Multiomics single cell data integration and regulatory reasoning based on graph linked embedding

Front content

Single cell multiomics

For single cell multiomics （Single Cell Multi-Omic）, Single cell sequencing counting has been developed so far , From the beginning scRNA-seq,scDNA-seq, Up to now scATAC-seq, Single cell methylation sequencing , Single cell proteome sequencing and other sequencing technologies , So that we have a better understanding of embryonic development , Brain Neuroscience , And cancer and so on , Really start from the cellular level , So that we can carry out research at the functional level of cells , So that we can better understand how genes affect individual traits by affecting the phenotype of cell subpopulations . For Reproductive Medicine , Precision medicine is of great significance .

However , As far as the vast majority of current sequencing technologies are concerned , After collecting data from a batch of cells, only one dimension of information can be obtained , such as , In the use of scRNA-seq When , We can only get the gene expression data of these cells , But I don't know its DNA Methylation modification or proteomic data . But often , Get a cell （ Or a subpopulation of cells ） Multiple omics information is important , This means that we can establish links between different omics data , Better depict the functions of cells and their internal regulatory processes . Combine several of these dimensions of data into a multinomial analysis of the same single cell , It will have an important impact in the fields of basic biology and biomedicine .

Multi source heterogeneous data

Multi source heterogeneous data , That is, data from different sources with different feature types but describing the same object , Concept and multimodality of multi-source heterogeneous data （multi-modal） similar , But multi-source heterogeneous data includes more data types , In the field of information , Mode can be understood as the existence of data format , For example, text , Audio , Images , Video and other formats . When there are multiple modes at the same time, it is called multimodal , For example, as a multimedia video, it can be decomposed into a variety of single-mode data , Such as images , Voice and text .

In fact, the integration of single cell multiomics data is very similar to a multi-source heterogeneous data fusion problem , such as , When we have hematopoietic stem cells scRNA-seq And scATAC-seq Data with two different characteristics , How to integrate these two types of data according to their potential cell subgroup types , For example, we integrate and cluster these two types of data , Suppose they belong to the same T Cell subsets , Chromatin open information belonging to this subgroup （scATAC-seq） And gene expression information （scRNA-seq） Will be divided together .

But it is unreasonable to directly use the characteristics of these two data , Because the characteristics of these two types of data are inconsistent , Therefore, we need to learn technology through certain representations , Get all the samples in the same space （ manifold ） In the vector , Define each cell again （ sample ） The distance between them can be used for subsequent clustering integration .

Single cell data integration is a multi-source data fusion problem , There are many sources , It means more than one experiments Or technology （ batch ）, When these experiments The resulting data has the same characteristics , That is, isomorphic or similar data , That is, the well-known problem of removing batch effect , For example, for different sources （ Sequencing platform , laboratory ） The resulting gene expression profile data , because ” Different sources ” Cause noise , Therefore, it is necessary to perform batch correction on the expression profile data .

Compared with multi-source isomorphic data fusion （ Remove the batch effect ）, Multi source heterogeneous data integration is a more extensive and difficult task . An important problem is how to embed different features of two datasets into the same manifold space , It makes it possible to measure , The corresponding cell to cell distance .

---

integration An important assumption of is ： Even from different sources , Characteristic of different types dataset, Their potential cell subpopulations are roughly the same , So these dataset（ At least part of the information ） It is possible to make connections , Because sharing information about the same object . But at the same time integration And hope to ensure that every dataset Truly unique information can also be preserved , For example dataset A There is a certain cell type that does not belong to dataset B, So in integration, After clustering , These belong only to dataset A Cell types in should not be associated with dataset B Of the cells have any overlap , Otherwise, it is over correction （over-correct）.

Integration of homogeneous data and heterogeneous data , All hope ：

• As close as possible to cells from the same cell subpopulation in different data sets , That is, they are as close as possible in the manifold space we want ;
• As far as possible, specific cell subpopulation information in different data sets should be retained ;

Pay attention to Paper reading notes - utilize Scanorama Efficient integration of heterogeneous single cell transcriptome Heterogeneity in , It is better to refer to the multi-source isomorphic data described in this chapter . A broad sense , The existing scRNA-seq The integration method can also integrate multi omics data , Because we can assume that the dimension of heterogeneous data sets is reduced to embedding The representation of is the same feature space , Then we can use isomorphic integration to integrate these embedding data .

---

Abstract

Despite the emergence of experimental methods to simultaneously measure multiple omics modes in a single cell , But most single-cell datasets contain only one mode . A major obstacle to integrating omics data from multiple modalities is , Different omics data usually have different feature spaces . ad locum , We put forward a proposal called GLUE（graph-linked unified embedding） Computing framework of , The framework bridges the modal gap by explicitly modeling interactions across omics . The benchmark test of the system shows that ,GLUE For single-cell heterogeneous multiomics data , More accurate than the most advanced work 、 More robust and scalable . We will GLUE Applied to a variety of challenging tasks , Including the integration of three groups 、 Regulatory reasoning and the construction of a multiomic human cell map of millions of cells ,GLUE Comments that can correct previous data errors .GLUE Modular design , Flexible expansion and enhancement for new analysis tasks .

Main

Recent technological advances in single cell sequencing have enabled us to mine maps through multiomics data , For example, chromatin accessibility chromatin accessibility（scATAC-seq）,DNA Methylation （snmC-seq,sci-MET） And single cell transcriptome single cell transcriptome（scRNA-seq）, It provides an opportunity to reveal the functions of different cell types . Although recently, there have been methods to analyze multiomics data at the same time , But different omics are usually measured independently , And produce mismatched data , This requires us to develop efficient multi group integration technology .

