Lume Test Plan — Time-Resolved TLF for Bacterial & Source Discrimination

A research test plan to evaluate whether the Lume sensor's nanosecond-scale fluorescence-decay capability and 2,600-condition LED × SiPM-bias sweep can resolve information beyond a single integrated TLF reading — up to and including bacterial source-class and species discrimination.

Draft 0.2 · CU Boulder Mortenson Center / Virridy · NSF ASCEND Engine

Contents

1. Background

1.1 What's actually known about lifetime-based species discrimination

The original demonstration is Dalterio et al. 1986[1], who measured biexponential protein-tryptophan decays from five live bacterial species (S. epidermidis, P. fluorescens, E. cloacae, E. coli, B. subtilis) under 290 nm excitation and 330–340 nm emission. Each species fit a sum of two exponentials with characteristic parameters — but the inter-species spread was small: τ1 = 1.93–2.27 ns and τ2 = 5.08–6.02 ns across the panel. The authors themselves noted that this narrow range "may limit the use of these parameters alone for the identification of bacteria." Their positive species claim required combining steady-state spectra with the lifetime parameters, not lifetime alone.

The clearest modern follow-up is Awad et al. 2014[2], who used fiber-coupled TCSPC on bacterial colonies and reported >98% cross-validation accuracy classifying S. aureus, P. aeruginosa, S. typhi, and K. pneumoniae. Their three-exponential fits gave τ3 separations as wide as 1.5 ns — but the discriminating signal was attributed primarily to NADH rather than tryptophan. Bhattacharjee et al. 2017[3] reproduced this on a panel that overlaps Dalterio's (E. coli, S. enterica, P. aeruginosa, B. subtilis, S. epidermidis) using two-photon FLIM with phasor analysis, again attributing discrimination chiefly to NADH bound/free ratios. The general review Ammor 2007[4] is explicit: intrinsic-fluorescence bacterial classification works in practice through multivariate spectral chemometrics (PCA, FDA, EEM-PARAFAC), not through τ alone.

Spectrally, NADH excites at ~340 nm and emits at ~460 nm[5]; the Lume's 275 nm excitation / 340 nm emission window collects tryptophan, not NADH. The strongest published lifetime-discrimination results (Awad, Bhattacharjee) therefore live in a spectral window the current Lume does not see. The Trp-only effect sizes are correspondingly narrower, and adding a NADH channel is the highest-leverage future hardware path.

1.2 Bulk TLF in environmental water — what the field has learned

Steady-state, intensity-only TLF for E. coli prediction in surface and groundwater is mature. Sorensen et al. 2018[6] defined a tryptophan-equivalent threshold of ~1.3 ppb for thermotolerant-coliform contamination across India, Malawi, South Africa, and Zambia (Spearman ρ = 0.80 with TTC, ~15% false-negative / ~18% false-positive). Nowicki et al. 2019[7] and Heibati et al. 2017 reported ρ = 0.71–0.77 against E. coli in groundwater; turbidity alone gave only ρ = 0.48. Khamis et al. 2015[8] quantified the dominant confounders: TLF declines ~1%/°C with temperature, is amplified by clay turbidity below ~150 NTU and silt below ~650 NTU, and is subject to inner-filter effects at high DOM. The canonical confounder catalog is Carstea, Bridgeman, Baker & Reynolds 2016[9].

The single most important caveat for the field application: Cumberland & Baker 2007[10] showed that 32–86% of TLF signal in environmental waters survives 0.2 µm filtration, and Sorensen et al. 2020[11] reinforced this in groundwater: most of the field-measurable TLF is dissolved or sub-cellular, not intact bacterial cells. Any lifetime-based species-discriminator claim has to address what fraction of the signal can plausibly be cellular at field-relevant DOM background. In practice this argues for size-fractionation as a deliberate part of Stages 1–3, and against any framing that implies the bulk TLF lifetime is a single "bacterial fingerprint."

1.3 What's novel here

To our knowledge no field-deployed lifetime-resolving TLF instrument has been published for water-quality monitoring. The Sorensen / BGS Proteus and Chelsea UviLux portable instruments are intensity-only[12]; the FLIM bacterial-discrimination literature (Awad, Bhattacharjee, Shofner-Brandes 2024[13]) is laboratory two-photon work. The novel contribution proposed here is bringing TCSPC-like discrimination capability into a continuous, in-situ, low-cost field sensor — a category gap that exists regardless of how strong the species discrimination ultimately turns out to be.

