regression on the remaining variance from each gene. The gene with the highest corrected variance is selected.

[0035] For the scATAC sequencing dataset, these values ​​may be normalized using TF-IDF, and the most frequently occurring peaks, i.e., the high mean normalized values ​​among cells, are selected. The user can decide to select the top "n" HGVs or the most frequently occurring peaks, where "n" is any number of features. Dimensionality Reduction The method may further include a step of performing dimensionality reduction on the single-cell genomics dataset before generating the cell-cell neighborhood graph. Dimensionality reduction may be performed using incremental principal component analysis (PCA).

[0036] In one embodiment, normalized data is subjected to dimensionality reduction. Dimensionality reduction techniques reduce the feature space, which helps amplify the signal-to-noise ratio. Dimensionality reduction also serves to reduce data redundancy. PCA can be applied, for example, to single-cell RNA data. In the case of scATAC sequencing, the method of this disclosure may apply an incrementally updatable version of latent semantic indexing technique. The user can configure the number of dimensions to reduce and retain for further steps. In the case of single-cell RNA sequencing, prior to subjecting the data to PCA, the data may be feature-wise scaled so that it has a zero mean and a unit standard deviation for each gene. The data may or may not be logarithmically transformed before this step. Generating cell-neighborhood graphs Cell-cell neighborhood graphs can be generated in various ways. In a preferred embodiment, the step of generating a cell-cell neighborhood graph from a single-cell genomics dataset includes implementing a k-nearest neighbors (KNN) algorithm. In such an algorithm, a graph is created and cells are represented as vertices. Vertices are connected via edges if at least one of them is a k-nearest neighbor of another vertex. To determine how close "neighboring" cells are to each other, it is useful to compare quantified genetic features with each other, preferably the quantified genetic features are normalized.

[0037] In one example, the nearest neighbors of each of k cells are identified, where k is an arbitrary positive integer given by the user. The KNN for each cell is used to create a graph data structure for the cells. In the graph, each vertex is a cell, and these vertices are connected by edges. Two vertices are connected if at least one of them is the nearest neighbor of another vertex. The edge weights in the graph are inversely proportional to the distance between cells. Euclidean distance or other suitable metrics may be used to calculate the distance between cells. Identifying the KNN for each cell by comparing each cell pair is not scalable to a large number of cells. Therefore, in one embodiment, after a cell index is created, the nearest neighbors are queried across these cells. The distances between KNN neighbors are converted to a continuous form using a Gaussian nucleus. These values ​​are treated as edge weights for the KNN graph. Other alternatives are possible for computing the graph structure of the data; for example, a shared nearest neighbor graph can be created by computing common neighbors in each cell pair, which are each other's KNNs. The method of this disclosure is not limited to the approach taken to create the cell-cell graph. In one embodiment, a roughly equivalent version of KNN, referred to as HNSWlib, is used to generate a cell-cell neighborhood graph from single-cell genomics.

[0038] Figure 4 shows an example of a cell-cell neighbor graph (300). In this example, each vertex (301) of the graph is a cell (301). The vertices (301) are connected by edges (302). The edges (302) indicate that the connected cell (302) is the k-nearest neighbor. Cell-neighborhood graph division of cells into seed cells and non-seed cells The step of dividing cells in a cell-cell neighborhood graph into seed cells and non-seed cells can be implemented in several ways.

[0039] In a preferred embodiment, this step is: -A step of clustering cells in a cell-cell neighborhood graph to obtain clusters of cells of approximately uniform size, -The step of dividing cells into seed cells and non-seed cells by selecting at least one cell from each cell cluster and classifying the selected cell as a seed cell, It is equipped with.

[0040] The step of clustering cells in a cell-neighborhood graph may be based on, for example, the Louvain algorithm, the Leiden algorithm, or the Paris algorithm. In this step, the cell-neighborhood graph may be converted into a dendrogram structure. The step of clustering cells in a cell-neighborhood graph may include limiting the number of cells in each cluster to a minimum and maximum number of cells. More specifically, the step of clustering cells in a cell-neighborhood graph may include converting the cell-neighborhood graph into a dendrogram structure, where each cell is represented as a leaf node, and the leaves are connected to each other through branch point nodes in the dendrogram structure, and the step of clustering cells in a cell-neighborhood graph includes partitioning the leaf nodes and branch point nodes into groups of cells, these groups having an upper and lower limit. The upper limit is preferably set to a relatively small number so that each group of cells has only a small fraction of the total number of cells, leading to "micro-clustering" where each group / partition of cells is highly uniform. The upper limit may be a number less than or equal to 300, preferably less than or equal to 200, and more preferably less than or equal to 100. The lower limit may be, for example, 10. In one example, the lower limit is 10 and the upper limit is 200.

