Single cell RNA-seq analysis in Javascript

Overview

This repository contains scran.js, a Javascript library for single-cell RNA-seq (scRNA-seq) analysis in the browser. The various calculations are performed directly by the client, allowing us to take advantage of the ubiquity of the browser as a standalone analysis platform. Users can then directly analyze their data without needing to manage any dependencies or pay for access to a backend. scran.js is heavily inspired by the scran R package and contains most of its related methods. Indeed, much of the implementation in this repository is taken directly from scran and its related R packages.

Analysis overview

scran.js implements key steps in a typical scRNA-seq analysis:

• Quality control (QC) to remove low-quality cells. This is done based on detection of outliers on QC metrics like the number of detected genes.
• Normalization and log-transformation, to remove biases and mitigate the mean-variance trend. We use scaling normalization with size factors defined from the library size for each cell.
• Feature selection to identify highly variable genes. This is based on residuals from a trend fitted to the means and variances of the log-normalized data for each gene.
• Principal components analysis (PCA) on the highly variable genes, to compress and denoise the data. We use an approximate method to quickly obtain the top few PCs.
• Clustering on the cells using the top PCs. This can either use k-means or community detection on a shared nearest neighbor graph.
• Dimensionality reduction with t-stochastic neighbor embedding (t-SNE), again using the top PCs.
• Marker detection using a variety of effect sizes such as Cohen's d and the area under the curve (AUC).
• Cell type annotation with a port of the SingleR algorithm.
• Batch correction via the mutual nearest neighbors method.
• Gene set enrichment tests and calculation of per-cell gene set activity scores.

The theory behind these methods is described in more detail in the Orchestrating Single Cell Analysis with Bioconductor book. All steps are implemented in C++ and compiled to WebAssembly for near-native performance - see the developer notes for details.

Quick start

scran.js is available as an npm package, so installation can be performed via the usual procedure:

npm install scran.js

Then you can import and initialize the library using the standard ES6 notation. Before any actual analysis steps are performed, the initialize() function must be run and its promise resolved; we suggest using a top-level await for convenience.

import * as scran from "scran.js";
await scran.initialize({ numberOfThreads: 4 }); // for old Node versions, set localFile: true

After that, you can run the remaining steps synchronously - for example, using the Node.js API:

// Reading in the count matrix.
import * as fs from "fs";
let buffer = fs.readFileSync("matrix.mtx.gz");
let mat = scran.initializeSparseMatrixFromMatrixMarketBuffer(buffer);

Basic analyses

The code chunk below implements a basic analysis, starting from count matrix loading and terminating at the markers for each cluster. This uses the Node.js API to read in one of our example Matrix Market files, but the same approach can be used on the buffer created from a File input in the browser.

import * as scran from "scran.js";
await scran.initialize({ numberOfThreads: 4 });

// Reading in the count matrix.
let mat = scran.initializeSparseMatrixFromMatrixMarket("matrix.mtx.gz");

// Performing QC.
let qc_metrics = scran.perCellRnaQcMetrics(mat, [ /* specify mito subset here */ ]);
let qc_thresholds = scran.suggestRnaQcFilters(qc_metrics);
let filtered = scran.filterCells(mat, qc_thresholds.filter(qc_metrics));

// Log-normalizing.
let normalized = scran.logNormCounts(filtered);

// Modelling per-gene variance and selecting top HVGs. 
let varmodel = scran.modelGeneVariances(normalized);
let hvgs = scran.chooseHvgs(varmodel, { number: 4000 });

// Performing the PCA.
let pcs = scran.runPca(normalized, { features: hvgs });

// Building the neighbor search index on the PCs.
let index = scran.buildNeighborSearchIndex(pcs);

// Performing the clustering. 
let cluster_graph = scran.buildSnnGraph(index, { neighbors: 10 });
let clustering = scran.clusterSnnGraph(cluster_graph);

// Performing the t-SNE and UMAP.
let tsne_res = scran.runTsne(index);
let umap_res = scran.runUmap(index);

// Detecting some markers.
let markers = scran.scoreMarkers(normalized, clustering.membership());

More documentation

Reference documentation for the Javascript API is available here.

Specific topics are discussed in more detail below:

• Manually free memory
• Use k-means clustering
• Annotate cell types against references

Developer notes are also available here.

Links

Check out kana to see how scran.js is used in an interactive scRNA-seq analysis application.

The scran.chan R package and scran executable are based on the same C++ libraries and allow the same analysis to be performed in different environments.