Toward rapid infectious disease diagnosis with advances in surface-enhanced Raman spectroscopy

As of June 18, 2020, the COVID-19 pandemic has infected over 8 million individuals and claimed the lives of nearly 450 thousand individuals worldwide. The emergence of COVID-19 has underscored the importance of rapid, sensitive, and specific diagnostics for tracking and containing an infectious disease.¹ For many infectious diseases, ranging from the viral infections COVID-19 and influenza, to the bacterial diseases tuberculosis and tetanus, nucleotide detection schemes, such as polymerase chain reaction (PCR),^2,3 are the current standard. Here, nucleotides including DNA, RNA, and miRNA are amplified and then detected by fluorescence. Though highly sensitive and specific, PCR requires prior knowledge of strains, sophisticated infrastructure (including thermal cyclers), and costly reagents (including primers, aptamers, polymerases, buffers, and other enzymes). PCR protocols can be time-consuming, requiring several hours to run cycles.⁴ It is therefore challenging to translate PCR across diverse clinical settings, as highlighted by COVID-19.^5–7 To complement PCR, antigen detection based rapid tests and immunoassays such as fluorescence-based ELISA (an “enzyme-linked immunosorbent assay”) have gained traction.⁸ However, these immunoassays are dependent on specific target antigen/antibody binding for each infectious agent and therefore require knowledge of which pathogens might be present; fluorescent based immunoassays are also influenced by low sensitivity, higher limits of detection (LOD), long assay times in the case of ELISA, and fluorophore bleaching and blinking. Other spectroscopic techniques, such as matrix-assisted laser desorption/ionization time-of-flight (MALDI-TOF) mass spectrometry⁹ promise high specificity, but require culturing of infectious agents prior to interrogation, as well as significant capital investment.

In contrast to these state-of-the-art techniques, Raman spectroscopy promises sensitivity, specificity, and speed without the need for unique sample identification tags, such as primers, antibodies, or antigens. In Raman scattering, photons inelastically scatter from molecules; the energy exchanged between photons and the vibrational modes of the molecules gives rise to photons with either lower energy (a Stokes Raman shift) or higher energy (an anti-Stokes Raman shift), as shown in Fig. 2(a). Raman spectroscopy is highly promising for biomedical applications. Foremost, it can be used as a label-free technique that relies on the intrinsic signals from the targeted sample and can, therefore, directly identify molecules, viruses, pathogens, or diseased cells without the use of markers. Alternatively, Raman labels can be used for specific targeting of biological samples. Unlike conventional fluorescent labels, Raman reporters do not suffer from photobleaching. Raman spectral peak widths are also typically much narrower than fluorescence peaks, enabling higher-order multiplexing to uniquely identify multiple biomarkers simultaneously. The technique is also generally less affected by water absorption than infrared spectroscopy, critical for probing biological specimens.¹⁰ Finally, Raman spectroscopes can be made portable,¹¹ with the potential for personalized medicine and point-of-care applications.

Though only one in 10⁶ photons is scattered inelastically, Raman signals can be significantly enhanced using a variety of substrates that control light–matter interactions, an approach termed surface-enhanced Raman spectroscopy (SERS). Generally, SERS relies on metallic substrates that, by virtue of their plasmonic resonances and chemical effects (such as charge transfer ability), enhance both the local electric field from the incident light and the Raman scattered light from the sample;^12–17 resonant dielectric substrates and metasurfaces have also gained traction as SERS substrates.^18,19 SERS can readily detect clinically-relevant biomarker molecules including antibodies, nucleic acids, and peptides,^20–23 even at the single molecule level.^24–31 Such molecular detection schemes have been applied to applications as diverse as tumor margin detection³² and monitoring of chronic hypertension and diabetes.³³ 