[0041] An example of a clustering process is described below. A cell-cell neighborhood graph of cells may be subjected to clustering to obtain clusters of cells of uniform size. Multiple clustering strategies exist for graph data structures using community clustering methods. Preferably, a clustering method is selected that has an operating time that scales linearly with respect to the number of cells present in the graph. This algorithm can transform the graph data structure into a dendrogram structure. In one embodiment, the number of cells in each cluster can be limited to predetermined upper and lower limits. Since the algorithm can be a greedy bottom-up approach, it can start the integration from the leaf nodes. In the dendrogram, each cell represents a leaf node, and these leaves are connected to each other through branch point nodes. Each branch point node is divided into two nodes that can be a branch point, a leaf node, or a combination of both. The branch point nodes eventually merge into an upstream root node. First, all cells are ranked based on how many branch point nodes there are between them and the root node. These ranks determine the order in which cells are considered as candidates for cluster aggregation. The reason behind this is that clustering of cells is prioritized based on how deep the dendrogram is at that point. The more branch points there are, the more subtle the differences between cells are. The algorithm starts with the highest-ranked node and then moves to the parent branch point node of that cell, with the other child nodes of that branch point node automatically being included in that cluster. The algorithm then increases the cluster size by including upstream branch points and continues moving upstream to branch points until it exceeds a predetermined threshold at which the cluster freezes and the algorithm moves to the next leaf node (cell) in the rank list that is not included in any cluster. Another condition for stopping upstream traversal is when the branch point's siblings are already included in the cluster.Accordingly, in one embodiment of the method of the present disclosure for extracting a subsample of cells from multiple cells in a single-cell genomics dataset, each cell is represented as a leaf node, the leaves are connected to each other through branch point nodes in a dendrogram structure, and the step of clustering cells in a cell-cell neighborhood graph comprises ranking each cell based on how many branch point nodes exist between the cells and root nodes in the dendrogram structure, and performing clustering based on this ranking. Once all leaf nodes are included in a cluster, the next step is to merge the clusters so that each cluster has at least the number of cells configured by the user. Clusters are considered in the order they were created, their size is less than a cutoff, and they are merged with clusters in sibling branch points (i.e., clusters with the most recent common ancestor). Merging is repeated until a minimum cluster size is obtained. If there are more than one cluster of sibling branch points, the cluster with the shortest distance to a given cluster is considered for merging, provided that its distance is less than or equal to a predetermined threshold. This dendrogram creation and clustering strategy, including bottom-up balanced clustering, provides clusters of cells that are highly similar and closely connected within the graph.

[0042] Alternatively, a dynamic sampling approach may be used in the step of clustering cells in a cell-neighborhood graph. This approach can also start from a dendrogram. Similar to the approach described above, the objective is to narrow down the cells in each group to a pool of cells from which seed cells can be randomly selected. In one embodiment, the step of clustering cells in a cell-neighborhood graph includes performing a first compartmentalization and determining the number of subcompartments for each compartment in the first compartmentalization based on measurements of the strength of cell interconnections with each compartment in the first compartmentalization. Once a suitable number of subcompartments has been determined, the actual step of performing subcompartmentalization within the compartments in the first compartmentalization can be carried out. In a further step, the cells within each compartment in the first compartmentalization may then be converted into a dendrogram structure showing hierarchical relationships between cells. Further examples of cell-neighborhood graphs showing the division of cells into seed and non-seed cells. For example, a seed cell is identified by calculating two metrics for each individual cell (referred to as a node in the neighborhood graph): n-neighborhood degree (NND) and neighborhood connectivity (NC). The degree of a node is calculated as the total number of other nodes to which this particular node is connected. The degree of one neighborhood is the sum of the degrees of all nodes connected to a given node. Thus, NND is calculated by iterating through neighborhoods over n-step distances to obtain the connectivity density around a given node in the graph. The second metric is neighborhood connectivity, which is obtained when a given number of connections are shared by many or few nodes. To calculate the NC for each cell, the sum of the nearest-neighbor distances (Jackard distances) shared between the node and all of its neighbors is calculated. Thus, a node will have a high value for neighborhood connectivity if it is connected to other nodes that are strongly connected to each other.

[0043] For the next step, the algorithm uses cell partitioning. Here, the median NND and NC are calculated for each cell cluster, and this median is used to adjust the ramping rate for each cluster. A higher median NND results in a lower sampling rate, and a higher NC results in a lower sampling rate, or vice versa. Based on the sampling rate, the number of cells sampled from each cluster is determined. Each cluster is then subclustered, with the number of subclusters being equal to the number of cells sampled, and one cell is sampled from each nuclear subcluster.

