Bioinfo, bench to analysis · an interactive tutorial
中EN
⭑ Try OmicOne

Glossary

Three things for every metric: which bench step produces it, what it means in the data, and its usual range.

Single-cell RNA-seq

Viability

From the bench
The percentage of live cells measured before loading (trypan blue or fluorescent dyes such as AO/PI); it reflects membrane damage from dissociation, freeze–thaw and waiting.
In the data
Low viability means more broken cells and more free RNA: more cells with high percent.mt, a stronger ambient background and a lower Fraction Reads in Cells.
Typical range
Higher is better; many core facilities ask for roughly 70–90% or more, depending on the facility and sample type. Cryopreserved samples are usually lower than fresh ones.
Chapters
Dissociation

Dissociation stress percent.diss

From the bench
Cells keep transcribing during 37 °C enzymatic digestion and mechanical handling, so immediate-early and heat-shock genes (FOS, JUN, EGR1, HSPA1A...) are rapidly induced. This is caused by the protocol, not the tissue's original state.
In the data
A set of cells shares high stress-gene expression and may form its own cluster or drive a principal component. Score it with a stress gene set (for example the list from van den Brink et al., 2017, often stored as percent.diss or a module score).
Typical range
No universal cut-off. Shorter handling, cold-active protease dissociation or single-nucleus RNA-seq (snRNA-seq) reduce it.
Chapters
Dissociation · Clustering

Mitochondrial fraction percent.mt

From the bench
When the membrane breaks during dissociation, freeze–thaw or waiting, cytoplasmic mRNA leaks out while transcripts inside the mitochondria stay, so the fraction rises; apoptotic and stressed cells also rise.
In the data
The share of a cell's UMIs from mitochondrial genes (MT- in human, mt- in mouse). High values usually come with low UMI and gene counts and flag broken or dying cells.
Typical range
The Seurat pbmc3k tutorial keeps cells below 5%; 5–10% is common for PBMCs; metabolically active tissue (heart, kidney, muscle) has a higher baseline; single-nucleus data are usually very low, and high values there point to cytoplasmic carry-over. Set the cut-off per tissue and per sample (e.g. median + 3 MAD).
Chapters
Dissociation · Quality control

Ribosomal protein fraction percent.ribo

From the bench
No single bench source: mostly set by cell type and state (resting lymphocytes, for example, run high), and also affected by ambient RNA and cell damage.
In the data
The share of UMIs from ribosomal protein genes (RPS and RPL). Used to describe and troubleshoot, rarely as a hard filter.
Typical range
Varies widely by cell type; no universal cut-off. Investigate when a whole cluster is unusually high or low.
Chapters
Quality control

Haemoglobin fraction percent.hb

From the bench
Red blood cells not fully removed (density-gradient separation or red-cell lysis of PBMCs), or haemoglobin mRNA released by lysed red cells that ends up in the suspension as ambient RNA.
In the data
The share of UMIs from HBA1, HBA2, HBB and related genes. A few very high cells are likely leftover red cells (or erythroid precursors); a raised level in every cell points to ambient contamination.
Typical range
Should be very low in PBMCs and most tissues; higher in whole blood, bone marrow and haemorrhagic tissue, so judge by sample type.
Chapters
Dissociation · Quality control

Ambient RNA (soup)

From the bench
Cells that burst during dissociation, centrifugation and waiting release mRNA into the suspension; this free RNA is packed into every droplet along with the cells, empty droplets included.
In the data
A background layer added to every cell, with the same profile as the empty droplets; abundant genes show up in cells that should not express them (LYZ or HBB in PBMC T cells). SoupX, CellBender or DecontX estimate and remove it; they need the raw (unfiltered) matrix.
Typical range
Varies widely with sample quality. A low Fraction Reads in Cells and a raised empty-droplet plateau on the barcode rank plot both signal a lot of ambient RNA.
Chapters
Dissociation · Cell Ranger · Marker genes and cell-type annotation

GEM (Gel Bead-in-EMulsion) GEM

From the bench
At the junction of a microfluidic chip, the cell suspension (with reverse-transcription reagents), gel beads and oil meet and form water-in-oil droplets. Each droplet is its own reaction chamber: the bead dissolves and releases primers, the cell lyses and reverse transcription runs inside.
In the data
One barcode in the data corresponds to one droplet, not one cell: empty droplets and droplets holding two cells become empty barcodes and doublets.
Typical range
Cells are loaded dilute to limit doublets, so most droplets hold no cell. The number of droplets per run depends on the chip and kit.
Chapters
Droplets

Gel bead

From the bench
A dissolvable hydrogel bead carrying many copies of one primer design: Illumina Read 1 primer sequence + cell barcode + UMI + poly(dT). Every oligo on a bead has the same barcode and a different UMI.
In the data
One bead is one barcode. Beads are deformable and close-packed, so they enter droplets nearly one at a time and most droplets get exactly one bead; barcodes and droplets correspond almost one to one. The occasional droplet with two beads splits one cell across two barcodes.
Typical range
Beads enter droplets almost regularly, not by Poisson statistics; cells enter at random. The exact bead occupancy depends on the chip version.
Chapters
Droplets · Cell barcodes, UMIs and reverse transcription

Doublet (multiplet) doublet_score

From the bench
Cells enter droplets at random (Poisson): the more you load, the more often two or more cells share a droplet. Clumps that were not broken up in the suspension also make doublets.
In the data
One barcode holds two cells' transcripts: more UMIs and genes, and markers of two cell types at once (T cell + monocyte). Heterotypic doublets are found by simulation methods such as scDblFinder or DoubletFinder; homotypic doublets (two cells of one type) are nearly invisible in expression. Run per sample (per capture channel).
Typical range
Rises roughly linearly with cells recovered; the classic 3′ v3/v3.1 user guide gives about 0.8% per 1,000 recovered cells (about 7–8% at 10,000), but check the guide for your kit.
Chapters
Droplets · Quality control · Marker genes and cell-type annotation

