Shining light of one color on a sample and recording the colors it glows back is a cheap, fast measurement. If you repeat this across many excitation colors, you get an excitation-emission matrix (EEM): a two-dimensional "fluorescence fingerprint" in which each molecule type, such as protein-like tryptophan or the metabolic cofactor NADH, lights up at its own coordinates. Bacteria carry such intrinsic fluorophores, so in principle an EEM could identify a pathogen with no reagents, antibodies, or amplification. In practice a full EEM takes minutes to scan on a benchtop instrument. More importantly, the brightest bacterial fluorophores (NADH and flavins) report how the cell is metabolizing right now, not only what species it is. A fingerprint that shifts with growth phase or sample fluid is a poor basis for diagnosis.
This note proposes to learn a handful of excitation-emission measurements, on the order of three to eight exposures, chosen explicitly so that species identity is preserved while growth phase, suspension fluid, and turbidity are held out. The claim is that a measurement design optimized on clean, single-condition data, which is how existing learned-illumination work proceeds, will pick out exactly the metabolic signal that fails to transfer. Designing the measurement against those nuisance factors should recover most of the accuracy of a full EEM at a small fraction of its acquisition time. If that holds, the result is the specification for a cheap LED-and-filter reader for reagent-free species identification. If it fails, the study yields a clear negative result on how much taxonomic information bacterial autofluorescence actually carries once metabolic state is controlled.
Core Claim
The closest prior work is Coded Illumination and Imaging for Fluorescence Based Classification. It models an RGB camera's reading as an illumination spectrum passed through a sample's EEM and then through the camera's color responses. It jointly learns the illumination spectrum and the channel weights with a linear support vector machine, and reaches 91.17% on Acacia versus Clover honey where standard lights gave roughly 50%. That work supplies the method we reuse: the forward model and the idea of optimizing the light itself. It only ever optimized for separability on bright, transparent liquids with no nuisance variation. The Feasibility of Automated Identification of Six Algae Types Using Feed-Forward Neural Networks and Fluorescence-Based Spectral-Morphological Features shows that a few fixed visible excitation bands classify pigmented algae at 95.7% from spectra alone, against 53.0% from shape alone. That result motivates the approach, but it relies on strong chlorophyll and phycocyanin signals and omits the ultraviolet range where non-pigmented bacteria fluoresce.
The constraining evidence comes from vibrational spectroscopy. Towards high-accuracy bacterial taxonomy identification using phenotypic single-cell Raman spectroscopy data found that accounting for growth stage improved Raman taxonomy by at most 3%, with the choice of algorithm mattering more. That paper also reports degradation when transferring to real environmental samples. A skeptic could read the 3% figure as evidence that growth phase is a minor nuisance. We do not assume it carries over. Raman sees bond vibrations of bulk biomass, whereas fluorescence is dominated by a few redox-active cofactors. Environmental EEM studies show microbial communities reshaping their protein-like fluorescence across metabolic transitions (Biochemical evolution of dissolved organic matter during snow metamorphism across the ablation season for a glacier on the central Tibetan Plateau). Whether growth phase dominates bacterial EEMs is the first thing we measure.
Concretely, we model measurement $k$ as $y_k = s_k^\top X r_k + \epsilon_k$. Here $X$ is the sample's EEM sampled on an excitation-by-emission grid, $s_k$ is a non-negative excitation weighting (an LED mixture or a programmable source), $r_k$ is a non-negative emission weighting (a filter or camera channel), and $\epsilon_k$ is shot and read noise at a fixed photon budget. We choose the $K$ code pairs and a classifier $h$ to minimize the worst-group loss over nuisance groups $g$, which are combinations of growth phase, fluid, and concentration:
$$\min_{\{s_k, r_k\}_{k=1}^{K},\, h} \; \max_{g} \; \mathbb{E}_{(X, c) \sim \mathcal{D}_g} \, \ell\big(h(y_1, \dots, y_K),\, c\big),$$
where $c$ is the species label and $\ell$ is cross-entropy. The new element is the max over nuisance groups. Asano et al. optimize the inner expectation over a single pooled distribution.
