Can fluorescence hyperspectral imaging see a seedling respond to bacteria?

In 2016 we imaged Arabidopsis seedlings by fluorescence hyperspectral imaging before and after treatment with a mock solution, the flagellin epitope flg22 or Pseudomonas syringae. Five genotypes were imaged: Col-0, two fls2 flagellin-receptor mutants and two cml24 Ca2+-sensor mutants. Our reanalysis finds a pathogen signature from 1 hour after inoculation. It also finds that the flg22 signal in this experiment does not depend on FLS2, so it is not perception.

488 nm excitation · 500–800 nm emission400 image cubes · 396 GB 2,486 hand-drawn seedling spectraPython 3.14 MIT code · CC-BY 4.0 data
Manuscript draft (PDF) → Slides (PDF) Tables Code

1 h

Earliest pathogen effect: red emission peak 1.40 nm broader than in Control plates, relative to each plate's own baseline (q = 0.018). It persists to 24 h, joined by more emission below 640 nm from 2 h.

AUC 1.00

Leave-plate-out classification of infected vs Control plates from change-from-baseline spectra at 1, 2 and 24 h (exact permutation p = 0.029).

21% → 4%

The 2016 summaries' intensity drop after treatment shrinks to within 4% once each cube is divided by its exposure time.

AUC 1.00 at hour 0

Plates were separable before treatment, so only change from baseline is evidence. The flg22 signal is as strong in fls2 as in Col-0.

The experiment

Seedlings on square agar plates were imaged in the Gilroy Lab (UW-Madison Botany) with a MacroPhor fluorescence hyperspectral line-scan imager (Middleton Spectral Vision), using 488 nm excitation and recording 324 bands from 400 to 809 nm. Each plate held five genotype lanes and was imaged as three row-scans. Four plates per treatment were imaged before treatment and at 1, 2, 3, 24 and 48 h. An independent November 2016 replicate used 405 nm excitation with Col-0 and both fls2 lines. Facts still to be confirmed (pathovar/strain, flg22 concentration, seedling age) are listed openly in FACTS_PENDING.md.

The decisive control is genetic. An flg22 signature that reflects perception must disappear in fls2 mutants, which lack the flagellin receptor. The pathogen is still perceived through other bacterial patterns, so its signature need not.

What we had to fix first

ProblemWhat we did
Exposure time changed between timepoints (0.305 s at hour 0, 0.203 s at 1–3 h, 0.22–0.27 s later).Divided every intensity by its own exposure, and based the primary analysis on exposure-independent spectral shape.
Hour-0 scans saturated the 12-bit detector (up to 13% of a cube's seedling pixels).Excluded saturated pixels.
Plates differed before treatment (classifier AUC 1.00 at hour 0).Analysed only change from each plate's own hour-0 baseline, with the plate as the experimental unit.
83 hand-drawn spectra, all on Pathogen plates, were exported without dark subtraction (a flat ~109-count offset that mimics broad autofluorescence).Subtracted each spectrum's own mean below the 488 nm excitation line (420–470 nm), where nothing can fluoresce.
Hour-0 plates were scanned "inverted", apart from a few later rescans.Excluded the two plates with mixed hour-0 scans from baseline analyses.
Plates were not always placed in the same orientation.Found each plate's genotype-lane direction from the 2016 hand-drawn ROIs.
November hand ROIs cover one plate per treatment.Treated set-B treatment effects as not estimable from ROIs; plate-replicated analysis comes from the raw cubes.
Three panels: exposure time per timepoint, raw intensity falling after hour 0, and exposure-corrected intensity flat
Exposure time changed between timepoints (a; plate-mean exposure of the Control plates, so hour 0 mixes 0.305 s scans with the 0.203 s rescan of Control-4). The 2016 summary graphs plotted raw intensity (b), so part of their apparent fall after hour 0 is a camera setting. Dividing by each cube's exposure (c) removes that step. Control plates only, set A, median 683-687 nm emission of the 2016 hand-drawn ROIs. Hour 24 mixed exposures within plates (0.222-0.268 s), so its plate-mean exposure is used. Data: results/roi_spectra_A.parquet, metadata/cube_manifest.csv; table: tables/fig1_exposure_artefact.csv.

