Colony picking: a camera decides, a pin commits

A colony is a visible pile of cells that grew, in principle, from one founder cell. Picking is the act of converting that optical claim into a physical one: decide from an image which spots on a lawn are single, healthy colonies of the wanted kind, then touch each with a pin so that cells adhere and can be re-grown elsewhere. The whole machine exists because the deciding and the touching demand different senses — one optical, one mechanical — and because a person doing both for thousands of colonies a day makes systematic errors of fatigue. Automation changes the error rate, not the error types.

The decision is a classification on a boundary

Recognition starts with imaging — transmitted, oblique and fluorescent light in combination — and segmentation of colonies against the agar, against each other and against the agar’s own imperfections. The classifier’s hard cases are not random: bubbles and scratches look like colonies to any rule built on “small bright disc”; two adjacent colonies merge into one blob that passes a size filter; a genuine but slow-growing colony is still under the size threshold. There is also a window: pick too early and the colony is too small to yield cells, too late and neighbours have fused with it. Every threshold therefore trades two errors that do not cost the same. A false negative merely loses a candidate — recoverable by relaxing criteria on the next plate. A false positive propagates: a merged or contaminated pick grows on as if it were a clone, and no downstream step sees the mixture unless someone re-verifies. Purity is decided at the moment of the pick and then taken on trust.

Identity is the deeper half of the error budget

The second failure mode is not picking the wrong object but recording the wrong place: a pick delivered to well 73 instead of 74 mislabels everything ever measured on that plate. So the verification loop matters as much as the picker. Mature systems re-image the source plate after the pass — a picked colony should be visibly gone, and the destination array should show growth where records claim it — and track every plate by barcode rather than by deck position. Identity errors are cheaper to prevent than purity errors, which is why the machinery concentrates there; but re-imaging costs a second imaging pass per plate, so it is bought, not assumed.

The sterility tax and the certainty budget

Every pick moves cells from a dense lawn to a fresh medium through a tool that must be clean again before the next colony, because the tool’s previous load is the perfect vector for cross-contamination. The plate is also open — lid off — while the head works, so the whole cycle runs under filtered air, and condensation on the agar defeats imaging and splashes the lawn. Per-colony time is dominated by these disciplines, not by motion: sterilising the pin (heating it, or pressing it into a disposable pad), the guarded approach, the touch. Throughput is simply the reciprocal of the tax, and raising it by shortening the tax spends certainty. The trade is fundamental because the classifier operates where distributions overlap — debris against slow growers, fusions against large singles — and moving the threshold only chooses which error the process will produce more of.