PCLDA: A cell annotation tool using scRNA-seq data based on simple statistics methods

doi:10.21203/rs.3.rs-3219242/v1

Download PDF

Research Article

PCLDA: A cell annotation tool using scRNA-seq data based on simple statistics methods

https://doi.org/10.21203/rs.3.rs-3219242/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

With the rise of single-cell transcriptome sequencing technology, more and more studies are focusing on single-cell-based disease diagnosis and treatment. Cell type annotation is the first and most critical step in analyzing single-cell genomic data. Traditional marker-genes-based annotation approaches require a lot of domain knowledge and subjective human decisions, which makes annotation time-consuming and generate inconsistent cell identities. In the past few years, multiple automated cell type identification tools have been developed, leveraging large amounts of accumulated reference cells. All these methods are extensions or revisions of vanilla supervised machine learning methods. However, complex models have four potential disadvantages (1) they may require more model assumptions which may not hold in real-world problems, (2) they may involve many model parameters to be tuned, (3) they may be harder to interpret, (4) they may require more computational resources. In this work, we propose PCLDA, a method based on the simplest statistical models, including principal component analysis and linear discriminant analysis, which do not suffer the problems mentioned above. We show PCLDA’s performance is not inferior to the fancier methods in real data. The key message we promote in this work is to use simple statistics if it can solve the problem, avoiding unnecessary complications.

single-cell genomics

single-cell annotation

linear discriminant analysis

simple statistics

Single-cell RNA sequencing (scRNA-seq) is a high-throughput sequencing technology used to study gene expression at the single-cell level ¹. This technology enables the measurement of gene expression profiles of individual cells in a population, rather than the average gene expression of a large group of cells as measured using the older (bulk) RNA-sequencing technology ². Investigating gene expression profiles at the cell level is important because cell types can have distinct gene expression profiles, reflecting their different functions, states, and lineages ³. The scRNA-seq was widely applied to identify the transcriptome characteristics of various cell types within highly organized tissues and to reveal the heterogeneity and dynamics of tissues, organisms, and complex diseases⁴.

Cell type annotation is a process of labelling the cell type of each cell according to its gene expression profile. It is always the first and most crucial step in analyzing scRNA-Seq data. Annotating cell types in single-cell genomic data can provide a comprehensive and accurate catalogue of different cell types present in a particular tissue or organism. This information can be used to study cellular diversity, understand cellular function, and aid in disease diagnosis and treatment. Accurate annotation of single cells has important implications for downstream analyses, such as reconstructing cells’ pseudo trajectories⁵, identifying feature genes that are differentially expressed in each cell type⁶, and revealing unknown cell states for further understanding heterogeneity⁷. However, annotating cell types in scRNA-Seq data is challenging since the data is complex, characterized by high dimensionality, high noise, missing data and sparsity. To address these challenges, many tools for cell type annotation have been developed in recent years, but the scientific community has not yet agreed on a standard workflow for it.

There are two major types of cell type annotation methods: manual annotation and automated annotation. Manual annotation clusters cells according to their gene expression profile and identify feature genes of each cluster which are differentially expressed in each cluster⁸, and then compares these feature genes against well-known reference markers for cell types or common annotation databases such as cellMarker⁹, PanglaoDB¹⁰, and CancerSEA¹¹. It is not only lab-intensive and time-consuming, but also the results are highly relied on the relevant background and effort of researchers. In contrast, automated annotation is a fast and effective method of annotating cells by computer algorithms or prior biological knowledge. Automatic annotation can be divided into three main categories¹²; (1) marker-based methods rely on predefined "marker genes", (2) reference-based or correlation-based methods rely on reference datasets with known cell labels, and (3) supervised classification-based methods build supervised machine learning models using labelled training data, and apply the model to classify cells in test data, which requires training and test data are similar to each other.