[0044] These sampled cells are referred to as "seed" cells.

[0045] Figure 7 visualizes a dendrogram containing a cluster of cells (401). Figure 7A shows the cluster obtained by applying the straight-cut algorithm. In this example, it can be seen that the clusters (401) have variations in size (i.e., the number of cells in the cluster). Each circle (401) represents a cluster of cells. The size of the circle indicates the number of cells in the cluster. The dendrogram has a root node (403). The dendrogram has numerous branch points (404). The branch points, their structure, and their clusters can be said to provide information about how similar the clusters (401) are. The more similar the clusters are, the closer these shared branch points are.

[0046] Figure 7B shows the clusters obtained by applying the balanced cut algorithm. This algorithm produces clusters of similar size to the straight cut algorithm. The figure also shows how a large cluster (402) can be broken down into smaller clusters (401) within the balanced cut algorithm.

[0047] Figure 8 shows a further example of clustering, with Figure 8A illustrating an example of how cells are divided into clusters (numbered 1-15). Cells are numbered based on their cluster identity. By applying the algorithm, a dendrogram of cells is created in the case of this Paris algorithm, which is shown in Figure 8B. Cells forming a cluster are merged into a single node (numbered circle in Figure 8B). The size of this circle indicates the number of cells in the cluster. The unlabeled node at the top of the dendrogram is the root of the dendrogram.

[0048] Once clusters are created, the method of the present disclosure may, in a further step, select at least one cell from each cluster of cells and classify the selected cell as a seed cell. Cells that are not selected are classified as non-seed cells.

[0049] Preferably, at least one cell must be selected from each cluster. Cells can be selected randomly or based on some measurement, such as how central they are within the cluster. Seed cells acquire cellular heterogeneity. Acquiring graph topology can be important for ensuring that differentiating trajectories and heterogeneity are acquired in sparse populations. Prize money setting A method of extracting a subsample of cells from multiple cells in a single-cell genomics dataset, as described in this disclosure, preferably involves traversing a cell-cell neighborhood graph using the PCST algorithm to obtain a subsample of cells. Before the PCST algorithm is performed, seed cells and non-seed cells in the cell-cell neighborhood graph must be assigned a prize. One way to set the prizes is to assign a single first prize to all seed cells and a single second prize to all non-seed cells. The single first prize is typically higher than the single second prize. The single first prize may be any high prize. This value may be derived from the number of cells in the graph. The required value increases with the number of cells present in the graph. Extremely high values ​​typically do not adversely affect the algorithm. The sole purpose of setting the prizes is to ensure that cells are considered to be traversed in later steps. Non-seed cells may be given a value of 0, meaning that the traversing algorithm will not be incentivized to include those cells.

[0050] In an alternative embodiment, the rewards can instead be configured individually for each seed cell and non-seed cell. For example, the reward for seed cells can be set to be proportional to their in-degree, i.e., the number of other cells to which a given cell is connected in the graph. Furthermore, manifold coverage can be smoothed by assigning a value greater than 0 to non-seed cells in sparse regions of the graph, based on the concept of cell neighborhood degree (i.e., the cumulative degree of cells adjacent to a given cell in the graph).

[0051] We can choose not to set penalties on graph edges, or to set very low penalties, thereby ensuring that seed cells located further away from other seed cells can be easily and reliably acquired during graph traversal. In our prototype, we use 1-edge weights in the KNN graph as penalties for edges. Since these edge weights have boundaries within the range of 0 and 1, the penalties are also within these boundaries. Alternatively, we can choose to set edge penalties based on neighborhood density so that traversal is encouraged or not encouraged in different parts of the graph.

[0052] Figure 5 shows an example of a cell-cell neighborhood graph (300). In this example, cells are divided into seed cells (303) and non-seed cells (304). Preferably, each seed cell (303) (larger dots in the figure) is assigned a first prize selected from a first score set. Preferably, each non-seed cell (304) (smaller dots in the figure) is assigned a second prize selected from a second score set. Edges (302) between cells (303, 304) are assigned penalties selected from a third score set. In this example, the width of the edges visualizes these penalties. For example, edge 302a is thicker than edge 302b, indicating that edge 302a has a higher penalty than edge 302b. Further examples of prize money settings There are different approaches to prize allocation. For example, all seed cells are assigned a fixed prize value. For instance, an initial value of 10 might be used. The edge penalty Ep for each edge can then be calculated as follows:

