Artificial intelligence–coupled plasmonic infrared sensor for detection of structural protein biomarkers in neurodegenerative diseases

Abstract

Diagnosis of neurodegenerative disorders (NDDs) including Parkinson’s disease and Alzheimer’s disease is challenging owing to the lack of tools to detect preclinical biomarkers. The misfolding of proteins into oligomeric and fibrillar aggregates plays an important role in the development and progression of NDDs, thus underscoring the need for structural biomarker–based diagnostics. We developed an immunoassay-coupled nanoplasmonic infrared metasurface sensor that detects proteins linked to NDDs, such as alpha-synuclein, with specificity and differentiates the distinct structural species using their unique absorption signatures. We augmented the sensor with an artificial neural network enabling unprecedented quantitative prediction of oligomeric and fibrillar protein aggregates in their mixture. The microfluidic integrated sensor can retrieve time-resolved absorbance fingerprints in the presence of a complex biomatrix and is capable of multiplexing for the simultaneous monitoring of multiple pathology-associated biomarkers. Thus, our sensor is a promising candidate for the clinical diagnosis of NDDs, disease monitoring, and evaluation of novel therapies.

AI-coupled plasmonic ImmunoSEIRA sensor performs multiplexed structural protein biomarker detection in neurodegenerative diseases.

INTRODUCTION

Neurodegenerative diseases (NDDs) including Parkinson’s disease (PD), Alzheimer’s disease (AD), dementia with Lewy bodies, amyotrophic lateral sclerosis (ALS), and Huntington’s disease are a set of heterogeneous disorders characterized by the progressive structural and functional degradation of the nerve cells and accumulation of misfolded and aggregated proteins in the affected brain regions (1, 2). These disorders are increasing in prevalence, especially in aging societies, and pose huge economic and health burdens. Unfortunately, there are no effective treatments to prevent or slow the progression of these devastating diseases. Although most NDDs start 10 to 15 years before the manifestation of the clinical diagnostic symptoms, we still lack a reliable diagnostic method for early detection or monitoring of disease progression, thus precluding any effort for early intervention. Furthermore, NDDs are often misdiagnosed because of their clinical heterogeneity, and overlapping symptoms and brain pathologies (3). These observations underscore the need for reliable diagnostic biomarkers and techniques for the early detection of the NDDs, monitoring of disease progression, patient stratification, and investigation of the efficacy of new treatments and disease-modifying strategies.

Although distinct proteins are considered to be involved in different NDDs, protein misfolding into fibrillary aggregates via oligomeric species formation is found to be a common mechanism shared by NDDs and most proteinopathies. For example, the pathological hallmark of PD is the presence of β sheet–rich aggregated species of the alpha-synuclein protein (aSyn; 14.5 kDa) (4, 5) in Lewy bodies (LBs) and Lewy neurites (6–8). Under physiological conditions, aSyn exists in equilibrium between a disordered conformation and α-helical–rich membrane-bound state. In pathological conditions, these two states undergo structural changes and self-assemble to form intermediary oligomers characterized by mixed secondary structure contents, which ultimately convert to β sheet–rich fibrillar aggregates (9, 10). In the case of AD, similar structural changes are observed for the tau proteins and amyloid-beta peptides (11, 12). Hence, identifying and detecting protein aggregates and their structural changes are crucial to diagnosing NDDs and understanding their pathologies. Recently, two drugs—aducanumab (Aduhelm) and lecanemab (Leqembi)—which are monoclonal antibodies preferentially targeting amyloid fibrils and oligomers, respectively, received accelerated Food and Drug Administration approval for early AD treatment. This not only shows the potential of the structural biomarkers to guide and assess future therapies but also, at the same time, underlines the importance of developing effective methods to detect and quantify the level and structural properties of different protein aggregates (13). Moreover, the compelling fact that these NDD-related proteins and their different structural forms are found in the body fluids like cerebrospinal fluid (CSF) (14, 15), blood (16, 17), and saliva (18, 19) provide a path for creating minimally invasive detection tools based on the structural biomarker criteria.

Gold standard methods currently used for biomarker identification and quantification in NDD research, such as mass spectrometry (MS) and enzyme-linked immunosorbent assay (ELISA), focus on quantifying the level of target proteins but insensitive to changes in their structural states and are thus not able to discriminate between different structural forms (20, 21). Progress in proteomic methods such as limited proteolysis-based MS (LiP-MS) shows promise in identifying even the structural alterations of proteins in complex samples on a global scale (22, 23). However, these MS-based techniques are in the early stages of development, and integration of them into clinical practices remains challenging. ELISA, on the other hand, is an antibody-based detection method, and there are ongoing biochemical approaches focused on developing antibodies specific either to the oligomers (18, 24) or to the fibrils (25–27) as a solution to structural insensitivity. But recent validation results (28) indicate that none of the used antibodies are specific to only one form of aSyn aggregate or natively unstructured monomers. Spectroscopic methods, including nuclear magnetic resonance and circular dichroism (CD), can probe the structural information of proteins and have been used to characterize NDD-related protein biomarkers (29–31). However, the need for laborious homogeneous sample preparations, the requirement of high protein concentration, and large sample volumes, as well as the inability to monitor the proteins in their physiological conditions, limit their use as a clinical diagnosis method. Recent advances have enabled the development of structural-based diagnostic assays such as protein misfolding cyclic amplification and real-time quaking-induced conversion (32, 33), which are shown to differentiate patients with PD and AD from healthy individuals with more than 90% accuracy. However, they do not allow the assessment of the actual amounts of aggregates or the ratio of oligomers to fibrils in body fluids. Moreover, they lack the capability to enable prodromal diagnosis, monitor disease progression, and differentiate among the NDDs due to their overlapping pathologies. Incorporating multiple biomarkers associated with neuropathology and neurodegeneration could enable more accurate diagnostics.

