BASEL screen — landmarks across 69 phages and 5 doses

Each column is one record. Phages run left to right in ascending mean deviation onset; within each phage the five doses run in ascending multiplicity, so dose 5 (most dilute) is leftmost of its group and dose 1 (most concentrated) rightmost. The sawtooth within a band is the dose response.

Phage amplification rate

A rate, not a time, so it gets its own axes. One point per phage, 69 across. Published is the rate Mulla et al. report for that phage. From collapse and from DO are derived here by their logic rather than their code: if phage amplifies exponentially, starting with less of it delays the culture's turn-over in proportion to the log of the shortfall, so the reciprocal of the slope of collapse time against dilution step is a rate. The dilution factor between plates is not given in the supplement, so these are rates up to one common constant — their ordering is meaningful, their absolute values are not.

Two of these come from curve shape, with no landmark and no dilution series. Steepest fall is the largest negative slope of ln(signal), in ln units per hour: how fast the culture is being destroyed at its worst moment. Descent acceleration is the physically motivated one: if bacteria die at a rate set by how much phage is present, and phage grow exponentially at rate r, then the killing rate itself rises exponentially at r, so plotting ln(−slope) against time through the descent should give a straight line whose slope is the amplification rate. It fits reasonably — median R² 0.70 — so the model is not obviously wrong.

None of them reproduces the published rate, and the gap is in spread rather than rank. Ranked against the published rate: steepest fall +0.568, fall-to-rise ratio +0.367, collapse time +0.296, descent acceleration +0.210, rise-to-fall interval −0.525. Curve shape does carry more of their signal than collapse time does — which is what one would expect, since PHORCE fits the curve — but the decisive difference is elsewhere. Their rates span 5,289-fold across the collection. Ours span 10-fold (collapse), 8-fold (steepest fall), 26-fold (descent acceleration). The ratio between their rate and ours varies 6,651-fold, so this is not one quantity in different units.

Landmarks against one another

The strongest two dozen pairings among the twelve per-record quantities, as a scatter with its rank correlation. Sorted by rank correlation, so the informative pairings come first; pairs sharing fewer than 20 records are omitted. Click any panel to enlarge it. Points are single records, so a phage contributes up to five.

What that leaves. Their published rate is in ml⁻¹ h⁻¹, a second-order constant carrying a volume term; every estimator here is a first-order rate in h⁻¹. A per-volume rate can express differences in adsorption efficiency that a curve's own timing cannot, which would explain a wider spread. Whether that extra range is information their model extracts or structure their model imposes is the question the paper turns on, and nothing here settles it.

These do not agree, and that is the finding. Our rate derived from their collapse times ranks phages only weakly like their published rate — Spearman +0.30 across all five doses, +0.43 over doses 2 to 5 — even though collapse time really is close to linear in dilution step (median R² 0.83 to 0.91). So the disagreement is not that the data are noisy; it is that PHORCE derives its rate by fitting a model to the whole luminescence curve, and the reciprocal-slope shortcut used here is not the same estimator. Any critique of their amplification rates has to reckon with that before it can rest on ours.

Collapse times are theirs, not ours. Mulla Y, Muller J, Trimcev D, Bollenbach T. Extreme diversity of phage amplification rates and phage–antibiotic interactions revealed by PHORCE. PLoS Biology 23(4): e3003065, 2025. doi.org/10.1371/journal.pbio.3003065. The luminescence curves and the collapse time t_col for each well are from their Supplementary Fig. 6; the phage identities from their S12 table. DO and OD_max are computed here from their data.

How OD_max is found. It is the interpolated time at which the smoothed d(signal)/dt crosses zero — the calculus definition the Deviation calculator uses, not the highest single reading. Where no zero crossing exists the naive maximum stands in, which happens on 51 of these 345 records; the other 294 are the calculus value. Note that the signal here is luminescence rather than optical density, so “OD_max” names the landmark rather than the units.

What to look for. Mulla's collapse time should fall as multiplicity rises, giving each band a downward slope from left to right. Where DO and OD_max flatten near the bottom while collapse stays high, the detectors have fired on early noise rather than on the culture turning over — the low-dose failure, visible as a band whose blue and grey points sit flat while the green ones climb.