Unravelling Electronic Structure and Molecular Vibrations of Proteins in Virus Using Novel Correlated Plasmon‐Enhanced Raman Spectroscopy With Machine Learning

Abstract

Proteins are believed to contain vital functional information of biosystems. In virus, structural proteins form the building blocks of the virus particle. Raman spectroscopy is a powerful technique to probe electronic structure at molecular vibration levels. However, due to weak Raman cross‐section, there is no reliable conventional Raman spectroscopy on E‐proteins in the virus. Herewith, a novel, non‐destructive and direct technique, named correlated plasmon‐enhanced Raman spectroscopy (CP‐ERS) is developed to directly probe electronic and molecular vibrations of proteins. Intriguingly, using CP‐ERS, new resonant quasielastic and inelastic electronic Raman scatterings and phonon excitations of E‐proteins in dengue virus (DENV) are discovered. The CP‐ERS is utilizing newly‐developed highly oriented gold‐quantum dots chips exhibiting low‐loss tunable correlated‐plasmons, high structural stability and reproducibility. By modifying E‐protein, anomalous glycosylation‐induced changes in CP‐ERS are observed. Moreover, CP‐ERS are used for training a machine learning algorithm, obtaining 100% accuracy in hold‐out and 93% mean accuracy in grouped 5‐fold cross‐validation. Our result reveals new resonant quasielastic and inelastic Raman scatterings and new phonons of E‐proteins in viruses and demonstrates a strategy in utilizing CP‐ERS with machine learning to directly measure and quantify electronic structure and structural and molecular vibrations of biological and solid state systems.

A novel, non‐destructive, stable, direct technique, named correlated plasmon‐enhanced Raman spectroscopy (CP‐ERS), which is based on a newly‐developed highly oriented single crystalline gold quantum dots (HOSG‐QDs) chip, is developed to measure and quantify molecular vibrations and electronic structure of analytes. The CP‐ERS on dengue virus strains reveals new quasielastic and inelastic electronic and vibrational Raman scatterings. Upon on‐resonance laser excitation, correlated‐plasmons in HOSG‐QDs chip excite and interact with the molecular and electronic structure of the virus, producing distinct spectral fingerprints for different dengue viruses. Machine learning model analysis allows accurate classification based on the CP‐ERS spectra.

1. Introduction

Dengue virus (DENV), a flavivirus from the Flaviviridae family, is a pressing global health concern that affects millions of people annually, with no approved antivirals and sub‐optimal vaccines available.^{[ 1} , ² , ³ , ^{4 ]} The DENV is a single‐stranded RNA virus, and mature DENV particles are spherical with a 50 nm diameter.^{[ 5 ]} Viruses, in general, depend on proteins, either structural or non‐structural for various functions.^{[ 6 ]} The structural proteins constitute the virus's structural components, while the non‐structural proteins primarily facilitate the intracellular life of the virus. In DENV, three structural proteins, envelope (E), membrane (M), and capsid (C), and 7 non‐structural proteins, named NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5, are generated after the protease enzyme cuts or processes a precursor polyprotein.^{[ 7} , ^{8 ]} The Envelope (E) protein consists of three envelope domains (ED): ED I at its center, ED II facilitating fusion, and ED III, the outermost domain. The E‐protein is a significant target for vaccine development due to its role in receptor binding and the presence of neutralizing antibody targets.^{[ 9 ]} It contains two key N‐glycosylation sites, N67 and N153. The N153‐linked glycans are believed to play a crucial role in maintaining the stability of the fusion peptide, affecting the pH threshold for membrane fusion, viral uncoating, and RNA release, while also influencing antibody binding and virus neutralization.^{[ 10 ]}

The recent implementation of molecular diagnostic techniques in viral detection has made significant progress.^{[ 11 ]} Commonly used techniques are such as reverse transcription polymerase chain reaction (RT‐PCR) and enzyme‐linked immunosorbent assay (ELISA). However, these techniques do not directly detect intact neither viral particles nor proteins; rather, they rely on modifications and external labelling strategies. The RT‐PCR, while highly sensitive, is exclusively designed for nucleic acid amplification and is limited to genome detection, making it unsuitable for direct protein or whole‐virus identification.^{[ 12 ]} Similarly, ELISA depends on the availability and specificity of antibodies in human serum, which introduces potential challenges such as cross‐reactivity and serotype‐dependent variations.^{[ 13 ]} Hence, their restricted ability to detect direct proteins or whole‐virus and our limited spectroscopic understanding of viral proteins necessitate molecular spectroscopy approaches for viral diagnostics and their functional understanding.