Infrared (IR) spectroscopy is sought as a promising method for NDD diagnostics since it provides chemical-specific and structural-sensitive detection in a label-free and noninvasive manner (34). The chemical specificity of the IR technique has been used for the spectrochemical analysis of body fluids and tissues for PD and AD (35–37). Particularly, among the protein absorption bands, the amide I absorption band offers a unique ability to identify different secondary structure motifs like disordered, α helix, β sheets, and β turns (38, 39). However, the low IR absorption cross section of the molecules and the overlap of water absorbance bands with that of proteins limit the use of the technique for conformational studies in the aqueous medium. Multiple reflection–based attenuated total reflection (ATR) overcomes these drawbacks to some extent. Recent works using this technique in combination with immunoassay showed promising results for the early detection of AD and ALS based on the secondary structural changes of amyloid-beta peptides and TDP-43 protein, respectively, present in body fluids (40–43). Surface-enhanced IR absorption (SEIRA) spectroscopy is an emerging method to expand the biosensing capabilities of IR spectroscopy (44). In SEIRA, engineered nanostructures supporting strong and localized electromagnetic fields at their resonant wavelength can be designed to overlap spectrally with the vibrational bands of the biomolecules (45). This amplifies the absorption signals of immobilized biomolecules by orders of magnitude, even in the aqueous medium, via the plasmonic internal reflection (PIR) effect (46). Earlier sensing applications have focused on detecting only the presence of proteins in a dry medium, and some studies extended its use to secondary structure analysis (47–49). Integration of SEIRA with microfluidics has advanced the capabilities to conduct in situ and in-flow experiments for extracting biomolecular interaction kinetics in real time. This also facilitated the protein sensing to be done in situ by directly capturing proteins on the chip surface for sensing and secondary structure identification (50, 51) and even for the real-time monitoring of protein secondary structure changes, but in a nonspecific manner (52). Despite these recent progresses, the potential of SEIRA to decode the structural information of proteins from disease-related biomarkers is yet to be explored. For such applications, it is desirable that the sensing method is highly specific to the target biomarkers from a small sample volume in a complex biomatrix, can distinguish between the pathological aggregates, and monitor a panel of complementary biomarkers.

Here, we introduce ImmunoSEIRA—an immunoassay coupled optofluidic SEIRA sensor capable of extracting unique structural fingerprints of different conformational species of one of the NDD biomarkers—aSyn, including its oligomeric and fibrillary aggregates. The sensor chip exploits a nanoplasmonic metasurface consisting of engineered gold nanorod arrays that are functionalized with antibodies for specific protein detection. The chip is fabricated in a two-dimensional (2D) microarray format and integrated with microfluidics to facilitate in situ capture and structural analysis of target protein in minute sample volumes. The near-field enhanced amide II and I absorption signals of our ImmunoSEIRA sensor are aided by artificial intelligence (AI) to quantitatively identify the percentage presence of aSyn oligomers and fibrils in their mixture, which is unprecedented. The different sensing elements of the microarray chip are selectively functionalized with two different types of antibodies specific to tau and aSyn for multiplexed biomarker detection using spectroscopic signals. Finally, we show that our structural biosensor can retrieve the pathological protein biomarker absorption signature from the complex biomatrix of human CSF. This brings us one step closer to expanding SEIRA sensors for clinical-based diagnostics of NDDs.

RESULTS

Realization of the optofluidic plasmonic SEIRA as a structural biomarker sensor

Figure 1A shows different structural aSyn species such as monomers, oligomers, and fibrils present on the pathway Lewy body formation. Figure 1 (B to E) presents an overview of our measurement setup and the detection principle of the plasmonic SEIRA as a structural biomarker sensor. The optofluidic sensor consists of a plasmonic chip placed in a polydimethylsiloxane (PDMS) part incorporating microfluidic channels. Both the IR illumination and the reflected signal are collected from the backside of the chip in PIR configuration to extract protein secondary structure signatures despite the overlap from aqueous buffer absorbance (Fig. 1B) (46). We combined SEIRA with immunoassay (ImmunoSEIRA) to selectively capture aSyn species for structural-based differentiation (Fig. 1C). The plasmonically enhanced absorption signatures from proteins in amide II and amide I bands (1500 to 1700 cm⁻¹) are retrieved by spectral reflection sensing (Fig. 1D). Uniquely, the amide I band (1600 to 1700 cm⁻¹) absorption contributed by C═O and C─N vibrational modes envelopes the contributions from individual secondary structure motifs that are variably present in different conformational states of the same protein, i.e., the disordered and α helix motifs have their structural fingerprints around 1643 to 1660 cm⁻¹, β turns absorb around 1667 to 1685 cm⁻¹, whereas the main component of the aggregate species, β sheets, has its absorption fingerprint in the lower wave numbers of 1615 to 1635 cm⁻¹ as well as in the higher wave numbers between 1688 and 1696 cm⁻¹ (Fig. 1E). This enables us to differentiate aSyn monomers from the pathological species of oligomers and fibrils with each having their own distinct spectral contributions that can be mapped using our technique.

Fig. 1. ImmunoSEIRA detection principle and the setup.

