Method for high-resolution measurement of concentrations of angiopoietin (ANG)-1 / 2 and soluble triggering receptor expressed on myeloid cells-1 (strem-1) biomarkers

Abstract

The present invention provides a method for near real-time high-resolution measurements of molecular (bio)markers selected from soluble triggering receptor expressed on myeloid cells-1 (sTREM-1), Angiopoietin-1 (Ang-1), and Angiopoietin-2 (Ang-2) based on surface-enhanced Raman spectroscopy (SERS). Using these measurements, the invention provides a means of monitoring progression of sepsis.

Description

[0001] METHOD FOR HIGH-RESOLUTION MEASUREMENT OF CONCENTRATIONS OF ANGIOPOIETIN (ANG)-1 / 2 AND SOLUBLE TRIGGERING RECEPTOR EXPRESSED ON MYELOID CELLS-1 (STREM-1) BIOMARKERS

[0002] FIELD

[0003] The disclosure relates to a high-resolution method that can accurately quantify sTREM-1 , Ang-1 , and Ang-2 biomarkers, which is useful, for example, for determination of sepsis state and associated treatment of a subject.

[0004] BACKGROUND

[0005] Sepsis, a substantial worldwide health challenge, is defined as a marked systemic host response to infection leading to multi-organ dysfunction. It can arise from bacterial, viral, fungal, or parasitic sources (Ref. 1 :Wright et al 2018, BMC Pediatrics. 18(1)). In certain cases, sepsis can advance to a stage where it is associated with severe inflammation, tissue coagulopathy (blood clotting issues), and organ dysfunction and death. It has an especially notable impact among maternal and newborn populations (Ref. 2: Briggs-Steinberg et al 2023, Pediatrics In Review, 44(1 ), 14-22). Sepsis begins with the activation of innate immune response mediated and endothelial cell activation / dysfunction (Ref. 3:Kim et al 2020, Infection & Chemotherapy, 52(1), 1 ).

[0006] Angiopoietins are involved in vascular development and angiogenesis, which regulate the integrity of the interface between neighboring endothelial cells. Four angiopoietins are known, named Angiopoietin-1 to Angiopoietin-4. Ang-1 is responsible for maintaining vascular quiescence in the adult endothelium (Ref. 4: Fiedler et al 2006, Trends in Immunology, 27(12), 552- 558). The Ang-1 stabilizing effect is antagonized by Ang-2, which is released in response to injury, hypoxia, and systemic infection due to bacterial, viral, fungal and protozoal sources (Ref. 5: Loughna et al 2001 , Matrix Biology, 20(5-6), 319-325). Increasing Ang-2 concentration and higher Ang-2 / Ang-1 ratios are thus associated with sepsis (Ref. 6: Yuan et al 2009, Molecular and Cellular Biology, 29(8), 2011-2022).

[0007] Furthermore, TREM-1 is another crucial mediator of sepsis. As an innate immune cell, TREM-1 can amplify the inflammatory responses to pathogens, thus promoting sepsis-induced immune dysregulation and organ dysfunction (Ref. 7:Cao et al 2017, Frontiers of Medicine, 11 (2), 169-177). Soluble TREM-1 (sTREM-1) is a particular form of TREM-1 , which can be directly tested in human bodily fluids. Both sTREM-1 and the Ang-1 / 2 complex present as sensitive, specific drug-targetable diagnostic and predictive biomarkers compared to traditional indicators like C-reactive protein (CRP) and procalcitonin (PCT) levels in sepsis. They have the function as an active mediator within pathways so that they can be seen as a “druggable” targets to enhance the prognosis of life-threatening infections (Ref. 8: Higgins et al 2016, Science Translational Medicine, 8(358); Ref. 9: Adly et al 2014, Cytokine, 65(2), 184-191 ; Ref. 10: McDonald et al 2018, Malaria Journal, 17(1 )).

[0008] For certain instances of severe sepsis and septic shock, no universal panacea exists. Consequently, the timely identification and diagnosis of sepsis is pivotal in curbing mortality. So far, sTREM-1 , Ang-1 , and Ang-2 biomarkers are quantified by enzyme-linked immunosorbent assay (ELISA), and Luminex assay measurements. The lower and upper limits of detection for each assay are as follows: Ang-1 (156.25 - 10,000 pg / mL; sensitivity 9.43 pg / mL; standard deviation 0.2; resolution 0.021 pg / mL; sample volume 50 pL; incubation time 3- 5 h), Ang-2 (54.69 - 3,500 pg / mL; sensitivity 17.1 pg / mL; standard deviation 0.2; resolution 0.012 pg / mL; sample volume 50 pL; incubation time 3-5 h) (Ref. 11 : Conroy et al 2009, Malaria Journal, 8(1 )) and sTREM-1 (31 .4 - 22870 pg / mL; sensitivity 31 pg / mL; standard deviation 0.25; resolution 0.008 pg / mL; sample volume 100 pL; incubation time 3.5 h) (Ref. 12: Balanza et al 2023, PLOS Global Public Health, 3(2)). However, commercial ELISA and Luminex assays provide low resolution at small concentrations, and require long incubation times, as well as requiring large amounts of sample. As such, the necessity persists for a rapid, sensitive, selective, and specific high-resolution technique for the early detection of sepsis, as it is pivotal in curbing mortality.

[0009] SUMMARY

[0010] The present disclosure provides a method for high-resolution measurements using SERS which involves using metallo-dielectric nanostructures with enhanced sensitivity which include plasmonic materials. The system exhibits high-resolution sensing biomarkers in uses machine learning classification pipelines for quantification of the amounts of biomarkers. The system involves biomarker and biomarker characteristic peak identification through decomposition or deconvolution of SERS spectra.

[0011] The present disclosoure provides a Surface Enhanced Raman (SERS) method of detecting biomarkers indicative of a state of sepsis in a patient, comprising: identifying at least one target antigen biomarker indicative of a state of sepsis to be screened for in a sample; functionalizing a plasmonic metal forming part of the nanostructured SERS active substrates with an antibody with specificity to the at least one target antigen biomarker; exposing the at least one functionalized SERS active substrate to the sample being screened for the absence or presence of the target antigen biomarker; illuminating the SERS active substrate with laser light of a preselected wavelength and recording a sample SERS spectra; and comparing the sample SERS spectra to a pre-collected set of biomarker SERS spectra of known biomarkers indicative of a state of sepsis at varying concentrations, and based on this comparison, determining whether the at least one target biomarker is absent from, or present in, the sample, and if present determining from the comparison of the SERS spectra a concentration of the at least one target biomarker in the sample.

[0012] The nanostructured SERS active substrates may include an array of bullseye structures with each bullseye structure comprised of a preselected number of concentric circular troughs with said preselected number concentric circular troughs being in a range from about one to about ten, a depth of said circular troughs being in a range from about 5 nm to about 500 nm, a width of said circular troughs being in a range from about 20 nm to about 500 nm, and wherein each bullseye structure has an adhesion layer coating a bottom of each trough and a plasmonic metal layer coating the adhesion layer coating in said troughs.

[0013] The preselected number of concentric circular troughs is greater than one with each trough having the same width and the same distance between each circular trough from a center of each bullseye structure towards the outermost peripheral trough, and wherein and wherein a separation between a middle of adjacent troughs is in a range from about 20 nm to about 500 nm.

[0014] A thickness of the plasmonic metal coating may be in a range from about 1 nm to about 100 nm.

[0015] The thickness of the adhesion metal coating may be in a range from about 1 nm to about 30 nm.

[0016] The plasmonic metal coating may be any one of gold, silver and copper.

[0017] The adhesion layer coating is any one of titanium, chromium, nickel, tungsten, niobium, chromium, or their oxide forms.

[0018] The substrate may be made of a material having a low index of refraction, low dielectric constant or low relative permittivity selected from any one of glass, silicon, polymer, dielectric and linear / non-linear optical materials.

[0019] The substrate is made of any one of silicon (100), silicon (111 ), polymethyl methacrylate (PMMA), polydimethylsiloxane (PDMS) and poly(N,N- dimethylacrylamide) (PDMA).

[0020] The nanostructured SERS active substrates include but are not limited to nano-crescents, nano-stars, nano-prisms, nano-grating, nano-pillars, nanochirped circular grating.

[0021] The method includes functionalizing a SERS active substrate with an antibody with specificity to the target antigen biomarker comprises depositing biotinylated PEG thiol (BAT) molecules onto the SERS active substrate which binds to locations on the plasmonic metal forming part of the SERS active substrate, and depositing a passivating agent onto the surface of the plasmonic metals which coated areas on the plasmonic metal free of the biotinylated PEG thiol (BAT) molecules to prevent non-specific binding of the target antigen biomarkers to the surfaces of the plasmonic metals.

[0022] The passivating agent may be any one of molecules of thiol polyethylene glycol) (Thiol-PEG), 1 -ethyl-3- (3_dimethylaminopropyl)carbodiimide (EDC) combined with N- hydroxysuccinimide (NHS).

[0023] The at least one target biomarker being tested for may be sTREM-1 , wherein the step of functionalizing at least one SERS active substrate further comprises deposition of sTREM-1 biotinylated antibodies which binds to the streptadivin and wherein, if sTREM-1 is present in the sample it will form complexes with the sTREM-1 biotinylated antibodies.

[0024] The at least one target biomarker being tested for may be Ang-1 , wherein the step of functionalizing at least one SERS active substrate further comprises deposition of Ang-1 biotinylated antibodies which binds to the streptadivin and wherein, if Ang-1 is present in the sample it will form complexes with the Ang-1 biotinylated antibodies.

[0025] The at least one target biomarker being tested for may be Ang-2, wherein the step of functionalizing at least one SERS active substrate further comprises deposition of Ang-2 biotinylated antibodies which binds to the of streptadivin and wherein, if Ang-2 is present in the sample it will form complexes with the Ang-2 biotinylated antibodies. The sample may be a sample of human bodily fluid which may be any one or combination of blood, plasma, serum, urine and saliva.

[0026] The method may include separate SERS active substrates functionalized for detection of each of sTREM-1 , Ang-1 and Ang-2 are integrated into the system, and wherein the concentrations of biomarkers are determined using machine learning classification and regression pipelines, which are trained to distinguish between any combination of sTREM-1 , Ang-1 , Ang-2 and predict their concentrations.

[0027] The classification and regression pipelines are configured specifically for spectral data and comprise: a) spectrum preprocessing, to remove or minimize background signal from radiation, noise, and / or SERS responses from other material, clean and validate the data to remove outliers or erroneous measurements, and standardize the data such that data from independent experiments can be compared; b) data augmentation algorithms to artificially generate new data samples from the existing data collected from real-world samples, to increase the number of diverse data samples used for training, validation, development, and / or testing of machine learning methods and thereby improve the performance of these methods; c) feature selection to extract a base set of the most informative features for biomarker discrimination from the dense spectral data collected, thereby reducing the dimensionality of the input data used by the machine learning methods and improving their computational time and space costs / complexity, and real-world generalizability; d) feature and data transformation methods for polynomial fitting or to place the spectra data into a desired feature space to improve machine learning performance and / or costs; e) outlier rejection or robustness techniques to mitigate and / or eliminate biased or erroneous data points or features within the data, to improve the robustness, accuracy, and reliability of the machine learning methods; and f) a machine learning classifier and / or regressor.

[0028] The regression pipeline predicts the concentration of sTREM-1 , Ang-1 , or Ang-2 from a spectrum containing only one biomarker.

[0029] The regression pipeline predicts the concentration of any combination of sTREM-1 , Ang-1 , and Ang-2 from a spectrum containing any mixture of these biomarkers.

[0030] The regression pipeline predicts the concentration of Ang-1 and subsequently predicts the concentration of Ang-2 using the biomarker relationship, and vice versa.

[0031] The method includes developing a methodology for correlating findings from a machine learning classification pipeline to spectral deconvolutions or decompositions, comprising: a) spectral deconvolution or decomposition into underlying Gaussian or Lorentzian functions; b) extraction of significant peaks and relevant peak properties from the deconvolution or decomposition; c) a machine learning classification pipeline, with feature selection or engineering functionalities to identify significant wavenumbers; d) comparison of significant wavenumbers from the machine learning classification model to extracted peaks and their properties from the deconvolution or decomposition.

[0032] The correlation is done between the spectral deconvolution or decomposition of Ang-1 , Ang-2, or sTREM and one or more machine learning classification pipelines including said biomarker for the purpose of identifying potential characteristic peaks.