For infectious disease detection, SERS can be extended beyond molecules to more complex targets such as viruses and bacteria.^{5–7,34–39} For example, SERS has been implemented to detect rhinovirus, influenza, and parainfluenza within several minutes with as little as 10² EID50/μl (50% egg infective dose per microliter) at 90% specificity; these accuracies are comparable to a quantitative reverse transcription PCR (RT-qPCR), with the added advantages of rapid diagnostics, rapid enrichment, and no prior knowledge of the target virus.^5,6,40,41 Likewise, SERS has been used for label-free human immunodeficiency virus (HIV) diagnosis (monitoring the p24 antigen in blood plasma), resulting in 97.5% sensitivity and 95% specificity with a limit of detection (LOD) of 95–100 copies/ml; viral load monitoring could also be achieved via specific Raman peak intensity changes.⁴² Beyond viral infections, SERS has been used to detect bacterial infections, including identifying bacterial species and strains, distinguishing between living and dead pathogens, and determining antibiotic susceptibility based on differences in spectral features.^43–46 More recently, the approach was used to analyze metabolites from mixed bacterial colonies, showcasing its utility in investigating intracellular signaling and intercolony interactions.⁴⁷ Importantly, SERS does not require sample amplification steps because of the strong, inherent SERS signal that can be detected, especially with the use of Raman reporters.^48,49 Raman reporters are small molecules with large Raman scattering cross sections and are used to label either the SERS substrates or pathogen of interest for sensitive and quantitative detection. In addition, for nanoscale pathogens such as viruses, which cannot be studied under light microscopes, SERS has many advantages that make it competitive for use in diagnostics, as summarized in Fig. 1 and Table I. Furthermore, unlike PCR and rapid tests, SERS enables antiviral resistance testing,⁵⁰ intracellular mutation tracking,⁵¹ and new strain identification.⁵² 

Despite these advances, Raman spectroscopy is generally not used clinically, apart from pharmaceutical applications such as drug development and drug quality assessment.⁵⁹ Medical diagnostics—whether in hospital, in clinical laboratories, or at the point-of-care—requires straightforward hardware and software that can be easily integrated into the workflow of physicians, microbiologists, and pathologists. Indeed, for Raman to translate to the clinic it must provide robust quantitative results with spectral reproducibility, rapid sample preparation, and streamlined low-cost instrumentation.^60–63 These constraints are even more pronounced for resource-limited settings, where diagnostic tools should meet the ASSURED (Affordable, Sensitive, Specific, User-friendly, Rapid and Robust, Equipment free, Deliverable to end-users) criteria set by the World Health Organization (WHO).⁶⁴ Raman spectrometers are sensitive and specific, but often lack a user-friendly element in their sample processing and spectral interpretation; in particular, isolating pathogens and other biomarkers from complex biological samples such as sputum, blood, saliva, and urine generally requires purification and isolation steps prior to Raman interrogation.⁶⁵ Once purified, Raman spectral datasets collected from patient samples can be hard to interpret into clinically useful information and require advanced data analysis techniques. Raman diagnostics can be rapid, but robustness can be a challenge. In particular, quantitative analysis is challenging due to variable signal enhancement, a by-product of nanoparticle (NP) aggregation, polydispersity, and/or analyte distribution on the SERS substrate. Raman spectral setups are also generally large, benchtop-scale platforms that require significant capital investment. However, recent work has shown that Raman spectra can be obtained using a mobile phone camera,⁶⁶ demonstrating its potential as a relatively low-cost, point-of-care tool.

Here, we show how emerging advances in nanophotonics, machine learning, and microfluidics and bioprinting can overcome these barriers to clinical adoption of Raman spectroscopy, especially for viral and bacterial infectious disease diagnostics. In particular, we focus on three complementary approaches, summarized in Fig. 2, that are quickly advancing Raman spectroscopy: (1) Design of data processing algorithms and SERS substrates that improve spectral interpretability. These advances are critical for interrogating patient samples that could contain a wide variety of pathogens (i.e., in developing a sensitive, specific assay for respiratory viruses or bacterial bloodstream infections). (2) Development of new capture techniques and affinity agents, such as aptamers and polymers. These advances are critical for rapid and sensitive determination of the presence or absence of a particular pathogen (for example, in deploying a binary test that determines if a patient is positive for coronavirus). (3) Development of microfluidic and bioprinting platforms for efficient clinical sample processing across a variety of complex fluids and infectious disease agents.

Label-free SERS can directly obtain the “spectroscopic fingerprint” of biological agents. It is, therefore, an attractive approach when patient samples need to be screened across a broad range of potential pathogens. Among the clinical applications, label-free SERS is poised to impact the diagnosis of bacterial infections—such as common blood-stream or urinary tract infections—where one or more of over 30 bacterial species and strains could be present, each with differing antibiotic susceptibilities and minimum inhibitory concentrations.^45,67 Label-free SERS could also revolutionize viral infection diagnosis—helping to distinguish between, for example, multiple influenza viruses, without needing to extract viral nucleic acids from patient samples.^68–72 However, challenges with distinguishing and interpreting Raman spectra can arise due to the complex nature of the infectious disease agent’s spectra. At the molecular level, different pathogens are composed of similar building blocks, namely, nucleic acids, amino acids, carbohydrates, and lipids. Therefore, Raman spectral differences between biological classes—such as those between seasonal coronavirus and SARS-CoV-2, H5N1 influenza and H3N2 influenza, and methicillin resistant and susceptible Staphylococcus aureus—may be subtle and difficult to discern.