Supervised classification methods are the most popular among all automated annotation methods due to their outstanding performance. Many general-purpose classification algorithms in machine learning were revised to develop cell annotation methods. For example, scANVI¹³ is based on kNN classifier¹⁴; scClassify¹⁵ is based on weighted kNN classifier; SingleCellNet¹⁶ is based on random forest¹⁷; Garnett¹⁸ is based on the elastic net method¹⁹; and scID²⁰ is a revised linear discriminant analysis (LDA) method²¹. More details about these automated annotation methods could refer to a recent review paper ²². Benchmark studies were conducted to compare the performance of these annotation methods ²³. These annotation tools modified the general-purpose classification algorithms to address particular challenges in annotating cells from single-cell genomic data, making the models more complex, which requires additional model assumptions that may not hold in real-world data. So, vanilla classifiers might outperform these modified classifiers in practice. This motivates us to consider using the simplest statistical models to develop a cell annotation tool, which may not outperform fancy methods in all data, but is expected to be more reliable due to its simplicity.

In this work, we aim to promote “simple statistics”. Simple statistics refers to the idea of using simple statistics to solve real and important problems, avoiding unnecessary complications in methods. We prefer simple methods; since, compared to complex methods, they often require fewer model assumptions, have fewer model parameters to be tuned, are easier to interpret and less computationally expensive, and are often more reliable (cross different data sets in real world) in their performance. So, we always prefer the simpler method when two methods have similar performances. This principle is well accepted in statistician communities. For example, several world-leading biostatisticians created a blog called “Simply Statistics” (https://simplystatistics.org/). They use this website to post ideas, discuss science/popular writing, link to articles that inspire us, and share advice with up-and-coming statisticians. Occam's razor²⁴, also known as the principle of parsimony or the law of parsimony (Latin: lex parsimoniae), is the problem-solving principle that "entities should not be multiplied beyond necessity". Particularly, in cell annotation methods, we found scID (modified LDA for single-cell data) is always outperformed by its vanilla version (LDA) in our experiments.

Cao ²⁵ conducted a benchmark study on binary cell phenotype classification using vanilla machine learning methods. They found LDA is much faster than all competitors, and its classification performance is only slightly worse than the top performers. LDA has a closed-form solution which is why it outperforms all competitor classifiers in speed. LDA map high dimensional predictors (tens of thousands of genes) into a low dimensional space (number of cell types minus one) and draw a linear boundary to define classes ²⁶. While some classifiers are designed for binary classification, LDA naturally handles multi-class classification problems. Good performance reported in the literature and the above-discussed good features motivate us to explore the performance of vanilla LDA for cell annotation and its combination with other simple statistical methods.

To this end, we proposed PCLDA, consisting of three simplest statistical methods. First, we used t-statistics for marginal gene screening to retain a set of (e.g., 1000) high rank genes. Then we employed principal component analysis (PCA) to further reduce dimension of the transcriptome data. Lastly, the PCA scores were utilized as input to train the LDA classifier. To illustrate the performance of the PCLDA, we compared its cell annotation results with several competitors on 10 widely used scRNA-seq datasets and 10 pairs (i.e., training and testing) of scRNA-seq data. We found that our simple statistical model showed robust performance with top tier accuracy.

1. Datasets

We compiled 10 real scRNA-seq datasets from previously published studies. These datasets differ by species, tissue contents and sequencing protocols (i.e. plate-based and droplet-based single-cell platforms). We used their annotated cell types as the ground truth for downstream analysis. The details of each dataset used in this study were listed as follows:

Mouse Pancreas:

The processed mouse pancreas dataset was generated by Baron et al. ²⁷ and downloaded from the GEO database with the accession number GSE84133. It consisted of 1886 cells and 14878 genes, and included 13 annotated cell types.

Human Pancreas:

The human pancreas dataset was collected from five studies^{27,28,29,30,31}. The processed scRNA-seq gene expression matrices would be found with their accession numbers of GSE84133, GSE85241, GSE81608, GSE83139, and GSE86469 in the GEO database, respectively. These datasets have 8569, 2285, 1492, 457, and 638 cells, 20125, 19054, 39851, 19950, and 19054 genes as well as 14, 13, 7, 8, and 13 cell types, respectively.

Tabula Muris(TM):

The Tabula Muris (TM)³² project generated two datasets of single-cell transcriptome from 20 mouse organs by using two different platforms (i.e. microfluid droplet-based 10X and fluorescence-activated cell sorting-based SMART-Seq2). These two mouse datasets contain 24622 and 20000 cells, 22252 and 17866 genes, and 37 and 32 cell types, respectively. The datasets were downloaded through the Chan Zuckerberg Biohub https://tabula-muris.ds.czbiohub.org/. Besides all organs, single cells of lung tissue were widely used for single cell annotation analysis. It includes 1563, 1303 cells, corresponding to 10 and 8 cell types, respectively. In this work we used both the full data and its lung subset.