[0033] The correlation is used to provide physical explainability, such as vibrational modes of the molecular structure of Ang-1 , Ang-2, or sTREM, to the machine learning classifier.

[0034] The findings from deconvolution or decomposition, or identified characteristic peaks, are used as primary features for the development of a machine learning classifier to distinguish between any combination of Ang-1 , Ang-2, and sTREM.

[0035] A further understanding of the functional and advantageous aspects of the disclosure can be realized by reference to the following detailed description and drawings.

[0036] BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The present disclosure to the method and apparatus for detecting target biomarkers Ang-1 , Ang-2 and sTREM-1 indicative of sepsis will now be described in reference being had to the accompanying drawings, in which: Figure 1A is a cross sectional side view showing a schematic illustration depicting a SERS substrate fabrication process from left to right: (i) spin-coating of a single layer of photoresist (a) (such as poly-methyl-methacrylate (PMMA A3)) onto a silicon wafer (100) substrate (b), (ii) e-beam lithography to pattern the photoresist and expose the silicon substrate in a preselected pattern, (iii) reactive ion etching of the exposed sections of the silicon substrate, (iv) plasmonic metal deposition with a plasmonic metal such as Ti / Au (c) to coat the exposed sections of the silicon substrate and the remaining photoresist, and (v) lift-off to remove the remaining photoresist (a).

[0038] Figure 1 B is a schematic illustration depicting a SERS substrate fabrication process and surface passivation for the capture of sTREM-1 (h) with steps (i) to (iv) identical to Figure 1A: (i) spin-coating of photoresist, (ii) e-beam lithography, (iii) reactive ion etching, (iv) plasmonic metal deposition, (v) lift-off of remaining photoresist and deposition of thiol polyethylene glycol) (Thiol- PEG) (d) and biotinylated PEG thiol (BAT) (e), (vi) deposition of streptadivin (f); (vii) deposition of sTREM-1 biotinylated antibodies (g), (viii) capture of sTREM- 1 (h) from delivered sample for sTREM-1 measurement.

[0039] Figure 1C is a schematic illustration depicting a SERS substrate fabrication process and surface passivation for the capture of Ang-1 (k) with steps (i) to (vi) identical to Figure 1B: (i) spin-coating of photoresist, (ii) e-beam lithography, (iii) reactive ion etching, (iv) plasmonic metal deposition, (v) lift-off of remaining photoresist and deposition of thiol polyethylene glycol) (Thiol- PEG) (d) and biotinylated PEG thiol (BAT) (e), (vi) deposition of streptadivin (f); (vii) deposition of Ang-1 biotinylated antibodies (j), (viii) capture of Ang-1 (k) from delivered sample for Ang-1 measurement. Figure 1 D is a schematic illustration depicting a SERS substrate fabrication process and surface passivation for the capture of Ang-2 (m) with steps (i) to (vi) identical to Figure 1B: (i) spin-coating of photoresist, (ii) e-beam lithography, (iii) reactive ion etching, (iv) plasmonic metal deposition, (v) lift-off and deposition of thiol polyethylene glycol) (Thiol-PEG) (d) and biotinylated PEG thiol (BAT) (e), (vi) deposition of streptadivin (f); (vii) deposition of Ang-2 biotinylated antibodies (I), (viii) capture of Ang-2 (m) from delivered sample for Ang-2 measurement.

[0040] Figure 2 shows an atomic force microscopy image of a bulls-eye nanostructured SERS substrate.

[0041] Figure 3 is a plot of Intensity (a.u.) versus Raman Shift (cm'1) showing the SERS spectrum of (a) sTREM-1 and (b) sTREM-1 , antibodies.

[0042] Figure 4 is a plot of Intensity (a.u.) versus Raman Shift (cm-1) showing the SERS spectrum of (a) Ang-1 and (b) Ang-1 , antibodies.

[0043] Figure 5 is a plot of Intensity (a.u.) versus Raman Shift (cm-1) showing the SERS spectrum of (a) Ang-2 and (b) Ang-2, antibodies.

[0044] Figure 6 shows the relationship between (a) sTREM-1 at 460 cm'1, (b) Ang-1 at 455 cm'1and (c) Ang-2 at 437 cm'1concentrations (pM) and their characteristic SERS peak intensities (a.u.).

[0045] Figure 7 shows the SERS spectra for (a) sTREM-1 with gold nanoparticles in solution and (b) gold nanoparticles in buffer solution.

[0046] Figure 8 is a schematic of an embodiment of a SERS system use for detection of biomarkers indicative of sepsis, comprising laser source, SERS substrate, detector and data capturing and processing. Figure 9 shows a schematic of sample delivery via a microfluidic circuit 10 to SERS-active regions on the substrate, with each region functionalized for one specific antigen; circuit 10 includes a sample inlet 12, microfluidic channels 14, array of target structures 16 functionalized for capture of one specific antigen, e.g., sTREM-1 , array of target structures 18 functionalized for capture of one specific antigen, e.g., Ang-1 , array of target structures 20 functionalized for capture of one specific antigen, e.g., Ang-2, sample collector 24, and substrate 26.

[0047] Figure 10 shows a confusion matrix displaying the true labels versus the predictions made on the test set by a multilayer perceptron classification model when distinguishing between the SERS spectra of Ang-1 , Ang-2, and their respective antibodies.

[0048] Figure 11A shows the process flowchart for the machine learning classification pipeline.

[0049] Figure 11B shows the process flowchart for machine learning regression pipeline 1.

[0050] Figure 11C shows the process flowchart for machine learning regression pipeline 2.

[0051] Figure 11D shows the process flowchart for machine learning regression pipeline 3.

[0052] Figure 12 shows a fitted curve using Eqn. (3) to experimental data from published data (Ref. 35:A. Leligdowicz et al 2021 , Nat Commun, 12(1 ), 6832) of sTREM_1 concentration versus probability of 7-day mortality.

[0053] Figure 13A shows the process flowchart for characteristic peak determination from deconvolution / decomposition and machine learning. Figure 13B shows the process flowchart for deconvolution / decomposition for the machine learning classification pipeline.

[0054] BRIEF DESCRIPTION OF THE TABLES

[0055] Table 1 shows geometrical Parameters WAU, Wspace for the patterns used herein.

[0056] Table 2 shows spectral assignments for SERS Spectra of R6G molecule.

[0057] Table 3 shows size or molecular weight of each component.

[0058] Table 4 shows predictions for Ang-1 , Ang-2, and sTREM-1 concentrations made on the test set by polynomial (Ang-1 , Ang-2) and linear regression (sTREM-1) models versus the actual concentrations.

[0059] Table 5 shows Ang-1 and Ang-2 concentrations and ratio (Ang-2 / Ang-1 ) predictions made on the test set by a convolutional neural network versus the actual concentrations, where elevated levels of Ang-2 and high values for the Ang-2 / Ang-1 ratio may potentially be used as predictors of worst-case prognoses for sepsis.

[0060] DETAILED DESCRIPTION

[0061] This disclosure provides a high-resolution method that can accurately quantify sTREM-1 , Ang-1 , and Ang-2 biomarkers, which is useful, for example, for determination of sepsis state and treatment of a subject. Although embodiments of the present invention are disclosed herein, the disclosed embodiments are merely exemplary and it should be understood that the invention relates to many alternative forms, including different shapes and sizes. Furthermore, the Figures are not drawn to scale and some features may be exaggerated or minimized to show details of particular features while related elements may have been eliminated to prevent obscuring novel aspects.

[0062] Therefore, specific structural and functional details disclosed herein are not to be interpreted as limiting but merely as a basis for the claims and as a representative basis for enabling someone skilled in the art to employ the present invention in a variety of manners.

[0063] As used herein, the terms “comprises”, “comprising”, “includes” and “including” are to be construed as being inclusive and open-ended, and not exclusive. Specifically, when used in this specification including claims, the terms “comprises”, “comprising”, “includes” and “including” and variations thereof mean the specified features, steps, or components are included. These terms are not to be interpreted to exclude the presence of other features, steps, or components.

[0064] As used herein, the terms “about” and “approximately”, when used in conjunction with ranges of dimensions, compositions of mixtures, or other physical properties or characteristics, are meant to cover slight variations that may exist in the upper and lower limits of the ranges of dimensions so as to not exclude embodiments where on average most of the dimensions are satisfied but where statistically dimensions may exist outside this region. It is not the intention to exclude embodiments such as these from the present invention.

[0065] As used herein, the coordinating conjunction “and / or” is meant to be a selection between a logical disjunction and a logical conjunction of the adjacent words, phrases, or clauses. Specifically, the phrase “X and / or Y” is meant to be interpreted as “one or both of X and Y” wherein X and Y are any word, phrase, or clause.

[0066] Raman spectroscopy is a promising technique desired for biomarker detection and measurement, which is an inelastic scattering process where photons or light incident on a sample transfer energy to or from the sample’s vibrational or rotational modes (Ref. 13: Orlando et al 2021 , Chemosensors, 9(9), 262). Individual bands in the Raman spectrum are characteristic of specific molecular vibrations. As a result, each analyte has its own unique Raman signature. For biomolecules, Raman spectroscopy is nondestructive without prior fluorescent or radioactive labeling, however, the low efficiency of Raman scattering hinders its applications in the detection at low concentrations. Surface enhanced Raman scattering (SERS) retains all the advantages of normal Raman spectroscopy while achieving significantly stronger signal intensity (the Raman signal increases by multiple folds of magnitude) (Ref. 14: Kazemi-Zanjani et al 2018, Advanced Optical Materials, 6(4), 1701136).

[0067] SERS is a process whereby the Raman signal is increased when a Raman-active molecule is spatially confined proximate to a strong local electromagnetic field generated upon excitation of the localized surface plasmon resonance (LSPR) on nanostructured metal surfaces. Accordingly, SERS possesses many desirable characteristics for biochemical analysis including high specificity, sensitivity, and being fast to acquire data. LSPR occurs when suitable wavelength electromagnetic radiation impinges on a noble metal nanostructure causing conduction electrons to oscillate collectively. The resonance oscillation is localized near the surface region of the nanostructure. Such resonance is advantageous in that the nanostructure is selectively excited at a particular photon absorption, which results in the generation of locally enhanced or amplified electromagnetic fields proximate to the nanostructures and which occurs in the visible and IR regions of the spectrum and can be measured by UV-visible-IR (200 to 3500 nm) extinction spectroscopy (Ref. 15: Cara et al 2020, Journal of Materials Chemistry C, 8(46), 16513-16519). SERS from metallic nanostructures increases the original Raman scattering intensity many orders of magnitude, which makes the Raman detection of low concentrations of biomolecules and in particular, biomarkers, practical.

[0068] In order to improve the SERS performance further, machine learning algorithms can be applied to acquired SERS spectra. Machine learning is a data-driven approach to creating algorithms (models) for tasks such as identification and prediction. Supervised learning is a common technique for developing machine learning models where the model attempts to approximate input-output mapping functions. These models learn to approximate the mapping by training on input data which has already been labelled with the expected output. Once trained, outputs to new data can be predicted through the mapping learned by the algorithm. In supervised classification, models aim to predict the correct class label from a predetermined set of classes using input data. In supervised regression, models aim to predict the value of a continuous, numerical variable using input data. SERS spectrum data has been used as input data for machine learning models in classifying materials and other biological samples.

[0069] In the case of sepsis detection using biomarkers, classification models can be employed for distinguishing between Ang-1 , Ang-2, sTREM-1 , and their respective antibodies on a substrate using SERS spectra. Furthermore, regression models can be used for identifying the existence of the various biomarkers or predicting their concentration within a biomarker mixture. The benefit of using machine learning in this context is the ability for the detection algorithms to learn to identify biomarkers within mixtures with little human input and to self-improve given more data or optimization. These features allow for biomarker detection to become more sensitive and selective compared to current techniques.

[0070] The present disclosure provides a high-resolution method for determining and measuring the presence and concentration of a biomarker of interest in a biological sample using SERS. As described in more detail below, the detection capabilities were improved in one or more ways, these improvements include, but are not limited to, an increase in signal intensity, enhanced sensitivity, higher resolution, and improved reproducibility (Ref. 16: Maccaferri et al 2021 , Nanoscale Advances, 3(3), 633-642).