In label-free Raman and SERS, advanced data analysis and prediction algorithms can assist in the interpretation and classification of spectra. Machine learning methods including principal component analysis (PCA)⁷³ and clustering⁴⁵ (both unsupervised learning methods) can discover unknown features in Raman/SERS spectral data and visualize spectral variations without a priori knowledge of classification groups. Meanwhile, supervised learning methods such as logistic regression, discriminant analysis,⁷⁴ decision trees,⁷⁵ support vector machines (SVMs),⁷⁶ and neural networks (NNs) can facilitate species and strain discrimination, even with hard-to-distinguish features. Note that, on high signal-to-noise-ratio (SNR) data, principal components are often correlated with signal differences in the Raman dataset. In datasets with high noise, however, the first principal components may be more indicative of an intrinsic dataset variation rather than differences in Raman or SERS signal between biological classes. Dimensionality reduction using PCA can reduce the number of parameters needed for supervised statistical learning methods. Ideally, machine learning training datasets are collected so that the data cover as much diversity in experimental conditions as possible. PCA paired with Raman and SERS has been implemented to interpret spectral data for distinguishing between species of infectious bacteria^65,77–79 and diagnosing viral infections such as HIV⁴² and hepatitis B virus (HBV).⁶⁸ 

To take advantage of the information-rich Raman signatures in diagnostics applications, advanced data analysis techniques such as neural networks (NNs) and support vector machines (SVMs) are becoming critical to Raman analysis.⁸⁰ However, a fully connected neural network (NN) structure can become unwieldy on high resolution (and, therefore, high-dimensional) Raman data, with large numbers of parameters that require large datasets to train. Convolutional neural networks (CNNs) can reduce the number of parameters necessary for high dimensional data by considering any repetitive features across the input; consequently, these deeper networks allow more complex, higher-order Raman features to be learned.⁸¹ NNs facilitate direct analysis of Raman spectra without preprocessing (i.e., artifact removal and background correction), eliminating possible misinterpretation due to signal preprocessing.⁸² CNNs can also learn low-level Raman and SERS features, such as peak shapes, trough shapes, and high-level nonlinear combinations, using these to map noisy data to a decision space where biological classes are more easily separated. Compared to statistical learning techniques,⁸³ CNNs perform better on low signal-to-noise data, therefore reducing the measurement time required for accurate identification.

As shown in Fig. 3, CNNs have been effectively implemented for predictions of bacterial type and antibiotic treatment choice across 31 bacterial species and strains.⁸³ Even with low-signal-to-noise data, over 99.6% accuracy was achieved with CNN-based models.⁸³ Furthermore, multivariate statistics and machine learning methods have also been shown to distinguish between genetically similar bacterial strains where disease diagnosis is clinically challenging due to serological cross-reactivity and clinical presentation similarities.⁸⁴ SVM was also implemented to identify the top 11 bacterial species known to cause urinary tract infections with 92% accuracy⁸⁵ and to distinguish between pathogenic and non-pathogenic Mycobacterium strains with 94% accuracy.⁸⁶ The synergy between Raman spectroscopy and machine learning also provides a unique opportunity to detect and identify viral infections. Recently, Raman-based machine learning was used to identify three different viruses—rhinovirus, influenza A virus, and human parainfluenza virus type 3 (HPIV 3)—in patient swabs with 93% accuracy.⁵ It was also used for hepatitis C virus (HCV) diagnosis with 95% accuracy⁸⁷ and HBV diagnosis with 98% accuracy.⁸⁸ We note that viral infection detection of RNA viruses (such as coronaviruses, influenza A, and HIV) poses a unique challenge, due to their potential for rapid mutation; such rapid evolution is responsible not only for the low effectiveness of vaccines existing for these viruses but also for the lack of vaccines for others. With emerging platforms for real-time Raman spectroscopy combined with machine learning, there is potential to quickly contain pandemics and also rapidly monitor viral mutations for rapid drug and vaccine development.^50–52,89

