The future of SERS in clinical diagnostics

This article builds on three earlier posts. The physics of Raman scattering and of the plasmonic enhancement is derived in Vibrational spectroscopy, plasmonics and SERS; the chemistry of label-free and labelled SERS — adsorption, Raman reporters, tags, hot spots — in Label-free and labelled SERS; and antibody capture layers in Putting antibodies on an SPR sensor. Here the question is different: not how SERS works, but what it takes for it to change a clinical decision. Numbers in square brackets refer to the references.

01What a clinic needs from a test

A diagnostic test earns its place when it answers a clinical question faster, cheaper, closer to the patient or more completely than what is already used — and does so reliably enough that a decision can rest on it. Today's workhorses set the bar: PCR for nucleic acids, immunoassays (ELISA, chemiluminescence and the lateral-flow strip) for proteins and antigens, culture and mass spectrometry for microbes, histology for tissue.

SERS brings four properties that none of these combine [1–4]:

• Sensitivity down to single molecules in the best hot spots, and picomolar to femtomolar detection limits in well-designed assays.
• Multiplexing from one excitation: Raman bands are 10–20 cm⁻¹ wide, so many reporters, or the many bands of one sample, can be read in one spectrum (the mathematics of multiplexing).
• Molecular specificity without a label: the spectrum is a vibrational fingerprint of what sits on the metal (Raman scattering).
• Portability: handheld Raman spectrometers and printed or colloidal substrates are cheap enough for the point of care.

The same reviews list the obstacles just as consistently: variable substrates, sample matrices that foul the surface, quantification that depends on rare hot spots, and validation studies that are too small. The rest of this article is about those two lists.

02Two roads: reporters and fingerprints

Clinical SERS follows the two roads described in detail in the article on label-free and labelled SERS:

• Labelled (indirect) SERS: an antibody, aptamer or other receptor recognises the target, and a nanoparticle carrying a bright Raman reporter reports it. The selectivity comes from biology, the signal from chemistry, and the format looks like an immunoassay with a better readout.
• Label-free (direct) SERS: the sample — serum, urine, saliva, breath, cells — is put on the metal, and its own spectrum is recorded. No receptor is needed, so one measurement can carry information about many molecules at once; but the selectivity must now come from the spectrum itself and from the analysis.

The two are at very different distances from the clinic, and they face different problems.

03Where tag-based SERS is close to the clinic

Tag-based assays inherit decades of immunoassay practice: the clinical question (is this protein above a cut-off?) and the validation path are familiar. An early landmark used gold particles carrying a nitro reporter and a detection antibody in a sandwich assay for prostate-specific antigen, reaching about 1 pg/mL in human serum [7] — the format and its mathematics are described in assay formats.

The most practical vehicle is the lateral-flow strip. Replacing the coloured gold particles of a rapid test with SERS tags keeps the cheap format but turns a yes/no line into a number, read by a portable Raman reader, and allows several targets on one line [8]. The detection limit of such assays is set less by tag brightness than by the blank — tags that stick where they should not — which is why the surface chemistry of non-specific binding matters as much as the plasmonics.

Two other uses are close to routine need. Therapeutic drug monitoring — measuring the blood level of drugs with a narrow therapeutic window such as some chemotherapeutics, immunosuppressants and antibiotics — needs fast, cheap, quantitative tests at the bedside, and SERS can measure many drugs directly [18]. The open issues are the same as everywhere: substrate-to-substrate variation and a method to validate for each drug.

04Seeing cancer during surgery

The most striking demonstrations of SERS in medicine are in imaging. Nanoparticles combining MRI, photoacoustic and Raman contrast delineated brain-tumour margins in mice, with the Raman signal revealing residual tumour at the edge of the resection [9]. Gold nanostars with a near-infrared resonant reporter reached a detection limit of 1.5 fM and imaged both large and microscopic tumours, including premalignant lesions, in several mouse cancer models without any targeting molecule [10]. Combining SERS with spatially offset Raman spectroscopy recovered signals of several tags through about 2 cm of tissue [12].