Research Program
No bacterial EEM benchmark appears in the material, so the first deliverable is a dataset. Using a benchtop spectrofluorometer, we will record full ultraviolet-to-visible EEMs, with excitation from about 240 to 500 nm and Rayleigh and Raman scatter bands masked. The samples are suspensions of roughly twenty clinically common species, chosen to overlap the pathogen list of Rapid identification of pathogenic bacteria using Raman spectroscopy and deep learning and including MRSA/MSSA pairs. Each species will be sampled at the four growth stages used by the single-cell Raman benchmark (lag, log, early and late stationary) and in three fluids: saline, culture medium, and pooled human urine. Urine is the natural first clinical target because urinary infections reach high bacterial loads, and urine has strong fluorescence of its own that acts as a hard nuisance. Each condition gets several biological replicates and a concentration series, so the minimum detectable load is measured rather than assumed.
The second step is analysis before any optimization. PARAFAC is the standard decomposition that splits a stack of EEMs into a few components, each with a fixed excitation shape, emission shape, and per-sample amount. Fitted to the full tensor, it tells us whether species identity lives in component amounts, which would be fragile under metabolic shifts, or in component shapes and minor components, which would be more robust. We will also report how much variance growth phase explains compared with species.
The third step is measurement design in simulation. We learn codes under the objective above for $K$ from 1 to 12, restricted to two hardware families: arbitrary non-negative spectra, as with a programmable source, and selections among a catalogue of commercial LEDs and bandpass filters. The fourth step builds the best low-$K$ design from LEDs and filters on a microwell reader and validates it on freshly grown isolates and on spiked urine from donors not used in training.
Baselines
The naive reference point is a single fixed measurement: one standard protein-region excitation-emission pair, or a two-channel ultraviolet reading like the WIBS-NEO configuration in The Observation and Characterisation of Fluorescent Bioaerosols Using Real-Time UV-LIF Spectrometry in Hong Kong from June to November 2018. Such instruments separate biological from non-biological particles but are not built for species identity. Beating this reference is necessary but uninteresting.
The upper reference is the full EEM fed to the best of three classifiers: PARAFAC scores with logistic regression; supervised dimensionality reduction (LDA) with a linear model, which won in the single-cell Raman benchmark at 96.1% taxonomy accuracy on its own data; and a small convolutional network over the EEM image. All three use the same splits and the same hyperparameter search budget. This tells us how much information exists before any compression. There is no reported result on Lacuna for bacterial species classification from EEMs, so this number is itself a finding.
The strongest existing method for the compression question is Asano-style learned coding: the same forward model and photon budget, codes optimized for average accuracy over pooled training data, with no nuisance term. It gets the identical code families, $K$ values, classifier class, and tuning budget as our method. On their own task they report 91.17% on honey; on our task no number exists, so we run it ourselves. A further baseline adapts the growth-stage-then-taxonomy hierarchy of the Raman benchmark to EEM features, which handles nuisance at the classifier rather than at the measurement.
The ablation that isolates the key ingredient is our method with the worst-group objective replaced by average loss, everything else fixed. The proposal beats the field if, on held-out growth stages and held-out urine donors, nuisance-aware codes at $K \le 6$ reach at least 90% of full-EEM worst-group accuracy while the ablation stays measurably lower, by more than the across-seed spread, at the same $K$. Raman results from Ho et al. (82.2% isolate-level accuracy over 30 isolates; 89% for MRSA versus MSSA) are on different samples and serve as context only, not as a head-to-head comparison.
Why not just…?
Why not just pick the best few LED bands greedily instead of learning codes? This is a real threat, because bacterial autofluorescence is spanned by a few fluorophores and discrete bands may capture it as well as spectrally shaped light. We therefore treat LED-catalogue selection as one of the two design families inside our method, not as a rival. The contribution we claim is the nuisance-aware objective, not coding per se. If greedy discrete selection under the worst-group objective matches continuous codes, we report that and build the cheaper LED device.