A technique that, in principle, can detect molecular vibrations such as phonons, is a surface‐enhanced Raman spectroscopy (SERS). The SERS has long been explored across various fields, including biological diagnosis and treatment^{[ 14} , ¹⁵ , ^{16 ]} food safety analysis,^{[ 17} , ^{18 ]} and environmental monitoring^{[ 19} , ^{20 ]} due to its molecular fingerprinting capability.^{[ 21} , ²² , ²³ , ²⁴ , ^{25 ]} The SERS allows vibrational and structural fingerprinting through the enhanced electric field in the closed vicinity of nanostructured plasmonic metal surfaces due to plasmon‐photon coupling.^{[ 26 ]} Although there has been various research in the field of SERS‐based virus detection, most of it is typically achieved by depositing metallic nanoparticles onto dielectric substrates.^{[ 27} , ²⁸ , ²⁹ , ³⁰ , ³¹ , ^{32 ]} However, these metallic nanoparticles‐based SERS chips have fundamental limitations and challenges. First, in conventional metals, including gold nanoparticles, strong absorption, such as from Drude response and interband transitions, and scattering, such as from electrons, charge carriers or impurities, occur simultaneously together with conventional plasmons (Figure S1a,b, Supporting Information). This yields in significant optical losses or high‐plasmonics‐loss,^{[ 32 ]} which reduce the overall efficiency of the plasmonic system and screen the Raman signal. Second, the synthesis of metallic nanoparticles requires multiple, complex step processes involving complex polymer coatings to act as a stabilizer, making nanoparticles oxidize easily (or no longer inert). This results in the SERS signal from nanoparticles themselves changing even within a minute (Figure S2a, Supporting Information). Third, the low reproducibility of these chips, stemming from the complexity involved in assembling metal nanoparticles, which hinders precise control over conventional plasmons and the optical properties of plasmonic systems^{[ 33 ]} (Figure S2b, Supporting Information). Due to these fundamental limitations, conventional SERS is not capable for quantitative analysis and comparison of data, even within the same chip, limiting their reliability and consistency in generating reliable data and practical applications. Furthermore, using the metallic nanoparticles, conventional SERS is not capable for quantitative analysis and comparison of data, even within the same chip, limiting their reliability and consistency in generating reliable data and practical applications. Therefore, while conventional SERS has been widely used, it has yet to be established as an effective tool for the detection and analysis of various analytes, including biological samples.

In this Article, we develop novel correlated‐plasmon enhanced Raman spectroscopy (CP‐ERS) based on newly‐developed highly oriented single‐crystalline gold quantum dots (HOSG‐QDs) chip. Using CP‐ERS supported with machine learning and theoretical calculations, we discover new resonant quasielastic and inelastic electronic Raman scatterings and phonons in DENV strains. The HOSG‐QDs chip exhibits tunable unconventional correlated plasmons due to quantum confinement and strong electronic correlations in the s‐band of low‐dimensional gold^{[ 34 ]} and is stable in the air for over a year, with high reproducibility on data collection and chip fabrications, enabling quantitative analysis. To acquire the Raman spectrum of the DENV, the virus is placed on the HOSG‐QDs chip using the drop‐casting method. A monochromatic laser source at an excitation wavelength of 633 nm (on‐resonance with correlated plasmon) or 785 nm (off‐resonance with correlated plasmon) is directed on the DENV in a backscattering configuration. In this configuration, the laser beam is incident on the sample from the up direction, and the scattered light is collected from the same direction. The on‐resonance laser excites the correlated plasmons in HOSG‐QDs chip, which interact with the molecular level of analytes, causing inelastic scattering of light. The scattered light contains Raman‐shifted photons with different energies from the incident light. They are collected and passed through a filter to remove the elastically scattered Rayleigh light. The filtered Raman‐scattered light is then dispersed by a spectrometer and detected using a charge‐coupled device detector. Measurements with two different photon excitation wavelengths, one on‐resonance and the other off‐resonance with correlated plasmons in HOSG‐QDs chip, are performed to examine the effect of incident energy on the CP‐ERS. In comparison, we also perform Raman spectroscopy using MgO as a chip at both photon excitation wavelengths. The HOSG‐QDs are grown epitaxially using ultra‐high‐vacuum molecular‐beam‐epitaxy pulsed‐laser‐deposition, followed by post‐annealing at various temperatures (350 °C, 450 °C, 550 °C, 650 °C, and 750 °C), which alters the energy of correlated plasmons. The HOSG‐QDs are further characterized using variable‐angle spectroscopic ellipsometry to get information on optical properties such as plasmonic excitations and mott‐like insulating state. We design two type DENV strains, namely DENV2 D2Y98P (wild type, WT)^{[ 35 ]} and DENV2 N153Q. The DENV2 N153Q differs from D2Y98P WT by one amino acid substitution, which results in the lack of the N153‐linked glycosylation on its E protein.^{[ 36 ]} Interestingly, the CP‐ERS shows anomalous glycosylation‐induced changes. A sample of media control without any virus is used as a negative control. Moreover, the CP‐ERS are used for training a machine learning algorithm and classified DENV2 strains and negative control samples with 100% accuracy. Details of the chip, characterization techniques, virus preparations and machine learning algorithm are described in the methods section.

To analyse the different properties of HOSG‐QDs chip, spectroscopic ellipsometry, atomic force microscopy analysis, and X‐ray diffraction measurements are conducted. The schematic illustration of spectroscopic ellipsometry measurement set‐up is shown in Figure 1a. The crystal structure and surface morphology of the HOSG‐QDs chip are analysed using synchrotron X‐ray diffraction and atomic force microscopy, respectively, as presented in Figure 1b). The θ‐2θ plot shows a dominant Au (111) peak at 38.12° and a minor Au (222) peak at 81.85°, indicating a high degree of orientation of HOSG‐QDs in a single plane. The MgO (002) peak is observed at 43.04°. The SE parameters are taken in four different incidence angles and fitted using combined Lorentz, Gaussian, and PSEMI‐Tri oscillators to obtain universal fitting with minimum mean squared error (for details, see the method section and Figure S3a,b, Supporting Information). We then obtain simultaneously the loss function and complex dielectric functions. Figure 1c,d present the loss function and the complex dielectric functions, ɛ(ω)  ≡ ɛ₁ (ω) + ɛ₂ (ω), respectively. Longitudinal electron oscillations, such as those generated by plasmons, can be observed in loss function.^{[ 37 ]} The loss function is fitted using Gaussian oscillators. From this fitting, we observe distinct correlated plasmons. Unlike conventional plasmon, which is due to the collective oscillation of free electrons in a metal, correlated plasmons emerge due to collective oscillations of correlated electrons. As a result, correlated plasmons have the following properties. First, the loss function shows a peak at ≈1.96 eV. Second, ɛ₁(ω) remains positive at this photon energy (Figure 1d). Third, ɛ₂(ω) shows a Mott‐like insulator peak at ≈1.82 eV and redshifts than that of correlated plasmon. Notably, all these confirm the existence of correlated plasmons in the HOSG‐QDs chip.