Empty droplet

From the bench
Cells are loaded dilute, so most droplets hold a bead and some suspension but no cell; the free RNA in that suspension is still captured and barcoded.
In the data
The many low-UMI barcodes below the knee of the barcode rank plot; their profile is the ambient RNA. Cell Ranger and EmptyDrops separate them from real cells by UMI count and by how much they differ from the ambient profile.
Typical range
The raw matrix (raw_feature_bc_matrix) holds far more barcodes than cells, mostly empty droplets; they are the input for estimating ambient RNA, so keep the raw matrix.
Chapters
Droplets · Cell Ranger

Cell barcode CB

From the bench
A fixed sequence on the gel-bead oligos, shared by every oligo on one bead and different between beads; reverse transcription writes it onto one end of the cDNA, so every molecule from one droplet carries the same cell address.
In the data
Cell Ranger uses it to assign reads to cells, matching it against a predefined whitelist and correcting up to one mismatch. One barcode is one droplet, not necessarily one cell (it may be empty or a doublet).
Typical range
3′ v3/v3.1: 16 bp; 3′ v2: 16 bp; 3′ v1 (used by the pbmc3k example): 14 bp.
Chapters
Cell barcodes, UMIs and reverse transcription · Library prep and sequencing · Cell Ranger

UMI (unique molecular identifier) UMI

From the bench
A random stretch next to the barcode on each gel-bead oligo, different on every oligo. Each mRNA molecule receives one UMI at reverse transcription, and all its PCR copies inherit it.
In the data
Reads with the same cell, gene and UMI are collapsed into one molecule, removing the bias of uneven PCR; the numbers in the count matrix are UMI counts. Cell Ranger merges a UMI one base away from a more abundant UMI as a sequencing error.
Typical range
3′ v3/v3.1: 12 bp (4¹² ≈ 16.8 million sequences); 3′ v2 and v1: 10 bp (about 1.05 million).
Chapters
Cell barcodes, UMIs and reverse transcription · Library prep and sequencing · Cell Ranger

poly(dT) capture poly(dT)

From the bench
A run of T at the end of the primer pairs with the mRNA's poly(A) tail and starts reverse transcription. A-rich stretches inside transcripts, including in introns, are sometimes primed too.
In the data
Only polyadenylated RNA is captured (most mRNA; not, for example, replication-dependent histone mRNA); in 3′ kits the reads pile up at the 3′ end of transcripts. Intronic reads from internal priming are a main source of the unspliced counts used in RNA velocity.
Typical range
A property of the chemistry, not of the sample; only part of a cell's mRNA is captured, which is why lowly expressed genes often read 0.
Chapters
Cell barcodes, UMIs and reverse transcription · Beyond

Sample index

From the bench
The final library PCR adds the P5/P7 adapters and short index sequences (i7, plus i5 in dual-index kits), so several libraries can share a sequencing run.
In the data
Demultiplexing (BCL Convert, bcl2fastq or cellranger mkfastq) assigns reads to samples by it; a wrong sample sheet or index hopping sends reads to the wrong sample.
Typical range
8 bp for single-index kits; 10 bp each for i7 and i5 in dual-index kits (depends on the kit). Unique dual indexes are advised when many samples share a patterned flow cell.
Chapters
Library prep and sequencing

Q30 bases Q30

From the bench
The base-calling quality of the sequencer, set by cluster density, reagents and flow-cell condition rather than by the sample.
In the data
The fraction of bases with Phred quality ≥ 30 (error rate ≤ 0.1%). Cell Ranger reports it separately for the barcode, UMI, RNA read and sample index.
Typical range
Depends on platform and read length; quality usually falls along the read. Cell Ranger warns below per-segment thresholds (see the docs for your version); a poor barcode segment directly lowers Valid Barcodes.
Chapters
Library prep and sequencing · Cell Ranger

Barcode rank plot (knee plot)

From the bench
Its shape is set by how many cells were loaded, their viability and the amount of ambient RNA.
In the data
All barcodes sorted by UMI count, plotted on log–log axes: the high plateau on the left is cells, the long tail after the drop is empty droplets. The steeper the cliff, the cleaner the separation.
Typical range
Read it qualitatively: a sharp knee and a high plateau are good; a gentle slope means many damaged cells or much ambient RNA, and the cell count is unreliable.
Chapters
Cell Ranger

Estimated Number of Cells Estimated Number of Cells

From the bench
Set by the number of cells loaded and the recovery efficiency: only part of the loaded cells are captured in droplets and yield barcoded cDNA.
In the data
The number of barcodes Cell Ranger calls as cells. Recent versions first take the clearly high-UMI barcodes, then use an EmptyDrops-like test to add lower-UMI barcodes whose profile differs from the ambient RNA.
Typical range
Should be close to the targeted recovery. Recovery depends on the kit, commonly around half to two-thirds of cells loaded; a large deviation suggests a counting error, clumps, or so much ambient RNA that cell calling fails.
Chapters
Droplets · Cell Ranger