(A) The misfolding pathway of aSyn protein shows the transition from unstructured monomers to β sheet–enriched fibrillar species via a heterogenous population of oligomers. (B) Schematic of the optofluidic setup used for backside-reflected SEIRA measurements. (C) The capture of all aSyn structural species using antibodies to form ImmunoSEIRA. (D) Amide I and amide II absorption bands are extracted from the captured proteins with a plasmonically enhanced SEIRA signal. (E) Within the amide I band, contributions from different structural motifs like β sheets, disordered, and β turns can be identified to extract structural information.

We fabricated the nanoplasmonic chip on an IR transparent calcium fluoride (CaF₂) substrate in a microarray format consisting of three parallel rows, where each row contains linearly arranged metasurface elements and gold mirrors for sensing and referencing, respectively (Fig. 2A). Each of these metasurface sensing elements is formed of periodically arranged unit cells consisting of plasmonic nanorods coupled with grating order and nanogaps (50). The dimensions of the nanorod length, the gap between the rods in the x direction, and the y-axis periodicity are L = 1.5 μm, G = 0.08 μm, and P_y = 3.2 μm, respectively (Fig. 2A). The sensor chip is then placed in a PDMS part (chipcell; Fig. 2B, left). A microfluidic flowcell is custom-made to continuously run the buffer and inject samples over the sensor, consisting of three independent channels with a width of 300 μm, depth of 30 μm, and length of 3.5 mm (micro-flowcell; Fig. 2B, right). The designed micro-flowcell with less than 35-nl channel volume helps to reduce the amount of required sample and amplifies the antibody-antigen interaction, thereby leading to higher binding performance. These independent channels are positioned such that when combined with the chipcell, the three channels overlap with the three respective rows of the sensing elements and mirrors.

Fig. 2. Components of the optofluidic ImmunoSEIRA sensor and TEM images of aSyn species.

(A) Nanoplasmonic sensor on an infrared transparent substrate with a three-row microarray design. Each row consists of gold mirrors and sensing elements. For example, the top row contains three gold mirrors at the two ends and the center and four sensing elements between them. Scanning electron microscopy images show a part of the sensor element with periodically arranged unit cells consisting of plasmonic nanorods designed with dimensions of L = 1.5 μm, G = 0.08 μm, and P_y = 3.2 μm. (B) The chipcell was fabricated according to the substrate shape to hold the chip, and the micro-flowcell follows the three-row design of the sensor. (C) The nanorod arrays are characterized using a simulation solver and compared with the measurement taken in an aqueouse medium using a Fourier transform infrared spectrometer to observe the resonance targeting the amide II and I bands around 1500 to 1700 cm⁻¹. a.u., arbitrary units. (D) In vitro synthesized pure species of aSyn monomers, oligomers, and fibrils (in the order from left to right) characterized by transmission electron microscopy show the differences in their morphologies.

The parameters of the nanorod arrays are optimized to overlap the plasmonic resonance with the amide II and amide I bands of the protein fingerprints, i.e., between 1500 and 1700 cm⁻¹, with the peak around 1600 cm⁻¹ in aqueous conditions. Such engineered nanostructures provide high sensitivity by enhancing the absorption signals from small quantities of protein on the surface compared to conventional IR spectroscopy methods, thereby opening up immense possibilities for IR-based diagnostic applications. The comparison between SEIRA and different existing IR methods in terms of sensitivity and the reference measurements without any SEIRA enhancement effect are shown in the Supplementary Materials (figs. S1 to S5). The design parameters are first evaluated using simulation based on a frequency domain solver, and then the optical response is validated through measurement with the optofluidic sensor (Fig. 2C). Details on the fabrication and design of the nanostructure and microfluidic parts are explained in Materials and Methods. In parallel, the generation and characterization of aSyn species are performed using well-established and optimized in-house protocols (53) (see Materials and Methods for detailed protocol and figs. S6 to S8). Figure 2D shows the transmission electron microscope (TEM) images of the three distinct aSyn species—monomers, oligomers, and fibrils. Monomers are disordered and smaller in dimensions, so they are not visible in the TEM grid (Fig. 2D, left). As reported previously, the unmodified oligomers exist as a mixture of different morphologies, such as spherical, tubular, and ring-like structures of different sizes (Fig. 2D, middle), and are characterized by mixed secondary structures. The fibril samples show fragmented filamentous structures, which are rich in cross β sheet structures (54), with an average length and width of 50 to 200 nm and 5 to 20 nm, respectively (Fig. 2D, right).

ImmunoSEIRA for the capture and spectral analysis of aSyn species

The spectral specificity of our sensor provides an inherent advantage for a structural biomarker–based diagnostic approach. But target selectivity becomes crucial in the clinical setting where body fluids contain numerous biomolecules. Hence, we leverage the strengths of immunoassay and spectroscopy together on the SEIRA sensor to capture aSyn species using well-characterized and validated aSyn-specific antibodies. The workflow of the immunoassay is a simple three-step procedure, as shown in Fig. 3A. The first step is thiolation. The plasmonic sensor chip is incubated in a mixture of N-hydroxysuccinimide (NHS)–activated carboxyl thiols and OH spacer thiols to form a uniformly spaced monolayer of the activated NHS esters on the plasmonic surface. This facilitates the covalent coupling of any molecules with free amino groups to attach to the surface. After overnight incubation, the chip is washed and mounted with the microfluidic parts of the optofluidic system to commence the continuous in situ spectroscopic measurement. We start flowing the running buffer [phosphate-buffered saline (PBS, 1×)] through the middle fluidic channel of the micro-flowcell component (Fig. 2A, green box) continuously. We focus on one of the sensing elements of the array while measuring a gold mirror intermittently as a reference. From these continuous reflectance spectra, we extract time-resolved absorbance signals over the amide II and I bands (1500 to 1700 cm⁻¹) by normalizing the measured spectrum at each time point to the baseline established during the thiolation step. The details of the spectral data acquisition are provided in Materials and Methods. The baseline correction procedure performed to retrieve the absorbance signals shown in the main text figures is explained in fig. S9. For monitoring target binding on the surface and the interaction kinetics, time-resolved absorbance signals are integrated over the amide range (1525 to 1650 cm⁻¹) to obtain the sensorgram, as shown in Fig 3B.