Figure 1.

Optical properties, surface morphology, and crystal structure of highly oriented single crystalline gold quantum dots on MgO (001) (HOSG‐QDs chip). a) A schematic illustration of spectroscopic ellipsometry, along with an optical model, is built to fit the spectroscopic ellipsometry parameters. b) Atomic force microscope image and θ‐2θ plot of HOSG‐QDs chip, post‐annealed at 350 °C. c) Loss function and d) real and imaginary parts of the complex dielectric function of the HOSG‐QDs chip obtained from spectroscopic ellipsometry.

Figure 2 shows the schematic of CP‐ERS process (Figure 2a) and the CP‐ERS of the HOSG‐QDs chip and Raman spectra of MgO with different excitation wavelengths (Figure 2b,c). We perform resonance Raman spectroscopy at 633 nm or 1.96 eV (on‐resonance with the correlated plasmon) and 785 nm or 1.56 eV (off‐resonance with the correlated plasmon). We observe a strong contrast between on and off‐resonance Raman spectroscopy. At on‐resonance, CP‐ERS shows a new resonant inelastic electronic Raman scattering from the HOSG‐QDs chip. Distinct phonons are observed in the range of 100 to 3000 cm⁻¹, including phonons below 1450 cm⁻¹ (≈0.179 eV ≈2.880 × 10⁻²⁰ J) and one phonon occurring at 1491 cm⁻¹ (≈0.179 eV ≈2.880 × 10⁻²⁰ J). Further, one phonon occurs at 1910 cm⁻¹ (≈0.236 eV ≈3.793 × 10⁻²⁰ J) and some above 2000 cm⁻¹; (≈0.247 eV ≈3.972 × 10⁻²⁰ J). Distinct phonons observed for the HOSG‐QDs chip are listed in Table S1 (Supporting Information). This provides a comprehensive view of the Raman signal of our chips. In MgO, the resonant inelastic electronic Raman scattering is absent, and several phonons are observed, listed in Table S1 (Supporting Information). In contrast, at off‐resonance, such a resonant inelastic electronic Raman scattering enhancement is absent, and the Raman spectrum of the HOSG‐QDs chip shows only phonon peaks attributed to MgO. This fully supports that the correlated plasmon plays a crucial role in enhancing the Raman signal of HOSG‐QDs chip.

Figure 2.

a) Schematic illustration of the CP‐ERS process. The enlarged circle highlights the structure of the Envelope (E) protein. In the DENV2 N153Q mutant, asparagine (Asn, N) at position 153 of the E protein is replaced by glutamine (Gln, Q). The red spheres inside the box represent the amino acids at position 153 of the E protein. The resonance Raman spectroscopy of highly oriented single crystalline gold quantum dots on MgO (001) (HOSG‐QDs chip) and MgO (001) chip. b) 633 nm (1.96 eV) and c) 785 nm (1.56 eV) on HOSG‐QDs chip and MgO.

Our main observation is new resonant quasielastic and inelastic electronic Raman scattering and phonon enhancements from DENV2 strains using CP‐ERS. Figure 3a shows the optical microscopy image of focused laser spots on the samples. The reproducibility and uniformity of droplet distribution are qualitatively verified under optical microscopy (Figure S4, Supporting Information). Figures 3b–d present the CP‐ERS of DENV2 N153Q, DENV2 WT, and negative control, respectively. We use in our assays two different DENV2 strains, DENV2 N153Q and DENV2 WT (c.f. Figure 2a). A sample made of culture media without any virus is included as the negative control. The media control contains Leibvotiz‐15 (L‐15) and 2% Fetal Calf Serum. The correct readout of CP‐ERS can reflect the specific capture of DENV2 WT and its mutant counterpart. In the presence of a 2.3 × 10⁴ plaque‐forming unit (pfu) concentration of DENV2 N153Q and DENV2 WT in the media control, CP‐ERSs reveal a strong contrast between virus samples and negative control. The first notable observation in the CP‐ERS on the DENV2 N153Q reveals new anomalous resonant quasielastic Raman scattering ≈0.02 eV and strong resonant inelastic electronic Raman scattering below ≈0.168 eV (or ≈1391 cm⁻¹ ≈2.691 × 10⁻²⁰ J) as shown in Figure 3b. The CP‐ERS on DENV2 WT shows a new resonant inelastic electronic Raman scattering at ≈0.3 eV (Figure 3c). The strength of resonant inelastic electronic Raman scattering of DENV2 WT below 1205 cm⁻¹ (≈0.149 eV ≈2.392 × 10⁻²⁰ J) and the quasielastic Raman scattering are reduced than that of DENV2 N153Q. The DENV2 N153Q mutant lacks the glycan motif at N153 on the E protein and exhibits a distinct CP‐ERS compared to DENV2 WT. This reveals glycosylation‐induced differences between the two strains. The negative control sample, on the other hand, shows very weak resonant inelastic electronic Raman scattering below 800 cm⁻¹ (≈0.099 eV ≈1.589 × 10⁻²⁰ J). Furthermore, in the Raman shift of 1646 cm⁻¹ to 3039 cm⁻¹, the Raman intensity is higher for DENV2 WT compared to its mutant counterpart, DENV2 N153Q. In contrast, the negative control exhibits an overall lower Raman intensity across the range of 100 to 3039 cm⁻¹ compared to DENV2 strains (Figure 3d). The resonant inelastic electronic Raman scattering is fitted using Voigt and Fano profiles to understand the underlying origin of this electronic background (see Supplementary Note S1 and Figure S5, Supporting Information).