[0071] In general, a metal coating can be disposed on a nanostructure, where the size of the nanostructure can be controlled, which can result in enhanced properties. An embodiment of the present disclosure includes a gold coated bullseye nanostructure SERS substrate (Table 1), which can have a depth of cavities about 50 nm to 100 nm, width of gold cavities about 50 nm to 1 urn and have a distance between cavities about 50 nm to 1 urn (see Figure 2).

[0072] Table 1

[0073] Figure 2 shows the atomic force microscope image for the bullseye nanostructure SERS substrate. As shown, each bullseye structure contains four concentric circles, which have constant width and constant distance in between from the centre towards the periphery with a channel etched into silicon wafer and filled ~40 nm above the surface using titanium and gold. While four (4) concentric circles are shown in Figure 2, it will be appreciated that the number of concentric circles can vary in a range from about one (1) to about ten (10).

[0074] Figures 1A to 1D show, for illustrative purposes, the fabrication and functionalization of two adjacent troughs in the bullseye structure shown in Figure 2.

[0075] The E-shaped in Figures 1A to 1D show cross-sectional views of the nanostructured substrate (with the top view resembling a bullseye as shown in Figure 2). The SERS activity is generated by the structure itself.

[0076] The present circular trough structures of the nanostructured SERS active substrates for immobilization of the antibodies with specificity to the target antigen biomarkers is advantageous for the following reasons. The dimensions of the troughs and the spacing between adjacent troughs can trap incident light striking the substrate. This can lead to optical resonance being established within the troughs which, when interacting with the antigen biomarkers increases the Raman signal. This leads to further enhancement of the Raman signal and increases the limit of detection for the biomarker. The use of up to ten (10) concentric circular troughs is advantageous because this can increase the region over which optical resonance is established, further increasing the Raman signal enhancement from the biomarker.

[0077] It has been demonstrated the significant enhancement of this technique. A common analyte for characterizing SERS enhancement, Rhodamine 6 G (R6G), for example, absorbs strongly onto gold via its chloride ion (Ref. 17: Pristinski et al 2006, Journal of Raman Spectroscopy, 37(7), 762-770). The normal Raman scattering of the pure R6G was not observable when the concentration of R6G was 10'5M. But in the presence of substrates, the Raman scattering of R6G can be observed distinctly. It is characteristic of Raman spectra of R6G and the assignment for the several peaks are listed in Table 2.

[0078] Wave number 608 765 1116, 1179 1306, 1575 1355, 1508,

[0079] (cm-1) 1652

[0080] Vibration mode C-C-C C-H C-H N-H C-C

[0081] Ring in Out-of- In-plane In-plane Stretching plane plane bend bend bend vibration

[0082] Table 2 To gain additional insight into Raman enhancement in this technique, the enhancement factor (EF) was calculated. These EFs were calculated by applying Eqn. (1 ): where, IsERs and I Raman denote to the highest peak (1504 cm'1) intensities collected over the SERS platform and on flat gold, respectively. NSERS and NRaman are the number of scattering molecules in the illuminated volume of the sample on the SERS substrate and over flat silicon substrate , respectively. The SERS spectra were collected at 785 nm laser wavelength (red laser). The laser power was maintained at 10% and the exposure time was kept at 10 s.

[0083] The method of Yue (Ref. 18: Yue et al 2020, Sensors & Actuators: B. Chemical, 322 128563) was followed to determine an enhancement factor. The radius of the laser spot was 11 pm, and the height of dried R6G was 5 pm. Enhancement factors ranging from 105to 108were determined.

[0084] The spectral resolution was also improved by this technique. Some peaks are not visible on the flat silicon but are very distinct on the bullseye nanostructure SERS, especially for those vibrational modes corresponding to aromatic ring bending and stretching, such as 608 cm'1, which comes from the aromatic ring in-plane bending (Ref. 19: Zhu et al 2018, Nanomaterials, 8(7), 520) at low concentration (1 O'3M to 10'8M). The reproducibility of the SERS substrate was further investigated by taking SERS spectra of R6G at concentration of 10'3M from 9 random locations on a single substrate The average relative standard deviation of the intensities (at 1361 cm'1) was 0.4% indicating that the substrates process good signal uniformity.

[0085] The enhancement of SERS is highly dependent on the interaction between adsorbed molecules and the metallic nanostructures on the substrates. Various types of nanostructures have been fabricated, including gold or silver nano-crescents, nano-stars, nano-prisms, nano-grating, and nano-pillars (Ref. 20: Li et al 2011 , Comprehensive Biotechnology, 125-139). SERS enhancement was observed in the nano-pillar and nano-grating, which was attributed to the LSPR coupling that in turn results from variations in surface nano-topography. Nano-gratings include many different structural variations, including but not limited to chirped gratings, and chirped circular gratings.

[0086] The present embodiment exhibits a specific / unique chirped circular grating / bullseye antenna (expansion as 4 concentric gold same-width rings separate with same distance between each ring). Within plasmonic gratings, altering the pitch presents opportunities for influencing the dispersion relation and tailoring the plasmonic response. Thus, the chirped circular grating shows a consistent cavity or groove with a range from 20 nm -1 pm, with the pitch being adjusted through variations in groove-groove separation.

[0087] The bullseye nanostructured SERS substrate features a large surface area (for better molecular adsorption) and a long edge length for the maximized total integration of multiple SERS tips compared to regular chirped grating, which resulted in even higher degree of field enhancement.

[0088] The nanostructures for capture and detection of the target sTREM-1 , Ang-1 , and Ang-2 biomarkers were constructed by the combination of electron beam lithography and reactive ion etching (see Figure 1A). Electron beam lithography is a highly precise method for defining the SERS nanostructures with high resolution (sub-10 nm) as compared with optical lithography owing to the shorter wavelength of electrons compared with ultraviolet photons. Reactive ion etching is a plasma etching process, wherein gaseous ions are accelerated normal to the patterned masked substrate surface to effect selective removal of material.

[0089] Preparation of SERS Substrates

[0090] Figure 1A (i) to (v) show the basic steps to prepare the SERS substrate prior to being functionalized for capture of the specific targets, as shown in Figures 1B, 1C and 1D. The bullseye patterns (Figure 2) were designed using LayoutEditor (Circular Utilities-> Boolean tool; B-A with delate A, B) in GDSII format, then the GDSII files were processed using BEAMER (GenISys), which can optimize the laser-beam exposure with proximity effect correction (PEC) to achieve high resolution, smooth pattern under electron-beam lithography (EBL). The sample for EBL was prepared from a 1 .5 cmx1 .5 cm silicon, which was cut from a 6-inch silicon wafer (100) (thickness of 375±15 gm, B-doped, p-type, supplied by SI-TECH, INC) following the correct orientation of the wafer using a high-quality diamond scribe pen. The sample was sonicated in 2-propanol (IPA) for 5 min, rinsed in acetone and deionized (DI) water for 1 min, then it was dried with nitrogen gas and baked at 180 °C for 5 min to remove water and cooled for 2 min. Those steps can make sure the top surface was clean enough to get photo resist coating. A single, photo resist layer of poly-methyl-methacrylate (PMMA, A3, 950 K 3% dissolved in anisole, supplied by MicroChem corp.) is spin-coated at a rate of 1000 rpm for 60 s to obtain a 300 nm film, and is subsequently baked at 180 °C for 60 s to remove any residual solvent. It will be appreciated that photoresists other than PMMA may be used as will be known to those skilled in the art along with the processing conditions associated with each.

[0091] The sample was loaded on an Electron Beam Lithography Holder and patterned by a modified transmission electron microscope (TEM) modified Raith EBPG 5000+ Electron Beam Lithography System at 100 keV and 10 nA beam current. The sample was exposed at 400 / C / cm2, which was achieved by varying the dose from 400 to 1000 / C / cm2. The exposed resists were developed for 60 s in isopropyl alcohol (MIBK: IPA) (1 :3) and rinsed for at least 15 s in IPA. All organic solvents were used as received unless specified otherwise, supplied by Sigma-Aldrich.

[0092] After EBL exposure, the sample is then exposed to SFe plasma in a reactive ion etching (RIE) chamber for 150 s at a RF power of 50 W, gas flow rate of 50 seem, at a chamber pressure of 30 mTorr using Oxford Instruments PlasmaPro Estrelas100 DRIE System (Si). Finally, a combination of 10 nm titanium (Ti) and 40 nm gold (Au) is deposited onto the sample in a E-beam evaporator (Angstrom Nexdep Electron Beam Evaporator) at the deposition rate of 0.5 A / s for Ti and 0.2 A / s for Au, then submersed in acetone for 24 hours to strip away the PMMA and lift-off the metal.

[0093] The SERS substrates were highly dependent on the interaction between adsorbed molecules and the surface of plasmonic nanostructures, therefore, SERS substrates were typically comprised of wafers and metallic coating. The metallic coating can be chosen from silver (plasmonic wavelength 390 nm), copper (plasmonic wavelength 560-570 nm) or gold (plasmonic wavelength 530 nm) since due to their LSPR properties, which covered a wide wavelength range in the visible and near-infrared regions of electromagnetic spectrum with good air stability and weak reactivity (Ref. 21 Pal et al 2020, Sensors and Actuators A: Physical, 314, 112225). In the case of the SERS biosensors for biological and medical applications, it is important to use an inert material, which does not change the properties of the biological molecules.

[0094] Therefore, gold being slightly more stable and unreactive is an appropriate material for this purpose. The material of wafer can also support surface oscillations of free electron (electromagnetic waves coupling) and trap lightwaves on the surface. These collective surface oscillations can concentrate electromagnetic fields on the nanoscale, enhancing local field strength in a particular direction by several orders of magnitude. Normal propagating electromagnetic waves can have constant phase and amplitude in the same plane. Even low inertia electrons can fail to keep up with high frequencies, the dependence on the material used can be described by the index of refraction

[0095] Eqn. (2):

[0096] „ . . , lightspeed refraction index = - medium wavelengthx frequency (2)

[0097] Thus, to achieve high frequency, the low index of refraction (or dielectric constant or relative permittivity) substrate layer may be utilized including but not limited to glass, silicon, polymer, dielectric, linear / non-linear optical materials (Ref. 22: Ciddor et al 1996, Applied Optics, 35(9), 1566).

[0098] In order to stabilize the interface between wafer and metallic coating, adhesion layers were used as a thin film (250 A - 400 A) of tungsten, niobium, chrome, or titanium or their oxide forms (dielectric material like TiO2 or Cr2O3) (Ref. 23: Colas et al 2015, Journal of Optics, 17(11), 114010). As an example, 10 nm of titanium was deposited onto the silicon substrate due to its high reactivity and chemical bonding with silicon, 40 nm of gold was deposited on top of titanium with nanostructures in an Angstrom Nexdep Electron Beam Evaporator at the deposition rate of 0.5 A / s for titanium and 0.2 A / s for gold, then the sample was submersed in acetone for 24 hours to remove the PM MA A3; lift-off the gold and titanium.

[0099] When the analyte of interest includes material such as protein / DNA / RNA biomarkers for disease and residing within a complex biological fluid, Coulomb forces became a major consideration vis-a-vis their adsorption characteristics on the SERS substrate. Herein, we investigated a method of surface charge reversal via self-assembled monolayers (SAMs) which allowed for modest detection of biomarkers, involving three standard steps: (i) coating a SERS substrate with a thiol polyethylene glycol) and biotinylated polyethylene glycol) thiol; (ii) coating streptavidin on the assembled biotinylated poly(ethylene glycol) thiol; and (iii) conjugating the coated streptavidin layer with molecule of synthetic biotinylated antibody of sTREM-1 , Ang-1 or Ang-2 (see Figures 1 B, 1C, and 1D, respectively). The first step created a linker on the plasmonic surface of SERS substrates, for example, gold films are produced by physical deposition techniques such as evaporation, meanwhile, element comprises of biotinylated polyethylene glycol chains grafted on the gold surface via thiol-Au chemistry.

[0100] Due to the large structure of Streptavidin (see Table 3) in the following step, it is necessary to use a blocking thiol to block the open sites of gold surface in the spaces between the biotinylated alkane thiol molecules. This prevents the nonspecific binding of other molecules onto the surface. The second step can ensure the immobilization of antibody with high specificity to target antigen, which is based on streptavidin / biotin assemblies with high stability of molecular assembly (Kd = 10'15) (Ref. 24: Ngo et al 2013, Journal of Colloid and Interface Science, 409, 59-65). The strong affinity of Streptavidinbiotin is based on the intra- and intermolecular interactions between tryptophan (Trp) residues and the non-polar side chain of streptavidin with the non-polar moieties of biotin. The third step ensured appropriate surface interaction is to imply antibodies specific to the protein of interest (Ref. 25: Ko et al 2021 , Nanotechnology, 32(50), 505207).