Beyond spectral data interpretation, machine learning algorithms can also facilitate a next-generation SERS substrate design. With label-free SERS diagnostics, it is crucial to have high electromagnetic field enhancements that are relatively uniform across the biomarker of interest. For viruses and bacteria, this means having signal enhancements that span several tens of nanometers, potentially up to micrometers. Metasurfaces enable tuning of local electromagnetic properties by their nanoscale architecture, when made of gold or silver films^90,91 or, more recently, using dielectric substrates.¹⁸ Furthermore, they promise uniform signal enhancement across a large area and, thus, more reproducible enhancement.^19,92–94 We note that while dielectric metasurfaces do not provide as high of Raman signal enhancement as metals (∼10³ compared to >10¹¹ for metallic/plasmonic substrates),^95,96 they do not induce localized sample heating and have significantly reduced absorption, particularly in the infrared region. Additionally, advances in dielectric “high-Quality-factor” metasurfaces can support extremely large electric field enhancements, with projected Raman enhancements of 10⁸–10⁹.^93,97–103 These metasurfaces are compatible with the conventional CMOS-based semiconductor fabrication processes, making their production scalable and cost effective.

Designing metasurfaces with the desired Raman optical properties can be a computationally intensive process, driven by empirical or trial and error approaches. To mitigate these design challenges, machine learning algorithms can be used for an optimization-driven design that enables the fabrication of rationally designed metasurfaces, tailored for specific properties of interest.^104–109 This approach also enables the selection of computationally efficient schemes for a metamaterial design.¹⁰⁴ For example, Jiang et al. used this approach to design silicon metagratings that deflect electromagnetic waves to a desired diffraction order.¹¹⁰ They used a generative adversarial network (GAN), a class of deep neural networks comprised of two competing neural networks, which can learn to mimic information in a dataset, as shown in Fig. 4. The algorithm was trained on images of 600 high performing, topologically complex metasurface designs and was able to produce thousands of other high performing metasurface architecture suggestions. The suggested architectures were unique and not the obvious design choice for diffractive metasurfaces. Therefore, machine learning based design approaches have potential for the discovery of new design principles for optimized metasurfaces.

For many infectious diseases—such as COVID-19, influenza A, or HIV—assays need not screen for a broad spectrum of pathogens, but rather detect the presence or absence of a particular pathogen. In this case, labeled-SERS approaches can enable higher sensitivity and specificity and allow for quantitative analysis. Across infectious diseases, the pathogen load of a patient sample varies depending on the infection stage, sample source (i.e., nasopharyngeal swab vs bronchotracheal lavage), and disease type; however, in each case, diagnostic approaches must detect low pathogen concentrations. For example, common respiratory infections such as influenza A, influenza B, and human rhinovirus (HRV) are associated with typical viral loads of 10⁷ copies/ml, 10⁵ copies/ml, and 10⁶ copies/ml, respectively.¹¹¹ For bacterial sepsis, mortality rates are high with just 1 cfu/ml (cfu = colony forming unit).⁶⁵ For other novel pathogens such as SARS-CoV-2, the relationship between viral load and disease severity is not fully known, emphasizing the need for a diagnostic approach to detect significantly lower concentrations.

Labeled-SERS enables rapid detection directly from complex samples such as whole blood, without the need for washing or amplification steps. This significantly improves sensitivity and specificity and allows for the multiplexed detection of several biomarkers. Labeled-SERS approaches have been implemented in several forms using affinity agents such as antibodies, aptamers, and polymers, highlighted in Table II, particularly for influenza diagnosis. While direct detection of SERS spectra from target analytes on affinity agents is possible, SERS signals can be further enhanced through the use of Raman reporters, such as resonant fluorescent dyes or aromatic thiols. These reporters are incorporated either onto the SERS substrate or into the affinity agent and facilitate pathogen detection and quantification, especially on samples with low pathogen concentrations. However, we note that conjugation with affinity agents can increase the potential for false positive results due to cross reactivity.