PBMC:

The PBMC dataset comprises four human PBMC scRNA-seq data from 10X genomics with different versions or generations of the protocols and was originally generated by Zheng et al.³³ These datasets could be downloaded from the 10X website directly: https://support.10xgenomics.com/single-cell-gene-expression/datasets/2.1.0/pbmc8k, https://support.10xgenomics.com/single-cell-vdj/datasets/2.2.0/vdjv1hspbmc5gex, https://github.com/10XGenomics/single-cell-3prime-paper and https://support.10xgenomics.com/single-cell-gene-expression/datasets. It consisted of 8089, 7378, 91649, and 2467 cells for these four datasets, respectively.

Mouse Brain:

The mouse brain datasets include the primary visual cortex (PVC) dataset by Tasic et al.³⁴, and the hypothalamic arcuate-median eminence complex (HArc-ME) dataset by Campbell et al.³⁵, which can be downloaded from NCBI’s GEO database with the accession number GSE71585 and GSE93374. It consisted of 1727 and 21086 cells, with 11 and 6 cell types for these two datasets.

A summary of the datasets used is shown in Table 1.

Table 1 Datasets used for benchmarking comparisons.

Comparison	Training/Testing	Study	Organism and tissue	Protocol	No. of cells	No.of cell types	Accession
Cross Validation	NA	Zheng et al	FACS-sorted PBMC	10X	2467	6	10X Genomics
		Zheng et al	PBMC	10X	91649	7	10X Genomics
		Wang et al	Human pancreas	SMARTer	457	8	GSE83139
		Tabula et al	Mouse Lung	10X	1563	10	GSE109774
		Tabula et al	Tabula Muris full	Smart-Seq2	24622	37	GSE109774
		Campbell et al	Mouse hypothalamic arcuate	Drop-seq	21086	11	GSE93374
		Segerstolpe et al	Human pancreas	SMART-seq2	3514	15	E-MTAB-5061
		Muraro et al	Human pancreas	CEL-seq2	2285	13	GSE85241
		Baron et al	Human pancreas	inDrop	8569	14	GSE84133
Cross Sample	Reference	Wang et al	Human pancreas	SMARTer	457	8	GSE83139
	Query	Xin et al	Human pancreas	SMARTer	1492	7	GSE81608
	Reference	Zheng et al	Human PBMC	10X	8089	9	10X Genomics
	Query	Zheng et al	Human PBMC	10X	7378	9	10X Genomics
	Reference	Zheng et al	PBMC	10X	91649	7	10X Genomics
	Query	Zheng et al	FACS-sorted PBMC	10X	2467	6	10X Genomics
Cross Platform	Reference	Muraro et al	Human pancreas	CEL-seq2	2285	13	GSE85241
	Query	Lawlor et al	Human pancreas	FluidigmC1	638	13	GSE86469
	Reference	Baron et al	Human pancreas	inDrop	8569	14	GSE84133
	Query	Xin et al	Human pancreas	SMARTer	1492	7	GSE81608
	Reference	Muraro et al	Human pancreas	CEL-seq2	2285	13	GSE85241
	Query	Baron et al	Human pancreas	inDrop	8569	14	GSE84133
	Reference	Muraro et al	Human pancreas	CEL-seq2	2285	13	GSE85241
	Query	Xin et al	Human pancreas	SMARTer	1492	7	GSE81608
	Reference	Campbell et al	Mouse hypothalamic arcuate	Drop-seq	21086	11	GSE93374
	Query	Tasic et al	Mouse primary visual cortex	SMART-seq	1727	6	GSE71585
	Reference	Tabula et al	Tabula Muris full	Smart-Seq2	24622	37	GSE109774
	Query	Tabula et al	Tabula Muris full	10X	20000	32	GSE109774
	Reference	Tabula et al	Mouse Lung	Smart-Seq2	1563	10	GSE109774
	Query	Tabula et al	Mouse Lung	10X	1303	8	GSE109774

2. A scRNA-seq annotation pipeline based on simplest statistical models

To illustrate the utility of simple statistics in the task of automated single-cell annotation, we adopted a PCA analysis followed by a traditional LDA classifier as the key components of our analysis pipeline. There are three main steps in our pipeline/workflow. We first normalized the datasets by performing Counts Per Million (CPM) for UMI counts-based datasets, and Transcripts Per Million (TPM) for full-length based datasets. We then applied a t-test and used the t-statistics to conduct the gene screening where cell-type specific gene markers were identified. In this screening, we ignore p-values and statistical significance and only use the rank of t-statistics to remove low rank genes and still keep a lot of gene. This step only removes genes surely useless for cell annotation to increase signal-to-noise ratio for downstream analysis.