Mean Reads per Cell Mean Reads per Cell

From the bench
Set by how many reads the library received on the sequencer (flow-cell output divided among the pooled libraries).
In the data
Total sequenced reads divided by the estimated number of cells. The numerator includes reads outside cells (empty droplets, ambient RNA), so with a lot of ambient RNA it overstates the depth cells actually received.
Typical range
10x recommends a minimum of about 20,000 read pairs per cell for 3′ gene-expression libraries (check your kit's guide); deeper sequencing finds more genes with diminishing returns, see sequencing saturation.
Chapters
Library prep and sequencing · Cell Ranger

Median Genes per Cell Median Genes per Cell

From the bench
Depends on the cells' RNA content and type, capture and reverse-transcription efficiency (chemistry version) and sequencing depth.
In the data
The median number of genes detected (UMI ≥ 1) per cell-associated barcode, i.e. the median of nFeature_RNA.
Typical range
Strongly depends on cell type, chemistry and depth: small, RNA-poor cells such as PBMCs score well below large epithelial or tumour cells, and newer chemistries above the v1 used for pbmc3k. Compare with past data of the same tissue and chemistry.
Chapters
Cell Ranger · Quality control

Median UMI Counts per Cell Median UMI Counts per Cell

From the bench
Reflects how many mRNA molecules per cell were captured and reverse-transcribed: cell size and RNA content, capture efficiency and sequencing depth together.
In the data
The median total UMI count per cell-associated barcode, i.e. the median of nCount_RNA.
Typical range
Depends on cell type, chemistry and depth; no universal pass mark. Compare with past data of the same tissue and chemistry.
Chapters
Cell Ranger · Quality control

Sequencing Saturation Sequencing Saturation

From the bench
Set by library complexity (how many distinct molecules were reverse-transcribed and amplified) together with depth: low complexity or deep sequencing means the same molecule is read again and again.
In the data
1 − (deduplicated UMIs ÷ reads): the fraction of reads that come from an already-seen molecule (same barcode, UMI and gene). 50% means each molecule was read twice on average.
Typical range
No universal pass mark: low saturation means more sequencing will still find new molecules; high saturation with few genes per cell means the library lacks complexity and more sequencing will not help much.
Chapters
Cell barcodes, UMIs and reverse transcription · Library prep and sequencing · Cell Ranger

Fraction Reads in Cells Fraction Reads in Cells

From the bench
Mostly reflects the amount of ambient RNA: the more RNA released by damage, dead cells and lysis, the more reads land on empty-droplet barcodes.
In the data
Of the reads with a valid barcode that map confidently to the transcriptome, the fraction whose barcode is called a cell. Low values mean many reads sit in the soup, or cell calling went wrong.
Typical range
10x expects roughly 70% or more, and Cell Ranger flags values below its threshold (see the docs for your version).
Chapters
Dissociation · Cell Ranger

Valid Barcodes Valid Barcodes

From the bench
Depends on the synthesis quality of the bead oligos and the sequencing quality of the barcode segment; it drops with a poor Read 1, an abnormal library structure or the wrong chemistry setting.
In the data
The fraction of reads whose barcode matches the whitelist exactly or after correcting one mismatch; the rest cannot be assigned to any cell.
Typical range
10x expects roughly 75% or more (see the Cell Ranger docs); if clearly lower, check the sequencing run and that the chemistry matches.
Chapters
Cell barcodes, UMIs and reverse transcription · Cell Ranger

Reads Mapped Confidently to Transcriptome Reads Mapped Confidently to Transcriptome

From the bench
Depends on RNA quality and library make-up: it drops with degraded RNA, genomic DNA or many intergenic reads, and collapses when the wrong species reference is used.
In the data
The fraction of Read 2 sequences mapping to a single gene; only these reads are used for UMI counting. From Cell Ranger 7, intronic reads are counted by default (include-introns).
Typical range
Usually high for good human and mouse samples; counting introns or not changes the value noticeably, so take care when comparing across Cell Ranger versions.
Chapters
Cell Ranger

UMIs per cell nCount_RNA

From the bench
The cell's RNA content (size, type, activity) × capture efficiency × sequencing depth; low in broken cells that leaked mRNA, lower still in empty droplets, high in doublets.
In the data
In the Seurat metadata, the sum of UMIs over all genes in a cell, also called library size.
Typical range
Depends on cell type and depth; QC often flags outliers as median ± several MADs on the log scale, per sample.
Chapters
Quality control · Normalization

Genes per cell nFeature_RNA

From the bench
Same origins as the UMI count, plus library complexity; low in empty droplets and broken cells, high in doublets.
In the data
The number of genes with at least one UMI in a cell.
Typical range
The Seurat pbmc3k tutorial keeps 200–2,500; other data, especially from newer chemistries, need cut-offs set for their chemistry, depth and tissue.
Chapters
Quality control

Normalization

From the bench
No direct bench source; it corrects the differences in totals caused by capture efficiency and sequencing depth, and in doing so also removes real differences in RNA content between cells.
In the data
Makes expression comparable between cells. LogNormalize: divide each cell's counts by its total, multiply by 10,000, take log1p; SCTransform: model depth with a regularised negative-binomial regression and keep the Pearson residuals.
Typical range
Seurat's default scale factor is 10,000 (CP10k). ScaleData before PCA then centres each gene and scales it to unit variance.
Chapters
Normalization

Highly variable genes (HVGs)

From the bench
No direct bench source; it reflects biological differences between cells (type, state) on top of technical variation (sampling noise, dissociation-stress genes, abundant ambient transcripts).
In the data
Genes whose variance stays high after accounting for the mean–variance trend (Seurat's vst method). PCA uses only these genes to emphasise biological signal.
Typical range
Seurat's default is 2,000; 2,000–3,000 is common. The number is a parameter, not a fact.
Chapters
Variable genes and PCA