Fig. 3. ImmunoSEIRA assay for the capture and spectral analysis of aSyn species.

(A) The used three-step immunoassay is shown as a schematic. At first, the sensor chip is incubated with N-hydroxysuccinimide–activated carboxyl thiols and spacer OH thiols overnight. Then, with the optofluidic configuration, the SYN-211 antibody is injected in situ and washed with the buffer to form a stable antibody layer. Last, the aSyn fibrils are injected and washed to selectively enrich the target on the functionalized sensor surface. (B) Time-resolved reflectance measurements are taken for each step starting from the thiol baseline, and the integrated absorbance signal over the amide region is used to extract the sensorgram, which shows the injection, binding, and stabilization of antibody as well as the successful capture of aSyn fibrils along with subsequent washing steps. (C) The absorbance spectrum as a function of time for each assay step is displayed. The distinct absorption signatures of the antibody and aSyn fibrils in the amide range can be visualized.

The measurement of the thiolated sensor in the buffer is continued for ~30 min to get a smooth and uniform baseline. In this time window, the time-resolved absorbance is zero as there is no change relative to the defined baseline (Fig. 3C, thiolated sensor). Next, we proceed to the second step—antibody immobilization. Here, we chose a sequence-specific antibody targeting the flexible C-terminal domain of the protein: SYN-211, as the capturing agent that binds to full-length aSyn irrespective of its conformation, thereby enabling unbiased structural analysis of the protein (55). The antibody solution is injected at ~30 min, and we note an increase in the sensorgram, indicating the binding of the antibody on the thiolated sensing element (Fig. 3B, blue region). Once the antibody solution completely flows over the surface, the running buffer is introduced back into the channels at ~75 min. The unbounded antibody is washed away as indicated by a small dip at the beginning of the buffer wash, and then we can observe a permanent increase in the sensorgram signal even after washing (~75 to 110 min). This indicates the immobilization of a sturdy capture layer of antibodies. The time-resolved absorbance spectra of this second step (between ~30 and 110 min) retrieve the amide II and I bands of the antibody (Fig. 3C, antibody immobilization). After we have obtained a steady signal for the antibody step, we progress to the final step of our immunoassay—aSyn capture. For demonstration in the figure, pure species of aSyn fibrils are used. Two hundred microliters of this target protein solution is injected around ~110 min at a concentration ensuring protein enrichment on the sensor surface. The increase in the sensorgram signal (Fig. 3B, yellow region) indicates the enrichment of aSyn fibrils on the sensor surface. After the course of the protein solution, the buffer is reintroduced to the channel for washing at ~140 min. The steady sensorgram signal (between ~140 and 250 min) conveys the successful capture and retention of the aSyn fibrils by the antibody layer even after the washing step. The time-resolved absorption spectra of this final step reveal the absorption signatures of aSyn fibrils from injection to stable binding (Fig. 3C, aSyn fibrils capture). We can clearly see its expected distinct spectral shape in the amide I region (1600 to 1700 cm⁻¹) with a peak around ~1630 cm⁻¹ and an off-shoulder around higher wave numbers, indicative of high β sheet content. The binding kinetics of aSyn oligomers and monomers and their respective absorbance spectra are provided in figs. S10 and S11, respectively. Figure S12 shows the results of the control experiment with and without a blocking step in the functionalization assay for the in-flow experiments using pure aSyn fibrils form. Figure S13 shows additional negative control experiments to verify the cross-reactivity of the SYN-211 antibody to other NDD-related proteins.

Conformational profiling discriminates the different structural species of the aSyn

Next, we sought to investigate the ability of our sensor to identify the secondary structure of the three aSyn species, i.e., monomers, oligomers, and fibrils, by conducting an in-depth analysis of their amide I band for conformational profiling. This is done independently using two conventional mathematical methods: second-derivative analysis and Fourier self-deconvolution (FSD) with curve fitting that helps to deconvolute the substantially overlapped component bands arising from multiple secondary structural motifs (39, 56). Using unbiased second-derivative analysis, we identified the peak positions of the individual contributing bands of different structural motifs. Using these peak positions as a reference, we then executed and optimized the FSD to extract quantitative data of the individual band contributions. Figure 4 shows the complete analysis of aSyn monomers (i), oligomers (ii), and fibrils (iii). We specifically focused on three main structural bands—disordered/α helix, β sheets, and β turns. For clarity, the original absorbance spectra of the monomers, oligomers, and fibrils used in this deconvolution analysis are shown separately in fig. S14.

Fig. 4. Conformational profiling of different aSyn species.