[0053]

number

[0054] Once the prize money for the seed cell and the penalties for all edges have been set, the PCST algorithm may be implemented, which will be explained and illustrated further. Traversing cell-cell neighbor graphs using a bounty-collecting Steiner tree algorithm Once a prize is set, the method then uses an implementation of a Prize-Collecting Steiner Tree (PCST). The cell-cell neighborhood graph, along with the set prize, is provided to the PCST algorithm. In one embodiment, the algorithm is configured to include as many seed cells as possible and as few non-seed cells as possible. In one embodiment, the PCST begins traversing the cell-cell neighborhood graph with the aim of including as few non-seed nodes as possible and all seed nodes. The cell-cell graph may already be considered cut at this point. This is most likely scenario when there are highly discrete sets in the dataset. In such a case, the PCST algorithm may be performed separately for each of the cut components of the graph. Thus, in one embodiment of the method of this disclosure, the cell-cell neighborhood graph includes a cut subgraph, and the PCST algorithm is performed for all subgraphs. As a result of the PCST graph traversal, subsamples of cells are obtained. These may include non-seed nodes so as to ensure that the subsamples of cells from the cell-cell neighborhood graph (or subgraphs if the graph was cut) are connected. As described above, cells are represented as vertices, which may be connected by edges, each representing a degree of similarity between cells. In one embodiment, a penalty is assigned to the edges between vertices, and the penalty represents the degree of dissimilarity between the cells at the vertices. The penalty can be taken into consideration when traversing the cell-cell neighborhood graph. In one embodiment of the method, iterative subsampling is performed by creating a new cell-cell graph using subsampled cells, applying clustering and seed identification steps, and then performing a new PCST search. The user may decide to perform subsampling iterations at the frequency required to obtain the desired downsampling size.

[0055] Figure 6 shows an example of a cell-cell neighborhood graph (300) relating to the method of the present disclosure for extracting a subset of cells from multiple cells in a single-cell genomics dataset, where a subsample of cells (303, 304) is extracted from multiple cells in the single-cell genomics dataset (i.e., all cells in graph (300)). In this visualization, the filled dots (303, 304) represent seed and non-seed cells included in the cell subsample. The open dots (305) represent dots not included in the cell subsample. The edges (302) seen in the figure are edges traversed by PCST to ensure that all cells in the graph are connected.

[0056] Figure 9 shows further examples of clustering and the resulting subsamples of cells after traversing a cell-cell neighborhood graph using the PCST algorithm. Figure 9A shows the results of clustering using a balanced cut approach, and the clusters and their sizes are visualized on the fused dendrogram in Figure 9B. Figure 9C shows the subsampled cells as a result of running PCST on the cell-cell neighborhood graph of cells with a set prize. All clusters are represented in the subsample, and even clusters of smaller numbers of cells are represented. Computer system The Disclosure further relates to a computer system for extracting cell subsamples from multiple cells in a single-cell genomics dataset. The system comprises a processing unit which may comprise any preferred single-processor or multi-processor central processing unit. The processing unit may operate using any preferred instruction set architecture, including but not limited to x86, ARMv6-v8, x86-64, and PowerPC, such as CISC and RISC-based instruction set architectures. The system further comprises memory which may comprise any preferred memory structure. Typically, the memory structure would comprise one or more random access memories (RAM) in the form of static random access memory (SRAM) and / or dynamic random access memory (DRAM). The processing unit is configured to perform any variation of the method of the Disclosure for extracting cell subsamples from multiple cells in a single-cell genomics dataset. A schematic diagram of one embodiment of the computer system (200) of the Disclosure is shown in Figure 3. The computer system (200) comprises a processing unit (201) and memory (202). A single-cell genomics dataset (204) is loaded into memory (202). After performing a method for extracting cell subsamples from multiple cells in the single-cell genomics dataset, the cell (203) subsamples from the cells of the single-cell genomics dataset are delivered by the computer system in any preferred format. In a preferred embodiment, the processing unit, - A step of loading a single-cell genomics dataset, represented in at least two dimensions, into memory, wherein information about each cell is represented in the first dimension and information about genetic characteristics is represented in the second dimension. - A step of generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of cells, and cells are represented as vertices in the cell-cell neighborhood graph. -A step of dividing the cells in the cell-neighborhood graph into seed cells and non-seed cells, - In a cell-cell neighborhood graph, the steps include assigning at least one first prize to seed cells and at least one second prize to non-seed cells, - Using a prize-collecting Steiner tree algorithm, the steps involve traversing the cell-cell neighborhood graph to obtain a subsample of cells, It is configured to implement the following: Further examples and results of implementing the method In one implementation, the method identifies landmark points (seed cells) in the graph and then uses the PCST algorithm to find paths connecting these seed cells (Figures 4-6). Using a PCST implementation with near-linear time complexity, rapid and scalable downsampling to millions of cells is achieved. The implementation found that downsampling to datasets of up to 4 million cells could be performed using less than 20GB of RAM (Figure 10A). Furthermore, downsampling took less than 3 minutes for a dataset of 1 million cells and 15 minutes for a dataset of 4 million cells (Figure 10B). UMAP utilization of four atlas-scale datasets shows that even with downsampling of up to 1% of the cells, cells belonging to all clusters across the UMAP space were sampled (Figure 11).