Principal component (PC) PC

From the bench
No direct bench source; the first PCs usually follow the main cell-type differences. A PC driven by mitochondrial genes, stress genes or depth means a technical factor is driving variation.
In the data
Linear combinations of gene expression, ordered by the variance they explain; the neighbour graph, clustering and UMAP are all computed on the leading PCs. The sign of a PC is arbitrary.
Typical range
Seurat computes 50 by default; the pbmc3k tutorial uses the first 10. An elbow plot helps; err on the side of more.
Chapters
Variable genes and PCA

k-nearest-neighbour graph (KNN / SNN)

From the bench
No direct bench source.
In the data
Each cell is linked to its k closest cells in PC space; Seurat then weights each link by how much the two cells' neighbourhoods overlap (Jaccard index), giving the SNN graph used for clustering.
Typical range
Seurat FindNeighbors uses k = 20 by default. A larger k smooths the graph and makes small groups easier to absorb.
Chapters
Clustering

Resolution resolution

From the bench
No direct bench source.
In the data
The parameter in Louvain / Leiden modularity optimisation: the larger it is, the more and finer the clusters. It does not correspond to any biological unit.
Typical range
The Seurat tutorial suggests 0.4–1.2 for data of about 3,000 cells; the pbmc3k tutorial uses 0.5. Larger datasets often need higher values.
Chapters
Clustering

UMAP UMAP

From the bench
No direct bench source; colouring the UMAP by sample, nCount_RNA or percent.mt shows whether a structure is technical.
In the data
A non-linear embedding that draws high-dimensional neighbour relations in 2-D: it keeps who neighbours whom, not distances between clusters, cluster sizes or densities. For visualisation only.
Typical range
Seurat RunUMAP defaults: n.neighbors = 30, min.dist = 0.3; a different seed or parameters change the picture.
Chapters
UMAP

t-SNE t-SNE

From the bench
No direct bench source.
In the data
Another non-linear embedding; it emphasises local neighbours even more than UMAP, and the relative placement of clusters means even less.
Typical range
The main parameter is perplexity (30 is common); also for visualisation only.
Chapters
UMAP

Marker gene

From the bench
No direct bench source; but ambient RNA and doublets make markers appear in clusters where they do not belong.
In the data
A gene expressed clearly higher in one cluster than in the other cells, used to infer the cluster's identity. Seurat's FindAllMarkers compares each cluster with all other cells, by default with a Wilcoxon rank-sum test.
Typical range
A good marker has a large avg_log2FC, a high pct.1 and a low pct.2 at the same time.
Chapters
Marker genes and cell-type annotation

Average log2 fold change avg_log2FC

From the bench
No direct bench source.
In the data
The base-2 log of the ratio of mean expression in a cluster to the rest; 1 means about two-fold higher. Seurat computes it on normalized data with a pseudocount.
Typical range
A filter around 0.25 is common (the default differs between Seurat versions); the value depends on normalization and pseudocount, so it is not comparable across tools or versions.
Chapters
Marker genes and cell-type annotation

Fraction expressing (pct.1 / pct.2) pct.1 / pct.2

From the bench
Shaped by capture efficiency: only part of the mRNA is captured, so lowly expressed genes read 0 in many cells.
In the data
pct.1 is the fraction of cells in the cluster with at least one UMI of the gene, pct.2 the fraction among all other cells. Ambient RNA inflates pct.2.
Typical range
0–1. An ideal marker has pct.1 near 1 and pct.2 near 0.
Chapters
Marker genes and cell-type annotation

Adjusted p-value p_val_adj

From the bench
No direct bench source.
In the data
The p-value after multiple-testing correction; Seurat's p_val_adj is Bonferroni over all genes in the dataset. Cell-level tests treat every cell as an independent sample, and the clusters were defined on the same data, so p-values come out tiny: use them to rank markers, not as evidence for differences between conditions.
Typical range
Extremely small p-values are routine in marker tables and do not by themselves mean a large biological effect.
Chapters
Marker genes and cell-type annotation

Pseudobulk

From the bench
Mirrors the biological replicates of the experimental design: each donor or sample is one independent unit.
In the data
Sum the counts of the cells of one cell type in each sample into one column per sample, then compare conditions with DESeq2, edgeR or limma; this avoids the false positives of treating cells as independent replicates.
Typical range
Power depends on the number of samples, not cells; with one or two samples per group, do not draw inferences.
Chapters
Marker genes and cell-type annotation · Beyond

Batch effect

From the bench
Systematic differences from different days, operators, dissociation protocols, chip lots, chemistry versions or sequencing runs.
In the data
One cell type splits into separate groups by sample. Harmony, Seurat CCA/RPCA or scVI align samples in a low-dimensional space; if condition and batch are fully confounded, no method can tell them apart.
Typical range
The best correction happens at the design stage: put samples from every group into every batch.
Chapters
Beyond

Pseudotime

From the bench
No direct bench source; it relies on the sample happening to contain all stages of a continuous process.
In the data
Each cell's distance from the root along a fitted path: how far it has progressed. It is an ordering by similarity, not real time, and the root must be set from biological knowledge.
Typical range
Usually scaled to 0–1 or arbitrary units; not comparable between trajectories.
Chapters
Beyond

RNA velocity

From the bench
3′ kits mainly capture mature mRNA; unspliced reads come largely from internal priming at A-rich intronic stretches and from nascent transcripts.
In the data
Compare each gene's unspliced and spliced counts: relatively more unspliced means the gene is switching on, fewer means switching off. The split must be counted separately from BAM or FASTQ (velocyto, STARsolo, alevin-fry); the standard count matrix does not contain it.
Typical range
Model assumptions (constant splicing and degradation rates, among others) often fail; the arrows on a 2-D plot are sensitive to the projection.
Chapters
Beyond