[0101] Chemical Size / molecular weight

[0102] Thiol polyethylene glycol) 376.53 Da

[0103] Biotinylated polyethylene glycol) thiol 5 kDa

[0104] Streptavidin 75 kDa

[0105] Human Ang-1 Biotinylated Antibody 56 kDa

[0106] Human Ang-1 56 kDa

[0107] Human Ang-2 Biotinylated Antibody 56 kDa

[0108] Human Ang-2 56 kDa

[0109] Human sTREM-1 Biotinylated Antibody 21 kDa

[0110] Human sTREM-1 21 kDa

[0111] Table 3 Preparation of SERS substrate for sTREM-1 Capture

[0112] The recipe to prepare the SERS substrate for sTREM-1 capture is as follows and illustrated in Figure 1 B. After completion of steps (i) to (v) in Figure 1A, the dried SERS substrate was immersed in 5x1 O'4M immersed thiol polyethylene glycol) (Thiol-PEG) and biotinylated PEG thiol (BAT) for 16 h with 9:1 ratio. After drying in air, 5x1 O'7M streptavidin (SA) solution that is prepared with Dulbecco's Phosphate Buffered Saline (DPBS) for 2 h. After rinsing with DPBS buffer (1X), the substrate was dipped into a triggering receptor expressed on myeloid cells- 1 (TREM-1) Biotinylated Antibody for 30 min. Then, the substrate was air dried (room temperature and 46% relative humidity). Drying of the samples was performed between each step at room temperature and 46% humidity, and all reactions were performed at standard room temperature and pressure.

[0113] Preparation of SERS substrate for Ang-1 Capture

[0114] The recipe to prepare the SERS substrate for Ang-1 capture is as follows and illustrated in Figure 1C. After completion of steps (i) to (v) in Figure 1A, the dried SERS substrate was immersed in 5x1 O'4M immersed thiol polyethylene glycol) (Thiol-PEG) and biotinylated PEG thiol (BAT) for 16 h with 9:1 ratio. After drying in air, 5x1 O'7M streptavidin (SA) solution that is prepared with Dulbecco's Phosphate Buffered Saline (DPBS) for 2 h. After rinsing with DPBS buffer (1X), the substrate was dipped into an Angiopoietin-1 (Ang-1 ) Biotinylated Antibody for 30 min. Then, the substrate was air dried (room temperature and 46% relative humidity). Drying of the samples was performed between each step at room temperature and 46% humidity, and all reactions were performed at standard room temperature and pressure. Preparation of SERS substrate for Ang-2 Capture

[0115] The recipe to prepare the SERS substrate for Ang-2 capture is as follows and illustrated in Figure 1D After completion of steps (i) to (v) in Figure 1A, the dried SERS substrate was immersed in 5x1 O'4M immersed thiol polyethylene glycol) (Thiol-PEG) and biotinylated PEG thiol (BAT) for 16 h with 9:1 ratio. After drying in air, 5x1 O'7M streptavidin (SA) solution that is prepared with Dulbecco's Phosphate Buffered Saline (DPBS) for 2 h. After rinsing with DPBS buffer (1X), the substrate was dipped into an Angiopoietin-2 Antibody for 30 min. Then, the substrate was air dried (room temperature and 46% relative humidity). Drying of the samples was performed between each step at room temperature and 46% humidity, and all reactions were performed at standard room temperature and pressure.

[0116] Finally, with the functionalized SERS substrates prepared as above, 10 / L Ang-1 , Ang-2, sTREM-1 solution (in PBS buffer or human serum) of different concentrations was dropped on the substrates. The Raman spectra were collected immediately after dropping for 10 times until surface was completely dried. The Raman spectra were recorded from Raman - Renishaw InVia with excitation wavelength at 785 nm. In both measurements, the laser beam was focused on a spot of ~2 urn diameter by microscope objective with a magnification of 20x with 0.18 mW urn'2. All spectra were collected at 60 s and calibrated with respect to the Raman peak of Si wafer at 520 cm'1. For the spot- to-spot Raman measurements, 3-4 points were set and repeated three times in different places with a wavelength of 200 cm-1to 1800 cm-1.

[0117] The near-infrared laser source is preferable in general bimolecular SERS detection because; (a) it can avoid the excitation of fluorescence from biomolecules; (b) it has a deeper penetration depth in biological material, and (c) low photon energy of near infrared laser minimizes photo thermal damage to biomolecules.

[0118] It is noted that in Figures 1B, 1C and 1 D, the thiol polyethylene glycol) (Thiol-PEG) blocker molecules and the biotinylated PEG thiol (BAT) with the biotin are bound in their own separate troughs which can happen naturally. However, it is noted that these molecules e) and d) can also be in the same trough.

[0119] Examples

[0120] The present process for SERS detection of Strem-1 , Ang-1 and Ang-2 will be illustrated using the following non-limiting examples presented below. Example 1 : SERS Spectra Analysis for sTREM-1

[0121] Referring to Figure 3, the SERS spectra for sTREM-1 (spectra (a)) and its antibody (spectra (b)) were collected at a concentration of 0.976 pM, and characteristic peaks appeared at 376.659 cm'1, 434.949 cm'1and 456.34 cm'1, which were absent from any of the previous steps. Those peaks appeared consistently even at low concentrations of sTREM-1 , eventually merging at concentrations below 0.00238 pM.

[0122] Example 2: SERS Spectra Analysis for Ang-1

[0123] Referring to Figure 4, the SERS spectra for Ang-1 (spectra (a)) and its antibody (spectra (b)) were collected at a concentration of 0.8 pM, and characteristic peaks appeared at 329.338 cm'1, 376.659 cm'1, 433.688 cm'1, and 455.084 cm'1, which were absent from any of the previous steps. Those peaks appeared consistently even at low concentrations of Ang-1 , eventually merging at concentrations below 0.01 pM.

[0124] Example 3: SERS Spectra Analysis for Ang-2

[0125] Referring to Figure 5, the SERS spectra for Ang-2 (spectra (a)) and its antibody (spectra (b)) were collected at a concentration of 1 pM, and characteristic peaks appeared at 402.076 cm-1and 437.47 cm-1, which were absent from any of the previous steps. Those peaks appeared consistently even at low concentrations of Ang-2, eventually merging at the concentrations below 0.1 pM.

[0126] Example 4: Biomarker Concentration vs. Characteristic Raman Peak Intensities

[0127] Figures 6a to 6c depict the correlation between each biomarker concentration and its characteristic Raman peak intensity. The band intensities were proportional to the concentration of biomarkers and a linear concentration dependence was observed. For sTREM-1 (Figure 6a), the linear concentration relationship is y = 0.9935x + 677.35 (R2= 0.94) in the range of 2.44 x 10'5- 0.0624 pM (equal to the range of 0.513 - 1310 pg / mL), where the range is completely at physiological level. For Ang-1 (Figure 6b), the linear concentration relationship is y = 693.12x + 141.42 (R2= 0.92) in the range of 0.1 -0.8 pM (equal to the range of 5600 - 44800 pg / mL) and for Ang-2 (Figure 6c), the linear concentration relationship is y = 259.66x +423.15 (R2= 0.94) in the range of 0.05 - 1 pM (equal to the range of 2800 - 56000 pg / mL).

[0128] Techniques for single molecule level detection and recognition of specific biomolecules are critical for early detection of sepsis. Optical spectroscopies are of foremost interest owing to their non-destructive nature, especially SERS for its added sensitivity and specificity. Under optical excitations of proper frequency, a metallic nanostructure / nanoparticle sustains surface plasmon resonance that results in highly enhanced local electromagnetic fields and distinct spectral extinction characteristics. Figure 7 shows unique signatures of sTREM-1 in a gold nanoparticle solution (spectra (a), where three new peaks appear in the range of 360 cm-1to 370 cm-1compared to two peaks around 450 cm'1of the pure gold nanoparticles in a buffer solution (spectra (b).

[0129] Here, gold nanoparticles (1-100 nm diameter size) were mixed with same amount (ratio = 1 : 1 ) of biomarker protein solution (for example, sTREM- 1 ), allowing for the characterization of the analyte molecule without digestion or extraction. The molecule could be differentiated based on the Raman marker bands, and qualitative detection is (can be) achieved.

[0130] SERS System

[0131] Figure 8 is a schematic of an exemplary SERS system comprised of an enclosed housing 30 within which is contained a laser source 32, an objective lens 34, a SERS substrate 38, a translation stage 40, on which the SERS substrate 38 is mounted, and a detector 46. The substrate includes the SERS active area with the sample on it. The substrate could include, for example, the microfluidic circuit shown in Figure 9 discussed below. Incident light 36 from the laser source 32 passes through the objective lens 34, striking the SERS substrate 38, atop of which is the sample. The scattered light 42 from the sample / substrate is collected at the detector 46, having passed through two filters 44 with the first filter being a notch filter to remove the elastic component of scattered light, and the second filter being a short / long pass filter for passing the anti-Stokes / Stokes radiation. Spectra are recorded and processed by computer 50, which also controls the laser source 32, detector 46 and translation stage 40 via electrical cables 48.

[0132] Biomarkers of Interest in Human Bodily Fluids Introduced Through Microfluidics

[0133] In some embodiments, methods are provided for sensing specific biomarkers for sepsis, such as sTREM-1 , Ang-1 , or Ang-2, from one or more human bodily fluids (e.g., blood, plasma and serum, saliva, and / or urine) (Ref. 26:Su et al., 2011 , Crit Care. 2011 ; 15(5): R250.; Ref. 27:Konvalinka et al., Clin Proteomics. 2016; 13: 16; Ref. 28:Jia et al., Open Life Sci. 2024; 19(1): 20220812). As used herein, the term 'bodily fluids' generally refers to naturally expressed bodily fluids, although some embodiments may include fluids collected intravenously (e.g., blood, plasma and serum). In the context of the disclosed embodiments, 'bodily fluids' refer to naturally expressed fluids such as saliva and urine.

[0134] A sample containing a bodily fluid is treated in a solution before it is measured. The solution ensures that biomarkers within the bodily fluid retain their antigenicity and cellular architecture. The solution can be buffered at a pH ranging from about 6.4 to 8.4, with a preferred range of about 7.2 to 7.6, and may include a blocking agent such as, but not limited to, Bovine Serum Albumin (BSA). The buffer may consist of phosphate-buffered saline (PBS) or Trisbuffered saline (TBS), and in some embodiments, may also contain fetal bovine serum. The treated sample is introduced to the microfluidic system such as that shown in Figure 9 for measurement.

[0135] Figure 9 shows a schematic of sample delivery via a microfluidic circuit 10 to SERS-active regions on the substrate, with each region functionalized for one specific antigen. Microfluidic circuit 10 when combined with the SERS system of Figure 8, is essentially the SERS substrate 38 shown in Figure 8. Microfluidic circuit 10 includes a sample inlet 12, microfluidic channels 14, an array of target structures 16 functionalized for capture of one specific antigen, e.g., sTREM-1 , an array of target structures 18 functionalized for capture of another specific antigen, e.g., Ang-1 , and array of target structures 20 functionalized for capture of another specific antigen, e.g., Ang-2. Microfluidic circuit 10 includes a sample collector 24, and substrate 26. The treated sample is introduced to the sample inlet 12 by, for example, pipette or pump, whereupon it travels under external force, for example pump, or on-chip force, for example, electrophoresis along the microfluidic channels 14, to the array of target structures 16, 18, 20 where some of the sample is retained due to surface friction and the remainder travels on to the sample collector 24 via the microfluidic channels 14. Each array of target structures 16, 18, 20 containing retained sample is measured separately using, for example, the techniques described above to acquire the Raman spectra of the sample.

[0136] In an embodiment, the SERS data collected as described above is passed into a machine learning classification pipeline to distinguish the biomarkers of interest from other substances, as illustrated in Figure 11 A. Specifically, these pipelines are trained to distinguish between Ang-1 , Ang-2, sTREM-1 and their respective antibodies or other materials such as thiol polyethylene glycol) / biotinylated polyethylene glycol) thiol, and streptavidin. Pipelines can be developed to classify spectral data between all categories or any subset. For example, one type of embodiment could be a classification pipeline for distinguishing SERS spectra belonging to Ang-1 from those belonging to Ang-2.