Each aSyn pure structural species, monomers, oligomers, and fibrils, are captured separately on different chips that are identically functionalized. The second-derivative analysis is used to identify the main peak contributions, and correspondingly, the FSD and curve fitting are done to identify contributions from individual structural motifs. (A) aSyn monomers have characteristic disordered structures evident from almost 70% absorption from those bands, with only 20% of β sheets present. (B) aSyn oligomers have 50% of their structure in β sheet configuration, and the rest are almost equally present as β turns and disordered states. (C) aSyn fibrils have the highest contribution from β sheets—58% among all species, with disordered motifs contributing 27% and the rest β turns.

aSyn monomers have a larger amount of disordered structures with almost 70% contribution and only 20% of β sheet outlining its native unstructured form. In fibrils, the dominant structure with almost 58% contribution is β sheets, which is acknowledged to be a characteristic of the fibrillar pathological species. The disordered content in fibrils is only 27%, and the rest is contributed by β turns that help to connect and form antiparallel β sheets. In the heterogenous population of unmodified oligomers, we found mostly the presence of β sheet with a 50% contribution and the same contribution of disordered structure as in fibrils (27%) but with the increased presence of β turns (23%). Using distinct spectroscopic absorption signals, we identified structural differences between all three species, despite the conformational similarities of oligomers and fibrils. Notably, we managed to decipher the structural components of oligomer species (Fig. 4B), which paves the way for the use of aSyn species–specific structure-based biomarkers in immunoassay-based diagnostic tools. This is important as the oligomeric forms of aSyn have been implicated to play important roles in aSyn toxicity and pathology spreading during disease progression (10, 57, 58). Since our method relies on the attachment of proteins on the surface by functionalization, it is crucial to verify that this does not alter the native conformations of immobilized proteins. For this purpose, we characterized the secondary structure of free-floating aSyn proteins in solution using CD. The secondary structure analysis results of ImmunoSEIRA show a good correlation with that of the CD data (presented as figs. S6 to S8).

AI-aided ImmunoSEIRA to quantitatively predict the percentage of aSyn oligomers and fibrils in a mixture

In a patient sample, it is likely that the three forms of aSyn species are present at different levels simultaneously. Although there are aSyn antibodies that bind preferentially to aggregated forms of aSyn, there are no antibodies that can selectively capture either oligomers or fibrils (28). This has hampered efforts to accurately quantify these aggregate species and/or correlate their levels to disease stages, symptoms, or rate of progression. Therefore, it is imperative to develop new methods that enable accurate quantification of each form of aSyn, including oligomers and fibrils. This would also enable accurate assessment of total aSyn in biological samples. The spectral decomposition approaches used in Fig. 4 are suitable for studying pure samples and cannot quantitatively predict the percentage composition of oligomers and fibrils in a mixture because they only provide the distribution of individual structural motifs (e.g., disordered and β sheet). This information is not sufficient to spectrally regress and associate these subbands quantitatively to different aggregates to reconstruct their percentage distribution. To tackle this challenging problem, we combined AI with spectral analysis using ImmunoSEIRA to accentuate the unique absorption signatures of oligomers and fibrils rather than their underlying individual secondary structure motifs. Machine learning and AI analysis, when combined with IR spectroscopy, have shown promise to unravel interesting applications in environmental monitoring (59–61), chemical sensing (62, 63), biosensing, diagnostics (64, 65), etc. Recently, AI has also been coupled with SEIRA, thus extending biological studies to enable dynamic monitoring of the interactions of proteins with other biomolecules, such as lipids, nucleic acids, or carbohydrates, when present simultaneously (66, 67). In this work, we combined AI with SEIRA to solve an even more complex problem—to discriminate among the different aggregation states of the same protein when present in a mixture.

To determine whether our combined AI-ImmunoSEIRA approach could distinguish oligomers from fibrils, we propose here a plausible scheme of total aggregate presence in a sample at different stages of the disease, with mostly oligomers in the beginning but with the presence of fibrils gradually increasing and eventually becoming the dominant species in the mixture (Fig. 5A). Based on this, we designed a sample set of different mixtures of titrated percentage combinations of aSyn oligomers and fibrils:100-0, 75-25, 60-40, 50-50, 40-60, 25-75, and 0-100% (oligomers-fibrils), respectively (Fig. 5B). We measured 3D (absorbance–wave number–time points) spectrograms of these different sample set using our optofluidic ImmunoSEIRA setup (Fig. 5C). The time-resolved spectral analysis generates a large set of data with close to 3.5 million data points, thus providing an ideal opportunity to use AI-DNN (deep neural network) models. We focused on the spectro-temporal range containing information on the amide II and I absorption bands that vary over time from the gradual binding of the aggregates over the surface up until stabilization. Each of such spectra contains 133 data points that correspond to time-varying absorbance in the wave number range of 1450 to 1700 cm⁻¹ with 4-cm⁻¹ spectral resolution. This data cube was then concatenated with labels assigned to each mixture ratio as 0, 25, 40, 50, 60, 75, and 100 in the order above that correlates to the percentage presence of fibrils (Table 1) to form the dataset quartet. We modeled a DNN based on a multilayer perceptron (MLP) model and applied it to the data quartet. The DNN was configured with 133 input nodes—each node for feeding the absorbance at a given wave number and a given time point and one output node to predict the label of the fed absorbance spectra. We split the entire dataset randomly into two, 80 and 20%. The 80% data quartet of each combination ratio was used for training and validating the neural network for the best possible prediction accuracy by performing cross-validation (CV) and parameter tuning for optimization. This was done with fivefold CV to find the optimum number of hidden layers and nodes that can yield the desired performance in terms of the least mean squared error for prediction. This training step yielded the DNN model consisting of 4 hidden layers with 26 nodes each, with an overall accuracy score of 94.66%. After this validation, the remaining 20% of the entire dataset of every combination ratio that has not been included in the training was fed to the network to evaluate the prediction accuracy for the regression percentages (i.e., the presence of oligomers and fibrils in percentage). From Fig. 5D, we can infer that, in this test, the network predicts the percentages of unknown sample ratio with excellent accuracy, and the average of the predicted percentage of fibrils in the mixtures matches the actual concentration percentage. Notably, in this DNN analysis, we used raw spectral data, including initial time points where the absorption signal is weak and noisy. Despite the low signal levels, without any smoothing and averaging, our AI-aided ImmunoSEIRA sensor is robust and structurally sensitive. This is crucial and could pave the way to detect and quantify aSyn oligomers and fibrils and correlate their ratios during disease progression in longitudinal studies where the physiological concentrations are expected to be low. Therefore, our novel approach can enable a more accurate understanding of the role played by the protein aggregates in the disease and their diagnostic value, as well as perform quantitative studies from a single sample without any manual filtration and additional steps (53).