Ligand–receptor analysis

From the bench
Dissociation destroys the cells' positions in the tissue, so the data cannot tell whether two cells were neighbours.
In the data
Scores sender–receiver pairs where the sender expresses a ligand and the receiver its receptor, using a database of known pairs (CellChat, CellPhoneDB, LIANA...). The output is a candidate list, database-dependent, and RNA is not protein.
Typical range
Needs validation by spatial, protein-level or functional experiments; comparing conditions needs biological replicates.
Chapters
Beyond

CNV inference

From the bench
No direct bench source; it needs enough normal cells (immune or stromal) captured in the same sample to serve as a reference.
In the data
Order genes by chromosome position, smooth expression along it, compare with normal reference cells: large gains and losses mark malignant cells (inferCNV, CopyKAT, Numbat...).
Typical range
Only large (arm-level) events are visible; reference cells are best taken from the same patient.
Chapters
Beyond

Whole-exome sequencing (WES)

Tumour purity

From the bench
Tumour tissue also contains stroma, immune cells and normal epithelium. Where the piece is taken, the pathologist's marked area and macrodissection set how much of the DNA comes from tumour cells.
In the data
The fraction of cells in the sample that are tumour cells. It scales VAFs down: a clonal heterozygous mutation in a diploid region has an expected VAF of about purity / 2.
Typical range
Pathology estimates (share of tumour nuclei) and computational ones (ABSOLUTE, FACETS, PureCN, Sequenza..., from VAFs and copy number) often disagree. At 20% purity a clonal heterozygous mutation is expected at only about 10% VAF, and sensitivity drops.
Chapters
Samples · Somatic variant calling · Beyond

Ploidy

From the bench
Not a bench artefact but a property of the tumour genome: chromosome gains, losses and whole-genome doubling.
In the data
The average copy number of the tumour cells. Together with purity it sets how VAF maps to the true clonal fraction; the two are usually estimated jointly.
Typical range
2 in normal cells; close to 4 after whole-genome doubling. A sizeable share of adult AML has a normal karyotype, with fewer copy-number changes than most solid tumours.
Chapters
Samples · Beyond

Matched normal

From the bench
Non-tumour DNA from the same patient, usually blood for solid tumours. In blood cancers the blood itself carries tumour cells, so another tissue is used: TCGA-LAML used skin biopsies (Ley et al., NEJM 2013).
In the data
Used to subtract germline variants: only variants present in the tumour and absent from the normal are somatic candidates. Tumour cells in the normal make true mutations look germline, and they get filtered.
Typical range
The normal is often sequenced less deeply than the tumour. Tumour-only calling relies on a panel of normals and population frequencies alone and has many more false positives.
Chapters
Samples · Somatic variant calling

Germline variant

From the bench
Inherited from the parents and present in every cell, so blood, skin and tumour all carry it.
In the data
VAF about 50% (heterozygous) or 100% (homozygous) in the normal. One person's exome differs from the reference at tens of thousands of germline sites, all of which somatic analysis must remove.
Typical range
A heterozygous site that drifts away from 50% in the tumour points to a copy-number change or loss of heterozygosity (LOH) there.
Chapters
Samples · Somatic variant calling

Somatic mutation

From the bench
Acquired after birth in one cell and passed only to its descendants, so it is present in the tumour (or one of its subclones) only.
In the data
A variant found in the tumour and not in the matched normal. Its VAF depends on purity, copy number and the fraction of tumour cells that carry it.
Typical range
Varies enormously by tumour type, from a handful to thousands per exome. TCGA-LAML has a median of about 9 non-synonymous mutations per sample.
Chapters
Samples · Somatic variant calling

FFPE artefact

From the bench
Formalin cross-links and fragments DNA and deaminates cytosine to uracil, which PCR copies as T, giving C>T (G>A on the other strand). Long fixation and old blocks make it worse.
In the data
Many low-VAF C>T calls with a read-orientation bias. Caught by orientation-bias filtering, a panel of normals and a look at the spectrum.
Typical range
Concentrated at low VAF. Uracil-DNA glycosylase (UDG) treatment before library prep removes much of it. Fresh-frozen tissue and blood are largely free of it.
Chapters
Samples · Somatic variant calling · Mutation spectra and signatures

DNA integrity

From the bench
Set by sample type and handling: blood and fresh-frozen DNA are long, FFPE DNA is short and damaged. Measured before library prep with a fragment analyser (e.g. TapeStation DIN) or a qPCR ratio of long to short amplicons.
In the data
Poor integrity shows up as shorter inserts, lower library complexity, more duplicates and more C>T artefacts.
Typical range
DIN runs from 1 to 10, higher is more intact; FFPE is often far lower. Acceptance thresholds are validated by each lab for its own workflow.
Chapters
Samples · Library prep and exome capture

Clonal haematopoiesis

From the bench
With age, blood often contains haematopoietic clones carrying somatic mutations (commonly DNMT3A, TET2, ASXL1). When blood is the normal, these mutations are present in it.
In the data
Blood cells infiltrating a tumour can carry them too, so they can be mistaken for tumour mutations, or real signal gets subtracted because the normal has it as well.
Typical range
More common with age and with deeper sequencing; usually at low VAF.
Chapters
Samples