[0057] Furthermore, for the quantitative analysis of downsampling, we analyzed the degree of connectivity between subsampled cells and other subsampled cells in the original neighbor graph. A high frequency of zero-degree values ​​indicates that many cells are disconnected from other subsampled cells, serving as a marker for unsatisfactory subsampling and indicating a lack of mediating cell states in the subsampled set. Comparing the number of disconnected cells between cells randomly subsampled from four atlas-scale datasets using this method (referred to as Scarf), 100% of Scarf subsampled cells showed non-zero-degree values ​​for all datasets, while random sampling resulted in non-zero-degree values ​​in 18.9% to 26.9% of the details (Figure 12).

[0058] Two primary objectives of downsampling are to reduce redundancy in the dataset and to preserve rare cell types / states. These two objectives can be accessed by calculating the change in the proportion of cells from each cluster after downsampling. In this example, the implementation was able to simultaneously increase the proportion of cells from fewer clusters while decreasing the proportion of cells from larger clusters (Figure 13). The proportion of cells from the smallest clusters in each dataset increased by between 8.13 and 16.82 times, while the proportion of cells from the largest clusters decreased by between 3.35 and 5.26 times. In comparison, random clustering did not show an increase in the proportion of smaller clusters by more than 1.5 times, nor a decrease in the proportion of larger clusters by more than 1.01 times. As a result, random clustering is less likely to sample rare clusters. For example, in the smallest clusters within the atlas-scale dataset, no cells were sampled in 20% (1M cell dataset), 40% (2M cell dataset), and 20% (4M cell dataset) (n=10) of the random sampling. In contrast, all clusters were sampled using Scarf, regardless of their intrinsic cluster size.

[0059] Furthermore, a further comparison between the implementation of this method and another sampling tool, GeoSketch, is presented in Figures 14 and 15. Two contrasting small datasets were used, consisting of either 10K PBMC cells of a different cell type (provided by 10X Genomics) or 3.5K pancreatic cells in a differentiated continuum. Visualization of progressively increasing levels of downsampling to each of these datasets showed that this method was able to acquire cells throughout UMAP. After 100 repetitions of downsampling using this method, GeoSketch, or random sampling, 100% of the downsampled cells selected by Scarf were non-zero in both the PBMC and pancreatic cell datasets, compared to 66.9% and 75.9% respectively for GeoSketch and 49.5% and 70.3% respectively for random sampling (Figures 14 and 15). Further details of the present invention Note 1. A computer implementation method for extracting cell subsamples from multiple cells in a single-cell genomics dataset, - A step of obtaining a single-cell genomics dataset represented in at least two dimensions, wherein information about each cell is represented in the first dimension and information about genetic characteristics is represented in the second dimension, - A step of generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of cells, and cells are represented as vertices in the cell-cell neighborhood graph. -A step of dividing the cells in the cell-neighborhood graph into seed cells and non-seed cells, - In a cell-cell neighborhood graph, the steps include assigning at least one first prize to seed cells and at least one second prize to non-seed cells, - A method comprising the steps of using a prize-collecting Steiner tree algorithm to traverse a cell-cell neighborhood graph and obtain a subsample of cells.

[0060] Appendix 2. The method according to Appendix 1, further comprising the step of performing dimensionality reduction of a single-cell genomics dataset prior to generating a cell-cell neighborhood graph.

[0061] Note 3. The step of performing dimensionality reduction on a single-cell genomics dataset is the method described in Note 2, including incremental principal component analysis.

[0062] Note 4. Information about each cell in the first dimension is provided in any one of the methods described in Note 1 to 3, including a cell identifier such as a cell barcode for each cell.

[0063] Note 5. Information regarding genetic features in the second dimension is the method according to any one of Notes 1 to 4, including annotated genes, and / or gene locations in coordinate form of the genome assembly, and / or the number of cleavage sites within identified accessible regions, and / or the number of cDNA fragments for each gene, and / or the degree of quantified genetic features such as DNA / RNA oligonucleotides, methylation on genomic loci, or any other epigenetic marks.

