Single-cell Quantification

Turns raw single-cell sequencing reads into a count matrix: a table with one row per cell, one column per gene, and in each box the number of times that gene was read in that cell. It is written as an AnnData .h5ad file — the format the Single-cell Embedding AI Tool reads.

Liatir drives this with simpleaf, which runs the whole chain for you: mapping with piscem, then cell detection, collation and UMI counting with alevin-fry. You run one step, not five.

Use it for

• 10x Genomics 3′ and 5′ single-cell RNA-seq;
• producing the .h5ad input the single-cell AI Tools and the Single-cell Viewer expect;
• re-quantifying a sample against a different annotation release.

Inputs

• Single-cell index — download a verified ready-made reference or build a custom one in Single-cell Reference Index.
• Reads R1 — the short read carrying the cell barcode and the UMI.
• Reads R2 — the read carrying the transcript sequence.
• Chemistry — the kit the library was made with. Getting this wrong makes almost nothing map, so check it first when a run comes back near zero.
• Which droplets are cells — see below.
• UMI resolution — how a read that could belong to more than one gene is counted.
• Threads.

Which droplets are cells

Most droplets in a single-cell run contain no cell. Three ways to decide:

Option	What it does	When to use it
Detect cells automatically	Finds the drop between real cells and empty droplets	The default. Needs nothing from you
Use my barcode list	Quantifies exactly the barcodes in a file you supply, one per line	You already know which cells you want
Keep a fixed number of cells	Takes the N barcodes with the most reads	You know roughly how many cells were loaded

UMI resolution

A UMI is a random tag attached to each original molecule, so the same molecule read ten times is counted once. When a read could have come from more than one gene, the resolution mode decides what happens:

• Standard (CellRanger-like) — the default. Whole counts, and what most published analyses use.
• Parsimony modes — whole counts, resolved differently.
• Shared reads (EM) modes — split an ambiguous read between the genes it could belong to, so counts can be fractional (1.4 copies of a gene).

Fractional counts are a legitimate estimate, but they are not raw counts. Liatir records that on the file, and a tool that requires raw counts will refuse it rather than produce a quiet wrong answer.

Outputs

• Count matrix (AnnData) — the .h5ad file.
• Run summary: cells, genes, percentage of reads mapped, percentage of barcodes kept, and how the matrix was made.

How to read the result

• Reads mapped below roughly 50% usually means the wrong chemistry, or R1 and R2 the wrong way round.
• Barcodes kept counts reads whose barcode matched a real cell, including those rescued from a single-letter sequencing error.
• Cells far from what you loaded suggests the cell-detection option needs changing.

What the counts include

Because the index is a spliced + intronic reference, the matrix holds spliced + unspliced + ambiguous counts together, and keeps the three separately as layers inside the same file. This is right for whole nuclei and standard for droplet single-cell work; it is stated in the run log so it is never an assumption.

Gene names

Genes are named by the identifiers in your annotation. An Ensembl or GENCODE release gives Ensembl IDs — which is what Geneformer expects. UCE and scGPT expect gene symbols and will report the mismatch instead of running on the wrong namespace. Symbols are stored alongside the identifiers when the annotation carried them.

Current limit

A sample split across several FASTQ files per read has to be concatenated before running.