2. The Lume capability

The Lume measures TLF using a UV LED (~275 nm) and a SiPM detector windowed at ~340 nm. Per measurement cycle, the firmware sweeps ~2,600 combinations of LED current/duration and SiPM bias voltage, recording the full time-resolved waveform for each combination.

Lume time-resolved waveform across a SiPM-bias sweep. A constant-power, constant-duration LED pulse (on at sample ~540, off at ~600) is held while SiPM bias is increased frame by frame. Baseline (sample 500–540) tracks the bias-dependent SiPM dark count. After LED-off at ~600, the signal decays through ~50 samples back toward baseline — the fluorescence-decay tail (convolved with the SiPM/electronics impulse response). Higher bias raises both signal and noise floors; the slower decay shoulder becomes increasingly visible. The plateau clamp near 3,700 ADC is likely amplifier/ADC saturation, not a real photon-rate ceiling.
~2,600(LED × bias) per sample
~50 nsvisible decay window (TBD vs. IRF)
12+ bitADC dynamic range per bin

SiPM-specific considerations

SiPMs are not single-photon TCSPC detectors in the strict sense; they are analog sums of microcell pulses[14]. Practical implications for lifetime work:

3. Hypothesis hierarchy

Three nested hypotheses, ranked by ambition. Failing a higher hypothesis still leaves the lower ones intact and useful. Effect-size targets are anchored to the published literature in §1.

LevelHypothesisOperational test
Floor Floor The Lume's decay-tail + bias-sweep signature can distinguish contamination categories in real water (raw sewage / animal-fecal / clean DOM background) better than a single integrated-TLF channel can. Multi-class classifier on labelled source samples with multi-marker ddPCR ground truth; balanced accuracy > 70% across 3+ classes.
Stretch Stretch The Lume can distinguish bacterial groups (e.g., gram-positive vs. gram-negative; coliform vs. enterococci vs. environmental) under controlled conditions. Pure-culture and spike-in studies; balanced accuracy > 75% across ≥ 3 groups under leave-one-replicate-out CV.
Reach Reach The Lume can distinguish bacterial species at Dalterio-level effect sizes (τ spread 0.3–1.0 ns; spectral cofactors permitted). Replication of Dalterio's 5-species panel; species cluster significantly in phasor / lifetime feature space at > 70% per-species accuracy under LOO-CV.

4. Test plan stages

Stage 0 — Hardware characterization Gate to all later stages

Site: CU Boulder Mortenson Center optics bench · Duration: ~3 weeks · Output: per-bias IRF, time-per-sample, afterpulsing & crosstalk corrections, lifetime-standard recovery

Quantify exactly what timescales the Lume resolves and what SiPM-specific artifacts must be corrected. Without this, every later stage's interpretation is ambiguous.

Procedures

Pass criterion (gate to Stage 1): recovered single-exponential τ for tryptophan within ±0.3 ns of the literature value, with IRF FWHM characterized at every (LED, bias) combination used in later stages.

Stage 1 — Pure-culture replication of Dalterio Reach hypothesis

Site: CU Boulder Mortenson Center microbiology lab · Duration: ~6 weeks · n: 5 species × 3 growth conditions × 5 replicates = 75 measurements + paired filtrate controls

Replicate the 1986 result on the Lume. If we can match Dalterio's effect sizes within factor-of-two error bars, the path to source discrimination is justified.

Organisms (matched to Dalterio's panel)

Procedure

  1. Grow each species in standard medium to mid-log phase.
  2. Wash cells in phosphate buffer to remove media autofluorescence.
  3. Resuspend at 106, 107, 108 CFU/mL.
  4. Acquire full Lume waveform (all 2,600 conditions) on each suspension; replicate ×5 across days.
  5. Critical addition: for each suspension, also acquire a 0.2 µm-filtered companion sample. The "cellular signature" is the (whole − filtered) difference; the filtered fraction sets the baseline DOM signal we'd see in the field[10].
  6. Run paired Colilert or plate counts for ground-truth concentration.
  7. (Strongly recommended) Acquire benchtop TCSPC reference decays for each species in parallel as ground truth on the Lume's lifetime fits. CU Boulder has multiple TCSPC systems available.