Figure 3.

a) Biological samples are prepared on chips by the drop casting method, and optical microscopy is used to focus the laser on DENV2 strains and negative control. Correlated plasmons enhanced Raman spectroscopy (CP‐ERS) of b) DENV2 N153Q, c) DENV2 WT, and d) negative control. e) Intrinsic CP‐ERS of DENV strains with subtracted negative control and intrinsic CP‐ERS of negative control with subtracted HOSG‐QDs chip.

It is worth highlighting that the energy range of the resonant quasielastic and inelastic Raman scattering in CP‐ERS of DENV strains is in the range of ≈10⁻¹⁹‐10⁻²⁰ J. The inactivation energies of various viral infections have been estimated using the Arrhenius equation, reporting values ranging from 10⁻¹⁹ to 10⁻²⁰ J.^{[ 38 ]} Among these, they reported the inactivation energy of Alkhurma hemorrhagic fever virus, a member of the Flaviviridae family, to be ≈10⁻¹⁹ J.^{[ 39 ]} The Alkhurma hemorrhagic fever virus belongs to the same viral family as the DENV and, hence, has similar chemical and physical properties.^{[ 40 ]} We argue that resonant quasielastic Raman scattering determines the inactivation energy of DENV (as further discussed below).

Raman phonons associated with amino acids such as asparagine, which is crucial for N‐linked glycosylation in DENV2 WT and glutamine, have been theoretically calculated.^{[ 41} , ^{42 ]} DENV2 N153Q and DENV2 WT exhibit new phonons (inset Figure 3b,c; Figure S6, Supporting Information), which are linked to amino acids of E proteins in DENV. Phonons, along with their assignments, are detailed in Table 1 . Phonon analysis is performed by fitting distinct peaks in CP‐ERS spectra using Lorentzian and Gaussian profiles. For the DENV2 N153Q and DENV2 WT phonons analysis refer to Figure S7a–c, Table S2, Figure S7d–g and Table S3 (Supporting Information).

Table 1.

Phonons assignment of DENV2 N153Q and DENV2 WT.

DENV2 N153Q	DENV2 WT	Theoretically calculated ([ 42 ] = L‐glutamine,[ 41 ] = L‐asparagine)	Assignment
203	199	209[ 42 ]	Bending skeleton deformation
298		291[ 41 ]	Torsion skeleton deformation
	300	296[ 42 ]	(H15 − N5 − C9 − C7) tors.
327	335	329[ 41 ]	C7 − C6 − N7 bend.
346, 371	368	343,[ 42 ] 360[ 42 ]	$C3-C2-N{H}_{3bend.}^{+}$
454	452	453[ 42 ]	C − C − NH 2bend.
479	481	476[ 42 ]	Bending skeleton deformation
529	525	521,[ 41 ] 527[ 42 ]	(H15 − N5 − C9 − C7) tors. 40%  − (C6 − C8 − O1) bend. 15%
596	596	596[ 41 ]	(C8 − C6) stre. 11%  +  (C6 − C8 − O1) bend. 49%
	627	623[ 41 ]	(C8 − C6 − N4) bend. 10%  − (C6 − C8  =  O1) bend. 15%  − (H15 − N5 − C9 − C7) tors. 38%
695	706	697[ 41 ]	(O3  =  C7 − N5 − C9) out  of  plan  vibr.
	748, 793	759,[ 41 ] 799[ 41 ]	(O2  =  C8  =  O1) bend.14% +  (O2  =  C6  =  O1 − C8) out  of  plane  vibr.44%
780		776[ 42 ]	NH 2twist.
818		805[ 42 ]	CH 2rock.
	853	839[ 42 ]	(O1  =  C8) stre. 11% +  (C8 − C6) stre. 29%  +  (O2 − C8 − O1) bend.35%
876, 931, 977		895,[ 42 ] 881,[ 42 ] 925,[ 42 ] 986,[ 42 ] 999[ 42 ]	CC stre.
1013	1006	1005[ 41 ]	(C7 − C6) stre. 15% −  (N4 − C6) stre. 27%  − (H13 − N4 − C6 − C7) tors.
1059		1051[ 42 ]	CN stre.
1085	1090	1086,[ 42 ] 1084[ 42 ]	CC stre. (N5 − C9) stre. 11% +  (H15 − N5 − C9) bend. 21%  −  (H13 − N4 − C6 − C7) tors. 28%
1168	1189	1164,[ 42 ] 1167[ 41 ]	$N{H}_{3rock.}^{+}{(C7-C6)}_{stre.}21\%+{(H10-C6-C7)}_{bend.}19\%+{(H13-N4-C6-C7)}_{tors.}19\%$
	1354	1354[ 41 ]	${(O2=C8=O1)}_{stre{.}_{symm.}}34\%-{(H10-C6-C8-O1)}_{tors.}17\%$
	1547	1532[ 41 ]	(H16 − N5 − H15) bend. 77%  −  (H15 − N5 − C9) bend. 10%
1607	1605	1603[ 42 ]	$N{H}_{3bend.}^{+}$
	1664	1668[ 41 ]	${(O2=C8=O1)}_{stre{.}_{asymm.}}73\%+{(H13-N4-H17)}_{bend.}11\%$
	1711	1694[ 41 ]	(C9  =  O3) stre. 76%
1753	1758		New
	1800		New
	2711, 2790	2752,[ 42 ] 2795[ 42 ]	$N{H}_{3stre.}^{+}$
	2862		New
	2924	2932[ 41 ]	(C7 − H11) stre.
	3010	2999[ 41 ]	(C6 − H10) stre.  +  (C7 − H11) stre.
	3119, 3183, 3329	3110,[ 41 ] 3343[ 41 ]	${(H13-N4-H17)}_{stre{.}_{symm}.}$
	3255	3255[ 42 ]	NH 2stre.
	3444	3453[ 41 ]	${(H15-N5-H16)}_{stre{.}_{symm}.}$