Fig. 5. AI-aided ImmunoSEIRA for accurate prediction of the quantitative presence of oligomers and fibrils distinctively from mixed samples.

(A) Proposed plausible scheme of total aggregate presence in a sample at different stages of the disease with mostly oligomeric aSyn in the beginning but with the gradual increase of fibrils presence and eventually becoming the dominant species in the mixture. (B) The bar plot of the titrated concentration mixes of oligomers and fibrils used in this study. (C) Three-dimensional spectrograms of all the mentioned mixtures are collected using the ImmunoSEIRA setup and labeled according to the presence of fibrils in each sample set. (D) The comparison of the actual fibrils percentage with the predicted value by the deep neural network (DNN) from the testing data shows excellent accuracy.

Table 1. Sample dataset with labels for AI analysis.

The molar percentages of aSyn oligomers and fibrils present in the mixed samples used for DNN analysis and their respective label fed to the DNN.

Percentage of oligomers (%)	Percentage of fibrils (%)	Label fed to DNN
100	0	0
75	25	25
60	40	40
50	50	50
40	60	60
25	75	75
0	100	100

Multiplexed ImmunoSEIRA for simultaneous detection of aSyn and tau proteins

Increasing evidence points to the need for incorporating multiple disease-relevant biomarkers to accurately diagnose and differentiate between different NDDs. One such additional biomarker of interest in aSyn and PD studies is the tau protein because phosphorylated and aggregated forms of tau often co-occur with aSyn pathology in PD and other NDDs. Several lines of evidence point to complex interactions between aSyn and tau in PD and suggest that they influence each other’s function and pathology formation (68). Furthermore, analysis of PD brain pathology also displayed the colocalization of tau and aSyn in LBs, which confirms tau as a complementary biomarker (69). Some biomarker studies have assessed the potential of simultaneously monitoring the levels of aSyn and tau in different forms and showed that it could improve the diagnosis of PD and synucleinopathies (70, 71). These studies are usually based on measuring the protein levels rather than the direct assessment of their conformations and require the use of multiple assays, which increases demands on the very precious human samples. This underscores the need for newly developed detection methods to be capable of screening several biomarkers in a multiplexed manner.

As a proof of concept, we assessed the potential of extending and multiplexing our ImmunoSEIRA microarray sensor by simultaneously detecting two different structural biomarker proteins—aSyn and tau from a single sample. As described in Fig. 2, the design of our optofluidic chip consists of three independent microfluidics channels, and, in each channel, there are linearly arrayed sensing elements and mirrors. Capturing these two different proteins simultaneously requires their respective sequence-specific antibodies to be selectively immobilized on the surface of the corresponding sensing regions. We performed spatially resolved bioprinting of the capture antibodies by using a noncontact and low-volume piezoelectric liquid micro-dispenser. Before antibody printing, the entire surface of the plasmonic chip is functionalized with activated NHS ester thiol and OH spacer thiol. We used the array in the middle fluidic channel and functionalized the first half of the array (four sensing elements and two mirrors) with SYN-211 antibody targeting aSyn and the second half of the array with HT7 antibody targeting tau protein (Fig. 6A). Antibody droplets of 450-pl volume were spotted on the respective regions uniformly within a matter of minutes with high spatial precision and at low antibody consumption down to few nanoliters per sensing element.

Fig. 6. Cross-reactivity experiment of the multiplexed ImmunoSEIRA for aSyn or tau fibrils injection.

(A) The schematic of the in-flow experiment shows blocking buffer injection followed by the aSyn or tau fibril injection in two separate experiments. (B) In the first experiment, the sensorgram obtained from the SYN-211 printed sensing element shows a permanent signal increase after the final washing step due to the specific binding of aSyn fibrils on the SYN-211 antibody, whereas the sensorgram from HT7-printed sensing element shows that the signal levels before aSyn fibril injection and after the final washing steps are similar, indicating no cross-reactivity of HT7 antibody to aSyn fibrils. (C) The retrieved absorbance (between 85 and 155 min) from SYN-211 shows the corresponding spectral signal increase and (D) the absorbance retrieved from HT7 for this relevant time window (between 85 and 155 min) shows no spectral signal increase. (E) In the second experiment with tau fibrils injection, the sensorgram obtained from SYN-211–printed sensor element shows that the signal levels before tau fibril injection and after the final washing steps are similar, indicating no cross-reactivity of the SYN-211 antibody to tau fibrils, but the sensorgram obtained from HT7-printed sensor element shows a permanent signal increase after the final washing step due to the specific binding of tau fibrils on HT7 antibodies. (F) The retrieved absorbance (between 80 and 150 min) shows no stable spectral signal increase for the SYN-211 sensing element, (G) whereas the absorbance retrieved from the HT7 printed sensing element for the same selected time window shows the corresponding spectral signal increase by the binding of tau protein.