In second step, we further performed a dimension reduction with the selected cell-type specific gene markers using PCA. On the one hand overfitting of the prediction algorithm can be prevented by further reducing the dimensionality of the training dataset by using PCA. On the other hand, PCA can further leverage multicollinearity by combining highly correlated variables into a set of uncorrelated variables. Furthermore, as in most cases, when applying automatic classification models to single cell annotation on unknown cells, the training dataset and testing dataset would be available. This motivated us to conduct PCA on pooled data from two genomic data, which is different from traditional model approach. That is the latter usually performs PCA on training dataset only (regardless of the testing datasets). By conducting PCA on pooled data, we could better capture the share structure between the two datasets and thus improve the annotation performance.

Next, we build the LDA model based on the output of PCA and the built LDA model will be used as a classifier for automatic cell annotation. In brief, LDA aims to find an optimal linear combination of features, in high-dimensional space, that could best separate two or more classes of samples. To accomplish this, LDA projects the data onto a lower-dimensional space, aiming to increase the separation between classes (between-class scatter) and decrease the variation within each class (within-class scatter). It thus implements both dimension reduction and linear classification simultaneously.

3. Performance evaluation

We reported the accuracy for the main comparison metric where we defined accuracy as the proportion of cells being correctly predicted (i.e. predicted label is consistent with its ground-truth label) over all cell types. To evaluate the performance, we designed both intra-dataset annotation and inter-dataset annotation comparisons. For the intra-dataset annotation, five-fold cross-validation was implemented, while for the inter-dataset annotation, cross-platforms comparisons were executed.

4. Comparison with other automated annotation approaches

We compared our PCLDA model with the other nine counterparts regarding the accuracy. We are particularly interested in comparisons with the scID where revised LDA was used. We adopted the default parameters to run all the competing methods.

We first explore the importance of the PCs involved in our frameworks. We implemented the workflow with PCs of 20, 50 to 500 with step of 50, and applied them to all the datasets. We noted that the models would achieve relatively stable performance when using 50 PCs or more (Fig. 1A). Particularly, for the pancreas datasets, the accuracy was 0.94 with 50 PCs, slightly increasing to 0.945 with more than 100 PCs. Similar observations were observed on PBMC and TM datasets. For instance, an accuracy of 0.88 was achieved with more than 50 PCs compared with 0.86 for the model with 20 PCs for the PMBC dataset. For the TM dataset, an improvement from 0.91 to 0.935 was observed when comparing the model with 50 PCs and the one with 20 PCs, while relatively stable results of 0.94 were reported when increasing the PCs to 100 and more. Thus, to enhance computational efficiency and adopt a more concise model, we have selected 50 as the default parameter in our approach, prioritizing speed and simplicity.

We further wondered the contribution of each step or each component in our PCLDA model. To do so, for different datasets, we used four different combinations to compare their accuracy (Fig. 1B): fitting LDA model only; T-test based gene screening and LDA model, PCA downscaling and LDA model, and as well as the PCLDA (i.e. combination of gene screening, PCA downscaling, and LDA model). We noted that the accuracy of PCLDA is the highest and most stable. Particularly, when fitting the LDA model only, several datasets were failed due to the memory issue. In addition, the collinearity issues were observed for some datasets when fitting the LDA model after gene screening but without PCA (i.e., Gene Screening + LDA in Fig. 1B), which consequentially leads to worse performance.

We then compared the classification accuracy of PCLDA with the other nine methods on the above-mentioned scRNA-seq datasets (Table 1). These benchmarking methods include two statistical metrics-based methods (SingleR³⁶, Scmap³⁷), two tree-based methods (CHETAH³⁸ and scClassify), three machine-learning-based methods (Granatt, SingleCellNet, scID, SCINA³⁹) and a semi-supervised learning based method (Seurat)⁴⁰ (Table 2 for the detailed models). We used overall accuracy as a performance measurement metric.