* stre = stretching, bend. = bending, rock. = rocking, tors. = torsion, symm. = symmetric, asymm. = asymmetric.

Figure 3e shows the difference in the CP‐ERS between DENV2 strains and the negative control, and the difference between the negative control and the HOSG‐QDs chip. This spectrum shows the intrinsic signals of DENV2 strains and negative control. We observe anomalous spectral weight transfer in DENV2 N153Q with respect to negative control in the 100 cm⁻¹ – 3000 cm⁻¹ Raman shift. The CP‐ERS of DENV2 N153Q is increasing dramatically below 1374 cm⁻¹ (≈0.17 eV ≈2.72 × 10⁻²⁰ J), while above 1374 cm⁻¹, the CP‐ERS is low. This shows the transfer of spectral weight from the high‐energy region to the low‐energy region in DENV2 N153Q. We attribute such a spectral weight as a fingerprint of the DENV2 N153Q. A spectral weight transfer also occurs in DENV2 WT at 1205 cm⁻¹ (≈0.15 eV or ≈2.40 × 10⁻²⁰ J), but it is less significant. Both DENV2 N153Q and DENV2 WT show resonant inelastic electronic Raman scattering up to 2424.19 cm⁻¹(≈0.30 eV or 4.8 × 10⁻²⁰ J). Resonant inelastic electronic Raman scattering in high‐energy regions from 1465.21 cm⁻¹ (0.18 eV, 2.89  × 10⁻²⁰ J) to 2424.19 cm⁻¹ is larger in DENV2 WT than DENV2 N153Q (shown in Figure 3e inset). After in the Raman shift 100 cm⁻¹ – 1465 cm⁻¹, the DENV2 N153Q experiences a substantial increase in resonant inelastic electronic Raman scattering while DENV2 WT remains consistent. Conversely, negative control is dominated by the reduction of CP‐ERS, which means negative control samples do not contribute to CP‐ERS.

In the DENV2 N153Q mutant, the asparagine (Asn, N) residue at position 153 of the E protein is replaced by glutamine (Gln, Q). The lack of glycan motifs at position 153 on the E protein in the mutant contributes to the deglycosylated induced anomalous changes in mutant’ CP‐ERS. The Asn, N has a shorter side chain with an amide group (‐CONH2) attached to a two‐carbon chain and Gln, Q has a longer side chain with an amide group (‐CONH2) attached to a three‐carbon chain. The higher density of the carbon chain in Gln, Q also contribute to the resonant quasielastic Raman scattering, resulting more enhancement in the lower Raman. Specific amino acids, particularly Asn residues at positions like Asn‐67 and Asn‐153, are crucial for N‐linked glycosylation, which impacts viral attachment, entry, and infectivity. The energy gap (difference of the HOMO and the LUMO) of L‐asparagine has been reported as 0.24 eV, which is used to represent intermolecular charge transfer⁴¹. Intriguingly, DENV2 WT shows the resonant inelastic electronic Raman scattering transitions in 0.2 to 0.35 eV energy (Figure 3e inset) and is not present in DENV2 N153Q. This can be directly related to the HOMU‐LUMO transitions (0.24 eV) of asparagine in DENV2 WT.

The vibrational modes detected by CP‐ERS, particularly those linked to amino acids in the E‐protein, reveal localised changes in molecular structure and electronic environment that are functionally significant in the viral life cycle. Specifically, glycosylation at N153 in DENV2 WT is known to modulate key viral processes such as host‐cell attachment, membrane fusion, and infectivity.^{[ 43} , ^{44 ]} The CP‐ERS on the DENV2 N153Q mutant, which lacks this glycan, reveals new quasielastic scattering, emergence of new phonon modes, and disappearance of phonon modes at the high Raman shift region (above 1800 cm⁻¹). These spectral changes are due to altered hydrogen bonding, polarity, and conformational dynamics around the glycosylation site, which are factors directly affecting viral‐receptor interactions. The CP‐ERS enables us directly to measure molecular‐level vibrational signatures that correlate with structural and functional changes in viral proteins. These spectral fingerprints are important for understanding viral mechanism and potentially distinguishing between infectious variants or glycosylation states.