Insert size MEDIAN_INSERT_SIZE

From the bench
Set by fragmentation (acoustic or enzymatic) and size selection: the length of genomic DNA between the two adapters. FFPE DNA is already broken, so inserts are shorter.
In the data
Inferred from where the two reads of a pair align. When inserts are shorter than the read length, the reads overlap or run into the adapter, and the overlapping bases must not be counted twice.
Typical range
Often about 150–250 bp in WES, depending on fragmentation and size selection.
Chapters
Library prep and exome capture · Alignment

Library complexity ESTIMATED_LIBRARY_SIZE

From the bench
The number of distinct original molecules in the library. Low DNA input, damaged (FFPE) DNA or poor ligation make it low.
In the data
With low complexity, more sequencing only yields more duplicates and effective depth stops rising. Picard reports an estimated library size.
Typical range
No universal threshold; compare across samples of the same kit and type.
Chapters
Library prep and exome capture · Sequencing and coverage depth

Duplicate rate PERCENT_DUPLICATION

From the bench
PCR copies one original molecule many times (PCR duplicates); on patterned flow cells neighbouring clusters can also read the same molecule (optical duplicates).
In the data
Read pairs with the same start positions at both ends are flagged as duplicates and only one is used for calling. Duplicates add no independent evidence and inflate depth.
Typical range
Depends on input, sample quality and depth: the deeper one library is sequenced, the higher it gets; low-input and FFPE libraries run higher. Compare within a batch and kit.
Chapters
Library prep and exome capture · Sequencing and coverage depth · Alignment

Capture size

From the bench
Fixed by the exome kit's probe design and listed in the kit's target BED file.
In the data
Total length of the target regions in Mb, and the denominator of TMB. A stricter denominator is the callable territory: the part covered well enough in both tumour and normal.
Typical range
About 30–60 Mb for common exome kits, depending on the kit, whether UTRs are included and whether padding is counted. maftools' tcgaCompare uses 35.8 Mb for TCGA by default; tmb() defaults to 50 Mb when none is given.
Chapters
Library prep and exome capture · Tumour mutational burden (TMB)

On-target rate PCT_SELECTED_BASES

From the bench
Reflects how efficient and specific hybrid capture was: probe design, hybridisation and wash stringency, insert size and multiplexed capture all affect it.
In the data
The share of reads or bases that fall on (or near) the targets. A low value means sequencing spent off target and too little effective depth.
Typical range
Varies widely with the kit and with how it is computed (near-target bases counted or not); compare with the same kit's history.
Chapters
Library prep and exome capture · Sequencing and coverage depth

Mean target coverage MEAN_TARGET_COVERAGE

From the bench
Set by the sequencing output allotted to the sample, the on-target rate and the duplicate rate.
In the data
How many reads cover an average target base, usually after removing duplicates and low-quality reads. An average hides unevenness.
Typical range
Often about 100–200× for tumour WES, less for the normal. The depth needed depends on purity and on the lowest VAF you want to detect.
Chapters
Sequencing and coverage depth

Target bases at ≥20× PCT_TARGET_BASES_20X

From the bench
Depends on both total output and capture uniformity: with poor uniformity some exons stay thin even when the mean is high.
In the data
The share of target bases covered at least 20 times: how much of the exome reaches a basic callable depth.
Typical range
20× is a common reference line; for low-VAF tumour calling also look at the 50× or 100× fractions. Requirements depend on the project.
Chapters
Sequencing and coverage depth

Coverage uniformity FOLD_80_BASE_PENALTY

From the bench
Very GC-rich and GC-poor regions capture and amplify less efficiently; probe design and library complexity also play a part.
In the data
How similar the depth is across exons. A common measure is Picard's FOLD_80_BASE_PENALTY: how many times more sequencing would bring 80% of target bases up to the mean; closer to 1 is more uniform.
Typical range
GC_DROPOUT and AT_DROPOUT report GC bias. Compare with the same kit's history.
Chapters
Library prep and exome capture · Sequencing and coverage depth

Base quality

From the bench
The confidence the sequencer assigns to each base call in each cycle. It falls towards the end of reads and with overloaded flow cells or reagent problems.
In the data
Phred scale Q = −10·log10(error rate): Q20 is a 1% error, Q30 is 0.1%. Low-quality bases count less, or not at all, in variant calling.
Typical range
Illumina uses the fraction of bases at ≥Q30 as a run metric; the specification depends on platform and read length.
Chapters
Sequencing and coverage depth · Alignment

Mapping quality (MAPQ) MAPQ

From the bench
Comes from the genome, not the bench: repeats, pseudogenes and gene families leave short reads without a unique home. Shorter inserts and reads make it worse.
In the data
Phred-scaled confidence that the alignment position is right. BWA gives 0 to reads that fit equally well in several places. Callers ignore low-MAPQ reads, so such regions are effectively uncallable.
Typical range
BWA-MEM caps at 60. Regions with close pseudogenes, such as PMS2 or CYP2D6, have persistently low MAPQ.
Chapters
Alignment · Somatic variant calling

Base quality score recalibration (BQSR)

From the bench
The base qualities a sequencer reports are systematically off, in ways that depend on cycle, neighbouring bases and lane.
In the data
GATK counts actual mismatches away from known variant sites (dbSNP, known indels), models them by read group, reported quality, cycle and dinucleotide context, and rewrites each base quality. The reads themselves are unchanged.
Typical range
A GATK Best Practices step between alignment and variant calling.
Chapters
Alignment

Variant allele fraction (VAF) t_alt_count / (t_ref_count + t_alt_count)

