The front-end of SingleCAnalyzer has been designed to provide a means to fully analyse scRNA-Seq data using the following two steps: 1) Setting input files and analysis parameters and 2) cluster determination and the execution of comparative analysis. In the first step, FASTQ/HDF5 files are uploaded or an ENA project identifier is provided by the user. Basic information regarding the species studied and type of sequencing performed, as well as optional parameters for the alignments of sequences can also be specified on the web. Once the files are uploaded, demultiplexed and aligned, users may perform further analysis including feature selection, empty droplet deletion, dimensional reduction, prediction of cellular type, analysis of trajectories/pseudotime and unsupervised clustering. These analyses can be performed and adjusted by selecting parameters in the ‘analysis parameters’ section (Figure 1).

Cluster determination and the execution of comparative analysis is accomplished through the website, which provides an interactive interface that allows the user to visualise cellular type prediction, pseudotime predictions or clustering results via six interconnected scatterplots generated using each dimensionality reduction technique. Point colour and type can be changed according to each analysis results. Users can also generate new representations of gene pairs and colour the points based on gene expression values. This interface specifies the most adequate aggrupation, cellular classification or time frame and is guided by the user’s knowledge regarding the samples studied. On the interface, the user can also launch a comparative analysis of all groups, or manually determine which groups should be compared. The comparative analysis includes an analysis of differentially expressed groups of genes, and the functional analysis of gene ontology categories and pathways.

Results are displayed in tabular form, which reveal the execution status of each computational process and provide a link to final results. These are provided as static reports and interactive web pages. Results regarding the quantification of gene expression values are provided with a table of quantification statistics and downloadable files that contain information for aligned reads regarding the number of reads generated per transcript and the number of transcripts per million (TPM). The quality control page descriptively reveals the distribution patterns of expression using box plots, reveals estimated numbers of expressed genes using a bar plot and represents the first two components of a PCA analysis. The clustering results page integrates dimensionality reduction, clustering, pseudotime and cellular classification results within self-explanatory interface which can also generate static reports that incorporate user modifications and launch comparisons between groups. Reports containing results are generated for each comparison, which include differential expression, functional enrichment and GSEA analysis. Differential expression results are summarised in a table which is linked to the following means to visualise data: MA plot, volcano plot, box plot, line chart and heatmap. Functional analyses are also summarised in tables and interactive visual means to represent data such as bar plots, networks and symmetric heatmaps are provided.

Supplementary Table S1 shows a comparison with 12 scRNA-Seq analysis platforms. The main features of the SingleCAnalyzer website are:

- scRNA-Seq analysis from raw FASTQ, HDF5 files or ENA project identifications

Figure 2 shows the analysis pipeline of SingleCAnalyzer. It integrates generally accepted tools used for the analysis of RNA-Seq data, which also perform well as computational resources. Supplementary Table S2 shows the computational time required to analyse nine scRNA-Seq public data sets. The complete analysis of 154 demultiplexed samples takes an average of 36 min, which allows for the real time analysis of low cell number scRNA-Seq experiments. The most time-consuming processes in the pipeline involves the upload, demultiplexing and alignment of samples, which are tasks that are performed in parallel. This parallelisation reduces the global analysis time by 57%, which makes the time requirement of our cloud infrastructure equal to virtual machine or local pipelines solutions. Moreover, SingleCAnalyzer does not store raw sequences or aligned files in order to avoid user disk space limitations, and the number of analysed samples of non-commercial cloud platforms.

The next steps of the pipeline include feature selection, empty droplets detection, dimensionality reduction, cellular type prediction, trajectory/pseudotime analysis and unsupervised clustering. SingleCAnalyzer applies gene filtering, which is based on user input parameters to avoid non-informative output, noise or drop out events. Afterward, six dimensionality reduction methods are applied to the data and samples are visualised using interactive scatter plots. Simultaneously, four unsupervised clustering algorithms are applied to produce nine possible clustering divisions for each method, a cellular type prediction method is executed for each training dataset, and a pseudotime analysis is performed (see methods). These cluster types can be mapped on interactive plots at the request of the user.

Based on the unsupervised or manually curated clusters produced, users can identify gene characteristics and the functions of each group by launching comparison analysis. This feature incorporates the differential expression analysis of groups and the functional analysis of gene ontologies and pathways. The analysis pipeline also processes quality control, clustering, differential expression and functional analysis results, and integrates them in an interactive and self-explanatory web interface.