The quasielastic and inelastic Raman scattering completely disappear at bare MgO, on virus on MgO, and off‐resonance (Figure 4 ). Most phonons are not visible, while a few of them are extremely weak, if any. This fully supports the importance of correlated‐plasmon with low‐loss to enhance electronic and molecular vibrations of proteins in virus, or otherwise they are not visible. Figure 4a illustrates the Raman spectroscopy of DENV2 WT and the negative control placed on a bare MgO (001). We also present the differences in Raman spectroscopy between DENV2 WT and MgO, as well as between the negative control and MgO. The loss function and ɛ₂ of MgO are detailed in Figure S8a,b (Supporting Information), revealing no plasmon and no absorption (insulator), respectively. When virus and negative control are placed on bare MgO, there is no observable resonant quasielastic and inelastic electronic Raman scattering enhancement. The negative control exhibits a slight decrease in the Raman spectroscopy, as indicated by the negative values observed in the difference between the spectra of the negative control and MgO. For DENV2 WT, the Raman spectroscopy remains consistent, with the difference in Raman signal between the DENV2 WT and MgO showing values close to zero. This reveals that the resonant quasielastic and inelastic electronic Raman scattering and phonon enhancement cannot be generated when the chip does not contain the necessarily correlated plasmons. Further, we perform off‐resonance CP‐ERS (excitation wavelength 785 nm or 1.56 eV) on DENV2 N153Q, DENV2 WT, and negative control samples prepared on HOSG‐QDs chip (Figure 4b). No distinct Raman signal is observed from the DENV2 N153Q, DENV2 WT, and negative control. Clearly that the resonant quasielastic and inelastic electronic Raman scatterings and phonons observed in the CP‐ERS spectra of DENV2 strains and negative control measured on the HOSG‐QDs chip are due to correlated plasmons. The correlated plasmons of the HOSG‐QDs chip couple with excitation wavelength, which further interact with the electric and molecular level of biomolecules and produce CP‐ERS. Therefore, it is important to have both photon energy that is resonant with correlated plasmon from the HOSG‐QDs chip.

Figure 4.

a) Raman signal of media control and DENV2 on a bare MgO (dielectric with no plasmons) is obtained using 633 nm excitation wavelength, including the Raman signal of DENV2 and negative control with MgO subtracted. b) Off‐resonance CP‐ERS of DENV2 N153Q, DENV2 WT, and negative control on the HOSG‐QDs chip are obtained using a 785 nm excitation wavelength (off‐resonance).

A machine learning integrated analysis is done to identify a virus's CP‐ERS signal through principal component analysis and by training an machine learning algorithm. A detailed workflow of principal component analysis and machine learning algorithm is explained in the methods section, Supplementary Note S2 (Supporting Information), and shown in Figure S9 (Supporting Information). For the PCA analysis, 20 CP‐ERS spectra from each class (DENV2 N153Q, DENV2 WT, and negative control) are taken. The principal component analysis scatter plot reveals three clearly distinguishable clusters represented by ellipses (Figure 5a). In each cluster, one dot represents one CP‐ERS spectra compressed into a lower‐dimensional space, the green ellipse corresponds to DENV2 N153Q, the blue to DENV2 WT, and the red to the negative control. The principal component analysis clearly differentiate CP‐ERS spectra of DENV strains from negative control, and it also clearly distinguishes DENV2 N153Q from the DENV2 WT strain. While there is a slight overlap among the ellipses, this may be attributed to shared components in the culture media present across all samples. The clear separation between DENV2 WT, DENV2 N153Q mutant, and negative control groups in the principal component analysis score plot arises primarily from variations in specific Raman spectral regions. These include quasielastic electronic Raman scattering in the low‐energy region (below ∼600 cm⁻¹), particularly dominant in DENV2 N153Q due to its altered electronic environment, and vibrational modes associated with amino acid residues and protein conformational changes, which are detailed in Table 1. Nonetheless, the distinct clustering accounts for the ability of the CP‐ERS to discriminate between different DENV strains and the negative control.

Figure 5.

Principal component analysis and a machine learning algorithm are trained to identify the different DENV strains and distinguish them from the negative control. a) Principal component analysis scatter plot, b) Confusion Matrix, c) Classification report of the multilayer perceptron algorithm for DENV2 N153Q, DENV2 WT, and negative control, labelled as DENV mutant, DENV, and Control, respectively, in the plots, and d) Summary of grouped 5‐fold cross‐validation scores.

In order to classify the DENV2 strains and negative control, several machine learnings algorithms, such as random forest, K‐nearest neighbour, logistic regression, support vector machine, and multilayer perceptron, are trained with the CP‐ERS datasets. The description of choosing multilayer perceptron as our final classification model and optimising the hyperparameters of all models are given in Figure S10 and Table S4 and Table S5 (Supporting Information). For each class, 20 CP‐ERS spectra are used, in total 60 datasets for three classes (DENV2 N153Q, DENV2 WT, and negative control), which are further divided into 80% (48) training data and 20% (12) testing datasets. The performance of the classification model is evaluated using a confusion matrix, which visualises the relationship between the true labels and the predicted outcomes (Figure 5b). The diagonal components of the confusion matrix represent correctly classified samples, i.e., true negatives and true positives. In contrast, classification errors appear in the off‐diagonal elements, false positives, when the model identifies negative samples as positive; second, false negatives, when the model identifies positive samples as negative. Various classification matrices, such as precision, f1‐score, and recall, are used to show the performance of the machine learning model. The trained multilayer perceptron model demonstrates 100% classification accuracy, correctly identifying all DENV2 N153Q, DENV2 WT, and negative control. All the predicted outcomes are located at the diagonal site in the confusion matrix. This highlights the model's strong predictive performance and the discriminative power of the CP‐ERS spectral features. The classification report, shown in Figure 5c, summarises the performance of the multilayer perceptron across the three classes (DENV2 N153Q, DENV2 WT, and negative control).

