Performance & Research

What is Tryptophan-Like Fluorescence?

Proteins fluoresce in the ultraviolet mainly through their tryptophan residues. Bacteria and planktonic algae excited at 290 nm emit with a maximum between 320 and 340 nm, and quenching and denaturation experiments assign that emission to protein tryptophan (Dalterio et al. 1986; Baek et al. 1988). In natural waters this region of the excitation-emission matrix is called Peak T, conventionally placed at 275 nm excitation and 340 nm emission (Coble 1996). Tryptophan-like fluorescence (TLF) is the intensity measured in that region. In situ instruments use broader bands, at 280 to 285 nm excitation and 350 to 365 nm emission in published sensors (Khamis et al. 2015); the Lume measures at 273 nm excitation and 350 nm emission. Tryptophan is one of several contributors in this window. Plant polyphenols and tannins fluoresce near Peak T (Maie et al. 2007), indole from bacterial breakdown of tryptophan fluoresces in the same region (Cumberland et al. 2012), and where protein-like material is scarce the apparent TLF can be entirely bleed-through from neighboring humic-like fluorescence (Sorensen et al. 2018a).

A TLF reading is an optical intensity, and turning it into a bacterial count requires a calibration. In laboratory cultures of E. coli, Bacillus subtilis and Pseudomonas aeruginosa, more than 75% of Peak T fluorescence was inside the cells, and the authors concluded that Peak T measures microbial activity and cannot by itself count bacteria (Fox et al. 2017, 2018). In natural waters the location of the signal varies. In African groundwater a median 96.9% of TLF passed a 0.22 µm filter (Sorensen, Carr et al. 2020), while in six contrasting freshwaters much of it was removed by a 1.2 µm filter (Baker et al. 2007). In lakes TLF tracks the rate of heterotrophic bacterial metabolism (Cammack et al. 2004), and across 469 surface waters and effluents in southwest England it correlated with five-day biochemical oxygen demand (Hudson et al. 2008). The Lume is calibrated against E. coli culture counts in CFU/100 mL, the unit in which recreational water limits are written.

TLF and E. coli rise together where sewage or animal waste is the main source of labile organic matter, because the same discharge delivers both. Combined sewer overflows carry median E. coli of about 5 × 10⁴ to 1.5 × 10⁶ CFU/100 mL (Madoux-Humery et al. 2013; Passerat et al. 2011), and wastewater fluoresces more intensely in the Peak T region than natural waters (Carstea et al. 2016). Tryptophan makes up about 1.1% of E. coli protein by weight (Neidhardt et al., BioNumbers 100642).

TLF is not specific to feces. Peak T fluorescence was present in every bacterial species examined in culture, including the non-enteric B. subtilis and P. aeruginosa (Fox et al. 2017), and a P. aeruginosa strain isolated from an urban river produced mainly tryptophan-like fluorescence (Elliott et al. 2006). Silage liquor, a plant waste, gave a higher ratio of tryptophan-like to fulvic-like fluorescence than cattle or pig slurry (Baker 2002). Fecal indicator bacteria are also a small share of the cells in a contaminated river. Rivers and lakes hold about 10⁶ bacterial cells per mL (Whitman et al. 1998); directly below an intense combined sewer overflow on the Seine, E. coli reached 2.9 × 10⁵ per 100 mL (Passerat et al. 2011), about 3 × 10³ per mL. In Dakar groundwater, total bacterial cells predicted TLF and thermotolerant coliforms did not (Sorensen, Diaw et al. 2020). TLF is a sensitive indicator of sewage and other labile, protein-rich organic inputs and of the microbial activity they support, and it correlates with fecal indicator bacteria where fecal sources dominate those inputs.

Humic-like fluorescence peaks at longer emission wavelengths: Peak A at 260 nm excitation and 380 to 460 nm emission, and Peak C at 350 nm excitation and 420 to 480 nm emission (Coble 1996). Peak A shares the ultraviolet excitation band used for TLF, and humic material still acts on the TLF channel. Its emission tail adds signal (Sorensen et al. 2018a). Dissolved organic matter absorbs strongly near 275 nm, and inner-filter losses exceed 5% once absorbance passes 0.042 in a 1 cm cell (Kothawala et al. 2013). Humic-like components from two natural waters quenched tryptophan fluorescence by 35% and 36% in titrations (Wang et al. 2015). Suspended sediment raised in situ TLF readings by 75% to 83% for silt and 7% to 21% for clay, and TLF falls with warming at about 4% to 5% per °C, more than humic-like fluorescence does (Khamis et al. 2015).

TLF persists longer than culturable fecal indicators, at least in groundwater. In Ugandan supplies monitored for 14 months, thermotolerant coliforms fell from a median 382 to 55 CFU/100 mL within 17 to 23 days while the ranking of sources by TLF stayed stable (Sorensen et al. 2021; see also Sorensen et al. 2015). TLF is also produced in place by microbial growth (Cammack et al. 2004; Fox et al. 2018), and in stored cattle and pig slurry it increased over 50 days (Baker 2002).