In calculation , Integrate unpaired multiomics data （ Also known as diagonal integration ） One of the main obstacles is that different omics have different characteristic spaces （ for example ,scATAC-seq The accessible chromatin regions in are related to scRNA-seq Genes in ）. The concise method is to transform multimodal data into a common feature space based on prior knowledge , Then apply the data integration method of single omics . This kind of clear “ Feature conversion ” It's easy , But it often leads to information loss . The algorithm based on coupling matrix decomposition avoids explicit transformation , But it can hardly process more than two omics data . Another option is to match cell data from different omics by nonlinear manifold alignment , This completely eliminates the need for prior knowledge , And it can reduce the loss of information between modes in theory ; However , This technique is mainly applied to data sets with a limited number of cell types and a relatively small number of cells .

The growing amount of data is another serious challenge . Recently developed sequencing techniques can usually obtain millions of cell scale data sets , The current integration method is only applicable to data sets with smaller data volume . In order to keep up with the growth of data volume , The design of the integration method should consider multi-scale .

Here it is , We proposed GLUE（graph-linked unified embedding）, This is a modular framework , It is used to integrate unpaired single cell multiomics data and realize regulatory reasoning at the same time . By explicitly modeling the interactions between the various omics ,GLUE The gap between the specific feature spaces of different omics is bridged in a biological intuitive way . System benchmarks and case studies show that ,GLUE For single cell multiomics data integration is accurate 、 Reliable and scalable . Besides ,GLUE Designed as a general framework , Allow easy expansion in a modular way .

Results

• chart 1：GLUE The architecture of . The unpaired three omics data were recorded as ${\text{X}}_1\in R^{N_1\times |V_1|},{\text{X}}_2\in R^{N_2\times |V_2|},{\text{X}}_3\in R^{N_3\times |V_3|}$, among ,$N_1,N_2,N_3$ Is the number of cells ,$V_1,V_2,V_3$ It is the feature set of each omics ,GLUE Low dimensional learning from each omics data using a omics specific variational self encoder embedding ${\text{U}}_1,{\text{U}}_2,{\text{U}}_3$. Dimensions of raw data and VAE The resulting distribution can remain different across different omics , but embedding Dimensions $m$ It should be shared . To link omics specific data spaces ,GLUE With guidance graph $G=(V,E)$ The form of the takes advantage of prior knowledge , Where nodes $V=V_1\cup V_2\cup V_3$ Is characteristic of different omics . Graph variational self encoder is based on prior knowledge guidance graph（the prior knowledge-based guidance graph） Learning characteristics of the group embedding $\text{V}=({\text{V}}_1^T,{\text{V}}_2^T,{\text{V}}_3^T{)}^T$, Then use this in the data decoder guidance graph, By interacting with cells embedding To reconstruct the data of omics by inner product , And effectively link the specific data space of omics , To ensure consistency embedding Direction . Last , Using the omics discriminator $D$ Aligning cells of different omics through antagonistic learning embedding.${\varphi }_1,{\varphi }_2,{\varphi }_3,{\varphi }_G$ Represents the learnable parameters in the data encoder and the graph encoder .${\theta }_1,{\theta }_2,{\theta }_3,{\theta }_G$ Represents the learnable parameters in the data decoder and the graph decoder .$\psi$ Represents the learnable parameter in the omics discriminator .
• Because it is a graph VAE, Therefore, the output control chart can be used as the result of control reasoning .

Inspired by previous research , We modeled the cell state as a low dimensional cell embedding via variational self coder learning . In view of their inherent differences in biological properties and analytical techniques , Each omics layer is equipped with a separate self encoder , The encoder customizes the probability model for the feature space specific to the omics layer .

Using previous biological knowledge , We recommend using knowledge-based graphs （guidance graph）, Define the regulatory role of features between the modeling cross organizational levels , To link feature spaces specific to the omics layer ; The vertices in the graph correspond to the features of different omics layers , Edges represent the regulatory role between features . for example , When integration scRNA-seq and scATAC-seq Data time , The apex is the gene （gene） And accessible chromatin regions （ namely ATAC peak）, An accessible region can be linked to its putative downstream gene . then , In the figure of the encoder feature embedding Under the guidance of , The multimodal alignment is carried out in the form of iterative optimization .

• chart 2： Consolidated performance .
• a： The score of biological conservatism and the score of omics integration of different integration methods ;
• b： Comprehensive scores of different methods ;
• c： Single cell level alignment error of different methods ;
• d： The performance of integration methods that depend on prior feature relations under different prior knowledge damage rates FOSCTTM Increasing trend ;
• e： Different integration methods on sub sample datasets of different sizes FOSCTTM value ;

• chart 3： The integration of three groups in mouse cortex . Colored by primitive cell types scRNA-seq（a）、snmC-seq（b） and scATAC-seq（c） Of embedding UMAP visualization . And “mPv” and “mSst” The aligned cells are highlighted with green circles . And “mNdnf” and “mVip” Aligned cells are highlighted with dark blue circles . And “mDL-3” The aligned cells are highlighted with light blue circles .
• d： All integrated cells embedding Of UMAP visualization , Colored by the omics layer .
• e： The significance of marker gene overlap for each cell type in all three omics layers .