To further validate the robustness and generalizability of the classification model, we conduct a grouped 5‐fold cross‐validation analysis. The dataset are divided into 5 groups, each containing four spectra from a single sample. In each fold of the cross‐validation, one entire group is used as the test set, while the remaining four groups are used as the training set. The grouped cross‐validation ensures that the data points within the same group remain entirely with either the training or testing set for each fold. This helps in preventing data leakage and provides a more realistic estimate of model performance when working with finite data sets. It also makes sure that all unique spectral features of CP‐ERS are contributing to the model training across different folds, thereby minimising the risk of over‐fitting and improving performance even on unseen new datasets. As shown in Figure 5d, while the model achieves 100% accuracy on a hold‐out test set, we observe an average classification accuracy of 93% using grouped 5‐fold cross‐validation. The relatively consistent performance across folds indicates that the model generalizes well and is not overfitting. The model is evaluated on raw and unprocessed data as well, on the raw data multilayer perceptron classifier achieved a mean accuracy of 55%. Hence to enhance classification accuracy and improve spectral clarity, preprocessing steps detailed in method section are implied. This provides proof of concept of CP‐ERS combined with machine learning for high‐accuracy viral classification for DENV2 WT and its deglycosylated mutant DENV2 N153Q. It is presently limited to one wild‐type DENV and its mutant with one glycosylation variant, and measurements are performed on laboratory‐prepared viral samples. Future studies aim to expand this approach by including multiple serotypes, additional glycosylation sites, and clinical samples, to enhance the generalizability and practical utility of the CP‐ERS and to generate database.

2. Conclusion

In conclusion, we report a novel CP‐ERS based on HOSG‐QDs chip and, using CP‐ERS, discover new resonant quasielastic and inelastic electronic Raman scatterings and phonon excitations of E‐proteins in DENV. The resonant quasielastic Raman scattering of L‐glutamine amino acid in DENV2 N153Q is very strong and plays an important role in its inactivation energy. The resonant inelastic electronic Raman scattering in 0.2 to 0.35 eV of DENV2 WT is related to the HOMO‐LUMO transitions of asparagine amino acid present in DENV2 WT. Intriguingly, supported with theoretical calculations, new phonons of different amino acids of E‐protein embedded in DENV are revealed. These unique resonant quasielastic and inelastic electronic scatterings and phonon excitations enable electronic and molecular fingerprinting and distinguish glycosylated‐induced changes in DENV strains that harbor subtle differences. Furthermore, the principal component analysis clearly shows three distinct clusters for each class, and a trained machine learning model achieved 100% accuracy in hold‐out test set and 93% mean accuracy in grouped cross‐validation in classifying complex CP‐ERS spectra of DENV2 strains and negative controls. This highlights the method's reliability and diagnostic potential. Our result reveals the importance of molecular‐level vibrations in the electronic structure of E‐proteins in viruses, which directly impact viral attachment, entry, and infectivity, and demonstrates a novel CP‐ERS with machine learning as a non‐destructive, stable, direct technique to measure and quantify molecular vibrations and electronic structure of biological and solid state systems.

3. Experimental Section

HOSG‐QDs/MgO(001) Sample Preparation

Each Highly‐Oriented Single‐crystalline Gold Quantum‐Dots (HOSG‐QDs) is prepared on a 1 cm × 1 cm MgO (001) substrate using a unique ultra‐high vacuum molecular beam epitaxy pulsed laser deposition (UHV‐MBE‐PLD system) equipped with a solid‐state Nd:YAG laser (266 nm laser output wavelength), and in situ reflection high‐energy electron diffraction (RHEED) for growth monitoring. Briefly, the substrate is first loaded as received into the UHV‐MBE‐PLD system with a base pressure less than 6 × 10⁻⁹ Torr. The Au target is pre‐laser‐ablated using 1000 pulses at 10 Hz frequency. Gold is subsequently pulsed laser deposited onto MgO (001) at 10 Hz at room temperature. The laser energy is fixed at ≈3.25 Jcm⁻² for all depositions. The sample is removed from the growth chamber and transferred to the heating chamber (base pressure of 10⁻¹⁰ mbar) for post‐anneal at various temperatures (350 °C, 450 °C, 550 °C, 650 °C and 750 °C) for 30 min. Heat is applied with resistive heating. The temperature is adjusted by increasing the applied resistor current in 0.2 A steps until the target current is reached. The time between each current step is determined by ensuring that the chamber pressure does not exceed 6 × 10⁻⁹ mbar at its peak pressure during heating.

Spectroscopic Ellipsometry Measurements

Spectroscopic ellipsometry measurements are performed using a custom‐designed J.A. Woolam V‐VASE ellipsometer. Measurements are conducted at room temperature at four different angles of incidence (40°, 50°, 60°, and 70°). An ellipsometry is a highly accurate technique for simultaneously investigating the complex dielectric function, loss function, and reflectivity. It works by measuring the amplitude ratio Ψ(ω) and phase difference Δ(ω) difference between the reflected p‐polarised (R_p) and s‐polarised light (R_s), which is expressed through $\tan \mathrm{\Psi }\exp \mathrm{\Delta }\equiv \frac{R_p}{R_s}$. Ellipsometer parameters Ψ(ω) and Δ(ω) are fitted using W‐VASE software.