[0064] Note 6. A single-cell genomics dataset is a single-cell RNA sequencing dataset, a single-cell ATAC sequencing dataset, a CITE sequencing dataset, or any similar method described in any one of Notes 1 to 5.

[0065] Note 7. A single-cell genomics dataset is a method described in any one of the methods in Notes 1 to 6, represented as a two-dimensional matrix.

[0066] Note 8. The step of generating a cell-cell neighbor graph from a single-cell genomics dataset is the method according to any one of Notes 1-7, which includes performing a k-nearest neighbor algorithm.

[0067] Note 9. The k-nearest neighbor algorithm is the method according to Note 8, which includes identifying the k nearest neighbors of each cell, where k is a positive integer, and creating a cell-cell neighbor graph, where cells are represented as vertices, and vertices are connected by edges if at least one of the vertices is the k nearest neighbor of another vertex.

[0068] Note 10. The step of dividing the cells in the cell-neighborhood graph into seed cells and non-seed cells is: -A step of clustering cells in a cell-cell neighborhood graph to obtain cell clusters of approximately uniform size, - The method according to any one of the appendices 1 to 9, comprising the step of dividing the cells into seed cells and non-seed cells by selecting at least one cell from each cell cluster and classifying the selected cell as a seed cell.

[0069] Note 11. The step of clustering cells in the cell-cell neighborhood graph is the method described in Note 10, which includes the Louvain algorithm, the Leiden algorithm, or the Paris algorithm.

[0070] Note 12. The step of clustering cells in a cell-neighborhood graph is the method described in Note 10 for converting a cell-neighborhood graph into a dendrogram structure.

[0071] Note 13. The method according to Note 10, wherein the step of clustering cells in a cell-cell neighborhood graph includes limiting the number of cells in each cluster to a minimum and maximum number of cells.

[0072] Note 14. The method according to Note 12, wherein each cell is represented as a leaf node, and the leaves are connected to each other through branch point nodes in a dendrogram structure, and the step of clustering cells in a cell-cell neighborhood graph includes ranking each cell based on how many branch point nodes exist between the cell and the root node in the dendrogram structure, and clustering based on the ranking.

[0073] Note 15. The method according to any one of Notes 1 to 14, wherein a single first prize is higher than a single second prize, and preferably the second prize is 0.

[0074] Appendix 16. The method according to any one of Appendix 1 to 14, wherein at least one first prize and at least one second prize are prizes that can be individually configured for individual seed cells and non-seed cells.

[0075] Note 17. Cells represented as vertices are connected by edges, and each edge represents the degree of similarity between cells, as described in any one of Notes 1 to 16.

[0076] Appendix 18. The method according to any one of Appendix 1 to 17, further comprising the step of assigning a penalty to an edge between vertices, wherein the penalty represents the degree of dissimilarity between cells of the vertices.

[0077] Note 19. The method according to any one of Notes 1 to 18, wherein the step of traversing the cell-cell neighborhood graph using a prize-collecting Steiner tree algorithm includes a step of including as many seed cells as possible and as few non-seed cells as possible.

[0078] Note 20. Cell subsamples are attached according to the method described in Note 19.

[0079] Note 21. The cell-cell neighborhood graph includes disconnected subgraphs, and the prize-collecting Steiner tree algorithm is performed for all subgraphs according to any one of the methods described in Notes 1-20.

[0080] Appendix 22. The method according to any one of the appendices 1 to 21, further comprising the step of loading a single-cell genomics dataset into computer memory in a memory-fitting partition before performing the other steps.

[0081] Appendix 23. The method according to any one of the appendices 1 to 22, further comprising the step of identifying an assay in a single-cell genomics dataset and sequence genetic features.

[0082] Appendix 24. The method according to any one of the appendices 1 to 23, further comprising the step of filtering and excluding cells, such as filtering and excluding cells having a number of features below a predetermined lower limit and / or above a predetermined upper limit, before generating a cell-cell neighborhood graph.

[0083] Appendix 25. The method according to any one of the appendices 1 to 24, further comprising the step of normalizing information about genetic characteristics before generating a cell-cell neighborhood graph.

[0084] Appendix 26. The method according to any one of the appendices 1 to 25, further comprising the step of performing feature selection by selecting a smaller subset of features from information about genetic features before generating a cell-cell neighborhood graph.

[0085] Note 27. A computer program that, when executed by a computing device or computing system, has instructions for the computing device or computing system to carry out a method for extracting cell subsamples from multiple cells in a single-cell genomics dataset described in any one of Notes 1 to 26.