[0137] Figure 11 A shows the machine learning pipeline for classifying the sepsis biomarkers from SERS data. This pipeline would be used to detect the presence of sepsis biomarkers in a sample. As shown in Figure 11 A, the classification pipeline consists of data preprocessing steps to ensure good quality data, feature selection or dimensionality reduction techniques to reduce complexity and cost, and the classifier module itself. First, raw SERS data collected as described above is preprocessed by subtracting the minimum value of the spectra, then subtracting the baseline spectra as estimated using algorithms such as asymmetric least squares, then normalizing the values of the spectra using the minimum and maximum values, then carrying out a feature selection technique to reduce the dimensionality of the data for machine learning.

[0138] Optionally, the pipeline can also include outlier rejection or other robustness techniques applied to the dataset to remove biased or unexpected datapoints, to improve the pipeline performance. Once the dataset has been preprocessed, it is then split into dedicated training, testing, and validation sets as required. Training and validation sets are used to train and evaluate the performance of a range of classification models to determine the model most suited for the task with regards to classification accuracy, sensitivity, and selectivity. After a model has been chosen, the training and validation sets are then used to train and optimize the model's parameters and hyperparameters to yield the highest accuracy, sensitivity, and specificity. As a final measure of model performance, the test set is fed into the model and the accuracy, sensitivity, and specificity is evaluated. If model performance is acceptable, the trained and optimized model is deployed for use on new SERS data obtained in real-time to classify which biomarker is in the sample. These steps are described in more detail in the following sections.

[0139] In general, preprocessing of SERS data is done by applying numerical methods for the removal of the baseline, background, cosmic rays, and noise before normalization (Ref. 29:Wahl et al 2020, Applied Spectroscopy, 74(4), 427-438). Minimum subtraction is done to centre the spectra about zero. Baseline subtraction can be carried out using the Asymmetric Least Squares algorithm, to remove background signals that are picked up by the machine. These steps can minimize background radiation, and fluorescence. Background radiation can be an effect of light sources that are difficult to shield or SERS from materials surrounding the sample (ex. PBS buffer or the microscope slide), and fluorescence is the result of molecules being excited to states of higher energy and emitting this energy as light.

[0140] The importance of preprocessing should not be understated, as faulty preprocessing may lead to erroneous conclusions (e.g., achieve smaller absolute error, better quality, smaller signal-to-noise) (Ref. 30: Oliveri et al 2019, Analytica Chimica Acta, 1058, 9-17). Finally, normalization can be performed by dividing the spectrum values by their norm, to ensure that data from various experiments can be compared (Ref. 31 : Palacky et al 2021 , Journal of Raman Spectroscopy, 42(7), 1528-1539).

[0141] As machine learning is a data-driven technique, it requires large amounts of good quality, labelled data to train. To ensure a large set of data, the cleaned spectra can be augmented to generate additional spectra samples. A simple technique for generating new spectra is to add Gaussian noise to existing spectra. Linear addition or subtraction of existing spectra can also yield new spectra.

[0142] Further, as large amounts of spectral data are required to train, feature selection or engineering methods can be applied to extract the most informative wavenumbers from the dataset for classification purposes. Popular methods for feature selection are dimensionality reduction techniques such as Principal Component Analysis or Linear Discriminant Analysis, or statistical techniques such as Analysis of Variance. By using only the most informative spectral data values, the time and computational complexity of the models can be reduced for application purposes.

[0143] Various machine learning classification models can be trained on processed spectral data. Popular models include Decision Trees, Support Vector Machines, K Nearest Neighbours, Logistic Regression, or Neural Network models. Using training and validation sets for training and evaluation of model performance, the best model from available options can be chosen for further optimization and eventual deployment.

[0144] Examples

[0145] The present embodiment is described by means of the examples presented below.

[0146] Example 1 : Classification of Ang-1, Ang-2, and Their Antibodies Using a Multilayer Perceptron

[0147] SERS spectra for Ang-1 (# samples = 20), Ang-2 (20), Ang-1 antibody (20) and Ang-2 antibody (17) were used to train a multilayer perceptron for the task of classifying spectra as one of these four categories, following the pipeline illustrated in Figure 11 A. Principal component analysis was used to obtain 6 features. A train / test split of 66% / 33% was employed. Figure 10 shows a confusion matrix displaying the true labels versus the predictions made on the test set by a multilayer perceptron classification model when distinguishing between the SERS spectra of Ang-1 , Ang-2, and their respective antibodies. Figure 10 illustrates the performance of the model on the test set, yielding an accuracy of 92.3%.

[0148] The SERS spectra regression pipeline consists of all steps mentioned in the classification pipeline of Figure 11 A, excluding the classifier. Further, depending on the model utilized as the regressor, the regression pipeline may include steps to transform the data into a different feature space. For example, when polynomial regression is utilized, transformation of the spectral features into polynomial features is necessary. Additionally, outlier rejection or other robustness techniques, such as Random Sample Consensus or the use of robust loss functions, can be used to improve the performance of the model by weighting potential outlier datapoints accordingly when performing the regression. These techniques allow for a more robust and accurate model. Regression models are generally evaluated using root mean squared error (RMSE) and R2goodness-of-fit metrics. Finally, various regression models can be employed for this pipeline, such as linear, polynomial, or Bayesian regressions or neural network models.

[0149] Figures 11B, 11C and 11D show the process flowcharts for the machine learning regression pipelines 1 , 2, and 3 respectively.

[0150] In an embodiment, the data obtained from SERS is passed into a machine learning regression pipeline to predict the percent, absolute, or relative concentration of Ang-1 , Ang-2, or sTREM-1 . In one pipeline, as shown in Figure 11 B, the dataset contains the spectral data of only one of the three biomarkers, in varying concentrations. In this case, the aim of the pipeline is to predict the concentration of that particular biomarker from the SERS spectrum. Here, all preprocessing, dataset splitting, and model development steps are similar to those described for Figure 11 A. However, the nature of the data used, the types of models tested for development, the evaluation of the models, and the application of the final model differ. In this pipeline, the dataset comprises of raw SERS data collected from a single biomarker at a range of concentrations such that the trained and optimized model can predict the concentration of that biomarker when fed new SERS data from a sample containing only that biomarker. In addition, as this is a regression task, the models used for development steps differ to those of the classification pipeline. In this pipeline, the regression models can include linear regressions or neural networks, as shown in Figure 11B.

[0151] Performance of regression models is evaluated using root mean squared error (RMSE) and R2goodness-of-fit metric. The application of this pipeline is to develop a single stand-alone module for predicting the concentration of an individual biomarker of interest in a sample containing only that biomarker.

[0152] In another pipeline, as shown in Figure 11C, the overall dataset contains data from two or more biomarkers, yet each spectrum itself contains the spectral data of only one biomarker. Again, the aim is to predict the concentration of one particular biomarker. Here, all preprocessing, dataset splitting, and model development steps are similar to those described for Figure 11 A. The types of regression models used in this pipeline and their evaluations are similar to those described for Figure 11B. However, the nature of the data used, and the application of the final model differ. In this pipeline, the dataset comprises of raw SERS data collected from any two or more of the biomarkers, with each individual measurement carried out on a sample containing strictly one of the biomarkers. The final model is then used on new SERS data from a sample containing an unknown, single biomarker to determine the biomarker and its concentration. The application of this pipeline is to develop a module that can determine both the biomarker and the concentration from a sample containing only one biomarker.

[0153] In another pipeline, as shown in Figure 11 D, the dataset contains data obtained from mixtures of two or more biomarkers, meaning each spectrum may have spectral data from multiple biomarkers. In this case, the aim of the regression pipeline is to predict the concentrations of all biomarkers present within the spectrum. Here, all preprocessing, dataset splitting, and model development steps are similar to those described for Figure 11 A. The types of regression models used in this pipeline and their evaluations are similar to those described for Figure 11 B. However, the nature of the data used, and the application of the final model differ. In this pipeline, the dataset comprises of raw SERS data collected from any two or more of the biomarkers, with each individual measurement carried out on a sample containing one or more biomarkers. The final model is then used on new SERS data from a sample containing an unknown number of biomarkers, to determine the biomarkers present and their respective concentrations. The application of this pipeline is to develop a comprehensive module that can analyze a sample with many potential biomarkers present and detect the existing biomarkers and their concentrations, which is most relevant for clinical blood samples.

[0154] Examples

[0155] The present embodiment is described by means of the examples presented below.

[0156] Example 1 : Concentration Prediction of Ang-1 using 2ndDegree Polynomial Regression

[0157] SERS spectra for Ang-1 at concentration molarities of 0.01 pM (# samples = 12), 0.1 pM (10), 0.2 pM (10), 0.5 pM (10), and 0.8 pM (9) were used to train a multivariate polynomial regression model of degree 2, for the task of determining the concentration of Ang-1 , following the pipeline presented in Figure 11B. The data was preprocessed via minimum subtraction, baseline subtraction, normalization, and polynomial data transformation. A train / test split of 66% / 33% was employed. The test set yielded an R2of 0.96 and a root mean squared error of 0.043. Table 4 illustrates sample predictions made by the pipeline from Figure 11 B on the test set.

[0158] Table 4

[0159] Example 2: Concentration Prediction of Ang-2 using 2ndDegree Polynomial Regression SERS spectra for Ang-2 at concentration molarities of 0.1 pM (# samples

[0160] = 11 ), 0.3 pM (9), 0.6 pM (6), and 0.8 pM (5) were used to train a multivariate polynomial regression model of degree 2, for the task of determining the concentration of Ang-2, following the pipeline presented in Figure 11B. The data was preprocessed via minimum subtraction, baseline subtraction, normalization, and polynomial data transformation. A train / test split of 90% / 10% was employed. The test set yielded an R2of 0.96 and a root mean squared error of 0.052. Table 4 illustrates sample predictions made by the pipeline from Figure 11 B on the test set.

[0161] Example 3: Concentration Prediction of sTREM-1 Using Linear Regression

[0162] SERS spectra for sTREM-1 at concentration molarities of 0.00048 pM (# samples = 10), 0.00238 pM (6), 0.0244 pM (14), 0.244 pM (15), and 0.976 pM (10) were used to train a multivariate linear regression model with Huber loss, for the task of determining the concentration of sTREM-1 , following the pipeline presented in Figure 11 B. The data was preprocessed via minimum subtraction, baseline subtraction, and normalization. A train / test split of 85% / 15% was employed. The test set yielded an R2of 0.88 and a root mean squared error of 0.102. Table 4 illustrates sample predictions made by the pipeline from Figure 11 B on the test set.

[0163] Example 4: Concentration Prediction of Ang-1 and Ang-2 Mixed Spectra Using a Convolutional Neural Network

[0164] SERS spectra for Ang-1 , Ang-2, and Ang-1 / Ang-2 mixtures at the following concentration molarities were used to train a convolutional neural network for the task of determining the concentration of both biomarkers within a mixture, following the pipeline presented in Figure 11D:

[0165] •Ang-1 : 0.01 pM (# samples = 12), 0.1 pM (10), 0.2 pM (10), and 0.5 pM (10) •Ang-2: 0.005 pM (# samples = 10), 0.01 pM (10), 0.05 pM (12), 0.1 pM (11), and 1 pM (13)

[0166] •Ang-1 / Ang-2: 0.01 pM 10.02 pM (10), 0.01 pM 10.1 pM (10), 0.01 pM 10.2 pM (10), and 0.01 pM I 0.5 pM (5)

[0167] The data was preprocessed via minimum subtraction, baseline subtraction, and normalization. A train / validation / test split of 68% / 22% / 10% was employed. The architecture consisted of two convolutional layers and four densely connected layers, all with the ReLU activation functions. The test set yielded a root mean squared error of 0.0217. Table 5 illustrates sample predictions made by the pipeline from Figure 11D on the test set.