For HOSG‐QDs on MgO, a two layers optical model is built to fit Ψ(ω) and Δ(ω) to obtain a complex dielectric function and LF of HOSG‐QDs. A bulk layer of MgO with 1 mm thickness is used as a substrate underneath the HOSG‐QDs layer. The HOSG‐QDs layer is fitted using combined Lorentz, Gaussian, and PSEMI‐Tri oscillators. The ɛ₂(ω) is fitted by including oscillator parameters such as Amplitude, energy position, and broadening in fit. For ɛ₁(ω), the magnitude and position of poles are fitted. Fitting is performed until the parameters reach convergence with the Levenberg‐Marquardt algorithm. The mean squared error for the spectroscopic ellipsometry fitting is <2.5. The obtained thickness of the HOSG‐QDs layer is ∼5.5 nm.

Further, loss function spectra are useful to acquire plasmonic excitation information obtained from the extracted ɛ(ω) using the following equation

$$lossfunction\equiv -\mathrm{Im}\left[{\stackrel{\sim }{\epsilon }}^{-1}\right]=\frac{{\epsilon }_2}{{\epsilon }_1^2+{\epsilon }_2^2}$$ (1)

Biological Sample Preparation

Dengue virus 2 D2Y98P wild type (DENV2 WT) and DENV2 N153Q strain are grown in C6/36 cells maintained in Leibvotiz‐15 media supplemented with 2% heat‐inactivated fetal calf serum. At day 4 post‐infection, supernatant containing viruses is harvested and centrifuged at 10 000 x g for 10 min at 4 °C. The concentrations of infectious virus particles are determined by plaque assay as described elsewhere.^{[ 45 ]} The viruses are then UV‐inactivated with a UV crosslinker at 5000 µJ cm⁻² twice. The inactivated viruses (2.3 × 10⁴ PFU) are then applied directly onto the HOSG‐QDs chip and air‐dried in the Biosafety cabinet overnight. A pure media control sample without any virus is also prepared on the same chips as a negative control.

Raman and CP‐ERS Measurements

All HOSG‐QD/MgO (001) are used directly as chips for CP‐ERS. A clean MgO (001) single crystal is also prepared as a control. Raman and CP‐ERS spectra are acquired using a u‐Raman‐Ci system (Technospex, Singapore), available with multiple laser wavelengths (785, 633, and 532 nm). A single‐mode, frequency‐stabilized 633 nm laser with 70 mW peak continuous wave power is chosen for resonant Raman spectroscopy. The 633 nm laser resonate with the correlated plasmon energy of HOSG‐QDs chip. For the comparison, off‐resonance 785 nm laser with 70 mW peak continuous wave power is also used. The spot size of the laser spot is <1 µm. The different spots on the sample are identified by focusing on the sample through a camera (Nikon Plan Flour, 0.50 numerical aperture). Before acquiring the Raman and CP‐ERS spectra, the focus is carefully adjusted and tested to minimise any fluctuations in the spectra caused by focusing inconsistencies. To assess potential photothermal artefacts in CP‐ERS spectra, spectra are recorded at two different laser powers, 7 and 20 mW (Figure S11, Supporting Information). Based on this comparison, 7 mW is selected for all subsequent measurements. All spectra are obtained with a photon flux of 7 mW and 1 s acquisition time with an averaging of 2 spectra through a 20x objective lens of the camera. Measurements are taken under ambient conditions; the instrument is calibrated using a standard reference material (polystyrene with a Raman peak at ∼1000 cm⁻¹) and optimised for accurate peak identification. The spectra are processed and decoded using uSoft software (Technospex). After data collection from uSoft, further processing is done using the normalization method in Origin software, with normalization performed relative to the Raman intensity in the range 3500 cm⁻¹ and above.

Principal Component Analysis and Machine Learning Model Building

For the evaluation of class‐specific clustering, principal component analysis is performed on 60 CP‐ERS datasets, 20 CP‐ERS spectra from each class, i.e., DENV2 N153Q, DENV2 WT, and the negative control. The entire CP‐ERS spectra, in the range 200–3500 cm⁻¹, are taken into analysis. To reduce the baseline contributions, the asymmetric least squares method is applied for baseline corrections to each spectrum. Following this correction, z‐score scaling via StandardScalar from the scikit‐learn library is used to normalise all CP‐ERS spectra. The principal component analysis is then conducted using a linear decomposition algorithm implemented in scikit‐learn's principal component analysis class, which utilises singular value decomposition to compute the two principal components (PC1 and PC2). A principal component analysis score plot is generated to show the distribution of samples across the reduced feature space.

To classify different DENV strains and negative control samples based on their CP‐ERS spectra, a supervised machine learning approach is employed. First, raw spectral data is stored into a file. Raman shift values are extracted and used as the independent variable, while the spectral intensities are assigned class labels: DENV2 WT, DENV N153Q, and negative control. To ensure comparability between features, spectral intensities are normalised using z‐score scaling via StandardScaler. The dataset is then split into training and testing subsets using an 80:20 ratio. We implemented and trained a multilayer perceptron neural network with a single hidden layer of 100 neurons and 500 maximum iterations.^{[ 46 ]} The model is computed using scikit‐learn's evaluation functions, and performance is assessed based on performance metrics, including accuracy, confusion matrix, and classification report (precision, recall, and f1‐score). To further ensure robustness and to prevent overfitting, a grouped 5‐fold cross‐validation analysis is also conducted. A machine learning model is trained and validated across 5 random data subsets, and the mean accuracy and standard deviation of the cross‐validation scores are recorded. This step confirmed the consistency and generalizability of the model's performance

Conflict of Interest

The authors declare no conflict of interest.

Supporting information

Supporting Information

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.