Analysis

Pass criterion (Reach): ≥ 70% per-species classification accuracy under LOO-CV on the (whole minus filtered) signal. Partial pass (50–70%) drops the program to Stretch.

Stage 2 — Source discrimination with multi-marker ddPCR ground truth Floor hypothesis

Site: Boulder area + Lab · Duration: ~8 weeks · n: ~30 grab samples per source class × 3+ classes

Move from defined cultures to defined source matrices. Tests whether the floor hypothesis holds in real environmental media with the dissolved-organic-matter and turbidity confounders that field samples carry.

Source classes

Ground-truth methods (upgraded from draft 0.1)

The current MST gold standard is multi-marker ddPCR Bayesian inference[19][20]. Use:

Pass criterion

Balanced classification accuracy ≥ 70% across the 3+ classes under leave-one-day-out CV, with confusion matrix interpretable in domain terms and with the Lume signature outperforming an integrated-TLF-only baseline.

Stage 3 — Mixed-source field deployment Stretch hypothesis

Site: Boulder Creek longitudinal transect · Duration: 12 weeks · Cadence: continuous Lume + weekly grab co-locations

The integration test: deploy continuously at sites with known mixed inputs and ask whether the Lume's high-dimensional signature recovers the source mix that the multi-marker ddPCR panel detects.

Sites (proposed)

  1. Boulder Creek above Eldorado Springs — clean reference.
  2. South Boulder Creek through urban Boulder — mixed urban runoff.
  3. Boulder Creek below 75th St. WWTP — treated effluent dominant.
  4. Boulder Creek through agricultural reaches downstream of Erie — mixed agricultural inputs.

Cadence

Analysis

Train Stage 2-derived classifier on field data; report Lume-derived source-mix posterior vs. ddPCR-derived ground truth at each weekly time point. Quantify temporal lead/lag of Lume signal vs. grab observations during storm events — the continuous channel's natural advantage.

5. Analysis methods

Per-waveform feature extraction

  1. Baseline subtract using pre-pulse samples per (LED, bias) combination.
  2. Afterpulsing/crosstalk correction using the Stage 0 SiPM transfer functions.
  3. IRF deconvolution using the Stage 0 per-bias IRF (Wiener or iterative reconvolution fit).
  4. Phasor transform[17][18]: Fourier transform the deconvolved decay; each (LED, bias) combination becomes a (G, S) point in phasor space. Pure species fall on the universal semicircle; mixtures fall on chords. This is now the primary representation, in light of the lit review — phasor is fit-free and demonstrably more robust than biexponential fitting in the FLIM bacterial-discrimination literature[3].
  5. Biexponential reconvolution fit as a secondary, model-based check that recovers (α1, τ1, α2, τ2) per condition. Useful for direct comparison with Dalterio numbers but more sensitive to IRF errors than phasor.
  6. Bias-sweep response curve. How the integrated TLF and recovered features change with bias is itself a fingerprint — saturation behavior is fluorophore-density dependent.
  7. Whole-minus-filtered residual: the cell-fraction signal isolated from the dissolved-DOM background[10][11].

Classifiers

Statistics

6. Risks & known limitations

Time resolution may be insufficient. Dalterio's inter-species spread is 0.3–1.0 ns. If the Lume's effective time-per-sample > 1 ns or per-bias IRF FWHM > 2 ns post-deconvolution, biexponential fits will not resolve species differences. Stage 0 explicitly measures this; failure here drops the program to Floor only.
Most TLF in environmental waters is extracellular. 32–86% of bulk TLF survives 0.2 µm filtration[10][11], meaning most of the field-measurable signal is dissolved Trp/peptides, not whole cells. The (whole minus filtered) residual approach in Stages 1–3 is the principled mitigation but may produce a small noisy difference on top of a large noisy total — SNR may be the binding constraint, not lifetime resolution.
Spectral window misses NADH. The strongest published lifetime-discrimination signal sits in the NADH window (~340/460 nm), not the Trp window (~280/340 nm) the Lume measures[2][3]. A second-channel hardware revision adding NADH detection is the highest-leverage future upgrade if Trp-only results disappoint.
SiPM-specific artifacts. Afterpulsing and crosstalk distort early-time bins; soft saturation distorts the LED-on portion at high power. These are well-characterized at the device level[14][15] but require explicit Stage 0 measurement on this hardware to set the operating window.
Sample preparation effects. Dalterio measured washed cell suspensions, not native water. Pure-culture work matches that; field work does not. Differences in cell physiology (stationary phase, biofilm fragments, viable-but-non-culturable cells) can shift lifetimes independent of taxonomy.