Antibodies are proteins that are naturally produced by immune cells (B-cell) to target and bind specific antigens as part of the immune response. They can be synthetically derived from either B-cell lines as monoclonal antibodies or from DNA recombination as recombinant antibody fragments using phage display technology; they have strong binding affinities with dissociation constants of 10⁻⁸ M–10⁻¹⁰ M.^120,121 SERS detection of HIV-1 virus-like particles has been achieved with Au nanoparticles functionalized with monoclonal antibody fragments against gp120, a common surface protein on HIV virus; here, viral particle concentrations as low as 35 fg/ml were detected within a few seconds.¹²² SERS-based whole virus detection has also been achieved with mosquito-borne flaviviruses using anti-flavivirus 4G2 antibody on Au nanoparticles.¹²³ In bacterial diagnosis, antibody-functionalized SERS-encoded nanoparticles have also been used to diagnose S. aureus, E. coli, P. aeruginosa, and S. agalactiae from blood and serum in millifluidic channels without the need for washing steps and with clinically relevant bacterial concentrations ranging from units to tens of cfu/ml.¹²⁴ In addition, an antibody against protein A (a bacterial surface protein) was used to sensitively detect S. aureus and S. pyogenes bacteria in prosthetic infections from joint fluids at a clinically relevant concentration of 3 × 10⁸ cfu/ml.¹²⁵ Antibodies provide specificity and tight binding affinities, but are often costly and complicated to produce; have unpredictable and uncontrollable orientations (when bound to plasmonic surfaces); are somewhat bulky in size (limiting packing density on SERS substrates and, hence, the achievable SERS enhancements); have a signal that interferes with the SERS signal; have a low thermal stability; and allow for false positive results from antigen cross-reactivity. New methods including reduction of the antibody’s disulfide bond and incorporation of single stranded DNA or synthetic peptides are tackling these challenges.^126–129 Recombinant antibody fragments have also been used to address some of the limitations of monoclonal antibodies as they are simpler and cheaper to produce, smaller in size, and relatively easily immobilized on SERS substrates.^130,131

Beyond antibodies, aptamers promise versatile, specific, and tight target binding and high packing density on SERS substrates. Aptamers are segments of nucleotides or peptides with specific binding affinity for a target molecule. They are 10× smaller than antibodies (5 kDa–10 kDa), enabling dense packing on SERS substrates; they also possess a range of binding affinities from K_D ∼ 10⁻³ M–10⁻¹² M and can be designed to bind more strongly than antibodies.^132,133 In addition, aptamers have a controllable orientation and higher selectivity as they three-dimensionally fold around targets.¹³⁴ Generally, the SERS signature from aptamer nucleotides also has less overlap with the target cell’s spectral range unlike antibodies, improving the SERS signal-to-noise ratio. Recently, influenza viruses were detected at concentrations of 10⁻⁴ hemagglutination units/probe using aptamer-based SERS targeting hemagglutinin on the virus.¹¹⁴ A similar approach also enables the detection of bacteria S. typhimurium¹³⁵ and S. aureus with a LOD of 1.5 cfu/ml, even in mixed samples with L. monocytogenes, S. flexneri, and E. coli.¹³⁶ Furthermore, B. anthracis spores were detected at concentrations as low as 10⁴ cfu/ml (a lethal dose is 10⁶ cfu/ml),¹³⁷ and the bioterror agent protein ricin was detected with a LOD of 10 ng/ml within 40 min.^138,139

Polymers are emerging affinity agents with extensive chemical functionalization properties. In contrast to antibodies, polymers have a lower binding affinity (KD of 10⁻⁶–10⁻⁵ M).^140,141 Molecularly imprinted polymers are cross-linked polymer systems resembling antibodies, but with lower cost and versatility.^140,141 Molecularly imprinted polymers have been implemented for small molecule detection including food borne toxin aflatoxin B1¹⁴² and pharmaceuticals such as the detection of antibiotics in food.¹⁴³ Besides acting as affinity agents, polymers, such as polyethylene glycol (PEG), the most versatile biocompatible polymer, are used to stabilize SERS metal particles for controlled synthesis with well defined end group incorporation such as thiols.¹⁴⁴ Another polymer, poly (N-isopropylacrylamide) (PNIAM), generated a porous structure, facilitating incorporation of Raman reporters via diffusion and trapping without the need for specific affinity to the metal.^145,146 Thus, with their low-cost of production, versatility in chemical moieties for targets, the number of binding sites, and the close distance from the SERS substrate, polymers may prove to be a powerful tool for pathogen diagnosis via labeled-SERS.