Why not just measure the full EEM and train a robust classifier with augmentation? It is the accuracy ceiling, and if minutes of scanning were acceptable it would win. The decisive comparison is how close our $K$-shot design comes to this ceiling under a matched photon budget. A large remaining gap would mean the compression is not worth it.
Why not sparse sampling plus tensor completion? Spectral Compressed Sensing via CANDECOMP/PARAFAC Decomposition of Incomplete Tensors reconstructs spectra from incomplete samples using a Bayesian low-rank decomposition. Reconstruction, however, spends measurements on recovering everything, including metabolic components that are nuisance for the diagnostic decision. We run it at equal $K$ with a classifier on the completed EEM; we expect task-optimized codes to win at small $K$, and the experiment settles it.
Why not Raman, or a reagent assay? Raman needs one second per cell on gold-coated substrates, and SERS depends on plasmonic substrates whose batch reproducibility Multiplex Detection and Quantification of Virus Co-Infections Using Label-free Surface-Enhanced Raman Spectroscopy and Deep Learning Algorithms leaves unresolved. Reagent tests such as Mercaptophenylboronic Acid‐Mediated Nanozyme Immunochromatographic Assay for Simultaneous Detection of Respiratory Bacteria and Virus reach far lower detection limits (17–21 cells/mL) than autofluorescence plausibly can. We do not claim to compete on sensitivity. The target is high-load, many-species identification with no consumables, where lateral flow needs one antibody line per target.
Proposed Experiments
The first decisive check is a variance decomposition on the full EEM dataset. If growth phase explains more EEM variance than species and the full-EEM classifier falls sharply when trained on log phase and tested on stationary phase, nuisance-aware design is necessary. If not, the Raman finding transfers and the simpler Asano-style objective suffices; we would report that honestly. The second check is accuracy against $K$ for every method, reported as worst-group and mean accuracy with leave-one-growth-stage-out and leave-one-urine-donor-out splits, plus MRSA versus MSSA as a fine-grained stress test. The third is a concentration sweep giving the minimum load at which the chosen design keeps its accuracy, compared against the 10^5 CFU/mL range relevant to urinary infection. The fourth is hardware validation: the built LED reader must reproduce the simulated accuracy within its confidence interval on fresh cultures, with acquisition under one second per well. Mixed-species samples are reserved as a final exploratory test of whether the per-measurement readings can be unmixed linearly, as PARAFAC unmixes overlapping organic matter in Geographical differentiation of riverine DOM composition and source apportionment: A case study of a riverine network of a mountainous stream, a Plain River, and an artificial canal.
References
- Coded Illumination and Imaging for Fluorescence Based Classification
- The Feasibility of Automated Identification of Six Algae Types Using Feed-Forward Neural Networks and Fluorescence-Based Spectral-Morphological Features
- Towards high-accuracy bacterial taxonomy identification using phenotypic single-cell Raman spectroscopy data
- Biochemical evolution of dissolved organic matter during snow metamorphism across the ablation season for a glacier on the central Tibetan Plateau
- Rapid identification of pathogenic bacteria using Raman spectroscopy and deep learning
- The Observation and Characterisation of Fluorescent Bioaerosols Using Real-Time UV-LIF Spectrometry in Hong Kong from June to November 2018
- Spectral Compressed Sensing via CANDECOMP/PARAFAC Decomposition of Incomplete Tensors
- Multiplex Detection and Quantification of Virus Co-Infections Using Label-free Surface-Enhanced Raman Spectroscopy and Deep Learning Algorithms
- Mercaptophenylboronic Acid‐Mediated Nanozyme Immunochromatographic Assay for Simultaneous Detection of Respiratory Bacteria and Virus
- Geographical differentiation of riverine DOM composition and source apportionment: A case study of a riverine network of a mountainous stream, a Plain River, and an artificial canal