After incubating the bioprinted chip at room temperature for approximately 2 hours and washing, it is mounted to the microfluidic parts for in situ measurements (see the detailed procedure in Materials and Methods). From the absorbance signals, we confirmed that the bioprinting of both SYN-211 and HT7 antibodies is highly uniform across all arrays (fig. S15). Consequently, we focused on one of the bioprinted functionalized sensing elements per antibody in the array and started time-resolved spectral measurements to establish the baseline signal (see Materials and Methods). Next, the sensor surface was blocked with milk buffer to eliminate nonspecific binding sites. Afterward, we checked the cross-reactivity of the platform by injecting the proteins, aSyn fibrils and tau fibrils, each in a different experiment, and evaluated the signal change. The schematics of both experiments are shown in Fig. 6A. In the first experiment with aSyn fibrils injection (after ~110 min, yellow region), we observe that the binding response and, thereby, the absorbance are increased in the SYN-211 spotted sensing element and the sensorgram signal stabilized even after washing (Fig. 6B, pink curve). On the other hand, the sensorgram signal for the HT7 spotted sensing element increased slightly with injection but, upon washing, reduced back to the previous baseline (Fig. 6B, blue curve). The binding of aSyn fibrils on the SYN-211-spotted array can also be observed with the retrieval of the distinct absorption signal during this injection step (Fig. 6C), and the absence of aSyn fibrils binding on the HT7 spotted array is confirmed with a lack of absorption signal increase in the same duration (Fig. 6D). In the second experiment with tau fibril injection (after ~100 min), we observed a slight increase in the sensorgram signal on the SYN-211 spotted sensing element (Fig. 6E, pink curve). But after washing (~130 min), the signal decreased back to the previous baseline, indicating that the protein passed over the sensor surface without binding. On the other hand, in the HT7-printed sensing element, a stable binding response was observed even after washing (Fig. 6E, blue curve), indicating the successful capture of tau protein by HT7 antibodies. This is also consistent with the lack of increase in the time-resolved absorbance spectra retrieved from the tau fibril injection step on the SYN-211 sensing element (Fig. 6F), whereas the binding of tau fibrils on the HT7 spotted array can be observed with the retrieval of the distinct absorption signal during this injection step (Fig. 6G). These results indicate that the surface is optimally blocked, and the chosen antibodies are specific with no notable affinity to the other biomarker. Last, for the validation of multiplexed detection of pathological tau and aSyn fibrils, we injected a sample containing the mixture of both biomarkers (Fig. 7A). Following the blocking and washing steps, we observed a stable binding response for both SYN-211– (Fig. 7B) and HT7-printed sensing elements (Fig. 7C) upon the mixture injection (~140 min) even after the final washing step. The retrieved time-resolved absorption spectra also show a stable signal increase and capture distinct signatures of aSyn and tau fibrils on the SYN-211– (Fig. 7D) and HT7-printed sensing elements (Fig. 7E), respectively.

Fig. 7. Multiplexed ImmunoSEIRA demonstration with simultaneous injection of aSyn and tau fibrils.

(A) The schematic of the in-flow experiment shows that blocking buffer injection is followed by the injection of a mixed sample containing both aSyn and tau fibrils. (B) A stable binding of aSyn fibrils on the SYN-211–printed sensing element during the mixture injection and the final washing step is indicated by the consistent signal increase in the sensorgram curve. (C) Similarly, stable binding of tau fibrils on the HT7-printed sensing element during the mixture injection and the final washing step is indicated by the consistent signal increase in the sensorgram. (D) The retrieved time-resolved absorption spectra (between 90 and 230 min) from the SYN-211 sensing element show a corresponding spectral increase by the binding of aSyn protein. (E) Absorbance retrieved from the HT7 printed sensing element shows a spectral increase for the same time window due to the specific binding of tau protein.

Retrieval of the absorption signature of aSyn fibrils in complex biomatrix

To take the developed sensor one step closer to clinical application, it is important to assess its detection abilities in a complex biomatrix. The transitioning of the operation from buffer to real and complex human body fluids is not straightforward. There are several differences between analyte studies in a buffer system to that in its native body fluid environment. For example, the physiological response of the analytes, such as their binding affinity and structural integrity, can be different. In addition, the presence of other biomolecular components in the body fluids can lead to biofouling, thus having an interfering effect with the target capture and detection. These factors need to be carefully assessed and dealt with for the successful application of clinical samples. To illustrate the feasibility of our sensor, we used healthy human CSF samples depleted of aSyn by an in-house optimized protocol as the biomatrix (see fig. S16 and Materials and Methods). We observed that milk buffer is a suitable blocking agent to minimize nonspecific binding from the molecules/proteins present in the biomatrix. The schematic of the experimental procedure is shown in Fig. 8A. After the antibody and blocking steps, we did a negative control by injecting aSyn-depleted human CSF (~290 min). Initially, the sensorgram signal is increased, indicating the presence of multiple nonspecific molecules present in it. But after the washing buffer was introduced (~340 min), the signal returned to the previous baseline, confirming that the surface is immune to biofouling and to the binding of other molecules in depleted CSF (Fig. 8B). Last, when we injected a 200-μl sample of aSyn fibrils (5 μM) spiked in depleted CSF matrix (~390 min) for enriching the protein capture on the sensor surface, a stable sensorgram signal increase is observed even after a washing step (~440 min), indicating specific binding of aSyn fibrils on the functionalized surface. We also show the absorbance spectra during the injection step in Fig. 8C, which demonstrates successful retrieval of the aSyn fibrils absorption signature in the presence of the complex biomatrix (see fig. S17 for comparison). Furthermore, we performed proof-of-concept experiments combining different strengths of ImmunoSEIRA, including multiplexing, AI-aided analysis, measuring aSyn aggregate mixture in different ratios and operation in complex biomatrix (see figs. S18 and S19), which paves the way to use our method in the scientific research on understanding NDDs and toward clinical studies.