[0086] Note 28. A computer system, Memory and It comprises a processing unit, and the processing unit is - A step of loading a single-cell genomics dataset, represented in at least two dimensions, into memory, wherein information about each cell is represented in the first dimension and information about genetic characteristics is represented in the second dimension. - A step of generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of cells, and cells are represented as vertices in the cell-cell neighborhood graph. -A step of dividing the cells in the cell-neighborhood graph into seed cells and non-seed cells, - In a cell-cell neighborhood graph, the steps include assigning at least one first prize to seed cells and at least one second prize to non-seed cells, - Using a prize-collecting Steiner tree algorithm, the steps involve traversing the cell-cell neighborhood graph to obtain a subsample of cells, A computer system configured to perform the following.

[0087] Note 29. A computer implementation method for extracting cell subsamples from multiple cells in a single-cell genomics dataset, - A step of obtaining a single-cell genomics dataset represented in at least two dimensions, wherein information about each cell is represented in the first dimension, and information about the genetic characteristics of the cell is represented in the second dimension. - A step of determining the similarity of the genetic characteristics of cells and generating a cell-cell neighborhood graph from a single-cell genomics dataset, wherein cells are represented as vertices in the cell-cell neighborhood graph, and the distance between vertices corresponds to the similarity. -A step in which some cells in the cell-neighborhood graph are designated as seed cells, and the remaining cells are designated as non-seed cells, - In a cell-cell neighborhood graph, a step is to assign at least one first prize to seed cells and at least one second prize to non-seed cells, wherein the first prize is higher than the second prize. - A method comprising the steps of extracting a subsample of cells by exposing a cell-cell neighborhood graph using a bounty-collecting Steiner tree algorithm. References The publication "Scarf: A Toolkit for Memory-Efficient Analysis of Large Single-Cell Genomic Data," published on 3 May 2021, by Parashar Dhapola, Johan Rodhe, Rasmus Olofzon, Thomas Bonald, Eva Erlandsson, Shamit Soneji, and Goeran Karlsson, doi:https: / / doi.org / 10.1101 / 2021.05.02.441899, https: / / www.biorxiv.org / content / 10.1101 / 2021.05.02.441899v1.full, is incorporated herein by reference in its entirety.

Claims

1. A computer implementation method for downsampling a single-cell genomics dataset by extracting a subsample of cells from multiple cells in the single-cell genomics dataset, wherein the cells in the single-cell genomics dataset are located on a data manifold, - A step of obtaining a single-cell genomics dataset represented in at least two dimensions, wherein information about each cell is represented in the first dimension and information about genetic characteristics is represented in the second dimension, - A step of generating a cell-cell neighborhood graph from the single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of the cells, and the cells are represented as vertices in the cell-cell neighborhood graph, the step of generating the cell-cell neighborhood graph from the single-cell genomics dataset comprising performing a k-nearest neighbor algorithm, the k-nearest neighbor algorithm determining the k nearest neighbors of each cell, where k is a positive integer, and generating a cell-cell neighborhood graph, wherein the cells are represented as vertices, and if at least one of the vertices is the k nearest neighbor of another vertex, the vertices are connected by edges. - A step of dividing the cells in the cell-cell neighborhood graph into seed cells and non-seed cells by clustering the cells in the cell-cell neighborhood graph and selecting at least one cell as a seed cell from each cluster, wherein the seed cells represent different cell populations in the single-cell genomics dataset, - In the cell-cell neighborhood graph, the steps include assigning at least one first prize to the seed cells and at least one second prize to the non-seed cells, A method comprising the steps of: applying a prize-collecting Steiner tree algorithm to the cell-cell neighborhood graph to select a connected subgraph, wherein the prize-collecting Steiner tree algorithm is configured to include as many seed cells as possible and as few non-seed cells as possible in the selected connected subgraph, and saving a data manifold by obtaining a subsample of cells consisting of cells corresponding to the vertices of the selected connected subgraph.

2. The method according to claim 1, wherein the step of clustering the cells in the cell-cell neighborhood graph includes limiting the number of cells in each cluster to a minimum and a maximum number of cells.

3. The method according to any one of claims 1 or 2, wherein the step of dividing the cells in the cell-cell neighborhood graph into seed cells and non-seed cells comprises calculating a first metric representing the connection density around a given node in the graph.

4. The method according to any one of claims 1 to 3, wherein the step of dividing the cells in the cell-cell neighborhood graph into seed cells and non-seed cells comprises calculating a second metric representing neighborhood connectivity, which indicates the degree to which connections are shared between many or few nodes.