A critical aspect of translating SERS to the clinic is sample preparation. Existing clinical assays, such as PCR, ELISA, and mass spectroscopy, generally require sample incubation and/or amplification due to low quantities of compounds of interest within a sample,¹⁴⁷ and purification to avoid background signal from other sample components. For SERS to replace these assays, it needs to offer either significantly increased sensitivity and specificity or easier sample preparation. Sections II and III described SERS substrates, affinity agents, and spectral analysis tools that are significantly increasing the specificity and sensitivity of the technique. Here, we describe advances in microfluidics and bioprinting that facilitate SERS sample preparation. Importantly, these methods provide advantages for both labeled and unlabeled SERS, bringing the technique much closer to clinical and point-of-care diagnostics.

Microfluidics is the manipulation of fluid flows operating in the low Reynolds number regime. Microflows in sub-millimeter channels experience diffusion-based mixing, resulting in consistent and predictable fluid behavior. Though solely relying on diffusion generally requires long mixing times, a number of active and passive mixing and separation mechanisms address the challenges of sample preparation, speed, and signal uniformity with clinical samples.^148,149 Many elements can be combined on a single device, creating lab-on-a-chip (LOC) designed to integrate sample preparation, bioreactions, and detection into one place.¹⁵⁰ For example, microfluidic chip components can be designed for fluid transport, fluid metering, fluid mixing, valving, and concentration or separation of molecules within picoliter and nanoliter volumes of fluid;¹⁴⁹ patterned impediments such as pillars also enable size-selective filtration and on-chip cell sorting,¹⁵¹ as highlighted by a recent report separating plasma from a whole blood sample.¹⁵² Most importantly, microfluidic chips allow for spatially defined detection areas, ideal for integration with multiplexed diagnostic techniques such as SERS.^153,154

Integrated microfluidic and optical spectroscopy platforms have recently gained traction for robust and reproducible SERS measurements.^113,155,156 Integration of Raman spectroscopy was made possible through the development of high temporal resolution spectroscopy systems matching the kilohertz range droplet generation frequencies.¹⁵⁷ These integrated systems improve SERS reproducibility, provide sensitive detection of small analyte concentrations in confined volumes within generated droplets,¹⁵⁸ and simplify sample preparation steps.¹⁵³ For example, a sensitive and wash-free microfluidic based immunoassay has been demonstrated for prostate-specific antigen detection with a LOD below 0.1 ng/ml, a limit below the clinical threshold for a cancer diagnosis.¹⁵³ This integrated setup removes noise from the sample by generating droplets containing free and bound anti-PSA antibody labeled plasmonic particles and magnetically separating them from the solution, which is the source of the background signal. Another microfluidics, SERS-based immunoassay was used for the rapid, automated, and quantitative detection of the H5N1 avian influenza virus. Here, by using Raman reporters and antibody labels, the assay achieved a limit of detection of 74 pg/ml, well below the 399 pg/ml achieved by the existent ELISA systems.¹¹³ 

In a different approach, antibody-conjugated silver nanoparticles and magnetic beads were used for on-chip, single-cell level cytokine production monitoring, where magnetic beads allow for controlled nanoparticle aggregation and antibody conjugation allows for specific binding to target cytokines and formation of immune-sandwiches, which then activate the signal of the Raman reporters. The cytokine complexes are then encapsulated in microdroplets and interrogated via SERS within the collection channel array located on the same chip.¹⁵⁹ Furthermore, the droplets are gas permeable, maintaining cell viability, which leads to a localized, intra-droplet rise in cytokine concentrations, allowing for a LOD of 1 fg/ml without the need for cell pre-culturing steps.¹⁵⁹ In addition, on-chip ultrasonication of microdroplets was used to homogenize colloidal particles mixing with the bacterial cell for label-free, culture-free, and reproducible single-cell level SERS.⁴⁶ Finally, we note that these integrated setups enable on-chip and real-time studies of dynamic processes at the single cell level,¹⁶⁰ such as monitoring cytokine secretion,¹⁵⁹ detecting glycans on cancer cells,¹⁶⁰ and detecting biomarkers of bacterial spores¹⁵⁸ from a range of samples including urine and blood.