Table 2

Compared machine learning models used for cell annotation.
Tool name	Language	Computational approach
Seurat	R	weighted-nearest neighbour
SingleR	R	Spearman
CHETAH	R, Shiny app	Spearman + confidence
SCINA	R	Bimodal distribution fitting to marker genes
singleCellNet	R	Random Forest
scID	R	LDA
Garnett	R	Elastic Net
scClassify	R, Shiny app	Weighted kNN classifier
scmap	R	Cosine distance based kNN

1. Intra-dataset annotation and performance comparison

We first tested the classification accuracy of ten methods (Table 2) on nine publicly available scRNA-seq datasets (Table 1). These datasets include two peripheral blood mononuclear cell (PBMC) datasets, four human pancreatic islet datasets, two Tabula Muris datasets ( TM full and TM lung), and one mouse brain dataset. To avoid potential bias, we used the 5-fold cross validation scheme to measure the averaged accuracy. The detailed comparison results were shown in Fig. 2. It demonstrated that most of the methods consistently showed strong performance across all datasets. Notably, PCLDA, Seurat, and singleR achieved remarkable average accuracies of 0.95, 0.96, and 0.945, respectively. In contrast, some methods, like scID, Garnett, and scmap, exhibited poor performance. Additionally, some methods, such as SCINA, CHETAH and SingleCellNet, perform decent on most datasets but struggled with the TM (full) data. We noted that the full TM dataset from the Smart-Seq2 platform contains 37 cell types (Table 1) and certain cell types have a very limited number of cells, making the classification task more challenging. As a results, those methods struggled to achieve high performance on the TM full datasets due to the increased complexity posed by these sparse cell type labels. The Garnett model even failed to train the model. Overall, in the intra-dataset cross-validation scenario, the PCLDA is top ranked in most of the datasets and has high stability with basically no outliers.

2. Inter-dataset annotation and performance comparison

To evaluate the annotation tools in a more realistic setting, we conducted inter-dataset performance evaluation on ten datasets/ten pairs. As shown in Table 1, three of the ten model testing scenarios are for the same protocol, and seven are for different protocols/platforms.

We performed within-protocol prediction to show how good the models could achieve when the query datasets were generated using the same protocol as the reference datasets, which would be the most common application. The PCLDA model has ranked the top 2 highest accuracies in all three datasets compared to other methods, with an average accuracy of 0.91, while the best competing methods have an average accuracy of 0.92 (Seuart) and 0.9 (SingleCellNet), respectively (Fig. 3). Particularly, when dealing with large PBMC datasets, which have 90k + cells, the PCLDA achieves a very similar performance to the top one, indicating that PCLDA is very effective in handling large data.

We then evaluated the model performance by comparing them on cross-protocol datasets. While not all the models perform well on all the datasets, a few methods (e.g. scID and Granett) even showed poor performance in most comparisons with accuracies lower than 0.7. In contrast, the PCLDA model consistently achieved decent performance and demonstrated the top tier rank in all seven comparisons, along with Seurat and SingleR. The average accuracy for PCLDA, Seurat and SingleR were 0.95, 0.95 and 0.946, respectively. Notably, when comparing models on lung tissue from the TM data, which contained only 10 cell types and had a relatively uniform distribution of cells among different cell types, the PCLDA stood out with the best performance, achieving an accuracy of 0.99. In contrast, some other methods failed to perform adequately in this specific scenario.

It is also worth mentioning that it is very challenging to predict the cell types for the TM full data. As we motioned in intra-dataset scenario, it has 37 cell types in TM full training (reference) set, many of which contain only a few cells, and the gene expressions within these cell types are highly correlated. Despite of these challenges, the PCLDA method exhibited significant improvement on this dataset with highest accuracy of 0.94. In contrast, many other methods struggled to perform well, underscoring the advantage of our approach in effectively distinguishing between highly similar cell populations.

Altogether, these results across the 19 benchmark datasets in the three scenarios demonstrated that the performance of PCLDA is very robust and stable. The model’s performance remains consistently high for both intra-dataset prediction (cross-validation scenario) and inter-dataset prediction (cross-sample and cross-protocol scenarios).

