Fragment analysis

Open a capillary fragment run (.fsa/.hid) and size its peaks against its own ladder: dye channels on an axis in bases, five sizing methods, and the ladder's own measured error beside every length.

What it doesCopy Anchor Link

A fragment run and a Sanger read come out of the same instrument, in the same file format, and are not the same document. A Sanger read is four dye channels the instrument has already interpreted: one base call per peak, with a confidence. A fragment run is dye channels and nothing else — no bases, no calls, no confidences. What it carries instead is a size standard: a set of DNA fragments of known length, run in their own dye channel alongside the sample.

That standard is the whole of the measurement. The instrument records where a peak appeared in scans — how far along the capillary run it was detected — and a scan is not a length. It depends on the polymer, the capillary, the temperature and the run itself, so the same fragment scanned on Tuesday and on Thursday appears at different numbers. Only the ladder converts one into the other, by giving you a set of points where the length is already known and asking what curve passes through them.

This page is about that curve, because choosing it is a scientific decision with published trade-offs and most software hides it.

What this computesCopy Anchor Link

• Peak detection — every local maximum in a dye channel above a height you set. A saturated peak reads as a flat plateau rather than a point, and is reported once, with its full width, rather than as one peak per sample of the plateau.
• Size calling — a peak's position in scans turned into a length in bases, by one of five curves fitted through the size standard's own fragments.
• The ladder channel — which dye carries the size standard, measured by matching the standard's recorded positions against every channel's peaks rather than assumed.
• Sizing accuracy — the run's own error in bases, by leave-one-out: each standard fragment is removed in turn and sized from the others, and the difference from its known length is the error. Reported as a root-mean-square and a worst case.
• Off-scale detection — which peaks saturated the detector, so their heights are a floor rather than a measurement.
• Peak resolution — two maxima closer together than a length difference the separation can resolve are reported as one peak, the taller, with the count of what was merged.
• Allele calling — each peak's locus and allele, from a marker kit's own ladder definition, with the deviation in bases between the measured length and the length the kit says that allele runs at.
• Sex typing from amelogenin — the X and Y reading of a ladder that names its two alleles positionally, given where the assay is the published six-base one, withheld with its reason where it is not, and always shown beside the label the kit itself declares.
• Off-ladder, ambiguous and no-locus designations — the three ways a peak inside or outside a locus can fail to be an allele, each reported with its reason rather than rounded to a label.
• Adenylation and stutter classification — a peak identified as a by-product of a taller one: which kind, which peak it belongs to, and the height ratio that classified it.
• Pull-up from a saturated peak — a small peak in one dye sitting inside a peak that saturated the detector in another, named as that fragment's light read through rather than called as an allele, with the dye it came from and its height ratio.
• Ladder calibration — an allelic ladder of the kit, run on your own instrument, labelled locus by locus as a whole (every allele, in order, matching the kit's declared spacing and relative peak heights), so a sample is called against the sizes your instrument gives rather than the manufacturer's reference run.
• Second peaks — a peak within tolerance of an allele a taller peak in the same locus already holds, reported rather than called a second copy of it.
• The repeat unit — each locus's, measured from its own ladder rather than tabulated.
• The genotype — the alleles called per locus, in channel order.
• The peak table — for each peak: its length, its dye, its height, its position in scans, and whether the method asked for could actually be applied to it.

The five sizing methods, and when each is wrongCopy Anchor Link

The four Applied Biosystems names them all; the fifth is the same reciprocal model fitted globally.

Local Southern uses the two standard fragments either side of each peak — four in all — and the reciprocal relationship between fragment length and mobility that Southern described in 1979. It is the default, and on the four real ladders here it is the most consistent of the five: the only one under 1.7 bases of leave-one-out error on all four. Its weakness is its locality: if one standard fragment near your peak runs anomalously, everything near it inherits that. It also cannot apply at all outside the standard's own range, where fewer than two fragments lie on one side, and there it falls back to a least-squares extrapolation — which the result says out loud, because a size produced by a different method than the one you chose is not a size you can compare.

Global Southern fits that same reciprocal curve once, to every standard fragment. One anomalous fragment barely moves it, which is its advantage. The cost is real and this page measures it rather than describing it: over a whole run the single curve fits less well than the local one, and on the four real ladders shipped with this app that is a leave-one-out error of 7.0 to 7.3 bases against 1.1 to 1.7 for the local methods — four to five times worse on every one of them.

Least squares, 2nd and 3rd order fit a quadratic or a cubic through the standard by regression. They compensate for a fragment that runs anomalously, which is exactly what they are for — and a cubic through a standard whose points nearly determine a quadratic is free to wander between them.

Cubic spline is forced through every standard fragment exactly. That sounds like the most faithful option and is the one to be most careful with: it reproduces an anomalous fragment perfectly instead of compensating for it, so a standard fragment that ran late drags every sample size near it late with it.

Reading the result, not just the numberCopy Anchor Link

Every size on this page carries the accuracy of the ruler that produced it. That figure is measured on your run, not quoted from a data sheet: a difference between two peaks smaller than the leave-one-out error is not a difference the run can resolve, however many decimal places the table shows.

The result also says when a size did not come from the method you chose. A peak outside the size standard's range is extrapolated, and extrapolated sizes are the least trustworthy numbers in the table; they are marked as such rather than printed like the rest.