[0168] Table 5

[0169] The calculations on protein markers (for example, sTREM-1) for diagnosis and prognosis of early onset sepsis have been reported (Ref. 32: Pontrelli et al 2016, Ital J Pediatr, 42(1 ), 44; Ref. 33: Chang et al 2022,

[0170] Front. Pediatr., 10, 929665; Ref. 34: Barber et al 2023, Current Opinion in Lipidology, (34)2, 70-81 ), however, the literatures tend to report a cut-off concentration related to certain chosen x-day mortality rate, where x could be 2- days, 7-days, 28-days, etc. In one embodiment, based on the probability of 7-day mortality as a function of sTREM-1 concentration (Ref. 35: A. Leligdowicz et al 2021 , Nat Commun, 12(1 ), 6832), a curve fit had been made using a logistic function (see Figure 12). In fitting the logistic function to that data set, an extrapolated curve from 38% to at least to 50%probability of 7-day mortality was assumed in order for the function to capture the sigmoid shape of the curve. The extrapolated curve was used to get a data point at 50% (which gives the centre of the x-axis data, x0).

[0171] The logistic function applied was:

[0172] Parameter values for the standard logistic function (Eqn. (3)) are:

[0173] Parameter (Eqn (3)) Value Standard error

[0174] A 60.1236 4.60696

[0175] Xo 1 ,117.24 73.1851 k 3.37299x1 O'33.704x1 O’4

[0176] The curve fit to experimental data provides a reference against which Raman measurements can be compared. Measurements described in Example 1 : SERS Spectra Analysis for sTREM-1 demonstrate that the characteristic peaks of sTREM-1 at sub-pM concentrations were detected and that a linear concentration dependence of characteristic band intensity was determined. This coincides with the lower concentration portion of the curve fit, providing a means to determine early onset of sepsis, to assess the degree of sepsis severity and to provide guidance on subsequent treatment protocol. In an embodiment, findings from the machine learning classification pipeline are correlated with the deconvolution or decomposition of the SERS spectra, to identify characteristic peaks of biomarkers or other subjects of interest, develop models, and provide model explainability, as described in Figures 13A and 13B

[0177] Figure 13A shows the process of correlating findings from the machine learning classification pipeline from Figure 11A to a decomposition or deconvolution of the spectra for a biomarker. The raw SERS data from a single biomarker is preprocessed up until the normalization step as described for Figure 11 A. Then, spectra are deconvolved or decomposed into a series of underlying functions. Alongside this, significant wavenumbers are determined from the machine learning classification pipeline (as shown in Figure 11 A). The properties from the deconvolution or decomposition are then compared to these significant wavenumbers to find spectral features of interest (e.g. peaks) and can be further related to the physical and molecular structure of the biomarker of interest.

[0178] Figure 13B shows the process of using the results of a spectral deconvolution or decomposition for a biomarker to inform the development of a machine learning classification model. The only difference in the process compared to Figure 11A is that the feature selection preprocessing step uses the results of the deconvolution or decomposition to inform which wavelengths or other significant spectral features to use as input to the classification model. Specifically, the SERS spectra are deconvolved or decomposed into a series of underlying functions, such as Gaussian or Lorentzian functions. Each underlying function contributes to a peak on the spectrum. The location, width, and other properties of this peak can be correlated to findings from machine learning or similar analysis to provide physical explanations. For example, significant wavenumbers found from analysis of variance (ANOVA) in datasets which contain the spectral data of two or more biomarkers can be compared to the wavenumbers at which peaks are found when doing a Lorentzian deconvolution. Wavenumbers that are considered significant by both ANOVA and spectral deconvolution may be considered characteristic peaks for a biomarker.

[0179] These characteristic peaks or peaks of interest are then related to known vibrational modes of molecular structures stored in a Raman database or library. For example, proteins which contain a S-S bridge in their structure yield peaks in the 500-550cm-1range due to the stretching of this bond under Raman stimulation. Relating the found peaks to the physical phenomenon underlying the SERS measurement and the structure of the target molecule provides real- world explanations for the classification models. This process is illustrated in Figure 13A and is further described in Example 1 below. In doing so, the machine learning classification models benefit from greater explainability and transparency. Further, the inverse can be done, wherein the peaks found from deconvolution or decomposition can be utilized as primary features in a machine learning model, as a form of dimensionality reduction or feature extraction. This is illustrated in Figure 13B.

[0180] Examples Example 1 : Correlation of ANOVA Wavenumbers to sTREM Spectral Deconvolution to Determine Characteristic Peaks

[0181] The process shown in Figure 13A was used to correlate ANOVA wavenumbers from a machine learning classification model to the spectral deconvolution of sTREM to determine potential characteristic peaks. A Lorentzian deconvolution of the SERS spectra for sTREM was performed to obtain wavenumber locations of significant peaks. SERS spectra for Ang-1 (# samples = 30), Ang-2 (30), sTREM (30), and healthy blood (20) were used to train 3 binary classification models: sTREM versus Ang-1 , sTREM versus Ang- 2, and sTREM versus healthy blood. Multilayer perceptrons were used for all classification models. The data was preprocessed via minimum subtraction, baseline subtraction, and normalization. A train / test split of 70% / 30% was employed. The top 10 most significant wavenumbers were extracted from ANOVA. All models performed with 100% accuracy on the test set.

[0182] The ANOVA wavenumbers used in these three models were directly compared to the spectral peaks obtained from the sTREM deconvolution. Wavenumbers which appeared to be significant in multiple classification models and simultaneously appeared as peaks in the deconvolution were determined to be potential characteristic peaks of sTREM. Multiple wavenumbers that are close in value and occur in multiple classification models and the deconvolution peaks may also be considered of interest. Peaks of interest occurred in the range of 1085-1098 cm-1, which could be correlated with the stretching vibration of C-C or C-0 bonds.

[0183] Example 2: Correlation of PCA Wavenumbers to sTREM Spectral Deconvolution to Determine Characteristic Peaks The process shown in Figure 13B was used to correlate PCA wavenumbers from a machine learning classification model to the spectral deconvolution of sTREM to determine potential characteristic peaks. A Lorentzian deconvolution of the SERS spectra for sTREM was performed to obtain wavenumber locations of significant peaks. SERS spectra for Ang-1 (# samples = 30), Ang-2 (30), sTREM (30), and healthy blood (20) were used to train 3 binary classification models: sTREM versus Ang-1 , sTREM versus Ang- 2, and sTREM versus healthy blood. Multilayer perceptrons were used for all classification models. The data was preprocessed via minimum subtraction, baseline subtraction, and normalization. A train / test split of 70% / 30% was employed. PCA was employed to reduce the spectra to 5 components. All models performed with 100% accuracy on the test set.

[0184] The wavenumber loadings for each PCA component were calculated. Wavenumbers with high loading values were compared to the spectral peaks obtained from the sTREM deconvolution. Wavenumbers which appeared to be significant in multiple classification models and simultaneously appeared as peaks in the deconvolution were determined to be potential characteristic peaks of sTREM. Multiple wavenumbers that are close in value and occur in multiple classification models and the deconvolution peaks may also be considered of interest. Peaks of interest occurred in the range of 1087-1098 cm'1, which could be correlated with the stretching vibration of C-C or C-0 bonds.

[0185] In embodiments, the concentrations of biomarkers are determined using machine learning classification pipelines, which are trained to distinguish between any combination of sTREM-1 , Ang-1 , Ang-2. the classification pipelines are developed specifically for spectral data and comprise: a) spectrum preprocessing, including but not limited to, minimum subtraction, baseline subtraction, and normalization; b) data augmentation algorithms used to increase number of spectra samples, including but not limited to, linear addition and subtraction of existing spectra and addition of Gaussian noise; c) feature selection or engineering methods applied to spectral data, including but not limited to, dimensionality reduction techniques and analysis of variance feature selection; and d) a machine learning classifier.

[0186] The machine learning classifier can be embodied by models including, but not limited to, Decision Trees, Support Vector Machines, K Nearest Neighbours, Logistic Regression, Deep Neural Networks, or Convolutional Neural Networks.

[0187] The measurement of biomarkers within a spectrum is facilitated via machine learning regression pipelines which are trained to predict the concentration of sTREM-1 , Ang-1 , or Ang-2. The regression pipeline predicts the concentration of sTREM-1 , Ang-1 , or Ang-2 from a spectrum containing only one biomarker. The regression pipeline predicts the concentration of any combination of sTREM-1 , Ang-1 , and Ang-2 from a spectrum containing any mixture of these biomarkers. In addition, the regression pipeline predicts the concentration of Ang-1 and subsequently predicts the concentration of Ang-2 using the biomarker relationship, and vice versa.

[0188] In an embodiment, the regression pipeline is developed specifically for spectral data, comprising: 1) Spectrum preprocessing, including but not limited to, minimum subtraction, baseline subtraction, and normalization; and

[0189] 2) Data augmentation algorithms used to increase number of spectra samples, including but not limited to, linear addition and subtraction of existing spectra and addition of Gaussian noise; and

[0190] 3) Feature selection or engineering methods applied to SERS spectra data, including but not limited to, dimensionality reduction techniques and analysis of variance feature selection; and

[0191] 4) Feature and data transformation methods for polynomial fitting; and

[0192] 5) Outlier rejection or robustness techniques, including but not limited to, Random Sample Consensus or the Huber loss function; and

[0193] 6) A machine learning regressor.

[0194] The machine learning regressor can be embodied by models including but not limited to Linear Regression, Generalized Linear Models, Polynomial Regression, Bayesian Regression, Neural Networks, or Convolutional Neural Networks.

[0195] The methodology is developed for correlating findings from a machine learning classification pipeline to spectral deconvolutions or decompositions, comprising:

[0196] 1) Spectral deconvolution or decomposition into underlying Gaussian or Lorentzian functions; and 2) Extraction of significant peaks and relevant peak properties from the deconvolution or decomposition; and

[0197] 3) A machine learning classification pipeline as described in claim 13, with feature selection or engineering functionalities to identify significant wavenumbers; and

[0198] 4) Comparison of significant wavenumbers from the machine learning classification model to extracted peaks and their properties from the deconvolution or decomposition.

[0199] The correlation is done between the spectral deconvolution or decomposition of Ang-1 , Ang-2, or sTREM and one or more machine learning classification pipelines including said biomarker for the purpose of identifying potential characteristic peaks.

[0200] The correlation is used to provide physical explainability, such as vibrational modes of the molecular structure of Ang-1 , Ang-2, or sTREM, to the machine learning classifier.

[0201] Findings from deconvolution or decomposition, or identified characteristic peaks, are used as primary features for the development of a machine learning classifier to distinguish between any combination of Ang-1 , Ang-2, and sTREM.

[0202] In an embodiment, separate SERS active substrates functionalized for detection of each of sTREM-1 , Ang-1 and Ang-2 are integrated into the system, and wherein the concentrations of biomarkers are determined using machine learning classification and regression pipelines, which are trained to distinguish between any combination of sTREM-1 , Ang-1 , Ang-2 and predict their concentrations.

[0203] In an embodiment the classification and regression pipelines are configured specifically for spectral data and comprise: a) spectrum preprocessing, to remove or minimize background signal from radiation, noise, and / or SERS responses from other material, clean and validate the data to remove outliers or erroneous measurements, and standardize the data such that data from independent experiments can be compared; b) data augmentation algorithms to artificially generate new data samples from the existing data collected from real-world samples, to increase the number of diverse data samples used for training, validation, development, and / or testing of machine learning methods and thereby improve the performance of these methods; c) feature selection to extract a base set of the most informative features for biomarker discrimination from the dense spectral data collected, thereby reducing the dimensionality of the input data used by the machine learning methods and improving their computational time and space costs / complexity, and real-world generalizability; d) feature and data transformation methods for polynomial fitting or to place the spectra data into a desired feature space to improve machine learning performance and / or costs; e) outlier rejection or robustness techniques to mitigate and / or eliminate biased or erroneous data points or features within the data, to improve the robustness, accuracy, and reliability of the machine learning methods; and f) a machine learning classifier and / or regressor. In an embodiment the regression pipeline predicts the concentration of sTREM-1 , Ang-1 , or Ang-2 from a spectrum containing only one biomarker.

[0204] In an embodiment the regression pipeline predicts the concentration of any combination of sTREM-1 , Ang-1 , and Ang-2 from a spectrum containing any mixture of these biomarkers.

[0205] In an embodiment the regression pipeline predicts the concentration of Ang-1 and subsequently predicts the concentration of Ang-2 using the biomarker relationship, and vice versa.