The clinic has already accepted Raman itself in this setting: a handheld Raman probe — without nanoparticles — distinguished brain cancer from normal brain during surgery in patients [11]. The step from there to SERS imaging is not the optics; it is the nanoparticles. An injected gold–silica particle is a drug as far as regulators are concerned, and its toxicity, clearance and long-term fate must be established in humans before it can be used. That makes imaging the most dramatic and, at the same time, the slowest of the clinical roads [5].

05Fingerprints of disease in biofluids, breath and pathogens

Label-free SERS of body fluids is where most clinical papers are published. Blood plasma or serum, urine, saliva and tears are mixed with silver or gold colloids or dropped on a substrate, and the spectra of patients and controls are compared. The spectra are dominated by a few small molecules that adsorb strongly and scatter well — in blood often purine metabolites such as uric acid and hypoxanthine — and differences between groups are usually found by multivariate analysis rather than by a single band [6]. The same review argues that the most realistic first role for biofluid SERS is low-cost screening, not definitive diagnosis.

Three recent lines show where the field is heading:

• Exosomes and liquid biopsy. SERS spectra of exosomes isolated from plasma, analysed by deep learning, separated lung-cancer patients from controls with an area under the ROC curve of about 0.91, also for stage I disease [15]. A follow-up study extended the approach to six cancer types in one test, reporting 90 % sensitivity at 94 % specificity and a prediction of the tissue of origin for most positive patients — in a retrospective study [16].
• Breath. A handheld breathalyser with a chip of three receptor-modified SERS sensors on silver nanocubes, read by a portable spectrometer, screened 501 people for COVID-19 at hospitals and airports and reported more than 95 % sensitivity and specificity, with a result in under five minutes [17].
• Pathogens. Deep learning on Raman spectra of bacteria identified 30 common pathogens and chose the right antibiotic treatment for clinical isolates with 99.7 % accuracy using ten spectra per isolate [13] — a Raman (not SERS) study, but the template for what SERS adds in speed and sensitivity. In our own work, a single measurement on a receptor-free SERS platform, analysed by machine learning, detected and classified five respiratory viruses [14].

06Receptor-free sensing: why removing the antibody is the hard part

Removing the receptor is attractive for clinical reasons. Antibodies and aptamers must be developed for every new target, are expensive, vary from batch to batch, need cold storage and can be blind to a new variant. A receptor-free sensor can, in principle, read a new pathogen or panel of markers by retraining a model, and one measurement contains information about everything that reached the surface.

But the receptor did a job, and removing it moves that job somewhere else:

• Selectivity moves to the surface chemistry. Without a capture layer, whatever adsorbs most strongly and scatters most efficiently dominates the spectrum — usually not the analyte of interest. The competition is described by the Langmuir isotherm with competing species (getting the molecule onto the metal).
• Specificity moves to the data analysis. The disease signal is a small change in a large, variable spectrum. A model must find it — and must be shown to find it, not differences between days, batches of substrate, spectrometers or the people who prepared the samples.
• Quantification moves to the statistics of hot spots. Spot-to-spot variation is large because a few sites dominate the signal; averaging and internal standards are needed (hot spots and internal standards).
• Interpretation becomes harder. A band assignment in a complex fluid is a hypothesis, and the bands a classifier uses should be checked against reference spectra (reading a biological spectrum).

Every one of these can be handled, but none can be skipped. The sections below are about how.

07The statistics every diagnostic study has to face

A diagnostic test is compared with a reference standard (PCR, culture, histology). Counting the four outcomes — true and false positives (TP, FP) and negatives (TN, FN) — gives the two properties of the test itself:

$$\text{sensitivity} = \frac{TP}{TP + FN}, \qquad \text{specificity} = \frac{TN}{TN + FP} .$$

Sensitivity is the fraction of diseased people the test finds; specificity the fraction of healthy people it correctly clears. A clinician, however, holds a result and wants the reverse: given a positive, how likely is disease? That is the positive predictive value, and by Bayes' theorem it depends on the prevalence \(p\) — the fraction of the tested population that actually has the disease:

$$\mathrm{PPV} = \frac{\mathrm{Se}\,p}{\mathrm{Se}\,p + (1 - \mathrm{Sp})(1 - p)}, \qquad \mathrm{NPV} = \frac{\mathrm{Sp}\,(1 - p)}{\mathrm{Sp}\,(1 - p) + (1 - \mathrm{Se})\,p} ,$$

where Se and Sp are the sensitivity and specificity as fractions and NPV is the negative predictive value, the probability that a negative result is truly negative. A test with 95 % sensitivity and specificity sounds excellent; screening a population with 1 % prevalence, it gives about five false positives for every true one.