7. Decision criteria

After...GO ifSTOP / pivot if
Stage 0 Tryptophan single-exponential τ recovered within ±0.3 ns of literature; per-bias IRF FWHM characterized; afterpulsing/crosstalk correction available. Recovered τ off by > 1 ns or IRF FWHM > 5 ns; pivot to non-lifetime feature engineering of the bias-sweep response only.
Stage 1 ≥ 3 of 5 species discriminable above noise on the (whole minus filtered) signal under LOO-CV. No species discrimination on whole-minus-filtered; fall back to Stage 2 (Floor hypothesis only) and consider NADH-channel hardware path.
Stage 2 ≥ 70% balanced accuracy across 3+ source classes and Lume signature outperforms integrated-TLF-only baseline. Lume not better than integrated-TLF baseline; conclude lifetime adds no information at field-relevant DOM.
Stage 3 Lume-derived source-mix posterior correlates (r > 0.6) with weekly ddPCR markers. No correlation; report negative result honestly.
Reporting commitment. Negative results from any stage are publishable and useful. The plan commits to publication of the IRF, time-resolution, and SiPM artifact characterization (Stage 0) regardless of outcome — this is itself a contribution to the time-resolved-fluorescence sensor literature, given no field-deployable lifetime-resolving TLF instrument has been published.

References

[1]Dalterio, R.A., Nelson, W.H., Britt, D., Sperry, J.F., Tanguay, J.F., Suib, S.L. (1986). The steady-state and decay characteristics of primary fluorescence from live bacteria. Applied Spectroscopy 40(1):86–90. opg.optica.org/as/abstract.cfm?uri=as-40-1-86

[2]Awad, F., Ramprasath, M., Mathivanan, N., Surianarayanan, M. (2014). Fluorescence lifetime imaging of bacterial colonies. ISRN Spectroscopy. PMC4897486

[3]Bhattacharjee, A., Datta, R., Gratton, E., Hochbaum, A.I. (2017). Metabolic fingerprinting of bacteria by fluorescence lifetime imaging microscopy. Scientific Reports 7:3743. nature.com/articles/s41598-017-04032-w

[4]Ammor, M.S. (2007). Recent advances in the use of intrinsic fluorescence for bacterial identification and characterization. Journal of Fluorescence 17(5):455–459. PMID 17624580

[5]Lakowicz, J.R. & Szmacinski, H. (1992). Fluorescence lifetime imaging of free and protein-bound NADH. PNAS 89(4):1271–1275. pnas.org/doi/10.1073/pnas.89.4.1271

[6]Sorensen, J.P.R., Nayebare, J., Carr, A.F., Lyness, R., Campos, L.C., Ciric, L., Goodall, T., Kulabako, R., Rwabulanga, E., Pedley, S., Taylor, R.G., Read, D.S. (2018). Tryptophan-like fluorescence as a high-level screening tool for detecting microbial contamination in drinking water. Science of the Total Environment. PMID 29890592

[7]Nowicki, S., et al. (2019). Tryptophan-like fluorescence as a measure of microbial contamination risk in groundwater. Science of the Total Environment 646:782–791.

[8]Khamis, K., Sorensen, J.P.R., Bradley, C., Hannah, D.M., Lapworth, D.J., Stevens, R. (2015). In situ tryptophan-like fluorometers: assessing turbidity and temperature effects for freshwater applications. Environmental Science: Processes & Impacts. PMID 25756677

[9]Carstea, E.M., Bridgeman, J., Baker, A., Reynolds, D.M. (2016). Fluorescence spectroscopy for wastewater monitoring: A review. Water Research 95:205–219. PMID 26999254

[10]Cumberland, S.A. & Baker, A. (2007). The freshwater dissolved organic matter fluorescence–total organic carbon relationship. Hydrological Processes. Foundational evidence that 32–86% of TLF survives 0.2 µm filtration in environmental waters.

