Droplet classification

1. 01Sorts the droplets of a two-target ddPCR well into the four populations — NN, NP, PN, PP — under THREE published methods that answer different questions, run on the same droplets so their disagreement is visible
2. 02Fixed thresholds: eight axis-aligned bounds drawn per class. Reproducible and auditable, the only method that does not move when the data does, and the only one whose numbers are your responsibility
3. 03Overlapping threshold boxes are REFUSED rather than warned about, because with them the nested tests decide a droplet’s class by which test ran first instead of by where the droplet sits
4. 04Clustering: four clusters fitted over both channels at once, then named by GEOMETRY rather than by discovery order — bounding boxes on a common scale, each name claiming the cluster nearest its corner, so no cluster takes a name another fits better
5. 05Nearest neighbours: each droplet takes the majority label of its k nearest labelled neighbours from a well you have already classified — the only method that can decline one droplet at a time
6. 06A vote floor for nearest neighbours: below it a droplet is reported as Rain rather than assigned, so the method’s confidence is a control instead of an assumption
7. 07Rain is a CLASS, not a failure — droplets between populations are counted and reported separately, never folded into a population
8. 08Clustering reports no uncalled droplets ever, and the panel says so rather than letting the zero read as a clean well: k-means partitions everything, so a droplet exactly between two populations is assigned as confidently as one in the middle of a cloud
9. 09Droplets per class with each as a share of the well — the counts a copy-number, linkage or fractional-abundance read consumes next
10. 10Up to FOUR targets in one well, told apart by amplitude: density gates find each target’s own cluster, every combination (up to sixteen populations) is placed from those, and rain is walked down the path between clusters
11. 11Copies per droplet for every named marker, from all the populations that carry it: -ln(1 - positives / total)
12. 12Markers are named as in a run template; an empty slot is a target the well does not carry, so its populations are never searched for and never shown as zeros

A port of the Bioconductor package twoddpcr 1.37.0 (Anthony Chiu, CRUK Manchester Institute, GPL-3), validated row for row against twoddpcr running in R over its own KRAS corpus. The up-to-four-target method is a port of the density member of the Bioconductor package ddPCRclust 1.32.0 (Benedikt Brink, Bielefeld University, Artistic-2.0), validated droplet for droplet against it in R over its own eight-well corpus. Kept separate from the droplet-gating port on purpose: that one ports ddpcr, asks whether a droplet is empty per channel from a fitted mixture, and ends in a concentration. This one asks which of four populations, jointly over both channels, and ends in a classification. The panel runs on @ports/discover, @ports/eds, @ports/melt, @ports/qpcr, @ports/rdml, @ports/workflow.

Everything runs client-side. Files are parsed in the browser and projects are saved on this machine, so an unpublished sequence is never uploaded.

Which of the three methods should I use?

None of them is a fallback for another. Fixed thresholds when the call must be auditable and repeatable across plates; clustering when you have no thresholds chosen for your assay and want the well to speak for itself; nearest neighbours when you have already labelled a comparable well and want the option of refusing individual droplets. Where they disagree is an honest measure of how much your answer rests on the method rather than on the data.

Why can clustering not report rain?

Because clustering partitions every droplet by definition — there is no "none of these" for it to return. That is a property of the method, so it is stated rather than left to look like a well with no ambiguous droplets.

What does the vote floor actually do?

It is a floor on the winning proportion among the k nearest labelled neighbours, not a probability model. Below it the droplet is called Rain. Setting it to zero disables it, which is what the upstream package does by default.

My reaction has three or four targets. Can I use this?

Yes, with the up-to-four-targets method. Name the markers the well carries and leave the others empty; it finds one cluster per named target and every combination of them, and reports copies per droplet for each marker.

Do I need a labelled well to use this?

Only for nearest neighbours. Fixed thresholds and clustering both work on a single unlabelled well.

Is this the tab that gives me a concentration?

No. Droplet gating ends in copies per microlitre; this ends in four counts. They are different questions on the same droplets, and they sit one tab apart.