Figure 1 · Interactive

Two more quantities appear in nearly every paper. Most classifiers output a score, and moving the decision threshold trades sensitivity against specificity; the ROC curve plots one against the other over all thresholds, and the area under it (AUC) is the probability that a randomly chosen patient scores higher than a randomly chosen control — 0.5 for a coin toss, 1 for a perfect test. An AUC summarises a model but does not choose an operating point; the clinical use decides which error is worse. And the analytical limit of detection — the lowest concentration distinguishable from a blank, \(3\sigma_\text{blank}\) divided by the calibration slope (derived here) — matters only relative to the clinical cut-off: an assay at femtomolar sensitivity for a marker whose decision level is nanomolar has gained nothing clinically, but may have gained new ways to be fooled by contamination.

08Machine learning without fooling yourself

Label-free clinical SERS depends on multivariate statistics and machine learning — PCA with discriminant analysis, support vector machines, random forests and, increasingly, deep neural networks [19, 20]. The most common way such studies overestimate their performance is not the choice of algorithm but the way the data are split.

Each patient usually contributes many spectra (several spots, several replicates). Spectra of one patient resemble each other for reasons that have nothing to do with the disease: their own biochemistry, the substrate piece, the day of measurement. If the spectra are shuffled and split at random into training and test sets, the same patient appears on both sides, and the model can score well by recognising the patient. The test then measures memory, not diagnosis. The figure shows how large the effect can be.

Figure 2 · Interactive

The remedies are well established and described in detail in a protocol for Raman chemometrics [19]:

• Split by patient (and, better, by day, substrate batch and instrument), never by spectrum; do every preprocessing step that learns from data — normalisation statistics, feature selection, hyperparameters — inside the training folds only.
• Keep an independent test set, ideally from another centre or time period, that is used once.
• Report confidence intervals; with 40 patients a 95 % accuracy has a wide uncertainty.
• Check what the model uses — loadings, saliency maps — against known chemistry; a classifier that relies on a substrate band is learning the substrate.
• Make preprocessing transparent and reproducible: baselines and normalisations change spectra enough to change a classifier. (This was one of the reasons for writing our own processing software.)

09Reproducibility and standardisation

Clinical laboratories run the same test on thousands of samples, on several instruments, for years. SERS has a reputation problem here, and the community has started to address it directly. The first large interlaboratory study of quantitative SERS involved 15 laboratories and 44 researchers measuring the same analyte with different substrates and instruments, and proposed a methodology for assessing reproducibility and trueness across laboratories [21]. A consensus paper by leading groups set out the key parameters and good analytical practice for reliable, quantitative SERS: how to report enhancement factors, controls, blanks and substrate characterisation [22].

For a clinical assay this translates into concrete requirements:

• Substrate quality control: lot-to-lot acceptance criteria, shelf-life data, and a reference measurement on every lot.
• Internal standards and calibrators measured with every sample, so drifts in laser power, focus and substrate activity cancel (why they help and when they do not).
• Instrument transfer: a model trained on one spectrometer must be calibrated or transferred to another — part of the modelling protocol in [19].
• Sample handling: collection tubes, anticoagulants, storage time and freeze–thaw cycles change biofluid spectra and must be fixed and reported.

Two developments may ease the quantification problem itself. Digital SERS counts single-molecule events instead of averaging intensities, turning the hot-spot lottery into Poisson statistics; a protocol for it reached picomolar quantification of antibiotics [23]. Uniform substrates and gap-enhanced tags narrow the spread of enhancement factors, so fewer spectra are needed for a reliable average.