From the bench
Set by what is in the tube: purity, copy number, the fraction of cells carrying the mutation, and artefacts such as FFPE damage.
In the data
Reads supporting the variant divided by all reads at the site. Germline heterozygous sites sit near 50%; a clonal heterozygous somatic mutation in a diploid region sits near purity / 2.
Typical range
Subject to binomial sampling: at 100×, a site at 25% VAF has a sampling standard deviation of about 4 percentage points. Copy-number changes move it away from purity / 2.
Chapters
Sequencing and coverage depth · Somatic variant calling · Beyond

Alt reads t_alt_count

From the bench
Set by depth, library complexity (how many independent molecules survive deduplication) and VAF.
In the data
The number of reads carrying the variant base. Candidates with too few are hard to tell from sequencing errors, so callers and filters require some minimum.
Typical range
By the binomial, 100× and 5% VAF give 5 alt reads on average; requiring at least 5 misses about half of true mutations.
Chapters
Sequencing and coverage depth · Somatic variant calling

Copy-number alteration

From the bench
Gains or losses of tumour genome segments, not a bench effect.
In the data
Changes the ratio of mutant to total copies and so the VAF. Copy number can be estimated from WES (CNVkit, FACETS, PureCN, Sequenza...), with more noise than from whole genomes.
Typical range
A mutation on the amplified allele sits above purity / 2; on the other allele it sits below.
Chapters
Somatic variant calling · Beyond

Loss of heterozygosity (LOH)

From the bench
Tumour cells lose one of the two parental alleles (by deletion, or by copy-neutral LOH).
In the data
Germline heterozygous sites stop sitting at 50% in the tumour. If the retained copy carries a somatic mutation, a clonal mutation can approach a VAF equal to the purity.
Typical range
Common at tumour suppressors: one allele mutated, the other lost.
Chapters
Somatic variant calling · Beyond

Panel of normals (PoN)

From the bench
Built from normal samples processed with the same kit and pipeline; it records artefacts this platform produces again and again.
In the data
Sites seen in several normals are flagged and Mutect2 stops calling them somatic. It catches systematic artefacts that one matched normal cannot reveal.
Typical range
Should match the kit and pipeline of the samples analysed; usually dozens of normals, and more is more stable.
Chapters
Somatic variant calling

Germline resource

From the bench
Not from this experiment: a summary of large population sequencing, such as gnomAD.
In the data
Gives each variant's allele frequency in the population. Mutect2 uses it as a prior: common variants are more likely germline. It matters most when the normal is shallow or missing.
Typical range
Population frequency is not VAF: one is a fraction of people, the other a fraction of reads in one sample.
Chapters
Somatic variant calling

Read-orientation bias

From the bench
Damage to one strand of the original DNA before library prep: cytosine deamination in FFPE (C>T) and guanine oxidation during acoustic shearing and handling (8-oxoG, G>T).
In the data
Such artefacts appear almost only in read pairs of one orientation (F1R2 or F2R1); true mutations show both. GATK learns the bias with LearnReadOrientationModel and FilterMutectCalls removes the calls.
Typical range
Strongest in FFPE samples and in libraries stored for a long time.
Chapters
Samples · Somatic variant calling

Cross-sample contamination

From the bench
DNA from another person gets into the sample during collection, extraction, pipetting or pooling. A sample swap is a different problem and is checked separately.
In the data
GATK's CalculateContamination estimates it from common germline SNPs: a small amount of the other allele at homozygous sites. Contamination creates a crop of low-VAF false somatic calls.
Typical range
Ideally close to 0. Whether tumour and normal come from the same person is checked by germline genotype concordance (e.g. Conpair, NGSCheckMate).
Chapters
Somatic variant calling

VEP

From the bench
A purely computational step.
In the data
Ensembl's Variant Effect Predictor: for each variant, which transcripts it hits and with what consequence (missense, stop gained, frameshift, splice...), plus annotations such as population frequency. vcf2maf runs it and picks one transcript per variant for the MAF.
Typical range
The protein change can differ between transcripts; check which transcript a study used before comparing.
Chapters
Annotation and the MAF file

MAF file

From the bench
A computed product: the table that remains after calling, filtering and annotation.
In the data
Mutation Annotation Format: tab-separated, one row per variant per sample. Key columns include Hugo_Symbol, Chromosome, Start_Position, Reference_Allele, Tumor_Seq_Allele2, Variant_Classification, Variant_Type and Tumor_Sample_Barcode.
Typical range
A MAF lists only samples that have variants: sequenced samples with none do not appear, so the denominator has to be kept separately.
Chapters
Annotation and the MAF file · Mutations and outcome

Variant classification Variant_Classification

From the bench
A computed annotation.
In the data
The MAF column that describes a variant's consequence; common values are Missense_Mutation, Nonsense_Mutation, Frame_Shift_Del/Ins, In_Frame_Del/Ins, Splice_Site and Silent.
Typical range
The mix of classes is itself a check: an unusual excess of Silent or non-coding classes can mean weak filtering.
Chapters
Annotation and the MAF file · The cohort landscape

Non-synonymous mutation

From the bench
A computed definition.
In the data
Coding mutations that change the protein. maftools counts 9 classes by default: Missense, Nonsense, Frame_Shift_Del/Ins, In_Frame_Del/Ins, Splice_Site, Translation_Start_Site and Nonstop_Mutation; Silent and the rest are kept aside.
Typical range
The definition sets TMB's numerator: some assays count synonymous mutations as well.
Chapters
Annotation and the MAF file · Tumour mutational burden (TMB)

Oncoplot

From the bench
An analysis figure, but each tile depends on capture, depth and filtering upstream: a poorly covered gene looks rarer than it is.
In the data
Genes as rows, samples as columns, tiles coloured by mutation type. Genes are sorted by frequency and samples are ordered so that co-occurrence and exclusivity form a staircase.
Typical range
The percentages on the right use the samples in the MAF as the denominator; sequenced samples with no mutation that are missing from the MAF inflate them.
Chapters
The cohort landscape