Positive relationships between TLF and E. coli or thermotolerant coliforms have been reported in groundwater in Africa and the UK (Sorensen et al. 2015, 2018a, 2021; Nowicki et al. 2019; Ward et al. 2020), in Alpine karst springs (Frank et al. 2018), in rivers in South Africa and England (Baker et al. 2015), in an urban river in Michigan (Danielsen et al. 2025), and in a US estuary (Harrison et al. 2024). The strength of the relationship varies by site. Across catchments, fluorescence and E. coli correlated at r = 0.74 over seven orders of magnitude, with more variable relationships within single catchments (Baker et al. 2015). In African drinking water supplies a fixed threshold classified thermotolerant coliform contamination with 15% false negatives and 18% false positives and could not resolve fewer than 10 CFU/100 mL (Sorensen et al. 2018b). Khamis et al. (2015) recommend site-specific calibration for long-term surface-water monitoring, and correcting raw in situ fluorescence improves its relationship with E. coli (Carstea et al. 2020).

A TLF reading needs no reagents and no incubation (Sorensen et al. 2015), so an in situ sensor reports within seconds, where a culture method such as Colilert-18 takes 18 hours. The Lume corrects TLF for temperature and turbidity before converting it to E. coli, and each site is calibrated against culture counts.

Lume Innovations

The Virridy Lume addresses many of the limitations identified in the TLF literature by coupling optical TLF sensing with machine learning models that account for environmental confounders. Rather than relying on static TLF thresholds, the Lume's algorithm adapts to site-specific conditions.

Key Innovations

Machine learning quantification: Unlike earlier TLF sensors that report relative fluorescent units, the Lume quantifies actual E. coli concentrations using gradient-boosted decision tree models that capture nonlinear patterns in noisy sensor data. The ML model was validated across freshwater and coastal environments, including saline settings where enterococci are the regulatory indicator.

Multi-parameter correction: Built-in turbidity and temperature sensors allow the algorithm to correct for environmental interference that confounds raw TLF readings.

Drinking water performance: Validated against 216 paired Compartment Bag Test (CBT) observations in Rwanda and Kenya. Binary classification at ≥10 CFU/100 mL achieves 85% balanced accuracy (sensitivity 76%, specificity 93%), approaching CBT's own inter-method reproducibility ceiling. WHO risk category agreement within ±1 tier reaches 97%.

Ocean transferability: In initial ocean deployments, the Lume achieved over 76% categorical accuracy using just six training samples, demonstrating transferability across microbial indicators and water matrices.

Patented Technology

Adaptive contamination detection: The Lume uses machine learning to learn normal conditions and trigger contamination alerts based on learned patterns rather than fixed thresholds. (US Patent 11,506,606 B2 — Bedell, Fankhauser, Sharpe, Wilson & Thomas)

Automated system-state classification: Time-series data from water infrastructure sensors are analyzed to classify system states and support operational decisions without manual inspection or rule-based logic. (US Patent 11,507,861 B2 — Wilson, Coyle, Thomas & Croshere)

Validation Details

216 paired sensor–CBT observations were collected across 3 Lume 1.2 sensors, with 204 used after documented exclusions (not-in-water filter, sensor faults, baseline transitions). The CBT (Aquagenx Compartment Bag Test) is a portable, field-deployable reference method with published sensitivity of 94.9% and specificity of 96.6% versus membrane filtration (Stauber et al. 2014). The estimation model is an 8-parameter right-censored linear regression (Tobit) on log₁₀(E. coli + 1), incorporating baseline-subtracted fluorescence, water temperature, turbidity proxy, and per-sensor calibration offsets.

Temporal Coverage

The Lume provides ~288 readings per day versus a single annual CBT grab sample — over 200× the temporal density. Continuous monitoring detects transient contamination events missed by periodic sampling. Gold Standard dMRV Pilot 14 (approved September 2025) authorizes Lume estimates to substitute for laboratory grab-sample requirements under SDWS Parameter 18.

Limitations

Cannot reliably distinguish 0 from 1–9 CFU: 73% balanced accuracy at the ≥1 CFU threshold. The Lume should not be used to certify zero-E. coli conformity.

Per-sensor calibration required: LOOCV agreement varies by sensor (85–98%). Each new sensor requires per-sensor calibration; quarterly CBT verification recommended.

Risk category, not precise count: R² = 0.52. Validated for categorical risk classification, not exact CFU concentrations.

Four-Panel Method Comparison

How does the Lume compare to both EPA-approved laboratory methods? We analyzed paired samples measured by Colilert (IDEXX), membrane filtration (MF), and the Lume sensor across regression, Bland-Altman agreement, and categorical classification frameworks. The fourth panel additionally refits the Lume regression directly against MF, isolating the contribution of reference-method choice from sensor performance.

Key Findings

The headline result: Sensor-to-reference agreement is bounded by reference-method reproducibility, not by Lume hardware. When the Colilert-trained Lume is tested against MF (a reference it was never trained on), R² drops to 0.514 — almost exactly the disagreement between the two lab methods themselves. When the Lume is instead calibrated directly against MF, performance jumps back to R² ≈ 0.87. Whichever culture method you hand the Lume as truth, it fits that method as well as or better than the two EPA-approved methods fit each other.

Why TLF Matters

Microbial water quality is most frequently assessed using E. coli as a risk indicator. However, relying exclusively on culture-based E. coli measurement is limiting: it is slow, expensive, requires trained personnel, and captures only a single point in time. Risk assessments can be significantly improved by integrating TLF as a complementary, continuously monitored parameter.

Advantages Over Traditional Methods

Known Limitations

TLF cannot be used as a direct proxy for E. coli on an individual sample basis. The TLF signal can be influenced by dissolved organic carbon (DOC), humic-like fluorescence (HLF), turbidity, and temperature. Standardization of TLF thresholds associated with different risk levels remains an active area of research. Performance is best in groundwater, which typically has low DOC, consistent temperature, and negligible turbidity.