[0206] In an embodiment a methodology is developed for correlating findings from a machine learning classification pipeline to spectral deconvolutions or decompositions, comprising: a) spectral deconvolution or decomposition into underlying Gaussian or Lorentzian functions; b) extraction of significant peaks and relevant peak properties from the deconvolution or decomposition; c) a machine learning classification pipeline, with feature selection or engineering functionalities to identify significant wavenumbers; d) comparison of significant wavenumbers from the machine learning classification model to extracted peaks and their properties from the deconvolution or decomposition.

[0207] The correlation is done between the spectral deconvolution or decomposition of Ang-1 , Ang-2, or sTREM and one or more machine learning classification pipelines including said biomarker for the purpose of identifying potential characteristic peaks. The correlation is used to provide physical explainability, such as vibrational modes of the molecular structure of Ang-1 , Ang-2, or sTREM, to the machine learning classifier.

[0208] The findings from deconvolution or decomposition, or identified characteristic peaks, are used as primary features for the development of a machine learning classifier to distinguish between any combination of Ang-1 , Ang-2, and sTREM.

[0209] Technical challenges encountered and overcome Fabrication Process

[0210] In Electron Beam Lithography (EBL), slight instability in beam exposure power can lead to defects in the patterns. In Reactive Ion Etching (RIE), chamber contamination can cause over-etching, potentially leading to the collapse of the patterns. Additionally, during e-beam deposition, the resulting layer may be several nanometers thinner on the vertical side walls compared to the top and bottom surfaces. These issues cause the fabricated SERS substrates to differ slightly from the designed structure, leading to lower-than- expected enhancement. To address the EBL issue, the present inventors tested various photoresist combinations to find the most tolerant one. For RIE, a 15- minute O2plasma cleaning step was introduced before the experiment. As an alternative to e-beam deposition, sputtering, a type of Physical Vapor Deposition (PVD), was used to achieve greater uniformity with the layer thickness.

[0211] Functionalization Step Biomarkers and their antibodies are extremely sensitive to temperature and pH levels, with excessive temperature or extreme pH causing these molecules to unfold or aggregate, thereby compromising their functionality. To stabilize these protein solutions, they should be stored at -80°C when not in use. For thawing, the proteins should be moved to -20°C overnight, then to 4°C for several hours. Following this, the proteins will be reconstituted and serially diluted in 10X PBS buffer, with ABS added if necessary. The prepared protein solution will then be applied to the SERS substrates and measured by Raman spectroscopy immediately. For transportation, the protein solution will be packaged in a temperature-controlled box with ice packs.

[0212] Raman Measurement

[0213] The selection of laser wavelength, power, and acquisition time is critical for obtaining accurate results. Wavelengths such as 532 nm and 633 nm often produce strong background fluorescence, making them less suitable for application to biomarker Raman spectroscopy. To minimize fluorescence from the plasmonic metal (e.g., gold) and achieve better signal-to-noise ratio, a 785 nm laser was used, as it is shifted further into the red, reducing fluorescence effects. To enhance signal clarity, higher laser power and longer acquisition times are generally preferred. However, excessive power or prolonged exposure can lead to spectrum oversaturation. Therefore, laser power is typically kept between 1-10%, with an acquisition time ranging from 10 to 60 seconds to balance signal strength and noise reduction.

[0214] Data Quantity

[0215] Machine learning models require large amounts of representative training data to perform reliably. To increase model accuracy and reliability, the present inventors took many more measurements of each biomarker at varying high and low concentrations.

[0216] Data Quality

[0217] To ensure the models were not heavily biased by noise, the present inventors performed a number of preprocessing steps to reduce noise / errors, remove background signals, and normalize the data.

[0218] Model Development

[0219] There are challenges with choosing which models / techniques to apply after trying well-known / common ones. The inventors found by using techniques on analyzing spectral data and signal processing from different fields (e.g., hyperspectral measurements in plant science, electrical signal processing, etc.).

[0220] REFERENCES

[0221] 1. Wright, J. K., Hayford, K., Tran, V., Al Kibria, G. M., Baqui, A., Manajjir, A., Mahmud, A., Begum, N., Siddiquee, M., Kain, K. C., & Farzin, A. (2018). Biomarkers of endothelial dysfunction predict sepsis mortality in young infants: A matched case-control study. BMC Pediatrics, 18(1). https: / / doi.Org / 10.1186 / s12887-018-1087-x

[0222] 2. Briggs-Steinberg, C., & Roth, P. (2023b). Early-onset sepsis in newborns. Pediatrics In Review, 44(1), 14-22. https: / / doi.Org / 10.1542 / pir.2020- 001164

[0223] 3. Kim, M.-H., & Choi, J.-H. (2020). An update on sepsis biomarkers. Infection & Chemotherapy, 52(1), 1. https: / / doi.Org / 10.3947 / ic.2020.52.1 .1

[0224] 4. Fiedler, U., & Augustin, H. G. (2006). Angiopoietins: A link between angiogenesis and inflammation. Trends in Immunology, 27(12), 552-558. https: / / doi.Org / 10.1016 / i.it.2006.10.004

[0225] 5. Loughna, S., & Sato, T. N. (2001). Angiopoietin and Tie Signaling Pathways in Vascular Development. Matrix Biology, 20(5-6), 319-325. https: / / doi.Org / 10.1016 / s0945-053x(01 )00149-4

[0226] 6. Yuan, H. T., Khankin, E. V., Karumanchi, S. A., & Parikh, S. M. (2009). Angiopoietin 2 is a partial agonist / antagonist of tie2 signaling in the endothelium. Molecular and Cellular Biology, 29(8), 2011-2022. https: / / doi.orq / 10.1128 / mcb.O1472-08

[0227] 7. Cao, C., Gu, J., & Zhang, J. (2017). Soluble triggering receptor expressed on myeloid cell-1 (strem-1 ): A potential biomarker for the diagnosis of infectious diseases. Frontiers of Medicine, 11(2), 169-177. https: / / doi.org / 10.1007 / s11684-017-0505-z

[0228] 8. Higgins, S. J., Purcell, L. A., Silver, K. L., Tran, V., Crowley, V., Hawkes, M., Conroy, A. L., Opoka, R. O., Hay, J. G., Quaggin, S. E., Thurston, G., Liles, W. C., & Kain, K. C. (2016). Dysregulation of Angiopoietin-1 plays a mechanistic role in the pathogenesis of cerebral malaria. Science Translational Medicine, 8(358). https: / / doi.Org / 10.1126 / scitranslmed.aaf6812

[0229] 9. Adly, A. A. M., Ismail, E. A., Andrawes, N. G., & El-Saadany, M. A. (2014). Circulating soluble triggering receptor expressed on myeloid cells- 1 (strem-1) as diagnostic and prognostic marker in neonatal sepsis. Cytokine, 65(2), 184-191 . https: / / doi.Org / 10.1016 / i.cyto.2013.11.004

[0230] 10. McDonald, C. R., Weckman, A., Richard-Greenblatt, M., Leligdowicz, A., & Kain, K. C. (2018). Integrated fever management: Disease severity markers to triage children with malaria and non-malarial febrile illness. Malaria Journal, 17(^. https: / / doi.org / 10.1186 / s12936-018-2488-x

[0231] 11. Conroy, A. L., Lafferty, E. I., Lovegrove, F. E., Krudsood, S., Tangpukdee, N., Liles, W. C., & Kain, K. C. (2009). Whole blood angiopoietin-1 and -2 levels discriminate cerebral and severe (non-cerebral) malaria from uncomplicated malaria. Malaria Journal, 8(1 ). https: / / doi.org / 10.1186 / 1475- 2875-8-295

[0232] 12. Balanza, N., Erice, C., Ngai, M., McDonald, C. R., Weckman, A. M., Wright, J., Richard-Greenblatt, M., Varo, R., Lopez-Varela, E., Sitoe, A., Vitorino, R, Bramugy, J., Lanaspa, M., Acacio, S., Madrid, L., Baro, B., Kain, K.

[0233] C., & Bassat, Q. (2023). Prognostic accuracy of biomarkers of immune and endothelial activation in Mozambican children hospitalized with pneumonia. PLOS Global Public Health, 3(2). https: / / doi.Org / 10.1371 / iournal.pgph.0001553

[0234] 13. Orlando, A., Franceschini, F., Muscas, C., Pidkova, S., Bartoli, M., Rovere, M., & Tagliaferro, A. (2021). A comprehensive review on Raman spectroscopy applications. Chemosensors, 9(9), 262. https: / / doi.org / 10.3390 / chemosensors9090262

[0235] 14. Kazemi-Zanjani, N., Shayegannia, M., Prinja, R., Montazeri, A. O., Mohammadzadeh, A., Dixon, K., Zhu, S., Selvaganapathy, P. R., Zavodni, A., Matsuura, N., & Kherani, N. P. (2018). Multiwavelength surface-enhanced Raman spectroscopy using rainbow trapping in width-graded plasmonic gratings. Advanced Optical Materials, 6(4), 1701136. https: / / doi.Org / 10.1002 / adom.201701136

[0236] 15. Cara, E., Mandrile, L., Sacco, A., Giovannozzi, A. M., Rossi, A. M., Celegato, F, De Leo, N., Hdnicke, P., Kayser, Y, Beckhoff, B., Marchi, D., Zoccante, A., Cossi, M., Laus, M., Boarino, L., & Ferrarese Lupi, F. (2020). Towards a traceable enhancement factor in surface-enhanced Raman spectroscopy. Journal of Materials Chemistry C, 8(46), 16513-16519. https: / / doi.org / 10.1039 / d0tc04364h

[0237] 16. Maccaferri, N., Barbillon, G., Koya, A. N., Lu, G., Acuna, G. P., & Garoli,

[0238] D. (2021). Recent advances in plasmonic nanocavities for single-molecule spectroscopy. Nanoscale Advances, 3(3), 633-642. https: / / doi.org / 10.1039 / d0na00715c

[0239] 17. Pristinski, D., Tan, S., Erol, M., Du, H., & Sukhishvili, S. (2006). In situ SERS study of Rhodamine 6G adsorbed on individually immobilized Ag nanoparticles. Journal of Raman Spectroscopy, 37(7), 762-770. https: / / doi.org / 10.1002 / irs.1496

[0240] 18. Yue, W., Gong, T., Long, X., Kravets, V., Gao, P., Pu, M. & Wang, C. (2020) Sensitive and reproducible surface-enhanced raman spectroscopy (SERS) with arrays of dimer-nanopillars. Sensors and Actuators B: Chemical, 322 128563 https: / / doi.Org / 10.1016 / i.snb.2020.128563

[0241] 19. Zhu, J., Lin, G., Wu, M., Chen, Z., Lu, P., & Wu, W (2018). Large-scale fabrication of ultrasensitive and uniform surface-enhanced Raman scattering substrates for the trace detection of pesticides. Nanomaterials, 8(7), 520. https: / / doi.org / 10.3390 / nano807052Q

[0242] 20. Li, K., Morton, K., Veres, T., & Cui, B. (2011). Nanoimprint lithography and its application in tissue engineering and Biosensing. Comprehensive Biotechnology, 125-139. https: / / doi.org / 10.1016 / b978-0-08-088504-9.0Q497-9

[0243] 21. Pal, P., Bonyar, A., Veres, M., Himics, L., Balazs, L., Juhasz, L., & Csarnovics, I. (2020). A generalized exponential relationship between the surface-enhanced Raman scattering (SERS) efficiency of gold / silver nanoisland arrangements and their non-dimensional interparticle distance / particle diameter ratio. Sensors and Actuators A: Physical, 314, 112225. https: / / doi.Org / 10.1016 / j.sna.2020.112225

[0244] 22. Ciddor, P. E. (1996). Refractive index of air: New equations for the visible and near infrared. Applied Optics, 35(9), 1566. https: / / doi.org / 10.1364 / ao.35.001566

[0245] 23. Colas, F., Barchiesi, D., Kessentini, S., Toury, T., & Chapelle, M. L. (2015). Comparison of adhesion layers of gold on silicate glasses for SERS detection. Journal of Optics, 17(11), 114010. https: / / doi.org / 10.1088 / 204Q- 8978 / 17 / 11 / 114010

[0246] 24. Ngo, Y. H., Then, W. L., Shen, W., & Garnier, G. (2013). Gold nanoparticles paper as a SERS bio-diagnostic platform. Journal of Colloid and Interface Science, 409, 59-65. https: / / d0i.0rg / l 0.1016 / j.jcis.2O13.07.051

[0247] 25. Ko, R. H., Shayegannia, M., Farid, S., & Kherani, N. P. (2021). Protein capture and SERS detection on multiwavelength rainbow-trapping width-graded nano-gratings. Nanotechnology, 32(50), 505207. https: / / doi.Org / 10.1088 / 1361 - 6528 / ac2842

[0248] 26. Su LX, Feng L, Zhang J, Xiao YJ, Jia YH, Yan P, Feng D, Xie LX. Diagnostic value of urine sTREM-1 for sepsis and relevant acute kidney injuries: a prospective study. Crit Care. 2011 ;15(5):R250. doi: 10.1186 / cc10508. Epub 2011 Oct 24. PMID: 22023777; PMCID: PMC3334801 .