SingleCAnalyzer was conceived as an agile project, and new scRNA-Seq analysis methods can be integrated within its analysis pipeline. Only generally accepted methods that have been demonstrated to generate reliable and reproducible results that require reasonable quantities of computational resources will be considered for addition to our cloud platform. The increasing development of computational methods will inspire the adaptation of the platform to meet the needs of researchers as scientific trends regarding scRNA-SEQ data analysis emerge.

Visualization is a key aspect on the interpretation of scRNA-Seq results (Cakir et al., 2020). Analysis pipelines performs scatter plots for the representation of dimensional reduction results where point colors represent clusters, cell types, gene expression or trajectory features of each cell (Kiselev et al., 2017; Lin et al., 2017; Stuart et al., 2019). These plots are adequate for publishing results, but not for explorative analyses. At present, new technologies based on JavaScript enable the generation of interactive graphs in a Web User Interface. They allow the connection between graphs and the use of HTML5 components which could control visualization aspects. SingleCAnalyzer adopts this technology to visualize information, interconnect graphs, show meta-information, calculate descriptive statistics, generate new graphs under user request and change the representation features interactively. SingleCAnalyzer includes six different graphical representations such as scatter plot, bar chart, heatmap, network, boxplot and density plot. These graphs are interactive, and the user can modify them by clicking on tables, html controls or other graphs.

The central result page is the representation of dimensional reduction and clustering of cells. It is composed by scatter plots where points represent cells, and the user can select the color and shape of points manually or by using clustering, cell population, pseudotime or gene expression results. The user can also explore group frequencies and define resulting groups based on meta-information or cell disposition on the graph. All graphs are interconnected, changes on graphical attributes or cell selections are synchronized on all displays. Cells can be located in all the graphs with a selection over one graph or by means of the locate samples menu. The application also allows the generation of new scatter plots which represent the expression of two genes in each cell.

Once the user defines the groups, he can launch a comparative expression analysis which results in two types of interactive reports. One is the differential expression report which are composed of interactive scatterplots, a boxplot, a line plot and a heatmap. All these graphs show information on mouse action and are connected with the table which summarizes the statistical analysis. They enable the comprehensive exploration of results and the query of information about expression changes of genes. The other report is the functional analysis which includes self-explanatory graphs such as bar plots, networks of terms and triangular heatmaps. Networks and heatmaps represent relations between gene sets which helps in the identification of related gene functions or pathways, while the bar plot shows the number of observed versus expected genes in each category.

Visualization features of SingleCAnalyzer enable the exploration and interpretation of results in an integrated platform which covers the main steps of scRNA-Seq analysis. The platform was presented and discussed at the VIZBI21 conference, where some improvements were suggested by attendants (VIZBI, 2022). Suggestions were focused on improving the usability of the platform and the adaptation of the analysis pipeline for their objectives. For example, an attendant required an adaptation for the analysis of RNA-Seq data which was developed and can be executed disabling the multiplexing process. SingleCAnalyzer is also distributed as a Docker machine and our graphical functions will be made public as R packages for its open use in analysis pipelines. All the representations performed with SingleCAnalyzer can be downloaded as graphical files ready for its inclusion in publications and analysis reports.

The SingleCAnalyzer website runs using LAMP architecture (Linux, Apache, MySQL and PHP). The front-end of the website was developed using PHP, HTML5, JavaScript, D3, JQuery, AJAX and CSS3. Its implementation was based on the RaNA-Seq project, which contains similar alignment, differential expression and functional analysis tools (Prieto and Barrios, 2019). The analysis pipeline can be executed by a task manager that runs the analysis processes using R, Python or Linux Bash Shell. It also balances the computational load on our high-performance computing cluster. The analysis pipeline integrates cutting-edge tools which rapidly and reliably analyse scRNA-Seq data. Figure 2 shows a flowchart of the pipeline used. We have optimised the analysis processes in our pipeline by harnessing computational clustering. Most of the tasks of analysis can be executed in real time. This optimisation has facilitated the development of an open and free cloud-based system.

Raw sequence files in FASTQ format can be demultiplexed with Alevin software (version 1.3.0) (Srivastava et al., 2019) or pre-processed using the Fastp tool (version 0.19.4) (Chen et al., 2018). Gene expression quantification of genes in the selected reference genome is performed using Alevin or Salmon software (Patro et al., 2017). The platform can be used to assess data generated from any organism. At present, we have downloaded the most popular genomes from Ensembl (164 genomes) and have incorporated their transcriptome indexes within our server (Cunningham et al., 2019). Quality control of samples is performed based on the alignment summary, descriptive statistics and the Alevin report of demultiplexed samples non-supervised clustering performed using AlevinQC package (version 1.4.0).