Mutual exclusivity and co-occurrence

From the bench
Not a bench effect but tumour evolution: hitting one pathway once is often enough.
In the data
Two genes mutated together less often (exclusive) or more often (co-occurring) than expected. maftools' somaticInteractions runs a Fisher's exact test per gene pair; correct for multiple testing.
Typical range
Confounded by burden: hypermutated samples carry many genes at once and inflate co-occurrence.
Chapters
The cohort landscape · Driver genes and hotspots

Tumour mutational burden (TMB) mut/Mb

From the bench
The denominator comes from the capture kit and coverage; the numerator depends on purity, depth and filtering. The same tumour can get different TMBs on different platforms.
In the data
Somatic mutations per megabase of callable territory. In WES, usually non-synonymous mutations divided by the capture size.
Typical range
Spans orders of magnitude across tumour types, with AML near the bottom. 10 mut/Mb comes from the FDA approval based on KEYNOTE-158 with FoundationOne CDx, a panel measure that does not transfer directly to WES non-synonymous TMB.
Chapters
Library prep and exome capture · Tumour mutational burden (TMB)

Transition/transversion ratio (Ti/Tv)

From the bench
FFPE deamination raises C>T (a transition) and 8-oxoG oxidation raises C>A (a transversion), so sample handling shifts the ratio.
In the data
Transitions (purine↔purine or pyrimidine↔pyrimidine: C>T, T>C) divided by transversions (the other four classes).
Typical range
Empirical values for germline SNVs: about 2.0–2.1 genome-wide and 2.8–3.0 in exomes. They do not apply to somatic mutations, whose Ti/Tv depends on the mutational processes.
Chapters
Mutation spectra and signatures

SBS96 spectrum

From the bench
Reflects the mutational processes (ageing, UV, tobacco, APOBEC, repair defects), plus artefacts such as FFPE and 8-oxoG damage.
In the data
The 6 pyrimidine-referenced substitution classes times 4 possible bases on each side (5′ and 3′): 6 × 16 = 96 categories.
Typical range
Needs the reference genome for the flanking bases; WES often yields few SNVs per sample, so single-sample spectra are noisy.
Chapters
Mutation spectra and signatures

Mutational signature

From the bench
Each mutational process leaves a characteristic pattern over the 96 categories; artefacts form patterns of their own.
In the data
Extracted from a cohort's spectra by non-negative matrix factorisation (NMF), or fitted from known signatures, then matched to the COSMIC catalogue by cosine similarity.
Typical range
E.g. SBS1 relates to spontaneous deamination of 5-methylcytosine (clock-like), SBS2/SBS13 to APOBEC, SBS4 to tobacco. A cosine similarity of about 0.85 or more is a common convention for a credible match.
Chapters
Mutation spectra and signatures

Mutation hotspot

From the bench
Not a bench effect. But some hotspots, such as the variable-length FLT3-ITD, are hard to detect in short reads, so detection depends on read length and insert size.
In the data
The same amino-acid position hit in many patients, usually the mark of a gain-of-function change and strong evidence of a driver.
Typical range
Classic AML examples: IDH1 R132, IDH2 R140/R172, DNMT3A R882, the NPM1 exon 12 frameshift insertion, FLT3 D835.
Chapters
Driver genes and hotspots

False discovery rate (FDR)

From the bench
A statistical concept.
In the data
When many genes are tested at once, the expected share of false positives among the results called significant. The Benjamini–Hochberg method turns p-values into q-values.
Typical range
Common thresholds are q below 0.05 or 0.1. Rank and highlight by q, not by the raw p.
Chapters
The cohort landscape · Driver genes and hotspots · Beyond

Wild-type

From the bench
Assumes the sample was actually sequenced and the gene was adequately covered; unsequenced or poorly covered samples are not wild-type.
In the data
A sequenced sample with no non-synonymous mutation in the gene. A sequenced sample with no MAF record at all is wild-type too, not missing.
Typical range
TCGA-LAML: 193 samples in the MAF, 200 rows in the clinical table; samples with clinical data but no MAF record count as wild-type in survival analysis.
Chapters
Mutations and outcome

Log-rank test

From the bench
A statistical concept.
In the data
Tests whether two or more Kaplan–Meier curves differ overall. It gives a p-value, not an effect size; for that use the hazard ratio (HR) from a Cox model.
Typical range
One row per patient; results are unstable when a group has fewer than 10 patients or 10 events.
Chapters
Mutations and outcome

Cancer cell fraction (CCF)

From the bench
Not measured directly; derived from VAF, purity and copy number.
In the data
The share of tumour cells carrying a mutation: close to 1 is clonal, clearly below 1 is subclonal. Common tools are PyClone, SciClone and ABSOLUTE.
Typical range
CCF = VAF × (purity × tumour copy number + 2 × (1 − purity)) / (purity × mutant copies). Wrong purity or copy number gives a wrong CCF.
Chapters
Beyond

MATH score

From the bench
Affected by purity and depth: shallower sequencing spreads VAFs by sampling noise and raises MATH.
In the data
Mutant-Allele Tumour Heterogeneity: 100 × MAD(VAF) / median(VAF), with the MAD scaled by 1.4826. More spread VAFs mean more heterogeneity (Mroz and Rocco 2013, head and neck cancer).
Typical range
Descriptive; compare only across samples of the same assay and pipeline; unstable with few mutations.
Chapters
Beyond