[0249] 27. Konvalinka A, Batruch I, Tokar T, Dimitromanolakis A, Reid S, Song X, Pei Y, Drabovich AP, Diamandis EP, Jurisica I, Scholey JW. Quantification of angiotensin Il-regulated proteins in urine of patients with polycystic and other chronic kidney diseases by selected reaction monitoring. Clin Proteomics. 2016 Aug 5;13:16. doi: 10.1186 / s12014-016-9117-x. PMID: 27499720; PMCID: PMC4974759.

[0250] 28. Jia L, Li X, Shen J, Teng Y, Zhang B, Zhang M, Gu Y, Xu H. Ang-1 , Ang- 2, and Tie2 are diagnostic biomarkers for Henoch-Schdnlein purpura and pediatric-onset systemic lupus erythematous. Open Life Sci. 2024 Feb

[0251] 8; 19(1 ):20220812. doi: 10.1515 / biol-2022-0812. PMID: 38465338; PMCID: PMC10921503.

[0252] 29. Wahl, J., Sjddahl, M., & Ramser, K. (2020). Single-step preprocessing of Raman spectra using Convolutional Neural Networks. Applied Spectroscopy, 74(4), 427-438. https: / / doi.Org / 10.1177 / 0003702819888949

[0253] 30. Oliveri, P., Malegori, C., Simonetti, R., & Casale, M. (2019). The impact of signal pre-processing on the final interpretation of Analytical Outcomes - A tutorial. Analytica Chimica Acta, 1058, 9-17. https: / / doi.Org / 10.1016 / j.aca.2O18.10.055

[0254] 31. Palacky, J., Mojzes, P., & Bok, J. (2011 ). SVD-based method for intensity normalization, background correction and solvent subtraction in Raman spectroscopy exploiting the properties of water stretching vibrations. Journal of Raman Spectroscopy, 42(7), 1528-1539. https: / / doi.org / 10.1002 / jrs.2896

[0255] 32. Pontrelli, G., De Crescenzo, F, Buzzetti, R., Calo Carducci, F, Jenkner, A., Amodio, D., De Luca, M., Chiurchiu, S., Davies, E. H., Simonetti, A., Ferretti, E., Della Corte, M., Gramatica, L., Livadiotti, S., & Rossi, P. (2016). Diagnostic value of soluble triggering receptor expressed on myeloid cells in paediatric sepsis: A systematic review. Italian Journal of Pediatrics, 42(1 ). https: / / doi.Org / 10.1186 / S13052-016-0242-y

[0256] 33. Chang, C., Gao, Q., Deng, G., Luo, K., & Zhu, H. (2022). Diagnostic and prognostic predictive values of triggering receptor expressed on myeloid cell-1 expression in neonatal sepsis: A meta-analysis and systematic review. Frontiers in Pediatrics, 10. https: / / doi.org / 10.3389 / fped.2022.929665

[0257] 34. Barber, G., Tanic, J., & Leligdowicz, A. (2023). Circulating protein and lipid markers of early sepsis diagnosis and prognosis: A scoping review. Current Opinion in Lipidology, 34(2), 70-81 . https: / / doi.org / 10.1097 / mol.0000000000000870

[0258] 35. Leligdowicz, A., Conroy, A. L., Hawkes, M., Richard-Greenblatt, M., Zhong, K., Opoka, R. O., Namasopo, S., Bell, D., Liles, W. C., da Costa, B. R., Jiini, P., & Kain, K. C. (2021). Risk-stratification of Febrile African children at risk of sepsis using strem-1 as basis for a rapid triage test. Nature Communications, 12(1 ). https: / / d0i.0rg / l 0.1038 / s41467-021-27215-6

Claims

THEREFORE WHAT IS CLAIMED IS:1 . A Surface Enhanced Raman (SERS) method of detecting biomarkers indicative of a state of sepsis in a patient, comprising: identifying at least one target antigen biomarker indicative of a state of sepsis to be screened for in a sample; functionalizing a plasmonic metal forming part of the nanostructured SERS active substrates with an antibody with specificity to the at least one target antigen biomarker; exposing the at least one functionalized SERS active substrate to the sample being screened for the absence or presence of the target antigen biomarker; illuminating the SERS active substrate with laser light of a preselected wavelength and recording a sample SERS spectra; and comparing the sample SERS spectra to a pre-collected set of biomarker SERS spectra of known biomarkers indicative of a state of sepsis at varying concentrations, and based on this comparison, determining whether the at least one target biomarker is absent from, or present in, the sample, and if present determining from the comparison of the SERS spectra a concentration of the at least one target biomarker in the sample.

2. The method according to claim 1 , wherein the nanostructured SERS active substrates include an array of bullseye structures with each bullseye structure comprised of a preselected number of concentric circular troughs with said preselected number concentric circular troughs being in a range from about one to about ten, a depth of said circular troughs being in a range from about 5 nm to about 500 nm, a width of said circular troughs being in a range from about 20 nm to about 500 nm, and wherein each bullseye structure has an adhesion layer coating a bottom of each trough and a plasmonic metal layer coating the adhesion layer coating in said troughs.

3. The method according to claim 2, wherein said preselected number of concentric circular troughs is greater than one with each trough having the same width and the same distance between each circular trough from a center of each bullseye structure towards the outermost peripheral trough, and wherein and wherein a separation between a middle of adjacent troughs is in a range from about 20 nm to about 500 nm.

4. The method according to claims 1 , 2 or 3, wherein a thickness of the plasmonic metal coating is in a range from about 1 nm to about 100 nm.

5. The method according to claims 2, 3 or 4, wherein a thickness of the adhesion metal coating is in a range from about 1 nm to about 30 nm.

6. The method according to any one of claims 2 to 5, wherein the plasmonic metal coating is any one of gold, silver and copper.

7. The method according to any one of claims 2 to 6, wherein the adhesion layer coating is any one of titanium, chromium, nickel, tungsten, niobium, chromium, or their oxide forms.

8. The method according to any one of claims 1 to 7, wherein the substrate is made of a material having a low index of refraction, low dielectric constant or low relative permittivity selected from any one of glass, silicon, polymer, dielectric and linear / non-linear optical materials.

9. The method according to any one of claims 1 to 7, wherein the substrate is made of any one of silicon (100), silicon (111), polymethyl methacrylate (PMMA), polydimethylsiloxane (PDMS) and poly(N,N- dimethylacrylamide) (PDMA).

10. The method of any one of claims 1 to 9, wherein the nanostructured SERS active substrates include but are not limited to nano-crescents, nanostars, nano-prisms, nano-grating, nano-pillars, nano- chirped circular grating.

11. The method according to claims 1 to 10, wherein functionalizing aSERS active substrate with an antibody with specificity to the target antigen biomarker comprises depositing biotinylated PEG thiol (BAT) molecules onto the SERS active substrate which binds to locations on the plasmonic metal forming part of the SERS active substrate, and depositing a passivating agent onto the surface of the plasmonic metals which coated areas on the plasmonic metal free of the biotinylated PEG thiol (BAT) molecules to prevent non-specific binding of the target antigen biomarkers to the surfaces of the plasmonic metals.

12. The method according to claim 11 , wherein the passivating agent are any one of molecules of thiol polyethylene glycol) (Thiol-PEG), 1 -ethyl-3- (3_dimethylaminopropyl)carbodiimide (EDC) combined with N- hydroxysuccinimide (NHS).

13. The method according to claims 11 or 12, wherein the at least one target biomarker being tested for is sTREM-1 , wherein the step of functionalizing at least one SERS active substrate further comprises deposition of sTREM-1 biotinylated antibodies which binds to the streptadivin and wherein, if sTREM-1 is present in the sample it will form complexes with the sTREM-1 biotinylated antibodies.

14. The method according to claims 11 or 12, wherein the at least one target biomarker being tested for is Ang-1 , wherein the step of functionalizing atleast one SERS active substrate further comprises deposition of Ang-1 biotinylated antibodies which binds to the streptadivin and wherein, if Ang-1 is present in the sample it will form complexes with the Ang-1 biotinylated antibodies.

15. The method according to claims 11 or 12, wherein the at least one target biomarker being tested for is Ang-2, wherein the step of functionalizing at least one SERS active substrate further comprises deposition of Ang-2 biotinylated antibodies which binds to the of streptadivin and wherein, if Ang-2 is present in the sample it will form complexes with the Ang-2 biotinylated antibodies.

16. The method according to any one of claims 1 to 15, wherein the sample is a sample of human bodily fluid.

17. The method according to claim 16, wherein the human bodily fluid is any one or combination of blood, plasma, serum, urine and saliva.

18. The method according to claims 13, 14 or 15, wherein separate SERS active substrates functionalized for detection of each of sTREM-1 , Ang-1 and Ang-2 are integrated into the system, and wherein the concentrations of biomarkers are determined using machine learning classification and regression pipelines, which are trained to distinguish between any combinationof sTREM-1 , Ang-1 , Ang-2 and predict their concentrations.

19. The method according to claim 18, wherein the classification and regression pipelines are configured specifically for spectral data and comprise: a) spectrum preprocessing, to remove or minimize background signal from radiation, noise, and / or SERS responses from other material, clean and validate the data to remove outliers or erroneous measurements, and standardize the data such that data from independent experiments can be compared; b) data augmentation algorithms to artificially generate new data samples from the existing data collected from real-world samples, to increase the number of diverse data samples used for training, validation, development, and / or testing of machine learning methods and thereby improve the performance of these methods; c) feature selection to extract a base set of the most informative features for biomarker discrimination from the dense spectral data collected, thereby reducing the dimensionality of the input data used by the machine learning methods and improving their computational time and space costs / complexity, and real-world generalizability; d) feature and data transformation methods for polynomial fitting or to place the spectra data into a desired feature space to improve machine learning performance and / or costs;e) outlier rejection or robustness techniques to mitigate and / or eliminate biased or erroneous data points or features within the data, to improve the robustness, accuracy, and reliability of the machine learning methods; and f) a machine learning classifier and / or regressor.

20. The method according to claim 18, wherein the regression pipeline predicts the concentration of sTREM-1 , Ang-1 , or Ang-2 from a spectrum containing only one biomarker.

21. The method according to claim 18, wherein the regression pipeline predicts the concentration of any combination of sTREM-1 , Ang-1 , and Ang-2 from a spectrum containing any mixture of these biomarkers.

22. The method according to claim 18, wherein the regression pipeline predicts the concentration of Ang-1 and subsequently predicts the concentration of Ang-2 using the biomarker relationship, and vice versa.

23. The method according to claim 18, wherein a methodology is developed for correlating findings from a machine learning classification pipeline to spectral deconvolutions or decompositions, comprising: a) spectral deconvolution or decomposition into underlying Gaussian or Lorentzian functions; b) extraction of significant peaks and relevant peak properties from the deconvolution or decomposition;c) a machine learning classification pipeline, with feature selection or engineering functionalities to identify significant wavenumbers; d) comparison of significant wavenumbers from the machine learning classification model to extracted peaks and their properties from the deconvolution or decomposition.

24. The method according to claim 23, wherein the correlation is done between the spectral deconvolution or decomposition of Ang-1 , Ang-2, or sTREM and one or more machine learning classification pipelines including said biomarker for the purpose of identifying potential characteristic peaks.

25. The method according to claim 23, wherein the correlation is used to provide physical explainability, such as vibrational modes of the molecular structure of Ang-1 , Ang-2, or sTREM, to the machine learning classifier.

26. The methods according to claims 18 and 23, wherein the findings from deconvolution or decomposition, or identified characteristic peaks, are used as primary features for the development of a machine learning classifier to distinguish between any combination of Ang-1 , Ang-2, and sTREM.

Images