In this study, we have implemented a workflow for single cell annotation using simple statistics. When comparing with other complex statistical and machine learning models, the simple statistics model achieved comparable performance to the top ones for cell annotation tasks with reference and query datasets from both the sample protocol and different protocols. In most cases, annotating cell types using reference data from the same protocol would be ideal. Our results showed that our model is also robust to cross protocol cell annotations where same protocol reference data is not available, supporting our hypothesis that simple models generally would be reliable.

As mentioned in the above section, our model contains three main steps: gene screening, PCA downscaling, and fitting LDA classification model. These are the relatively simple and necessary steps that we have validated. The fist step gene screening is to select significantly expressed genes, not simply to select genes with large variance. Without performing gene screening and directly perform PCA downscaling with raw data, genes with large variance will be more likely to be selected in the top PCs, but such PCs may not be helpful to distinguish cell types. Because some genes have large variance in all cell types, in this case the larger the variance is instead the greater the noise, and gene screening can avoid this risk. This step thus helps us to remove unrelated genes and focus on only a few hundred genes, which allows future work to save money and time by using targeted sequencing techniques. The second step PCA can further reduce the data dimensionality and extract the signal from data. Moreover, PCA takes advantage of multicollinearity and combines the highly correlated variables into a set of uncorrelated variables. Therefore, PCA can effectively eliminate multicollinearity between features. Lastly, in our approach, by combining a linear boundary of LDA and loading matrix of PCA, we learn a linear combination of genes as boundary, the absolute value of weights can interpret how important a gene is for the task of cell annotation.

The models will reach stable performance after providing a certain big number of top PCs which capture most of the cell type variation information, indicating its robustness once including sufficient feature genes and top variable PCs. LDA map high-dimensional genomics data into K-1 dimensions, where K is the number of cell types. Thus, choosing relatively larger number of feature genes and PCs would ensure its performance. On the other hand, we also suggest not to use too many PCs for the objective of being parsimonious.

We observed that several machine learning models showed quite variable performance on different datasets. For instance, elastic net based model (Garnett) didn’t show good performance on most of the datasets might be mainly due to its limited pre-selected marker genes. The scID approach, which is weighted LDA based, performed badly on most of the cross-protocol comparisons, indicating add more model parameters/assumptions doesn’t necessary improve the performance on simple tasks. We further noticed that when the number of cell types increased, most of the competitors failed to annotate cells accurately.

In summary, we found that the simple statistics methods can be applied to most real-word applications since it is not only fast, but also has reliable accuracy. In solving cell type annotation problems, we could consider the most basic machine learning algorithm as a preferred method over fancy methods. Through this study, we'd like to promote simple methods to our research community. We always prefer simple methods if they can solve real-world problems. Such methods are often faster, have fewer model assumptions (hence more reliable), and provide easier-to-interpret results.

Ethics approval and consent to participate

Not applicable

Consent for publication

Not applicable

Availability of data and materials

The datasets used and/or analysed during the current study are mainly obtained from published studies as indicated in the Method section and are available from the corresponding author on reasonable request.

Competing interests

The authors declare that they have no competing interests.

Funding

This work is supported in part by funds from Genome BC SIP7 (X.Z., L.X.), NRC Digital health and geospatial analytics program (X.S., L.X., X. Z.), the Natural Sciences and Engineering Research Council Discovery Grants (RGPIN-2021-03530 L.X.), and the Canada Research Chair # CRC-2021-00232 (X.Z.).

Authors' contributions

X.S. and X.Z. contributed to the study concept and design. K.B. collected the datasets, data processing and data analysis. K.B. wrote the first draft with input from X.S. and X.Z. X.Z. and X.S. supervised this research. All authors read and approved the final manuscript.

Acknowledgment

This research was enabled in part by support provided by WestGrid (www.westgrid.ca) and Compute Canada (www.computecanada.ca).