10From proof of concept to clinical use

Figure 3

Diagnostic accuracy studies have their own reporting standard, STARD, which lists what a paper must state for its results to be interpretable: how patients were recruited, the reference standard, blinding, flow of participants, and the uncertainty of the estimates [24]. Many SERS studies fall short of it — small, single-centre, case–control cohorts with clear-cut cases and healthy volunteers, which inflate accuracy compared with the mixed population a test meets in practice [4].

Beyond the evidence, a test has to fit a workflow. Who prepares the sample, how long it takes, what it costs per test, how a result is reported and acted on, and what happens when the substrate lot changes are questions that decide adoption as much as sensitivity does. SERS tests that reuse established formats — lateral-flow strips, cartridge readers, microplates — have the shortest path.

11What the next decade is likely to bring

Predictions are cheap; here are the directions the literature points to [1, 4, 8, 20], with the caveat that each still needs the evidence of the previous section.

• Point-of-care readers. Handheld spectrometers with smartphone-level interfaces, reading SERS lateral-flow strips or cartridges quantitatively and for several targets at once.
• Hybrid assays. A receptor to pull the target out of the matrix, combined with label-free reading of its spectrum — keeping selectivity while gaining structural information (for example on variants or modifications).
• Explainable, well-validated machine learning. Models trained on multicentre data, tested prospectively, with the spectral features they rely on linked to known biochemistry; spectral libraries and shared datasets to train them.
• Standard substrates and reference materials, so that a spectrum measured in one laboratory means the same in another, and digital SERS for robust quantification at very low concentrations.
• Microfluidics that integrate sample preparation — separating cells, concentrating exosomes or pathogens — with the SERS readout, removing the operator from the most variable step.
• Imaging and deep-tissue SERS (SESORS) for surgery and endoscopy, once biocompatible probes clear safety evaluation [5, 12].

12What a receptor-free SERS test has to prove

None of these is unique to SERS; they are what every diagnostic is asked. What is unique is the combination of sensitivity, multiplexing and molecular information that SERS can bring once it meets them. That is why we work on it, and why we try to hold our own receptor-free platforms to this list.