5. The method according to any one of claims 1 to 4, wherein the step of dividing the cells in the cell-neighborhood graph into seed cells and non-seed cells is further comprising selecting at least one cell from each cell cluster randomly or based on how central the cell is within the cell cluster.

6. The method according to any one of claims 1 to 5, wherein the step of clustering the cells in the cell-neighborhood graph is to convert the cell-neighborhood graph into a dendrogram structure, where each cell is represented as a leaf node, and the leaf nodes are connected to each other through branch point nodes in the dendrogram structure, and the step of clustering the cells in the cell-neighborhood graph is to rank each cell based on how many branch point nodes exist between the cell and the root node in the dendrogram structure, and cluster the cells based on the ranking.

7. The method according to any one of claims 1 to 6, wherein the step of clustering the cells in the cell-neighborhood graph is to convert the cell-neighborhood graph into a dendrogram structure, where each cell is represented as a leaf node, the leaf nodes are connected to each other through branch point nodes in the dendrogram structure, and the step of clustering the cells in the cell-neighborhood graph is to partition the leaf nodes and branch point nodes into groups of cells, the groups of cells having an upper limit and a lower limit.

8. The method according to claim 7, wherein the upper limit is 300 or less.

9. The method according to any one of claims 1 to 8, wherein the step of clustering the cells in the cell-cell neighborhood graph includes the step of performing a first compartmentalization, and for each compartment in the first compartmentalization, determining the number of subcompartments based on a measurement of the strength of the interconnections of cells in each compartment in the first compartmentalization.

10. The method according to claim 9, further comprising the step of performing sub-partitioning within each of the first partitions in the partitioning process.

11. The method according to claim 10, wherein the step of performing the sub-compartment determination includes the step of converting the cells in each compartment in the first compartment determination into a dendrogram structure showing the hierarchical relationships between the cells.

12. The method according to any one of claims 1 to 11, wherein a single first prize is higher than a single second prize, or the at least one first prize and the at least one second prize are individually configurable prizes for the individual seed cells and non-seed cells.

13. The method according to any one of claims 1 to 12, further comprising the step of assigning a penalty to the edge between the vertices, wherein the penalty represents the degree of dissimilarity between the cells of the vertex.

14. The method according to any one of claims 1 to 13, further comprising the step of performing single-cell genomics sequencing to obtain the single-cell genomics dataset represented in at least two dimensions.

15. The method according to any one of claims 1 to 14, further comprising the step of performing single-cell separation.

16. The method according to any one of claims 1 to 15, further comprising the step of loading the single-cell genomics dataset, which is represented in at least two dimensions, onto a digital medium of a computer system.

17. A computer system, Memory and It comprises a processing unit, and the processing unit is A step of loading a single-cell genomics dataset represented in at least two dimensions into the memory, wherein information about each cell is represented in a first dimension and information about its genetic characteristics is represented in a second dimension, A step of downsampling the single-cell genomics dataset by extracting a subsample of cells from multiple cells in the single-cell genomics dataset, wherein the cells in the single-cell genomics dataset are located on a data manifold. - A step of generating a cell-cell neighborhood graph from the single-cell genomics dataset, wherein the cell-cell neighborhood graph provides information about the similarity of the genetic features of the cells, and the cells are represented as vertices in the cell-cell neighborhood graph, the step of generating the cell-cell neighborhood graph from the single-cell genomics dataset comprising performing a k-nearest neighbor algorithm, the k-nearest neighbor algorithm determining the k nearest neighbors of each cell, where k is a positive integer, and generating a cell-cell neighborhood graph, wherein the cells are represented as vertices, and if at least one of the vertices is the k nearest neighbor of another vertex, the vertices are connected by edges. - A step of dividing the cells in the cell-cell neighborhood graph into seed cells and non-seed cells by clustering the cells in the cell-cell neighborhood graph and selecting at least one cell from each cluster as a seed cell, wherein the seed cells represent different cell populations in the single-cell genomics dataset, - In the cell-cell neighborhood graph, the steps include assigning at least one first prize to the seed cells and at least one second prize to the non-seed cells, - A step of applying a prize-collecting Steiner tree algorithm to the cell-cell neighborhood graph to select a connected subgraph, wherein the prize-collecting Steiner tree algorithm is configured to include as many seed cells as possible and as few non-seed cells as possible in the selected connected subgraph, and saving a data manifold by obtaining a subsample of cells consisting of cells corresponding to the vertices of the selected connected subgraph. A computer system configured to perform the following.

18. The computer system according to claim 17, further comprising a display, wherein the processing unit is configured to display a graphic representation of the cell subsample on the display.

Images