Nawy T. Single-cell sequencing. Nat Methods 2014;11(1):18–18; doi: 10.1038/nmeth.2771.
Gawad C, Koh W, Quake SR. Single-cell genome sequencing: current state of the science. Nat Rev Genet 2016;17(3):175–188; doi: 10.1038/nrg.2015.16.
Svensson V, Natarajan KN, Ly L-H, et al. Power analysis of single-cell RNA-sequencing experiments. Nature methods 2017;14(4):381–387; doi: 10.1038/nmeth.4220.
Jovic D, Liang X, Zeng H, et al. Single‐cell RNA sequencing technologies and applications: A brief overview. Clin Transl Medicine 2022;12(3):e694; doi: 10.1002/ctm2.694.
Rocha EL da, Rowe RG, Lundin V, et al. Reconstruction of complex single-cell trajectories using CellRouter. Nat Commun 2018;9(1):892; doi: 10.1038/s41467-018-03214-y.
Porcu E, Sadler MC, Lepik K, et al. Differentially expressed genes reflect disease-induced rather than disease-causing changes in the transcriptome. Nat Commun 2021;12(1):5647; doi: 10.1038/s41467-021-25805-y.
Baker WL, White CM, Cappelleri JC, et al. Understanding heterogeneity in meta‐analysis: the role of meta‐regression. Int J Clin Pract 2009;63(10):1426–1434; doi: 10.1111/j.1742-1241.2009.02168.x.
Satija R, Farrell JA, Gennert D, et al. Spatial reconstruction of single-cell gene expression. Nat Biotechnol 2015;33(5):495–502; doi: 10.1038/nbt.3192.
Zhang X, Lan Y, Xu J, et al. CellMarker: a manually curated resource of cell markers in human and mouse. Nucleic Acids Res 2018;47(Database issue):gky900-; doi: 10.1093/nar/gky900.
Franzén O, Gan L-M, Björkegren JLM. PanglaoDB: a web server for exploration of mouse and human single-cell RNA sequencing data. Database 2019;2019:baz046; doi: 10.1093/database/baz046.
Yuan H, Yan M, Zhang G, et al. CancerSEA: a cancer single-cell state atlas. Nucleic Acids Res 2018;47(Database issue):gky939-; doi: 10.1093/nar/gky939.
Abdelaal T, Michielsen L, Cats D, et al. A comparison of automatic cell identification methods for single-cell RNA sequencing data. Genome Biol 2019;20(1):194; doi: 10.1186/s13059-019-1795-z.
Xu C, Lopez R, Mehlman E, et al. Probabilistic harmonization and annotation of single‐cell transcriptomics data with deep generative models. Mol Syst Biol 2021;17(1):e9620; doi: 10.15252/msb.20209620.
Zhang Z. Introduction to machine learning: k-nearest neighbors. Ann Transl Medicine 2016;4(11):218–218; doi: 10.21037/atm.2016.03.37.
Lin Y, Cao Y, Kim HJ, et al. scClassify: hierarchical classification of cells. Biorxiv 2019;776948; doi: 10.1101/776948.
Tan Y, Cahan P. SingleCellNet: A Computational Tool to Classify Single Cell RNA-Seq Data Across Platforms and Across Species. Cell Syst 2019;9(2):207-213.e2; doi: 10.1016/j.cels.2019.06.004.
Breiman L. Random Forests. Mach Learn 2001;45(1):5–32; doi: 10.1023/a:1010933404324.
Pliner HA, Shendure J, Trapnell C. Supervised classification enables rapid annotation of cell atlases. Nat Methods 2019;16(10):983–986; doi: 10.1038/s41592-019-0535-3.
Mol CD, Vito ED, Rosasco L. Elastic-net regularization in learning theory. J Complexity 2009;25(2):201–230; doi: 10.1016/j.jco.2009.01.002.
Boufea K, Seth S, Batada NN. scID Uses Discriminant Analysis to Identify Transcriptionally Equivalent Cell Types across Single-Cell RNA-Seq Data with Batch Effect. Iscience 2020;23(3):100914; doi: 10.1016/j.isci.2020.100914.
Fletcher JM, Rice WJ, Ray RM. Linear Discriminant Function analysis in Neuropsychological Research: Some Uses and abuses. Cortex 1978;14(4):564–577; doi: 10.1016/s0010-9452(78)80031-8.
Pasquini G, Arias J, Schäfer P, et al. Automated methods for cell type annotation on scRNA-seq data. Comput Struct Biotechnology J 2021;19:961–969; doi: 10.1016/j.csbj.2021.01.015.