13References

1. J. Langer et al., “Present and future of surface-enhanced Raman scattering,” ACS Nano 14, 28–117 (2020).
2. C. Zong et al., “Surface-enhanced Raman spectroscopy for bioanalysis: reliability and challenges,” Chem. Rev. 118, 4946–4980 (2018).
3. D. Cialla-May, X.-S. Zheng, K. Weber and J. Popp, “Recent progress in surface-enhanced Raman spectroscopy for biological and biomedical applications: from cells to clinics,” Chem. Soc. Rev. 46, 3945–3961 (2017).
4. D. Cialla-May, A. Bonifacio, T. Bocklitz, A. Markin, N. Markina, S. Fornasaro, A. Dwivedi, T. Dib, E. Farnesi, C. Liu, A. Ghosh and J. Popp, “Biomedical SERS – the current state and future trends,” Chem. Soc. Rev. 53, 8957–8979 (2024).
5. S. Laing, L. E. Jamieson, K. Faulds and D. Graham, “Surface-enhanced Raman spectroscopy for in vivo biosensing,” Nat. Rev. Chem. 1, 0060 (2017).
6. A. Bonifacio, S. Cervo and V. Sergo, “Label-free surface-enhanced Raman spectroscopy of biofluids: fundamental aspects and diagnostic applications,” Anal. Bioanal. Chem. 407, 8265–8277 (2015).
7. D. S. Grubisha, R. J. Lipert, H.-Y. Park, J. Driskell and M. D. Porter, “Femtomolar detection of prostate-specific antigen: an immunoassay based on surface-enhanced Raman scattering and immunogold labels,” Anal. Chem. 75, 5936–5943 (2003).
8. S. Sloan-Dennison, E. O'Connor, J. W. Dear, D. Graham and K. Faulds, “Towards quantitative point of care detection using SERS lateral flow immunoassays,” Anal. Bioanal. Chem. 414, 4541–4549 (2022).
9. M. F. Kircher et al., “A brain tumor molecular imaging strategy using a new triple-modality MRI-photoacoustic-Raman nanoparticle,” Nat. Med. 18, 829–834 (2012).
10. S. Harmsen et al., “Surface-enhanced resonance Raman scattering nanostars for high-precision cancer imaging,” Sci. Transl. Med. 7, 271ra7 (2015).
11. M. Jermyn et al., “Intraoperative brain cancer detection with Raman spectroscopy in humans,” Sci. Transl. Med. 7, 274ra19 (2015).
12. N. Stone, M. Kerssens, G. R. Lloyd, K. Faulds, D. Graham and P. Matousek, “Surface enhanced spatially offset Raman spectroscopic (SESORS) imaging – the next dimension,” Chem. Sci. 2, 776–780 (2011).
13. C.-S. Ho, N. Jean, C. A. Hogan, L. Blackmon, S. S. Jeffrey, M. Holodniy, N. Banaei, A. A. E. Saleh, S. Ermon and J. Dionne, “Rapid identification of pathogenic bacteria using Raman spectroscopy and deep learning,” Nat. Commun. 10, 4927 (2019).
14. A. Yadav, R. Naik, E. Gupta, P. P. Roy and S. K. Srivastava, “Single-shot, receptor-free, rapid detection and classification of five respiratory viruses by machine learning integrated SERS sensing platform,” Biosens. Bioelectron. 279, 117394 (2025).
15. H. Shin et al., “Early-stage lung cancer diagnosis by deep learning-based spectroscopic analysis of circulating exosomes,” ACS Nano 14, 5435–5444 (2020).
16. H. Shin, B. H. Choi et al., “Single test-based diagnosis of multiple cancer types using Exosome-SERS-AI for early stage cancers,” Nat. Commun. 14, 1644 (2023).
17. S. X. Leong et al., “Noninvasive and point-of-care surface-enhanced Raman scattering (SERS)-based breathalyzer for mass screening of coronavirus disease 2019 (COVID-19) under 5 min,” ACS Nano 16, 2629–2639 (2022).
18. A. Jaworska, S. Fornasaro, V. Sergo and A. Bonifacio, “Potential of surface enhanced Raman spectroscopy (SERS) in therapeutic drug monitoring (TDM). A critical review,” Biosensors 6, 47 (2016).
19. S. Guo, J. Popp and T. Bocklitz, “Chemometric analysis in Raman spectroscopy from experimental design to machine learning–based modeling,” Nat. Protoc. 16, 5426–5459 (2021).
20. F. Lussier, V. Thibault, B. Charron, G. Q. Wallace and J.-F. Masson, “Deep learning and artificial intelligence methods for Raman and surface-enhanced Raman scattering,” TrAC Trends Anal. Chem. 124, 115796 (2020).
21. S. Fornasaro et al., “Surface enhanced Raman spectroscopy for quantitative analysis: results of a large-scale European multi-instrument interlaboratory study,” Anal. Chem. 92, 4053–4064 (2020).
22. S. E. J. Bell, G. Charron, E. Cortés, J. Kneipp, M. Lamy de la Chapelle, J. Langer, M. Procházka, V. Tran and S. Schlücker, “Towards reliable and quantitative surface-enhanced Raman scattering (SERS): from key parameters to good analytical practice,” Angew. Chem. Int. Ed. 59, 5454–5462 (2020).
23. C. D. L. de Albuquerque, R. G. Sobral-Filho, R. J. Poppi and A. G. Brolo, “Digital protocol for chemical analysis at ultralow concentrations by surface-enhanced Raman scattering,” Anal. Chem. 90, 1248–1254 (2018).
24. P. M. Bossuyt et al., “STARD 2015: an updated list of essential items for reporting diagnostic accuracy studies,” BMJ 351, h5527 (2015).

---

The simulations use the textbook models stated in their captions with synthetic data; performance figures quoted from studies are those reported by their authors. This article is a perspective, not clinical guidance. Corrections and questions are welcome at rakesh_n@cpqct.iitr.ac.in.

← Back to the Lab Notebook · Label-free and labelled SERS → · Vibrational spectroscopy, plasmonics and SERS →