[11]Sorensen, J.P.R., Diaw, M.T., Pouye, A., Roffo, R., Diongue, D.M.L., Faye, S.C., Faye, S., Cisse Faye, S., Re, V., Goodall, T., Lapworth, D.J., MacDonald, A.M., Read, D.S., Tijani, M.N., Taylor, R.G. (2020). In situ fluorescence as a faecal contamination indicator: characterising the dissolved nature of tryptophan-like fluorescence in groundwater. Scientific Reports 10:15379. nature.com/articles/s41598-020-72258-2

[12]British Geological Survey (2018). Tryptophan-Like Fluorescence: Methods, Applications and Field Experience. BGS Open Report OR/18/042. earthwise.bgs.ac.uk/index.php/OR/18/042_Tryptophan-Like_Fluorescence

[13]Shofner-Brandes, S., et al. (2024). Fluorescence lifetime imaging of bacteria in mineral substrates. Frontiers in Microbiology.

[14]Vinogradov, S. & Popova-Wackerle, K. (2022). Silicon photomultipliers for biophotonics: a review. Sensors. PMC9599212

[15]Krstajić, N., Levitt, J., Poland, S., Ameer-Beg, S., Henderson, R. (2016). 256×8 SPAD line sensor for time-resolved fluorescence spectroscopy. Optics Letters 41(4):673–676. opg.optica.org/ol/abstract.cfm?uri=ol-41-4-673

[16]Lakowicz, J.R. (2006). Principles of Fluorescence Spectroscopy, 3rd ed. Springer. Reference values for tryptophan, NATA, indole, and 2-aminopurine lifetimes; chapters on protein fluorescence, IRF deconvolution, phasor analysis.

[17]Digman, M.A., Caiolfa, V.R., Zamai, M., Gratton, E. (2008). The phasor approach to fluorescence lifetime imaging analysis. Biophysical Journal 94(2):L14–L16. PMC2157251

[18]Ranjit, S., Malacrida, L., Jameson, D.M., Gratton, E. (2018). Fit-free analysis of fluorescence lifetime imaging data using the phasor approach. Nature Protocols 13(9):1979–2004. nature.com/articles/s41596-018-0026-5

[19]Boehm, A.B., et al. (2013). Performance of forty-one microbial source tracking methods: a twenty-seven lab evaluation study. Water Research. researchgate.net/publication/251568043

[20]Cao, Y., et al. (2019). Bayesian-based application of multi-marker microbial source tracking. Environmental Science & Technology. acs.org/doi/10.1021/acs.est.9b03843

[21]Bernhard, A.E. & Field, K.G. (2000). A PCR assay to discriminate human and ruminant feces on the basis of host differences in Bacteroides-Prevotella 16S rRNA. Applied and Environmental Microbiology. Defines HF183 marker.

[22]Shanks, O.C., Kelty, C.A., Sivaganesan, M., Varma, M., Haugland, R.A. (2009). Quantitative PCR for genetic markers of human fecal pollution. Applied and Environmental Microbiology 75(17):5507–5513. Defines HumM2 marker.

[23]Stachler, E., Kelty, C., Sivaganesan, M., Li, X., Bibby, K., Shanks, O.C. (2017). Quantitative crAssphage PCR assays for human fecal pollution measurement. Environmental Science & Technology 51(16):9146–9154.

[24]Skala, M.C., et al. (2007). In vivo multiphoton microscopy of NADH and FAD redox states, fluorescence lifetimes, and cellular morphology in precancerous epithelia. PNAS 104(49):19494–19499. pnas.org/doi/10.1073/pnas.0708425104

[25]Stringari, C., et al. (2011). Phasor approach to fluorescence lifetime microscopy distinguishes different metabolic states of germ cells in a live tissue. PNAS 108(33):13582–13587. pnas.org/doi/10.1073/pnas.1108161108

[26]Berezin, M.Y. & Achilefu, S. (2010). Fluorescence lifetime measurements and biological imaging. Chemical Reviews 110(5):2641–2684.

[27]Knopp, W., Thomas, E., et al. (in prep). Lume validation in natural waters — CU Boulder Mortenson Center / Virridy, NSF ASCEND.