A peak that saturated the detector is still sized normally. Saturation flattens the top of a peak, which destroys its height as a measurement, but the position — and therefore the length — is still read. So an off-scale peak's size is usable and its height is not, and the two are labelled differently for that reason.

On the plot, each named peak carries its allele above it. Heights are in RFU (relative fluorescence units), titled up each dye's axis. An allelic ladder packs more alleles per base than a label can name side by side, so a name with no free line left above its peak is not drawn, and the dye's panel says how many were left off. Hovering any peak reads its dye, size, height, locus and allele.

From a length to an alleleCopy Anchor Link

A size is a measurement. An allele is a claim about which marker the fragment came from and how many repeats it carries, and making that claim needs the marker kit's own definition: which locus sits in which dye channel, over which range of lengths, and what length each of its ladder alleles runs at. Choose a kit and every peak is named at its locus, or says why it was not. Forty-seven kits are here, which is every one NCBI's OSIRIS publishes a ladder definition for, and each one carries its own loci, their dye channels, their length windows and every allele its ladder contains.

A kit that is not among them can be brought. A regional or in-house panel, or a newer version of a published one, is defined by a file in exactly the same form — an OSIRIS LadderInfo definition, which is what configures OSIRIS itself. Add it to the project and it is offered above the built-in kits, under the name the file itself declares for the kit rather than the name the file was saved under, and it names peaks the same way and to the same standard. Nothing about the calling changes: the loci, windows and ladder alleles all come from the definition either way.

A peak that is not an allele is told apart from one that is, by name. A length falling in no locus window is signal rather than an allele — primer peaks and the space between markers both look like this — and is reported as such. A peak inside a locus with no ladder allele close enough is off ladder, which is a finding rather than an error: a real variant the kit's ladder does not carry looks exactly like that, and so does a run whose sizing has drifted. A peak close enough to two ladder alleles at once is reported as ambiguous, with both, rather than resolved by a tie-break nobody could check.

Two by-products of the chemistry are named rather than called as extra alleles. Taq polymerase adds a non-templated adenine to most, not all, of a product, so a real fragment often appears as a pair of peaks one base apart — the incompletely adenylated form and the full one. And polymerase slippage in a repeat leaves a stutter peak one repeat unit short of its parent. Both are recognised by the same two facts: how far the peak is from a taller peak in the same locus, and how small it is relative to it. Both fractions are yours to set, and setting either to zero reports every peak as its own. Left alone they are the thresholds NCBI's OSIRIS uses.

A third is optical rather than chemical. A fragment bright enough to saturate the detector bleeds into neighboring dye channels, and a small peak appears there at the same length. Where the run itself records those scans as off scale, a peak in another dye inside the saturated one, at or below a set fraction of its height, is named pull-up. The fraction is the one NCBI's OSIRIS publishes for the kit's samples, shown beside the control and yours to change; zero turns it off.

One allele is one fragment. Where a peak is close enough to an allele that a taller peak in the same locus has already been called as, it is reported with that reason rather than listed as a second copy of the allele. A run producing many of those is telling you the tolerance is wider than its own sizing error.

Amelogenin reads as X and Y, and the reading comes from the assay rather than the kit. A kit definition declares amelogenin's two alleles positionally, as 1 and 2, and carries no mapping to those letters anywhere in the file. The mapping is published instead: the primer pair spanning amelogenin's first intron gives a 106-base product from X and a 112-base product from Y, because the X homologue carries the six-base deletion the test reads, so the shorter amplicon is X. A ladder whose two alleles sit six bases apart is reading that deletion, whatever its absolute sizes, and its peaks are named X and Y with the kit's own label kept beside them to check against.

A ladder that is not six bases wide keeps its positional labels, and the page says why. Some kits amplify a different part of amelogenin: their two alleles sit three or nine bases apart, so they are not reading that deletion and no published source settles which of the two is X. Those are reported as the kit spells them, with the reason, rather than translated by assumption.

ControlsCopy Anchor Link

Sizing

Alleles

What it does not doCopy Anchor Link

It does not interpret mixtures, subtract pull-up from the signal, or judge heterozygote balance. Each is a separate analysis. Pull-up is named where the detector saturated, which the run records; pull-up from a peak that stayed on scale is not detected, because nothing in the file marks the shared scans.

It does not apply a per-locus stutter ratio. Those tables are published per kit by their manufacturers; one threshold is applied everywhere, stated, and adjustable.

It does not compare runs. Each run is sized against its own standard, which is the correct way to size one, and the comparison of several runs on one axis is a different question with its own requirements.

Provenance of correctnessCopy Anchor Link

The sizing methods are implemented from Applied Biosystems' own DNA Fragment Analysis by Capillary Electrophoresis user guide, which names them and states each one's trade-off. Peak detection is checked against Fragman, the maintained R package for fragment analysis, run on the same real files. The interpolation and the regression are checked against R's own spline and lm. The sizing itself is checked against the one authority that cannot be argued with: the size standard's own fragments, whose lengths are known by construction. The genotype is checked against the published profile of control DNA 9947A, the US forensic reference standard, on a positive-control run for each of four kits, each called against its own allelic ladder at the thresholds OSIRIS publishes for that kit.