Gene filters based on the quantification of gene expression, which reduce the noise and computational costs are available on SingleCAnalyzer. The current version can filter genes with the lowest levels of expression or standard deviations. We have also integrated the function ‘FindVariableFeatures’ within the Seurat package (version 3.2.2), which can identify variably genes by considering the strong relationship between variability and expression level (Stuart et al., 2019). Moreover, the user can also perform further dimensionality reduction and clustering processes by analysing the principal components obtained via principal component analysis (PCA). The optimum number of components used for the analyses can be determined using the calc_npc function of the CIDR package (version 0.1.5) (Lin et al., 2017). Empty droplets can be detected and removed with the application of the DropletUtils tool (version 1.8.0) (Lun et al., 2019).

Interactive visualisation of samples in scatter plots requires a dimensionality reduction process, which is performed using the following methods: 1) PCA, which is generated with the prcomp function of the stats R package (version 4.0.3); 2) Classic multidimensional scaling (cMDS), which is performed with the cmdscale function of the stats R package using camberra as distance method; 3) Nonmetric multidimensional scaling (isoMDS), which is performed using the isoMDS function of the MASS R package (version 7.3); 4) t-distributed stochastic neighbor embedding (t-SNE), which is performed using the Rtsne function of the Rtsne R package (version 0.15); 5) Uniform manifold approximation and projection (UMAP), which is performed using the umap function of the uwot R package (version 0.1.9); 6) and Non-negative matrix factorisation (NMF), which is performed using the nnmf function of the NNLM package (version 0.4.3). Collectively, application of these methods provides users with a multi-perspective assessment of the relationships between data.

Determination of clusters within the interactive web interface is supported by the results provided by unsupervised clustering methods. At present, SingleCAnalyzer applies the following unsupervised clustering methods: 1) k-means, which is computed using the kmeans function of the stats R package (with iter_max = 15); 2) partition around medoids (PAM), which is computed using the pam function of the cluster R package (version 2.1.0); 3) hierarchical clustering, which is performed using the hclust function of the stats R package; 4) leiden clustering and pseudotime analysis, which is performed using Monocle3 R package (version 0.2.3) (Qiu et al., 2017). The user can specify input parameters such as the desired number of groups, the distance metric used by pam and hclust functions and the agglomeration parameter of hclust.

Trajectory and pseudotime analyses are performed using the Monocle3 R package (Qiu et al., 2017). It calculates possible trajectories between leiden clusters over the UMAP projection. The pipeline calculates the pseudotime prediction for each cluster centroid and a scale colour which represent the time is applied over the points when an origin cluster is selected. The function preprocess_cds uses PCA or LSI output based on user options with the following parameters: norm_method = log and scaling = true. The function reduce_dimension uses the following parameters: max_components = 2, reduction_method = UMAP, umap. metric = cosine, umap. min_dist = 0.1, umap. n_neighbors = 15L, umap. nn_method = annoy. The function cluster_cells uses the following parameters: k = 20, cluster_method = Leiden, nunm_iter = 2, partition_qval = 0.05. The function learn_graph uses use_partitium and close_loop as true.

Groups of samples can be compared by applying different methods to assess differential expression. Reviews of the use of methods have concluded that no single method outperforms the others under all circumstances, and suggest that it is necessary to determine the optimal method or pipeline for each analysis performed (Seyednasrollah et al., 2013; Soneson and Delorenzi, 2013). However, researchers have acknowledged that DESeq2 (version 1.28.1) (Love et al., 2014), EdgeR (version 3.30.3) (Robinson et al., 2010) and limma (version 3.44.3) (Law et al., 2014) are the most widely used methods and consistently performed well when their reliability was assessed. We have integrated all of the methods within a SingleCAnalyzer that can be adjusted to apply customised parameters to individual tests.

SingleCAnalyzer performs a functional enrichment analysis and a gene set enrichment analysis (GSEA) for each comparison result. The enrichment analysis is performed with the R package GOseq (version 1.40) (Young et al., 2010) and the GSEA is performed with the R package fgsea (version 1.14) (Korotkevich et al., 2016). Functional annotation database used by these methods was downloaded from the NCBI BioSystems repository (Geer et al., 2009). Resulting graphs are generated with the package RJSplot (version 2.6) (Barrios and Prieto, 2018).

Analyses can be launched as anonymous or registered users. Anonymous accounts are regularly deleted, and registered users can require the cancellation of their account. Data of registered users are protected by their personal password which is encrypted on our system. Users can freely download or delete their processed data and analysis results without any limitation. Raw data files uploaded by users (FASTQ, HDF5) are deleted once they are processed. This deletion avoids storage limitations and the presence of sequences in our system.