Microfluidic chips have also been integrated with SERS for virus detection and identification. VIRRION (virus capture with rapid Raman spectroscopy detection and identification) is a system designed for rapid, nondestructive capture, enrichment, and label-free optical detection of viral particles from clinical samples.⁵ The chip works by capturing viral particles within a forest of constructed, nitrogen-doped carbon nanotube (CNxCNT) arrays doped with gold nanoparticles. The CNxCNT are patterned in a microfluidic channel in multi-sectioned herringbone array to selectively capture viral particles. A patient’s sample is fed into the channel allowing the viruses to be captured in the CNxCNTs, while the filtrate is removed, negating any need for prior sample filtration. Raman spectroscopy is then performed on the captured particles, in situ, leading to particle identification, as shown in Fig. 5.^5,161

Matching the kilohertz range of droplet generation frequencies with spectra collection is a significant challenge in integrated Raman and microfluidic setups. To mitigate this challenge, either slowing the droplets via trapping¹⁵⁸ or compensating for lower temporal resolution at the spectra detector level via algorithms¹⁵⁷ can be performed. At the droplet trapping level, a method for spectral interrogation of thousands of stationary droplets at once was designed as a two-step process; a large area scan (760 × 760 µm² window) was performed followed by higher resolution remapping of smaller regions of interest, as shown in Fig. 6.¹⁶⁰ In a different approach, a 3D plasmonic trap made of silver was designed into the microfluidic chips during fabrication for encapsulating cells in these silver shells for SERS.¹⁵⁸ At the detector level, sub-millisecond time resolution was achieved via binning and cropping of the CCD sensor enabling real-time droplet spectral interrogation.¹⁵⁷ Advances in integrated optofluidic platforms for SERS promise simplified sample preparation and high-throughput and reproducible SERS from biological samples.

While microfluidics based droplet generation has been the predominant method employed for biological samples, several other methods such as inkjet printing and acoustic printing (collectively termed bioprinting) have been used to split and pattern biological samples. Though more frequently used for cell culturing, tissue engineering, biochemical assays, drug development, and creation of cell-based biosensors,^162,163 bioprinting methods provide several advantages over existing microfluidic droplet generation methods. For example, unlike on-chip microfluidics, these bioprinting platforms can pattern droplets on SERS substrates, print with a range of bioink viscosities, and eject droplets at accelerated print speeds with fine control over droplet resolution, diameter, and print directions without the need for cytotoxic surfactants that lower cell viability counts.^164,165

Though work on integrating bioprinting platforms with Raman is nascent, inkjet printing has recently been used in conjunction with SERS, paving the way for future single-cellular SERS interrogation.^166,167 In a method termed Inkjet Dispense SERS (ID-SERS), commercially available HP D300e or D100 thermal inkjet digital dispensers and the associated HP BioPattern software were used to pattern microdroplets of 1,2-bis(4-pyridyl)ethylene (BPE), with volumes of 28 pl onto a label-free SERS substrate with varying analyte concentrations.^166,167 The system was used to pattern a high density, 330 dots per inch (DPI), of droplets using an automated method that can be generalized to other substrates, providing the advantages of speed with subsecond drop-and-dry times and low sample volumes in the picoliter range.^166,167 This work highlights the significant potential of bioprinting droplet generation techniques for efficient sample preparation in diagnostic assays and paves the way for clinical translation of SERS.

Raman spectroscopy ranks among the most promising biomedical diagnostic tools. The technique provides a specific and sensitive fingerprint of biomarkers, with the potential for rapid, multiplexed analysis in a portable, low-cost platform. In the past, Raman spectroscopy has suffered from relatively low signal-to-noise ratios, complexity in spectral interpretation, and a lack of an efficient workflow from sample collection to spectral acquisition, challenging clinical translation. Today, advances in nanophotonic SERS substrates, machine learning, microfluidics, and bioprinting are enabling breakthroughs that can overcome each of these challenges.

In label-free SERS, machine learning promises to revolutionize Raman-based diagnostics. It has already enabled advanced spectral analysis of Raman signatures, leading to increased identification accuracy of large numbers of pathogenic targets. Importantly, these signal processing techniques promise to mitigate the cost and complexities typically associated with spectroscopy hardware by allowing for accurate results even with low-resolution or low signal-to-noise data. Beyond identification, machine learning techniques could also be applied to establish correlations between spectral information and genetic differences. For example, they could be used to monitor viral and bacterial genetic evolution and mutation or to study the real-time interaction of pathogens with drugs (antivirals or antibiotics). A comparison of spectral differences in the protein, nucleic acid, lipid, carbohydrate, and cholesterol Raman bands could also lead to a greater understanding of cell–drug interactions, guiding clinical treatment, as well as the development of new antivirals and antibiotics. Finally, machine learning could potentially be used to predict drug susceptibility and minimum inhibitory concentration, even without drug additives. In parallel, advanced machine-learning based optical design methods are enabling the introduction of optimized, unconventional optical substrates that can boost Raman sensitivity and clinical applicability. For example, Raman substrates with strong and uniform Raman enhancements could be designed and manufactured from low-cost and mass-produced materials, such as silicon and silicon dioxide.