Abdelaal T, Michielsen L, Cats D, et al. A comparison of automatic cell identification methods for single-cell RNA sequencing data. Genome Biol 2019;20(1):194; doi: 10.1186/s13059-019-1795-z.
Domingos P. The Role of Occam’s Razor in Knowledge Discovery. Data Min Knowl Disc 1999;3(4):409–425; doi: 10.1023/a:1009868929893.
Cao X, Xing L, Majd E, et al. A Systematic Evaluation of Supervised Machine Learning Algorithms for Cell Phenotype Classification Using Single-Cell RNA Sequencing Data. Frontiers Genetics 2022;13:836798; doi: 10.3389/fgene.2022.836798.
James G, Witten D, Hastie T, et al. An Introduction to Statistical Learning. (second. ed). Springer: New York; 2013.
Baron M, Veres A, Wolock SL, et al. A Single-Cell Transcriptomic Map of the Human and Mouse Pancreas Reveals Inter- and Intra-cell Population Structure. Cell Syst 2016;3(4):346-360.e4; doi: 10.1016/j.cels.2016.08.011.
Muraro MJ, Dharmadhikari G, Grün D, et al. A Single-Cell Transcriptome Atlas of the Human Pancreas. Cell Syst 2016;3(4):385-394.e3; doi: 10.1016/j.cels.2016.09.002.
Xin Y, Kim J, Okamoto H, et al. RNA Sequencing of Single Human Islet Cells Reveals Type 2 Diabetes Genes. Cell Metab 2016;24(4):608–615; doi: 10.1016/j.cmet.2016.08.018.
Wang YJ, Schug J, Won K-J, et al. Single-Cell Transcriptomics of the Human Endocrine Pancreas. Diabetes 2016;65(10):3028–3038; doi: 10.2337/db16-0405.
Lawlor N, George J, Bolisetty M, et al. Single-cell transcriptomes identify human islet cell signatures and reveal cell-type–specific expression changes in type 2 diabetes. Genome Res 2017;27(2):208–222; doi: 10.1101/gr.212720.116.
Schaum N, Karkanias J, Neff NF, et al. Single-cell transcriptomics of 20 mouse organs creates a Tabula Muris. Nature 2018;562(7727):367–372; doi: 10.1038/s41586-018-0590-4.
Zheng GXY, Terry JM, Belgrader P, et al. Massively parallel digital transcriptional profiling of single cells. Nat Commun 2017;8(1):14049; doi: 10.1038/ncomms14049.
Tasic B, Menon V, Nguyen TN, et al. Adult mouse cortical cell taxonomy revealed by single cell transcriptomics. Nat Neurosci 2016;19(2):335–346; doi: 10.1038/nn.4216.
Campbell JN, Macosko EZ, Fenselau H, et al. A molecular census of arcuate hypothalamus and median eminence cell types. Nat Neurosci 2017;20(3):484–496; doi: 10.1038/nn.4495.
Aran D, Looney AP, Liu L, et al. Reference-based analysis of lung single-cell sequencing reveals a transitional profibrotic macrophage. Nat Immunol 2019;20(2):163–172; doi: 10.1038/s41590-018-0276-y.
Kiselev VY, Yiu A, Hemberg M. scmap: projection of single-cell RNA-seq data across data sets. Nat Methods 2018;15(5):359–362; doi: 10.1038/nmeth.4644.
de Kanter JK, Lijnzaad P, Candelli T, et al. CHETAH: a selective, hierarchical cell type identification method for single-cell RNA sequencing. Nucleic Acids Res 2019;47(16):e95–e95; doi: 10.1093/nar/gkz543.
Zhang Z, Luo D, Zhong X, et al. SCINA: Semi-Supervised Analysis of Single Cells in Silico. Genes-basel 2019;10(7):531; doi: 10.3390/genes10070531.
Hao Y, Hao S, Andersen-Nissen E, et al. Integrated analysis of multimodal single-cell data. Cell 2021; doi: 10.1016/j.cell.2021.04.048.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

PCLDA: A cell annotation tool using scRNA-seq data based on simple statistics methods

Status:

Version 1

Abstract

Figures

Introduction

Materials and Methods

1. Datasets

2. A scRNA-seq annotation pipeline based on simplest statistical models

3. Performance evaluation

4. Comparison with other automated annotation approaches

Results

Discussion

Conclusions

Declarations

References

Additional Declarations

Status:

Version 1