Smart-seq2

Start a Smart-seq2 analysis

In a project, click New session and choose Smart-seq2 as the kind of analysis.

What to upload

The files your pipeline left, alone or in a .zip or .tar.gz:

• one quantification file per cell: featureCounts, HTSeq-count, STAR ReadsPerGene.out.tab, RSEM .genes.results (and .isoforms.results), salmon quant.sf or kallisto abundance.tsv. Each file is recognised by its header, and the cell is named after its file, or after its folder when the file has a generic name (salmon's A01/quant.sf);
• a featureCounts table of every BAM, one column per cell;
• a genes × cells count table, as GEO and ArrayExpress have them (a gene length column is recognised and kept apart);
• an AnnData, loom or Seurat file.

Add a table of the cells if you have one: a column with the cells' names and any of plate, well, sample, condition, donor or sorted population. Its columns become the cells' metadata. For salmon or kallisto transcript files, add a transcript-to-gene table (transcript ID, gene ID and optionally the gene name), unless the transcripts carry GENCODE's full names, which name their gene.

Plates and wells

Every cell is placed on its plate and well: from the cell table's plate and well columns, else from the cell's name (P1_A01, SS2_15_0048_H12 or Tabula Muris' A1.B003290.3_38_F.1.1), else each uploaded file is a plate. The metadata dialog then asks for the sample, condition and batch of each plate, one row per plate, unless the cell table already names the samples. ERCC and SIRV spike-ins are set apart from the genes: they measure the technical noise and can normalise, but they are not clustered or tested.

The workflow

Plate QC replaces the droplet steps (Initial QC, Remove Outliers and Remove Doublets). Wells hold one sorted cell each, so there are no doublets to look for, and cells fail together by plate. A well with almost no reads held no cell (a no-cell control or a missed sort) and is named as such. Any other cell is compared with the cells of its own plate, as scater's isOutlier does with a batch: a cell is removed when its reads or genes are more than three median absolute deviations below its plate's median, or its mitochondrial or spike-in share that far above it. Reads, genes and the spike-in share (often under a percent, and skewed) are compared on a log scale. A plate whose median is itself an outlier among the plates is judged against the others, so its poor wells cannot hide behind each other; a plate without spike-ins is not judged on them.

Normalize Data divides each cell by a size factor. Read counts depend on which genes a cell expresses: a cell type with a few very highly expressed genes (acinar cells, plasma cells) would look deeper than it is by library size. Pooled size factors (Lun, Bach & Marioni 2016, as scran's computeSumFactors) sum cells in pools of 21 to 101, compare each pool with its cluster's average cell, and solve each cell's factor from all the pools. Spike-in size factors scale every well to the ERCC added in the same amount, so differences in total RNA (cell size, cycling) stay in the data; they are compared within each plate, since plates often get another dilution or mix of the spike-ins. The figure's bottom panel shows each cell's relative log expression: after a good normalisation every cell's box sits on 0.

Technical Noise & Variable Genes keeps the genes that vary more than the noise does. With spike-ins, the noise is measured: all their variation is technical, so their squared coefficient of variation is fitted against their mean, and each gene is tested for variation beyond that fit plus a 50% biological coefficient of variation (Brennecke et al. 2013). The fit is made plate by plate and the plates' results combined, as scran's modelGeneVarWithSpikes does with a block, so plates with another spike-in mix do not add false noise. The genes kept are those that vary the most times over the noise at their own mean, which favours neither rare nor abundant genes. Without spike-ins, the trend is fitted to the genes themselves (scran's modelGeneVar).

From the PCA on, the single-cell steps run as they do in a single-cell RNA-seq analysis: integration, the UMAP, clustering, annotation, markers, pseudobulk, trajectories and pathways. The batch is the plate unless you set it in the metadata dialog. When each plate holds one condition, integrating over plates would remove the difference you are looking for: give plates processed together the same batch, or skip integration.

Plate Maps draws the plates with each well coloured by its cluster, cell type or a measure. A cell type that keeps to a few columns is often the sort layout; one that keeps to the edge wells, a plate artefact. The row, column and edge are tested within plates, so a plate effect does not pass for a position effect.

Isoform Usage uses what full-length reads give: which isoforms a gene's reads come from. It needs transcript-level files at upload (RSEM .isoforms.results, salmon quant.sf or kallisto abundance.tsv). A gene can switch isoforms between cell types without changing its expression, which gene-level tests do not see.

Step reference