For diagnostic assays aiming to identify the presence or absence of a specific pathogen, affinity agents and Raman reporters are very promising. These systems have enabled rapid detection of pathogens directly from complex samples such as whole blood, without the need for washing or amplification steps. Antibodies offer high binding affinity, but are bulky (limiting SERS enhancements), challenging to correctly orient on SERS substrates, and often interfere with the target pathogen SERS spectra (antibodies are made of proteins whose Raman peaks overlap with the pathogen spectral region). Aptamers and polymers offer more versatility, smaller sizes for larger packing density, minimal interference on the SERS signature, and generally a cheaper cost of production. Such labeled-SERS approaches promise targeted diagnosis of specific infectious agents at the point-of-care with lower cost and faster turnaround time.

The integration of microfluidics and bioprinting platforms with SERS has shown great potential for enabling clinical translation of Raman spectroscopy. Microfluidic platforms facilitate complex and efficient mixing, separation, parallelization, and multiplexing. Therefore, they can reduce the number of required sample preparation steps, decrease the required sample and reagent volumes, improve sample homogeneity, and increase throughput. However, they are limited to working in the microliter sample volume ranges, and generated droplets follow Poisson distributions, generating upwards of 90% empty droplets for guaranteed single-cell analysis.¹⁶⁵ Emerging bioprinting technologies, if used in conjunction with SERS, offer the potential to increase temporal resolution while limiting the use of surfactants and oils often necessary for microfluidic based droplet generation, mitigating challenges of cytotoxicity and lensing effects in Raman interrogation caused by these additives. These technologies also allow for patterning on SERS substrates, facilitating greater flexibility in SERS-based analyses.

Bioprinting technologies promise a wealth of untapped potential that may be ideal for infectious disease detection. Acoustic ejection, for example, is a nozzle-free printing method and, as such, avoids challenges of sample contamination, nozzle clogging, and temperature increase found in nozzle-based ejection methods, such as inkjet printing. Furthermore, acoustic transducer designs can manipulate acoustic waves, allowing for continuous particle and fluid mixing including particle dispersion and size selective ejection.^168–170 Pathogens such as bacteria and viral particles could even be efficiently delivered to the ejection site with acoustic printing, potentially achieving non-Poisson distributions of ejected particles, thereby increasing the number of single-cell or single-virion droplets.

We envision that these emerging nanophotonic SERS substrate, machine learning, microfluidic, and bioprinting platforms for SERS could readily be applied to clinical diagnostics. In particular, these advances enable streamlined sample preparation platforms with more robust and quantitative SERS with sensitivity and specificity matching clinical standards. We are particularly excited about prospects for identifying pathogens such as viruses and bacteria without requirements for sample cleaning or pathogen amplification, both with label-free and labeled-SERS.

We note that for clinical SERS adoption, Butler and colleagues have suggested a useful flowchart for optimizing the optical components, incident laser wavelength, sample preparation techniques, SERS substrate choice, and Raman data processing approaches for specific biological samples of interest.¹⁰ In parallel, organizations such as Raman4Clinics¹⁷¹ are actively working toward identifying the most appropriate entry to the biomedical device market in order to realize the outstanding qualities Raman spectroscopy offers as a sensitive and specific diagnostic tool. Synergy between these efforts and the research advances outlined in this paper should accelerate clinical translation of SERS. Particularly in the COVID-19 era, there is a pressing need for faster disease diagnosis and drug susceptibility testing. SERS could accelerate personalized patient care, from rapid diagnosis to prompt and accurate treatment.

Data sharing is not applicable to this article as no new data were created or analyzed in this study.

The authors gratefully acknowledge the support from the Stanford Catalyst for Collaborative Solutions under Grant No. 132114, the Gates Foundation under OPP 1113682, the Moore Inventors fellowship under Grant No. 6881, and the National Science Foundation under Grant No. 1905209. L.F.T. acknowledges support from the NIH Biotechnology Training Program under Grant No. T32GM008412, Agilent, Stanford EDGE, and Stanford DARE graduate fellowships.