Results

An early pathogen signature

Change from hour 0 in red-peak width, short-wavelength fraction and red/far-red ratio for Control, flg22 and Pathogen plates
Change from hour 0 in three emission-shape features, Col-0, set A. Points are means over plates (n = 3-4 plates per treatment; the plate is the treated unit) ± s.e.m.; each plate value is the mean of its Col-0 hand-ROI spectra minus the same plate's hour-0 mean. Time axis is categorical (1, 2, 3, 24, 48 h). Plates Control-4 and flg22-2 are excluded because their hour-0 scans were partly upright rescans. Data: results/roi_features.csv.gz; table: tables/fig3_timecourse.csv. Statistics: tables/stats_contrasts_baseline_adjusted.csv.
Change in normalised emission spectrum, treated minus Control, by hour
Treatment-induced change in emission-spectrum shape, Col-0, set A. Each spectrum is area-normalised over 650-800 nm (so exposure and seedling size cancel), averaged per plate, and expressed relative to the same plate at hour 0. Curves show treated minus Control at each time after treatment. No hand ROIs exist for Pathogen at 48 h. Data: results/roi_features.csv.gz; table: tables/fig2_spectral_change.csv.

Classification, and why baseline matters

Plate-level AUC by hour for absolute and change-from-baseline spectra
Can a plate's treatment be told from its seedlings' emission spectra? Logistic regression on the area-normalised 650-800 nm spectrum of Col-0 hand ROIs, evaluated by leave-one-plate-out cross-validation. AUC is computed on plate means (8 plates; 6 for the change-from-hour-0 variant). * exact plate-label permutation p ≤ 0.05. Hour 0 is before treatment: the high hour-0 AUC in (a) shows that plates differ before treatment, so only (b) is evidence of a treatment effect. Data: tables/model_auc.csv.

The flg22 signal is not FLS2-dependent

Col-0-trained signature applied to each genotype
Is the signature genotype-specific? A logistic model trained on Col-0 shape features (change from hour 0) is applied to every genotype on the same plates. Values are the mean predicted probability of 'treated' on treated plates minus that on Control plates (Col-0: out-of-fold). An FLS2-dependent flg22 signature would vanish in fls2-17 and fls2-101; it does not, so the set-A flg22 signal is a plate-level effect. Data: tables/model_specificity.csv.

Every seedling in the raw cubes

MCR components and plate-level changes from the raw cubes
Seedling-level analysis of the raw cubes. (a) Spectra of the three MCR-ALS components fitted to pixels from every set-A cube (results/mcr_components.csv). (b, c) Set A, all genotypes pooled per plate, change from the same plate at hour 0: short-wavelength fraction and exposure-corrected 685 nm emission; unlike the hand ROIs, these include Pathogen plates at 48 h. (d) Set B (405 nm excitation), 2-3 plates per treatment, Laser58 scans. Points are plate means ± s.e.m. Data: tables/seedling_plate_changes.csv; statistics: tables/seedling_stats.csv; table: tables/fig6_seedling_level.csv.

1,623 seedlings segmented from the July cubes reproduce the early signature without the 2016 annotations. 685 nm emission then falls (by -3,462 counts s−1 at 48 h) and seedlings shrink. Curve resolution finds no separate broad autofluorescence component. Bacteria-only scans do not fluoresce, so the signal is the plant's. In the November replicate, a deteriorating Control plate showed the largest short-wavelength rise, so that emission marks damage, not infection specifically. No CML24 effect replicated.

Spectrum explorer

Means of the 2016 hand-drawn ROI spectra (set A), each area-normalised over 650–800 nm so exposure and seedling size cancel. "Change from hour 0" subtracts the same treatment group's hour-0 mean. Data: spectra.json (from results/roi_features.csv.gz).

Code & data