Fig. 8. Absorption signature of pathological aSyn fibrils is accurately retrieved in the presence of complex biomatrix.

(A) Schematic of the assay steps for characterizing aSyn fibrils binding in human CSF biomatrix with in-flow SEIRA measurements. (B) The sensorgram highlights different stages of binding. During the injection of aSyn-depleted human CSF, the signal increases due to nonspecific biomolecule presence. But after the washing step, the signal goes back to the previous baseline, indicating that the surface blocking is successful. The signal level after the injection of aSyn fibrils spiked human CSF and the final washing step increased, indicating that protein binds to the surface even in the presence of a complex biomatrix. (C) The characteristic absorbance fingerprints of the aSyn fibrils are accurately retrieved from the time-resolved absorbance spectra of the final step of spiked injection.

DISCUSSION

In this study, the unique capabilities of SEIRA are explored for its potential as a diagnostic tool by structural analysis of protein biomarkers. The detection principle of SEIRA in PIR configuration provides extreme field confinement close to the nanoantennas, thereby overcoming the problem of water absorption obscuring the protein signal. This subsequently enabled straightforward and easy integration of SEIRA substrate with microfluidics to perform in situ detection from a small sample volume while maintaining the conformational integrity of the proteins, in contrast to other IR methods like grazing incidence reflection. We incorporated immunoassay with SEIRA (ImmunoSEIRA) for conformationally sensitive and label-free analysis of structural biomarkers of the NDD-relevant proteins aSyn and tau, by directly accessing their distinct chemical vibrational fingerprints. With this immunoSEIRA sensor, we performed comprehensive structural profiling of all three conformational species of aSyn associated with pathology formation in PD and other NDDs, including monomers, oligomers, and fibrils. We successfully quantified different structural motifs present in each of the conformational aSyn species and identified the differences and similarities between them. In an unprecedented manner, by combining ImmunoSEIRA with AI, we showed that we could not only differentiate between different aggregations states of aSyn with distinct conformational properties but also quantitatively predict the individual presence of oligomers and fibrils from mixed aSyn aggregate samples. This outcome is a major advancement from current assays, which depend on antibodies that cannot reliably distinguish between oligomers and fibrils. The AI-aided ImmunoSEIRA analysis is crucial for extensively profiling the quantitative presence of different conformational species of the same protein biomarker in patient body fluids. This can lead to developing a structural fingerprint map of protein variations occurring from the early prodromal stage until the late clinical stage to understand in-depth the potential role of aSyn in disease pathology and as a clinical biomarker. We also demonstrate the potential of expanding our sensor to form a multiplexed SEIRA microarray that can incorporate and detect multiple pathological markers of NDDs, e.g., tau and aSyn fibrils, simultaneously. Existing efforts for multiple biomarker detection using IR methods either use different sensors for each biomarker or regenerate the same sensor by repetitive and laborious functionalization steps (72, 73). Crucially, these approaches require the repetitive use of higher amounts of precious biofluids, which is not clinically feasible considering the total volume of extracted sample (e.g., a few ml of CSF) and the different types of tests and repetitions of measurements to be done (70). We managed to overcome these limitations by exploiting the inherent 2D microarray design format of our sensor and performed multiplexed detection of two protein biomarkers from a single low-volume sample injection of 200 μl, which could be brought down further in the future (74). These multiplexing capabilities provide unique opportunities for the assessment of multiple disease-relevant biomarkers, including multiple proteins linked to NDDs. Recent studies have shown that the presence of pathological aggregates of multiple proteins, including amyloid-beta, aSyn, tau, and TDP-43, is the norm rather than the exception in NDDs and that the relative distribution of aggregates of these proteins in the brain influences disease symptomology and rate of progression. One alternative use of our multiplexed SEIRA microarray is to use multiple antibodies of the same protein to capture the diversity of aggregates in biological samples. The availability of species-specific antibodies would substantially increase the diagnostic power of our AI-aided ImmunoSEIRA as it will allow more precise determination of the ratios of different species (monomers, oligomers, fibrils, or posttranslationally modified forms of these proteins), which could be combined with other biomarkers to enable more accurate assays for early detection and monitoring disease progression. Last, we evaluated the performance of our sensor for conformational sensing of aSyn fibrils when spiked in healthy human CSF sample and demonstrated its structural fingerprint retrieval even in the presence of compounded signals from the complex biomatrix. These observations represent proof-of-concept results paving the way to test and validate our ImmunoSEIRA assay in body fluids from cohorts of patients at different stages of PD and patients with different synucleinopathies.

Looking forward, we foresee that replacing invasive biofluids like CSF with noninvasive and easily accessible samples like blood serum or plasma can increase the adaptability of new biosensors in PD and NDD diagnostics. However, given that blood samples present a more complex biomatrix than CSF, optimization of surface functionalization becomes an important aspect of sensor development. Furthermore, optimization of the ImmunoSEIRA sensor, such as engineering highly sensitive resonant geometries and alternative optofluidic schemes based on pneumatic valves, is instrumental in achieving the required limit of detection that correlates to the protein biomarker levels in patient samples. At this stage, the sensitivity performance of the optimized sensor would be benchmarked with other in-solution compatible IR techniques like ATR. The multiplexing prospect can be expanded beyond two biomarkers by adapting the microarray design and the immunoassay protocol with robust capture agents not limited to antibodies but aptamers and small molecules offering minimal cross-reactivity (75, 76). We believe that the compatibility of our 2D microarray plasmonic chips with microfluidics brings an advantage for increasing the analysis throughput in future studies. The design of multiple sensing elements within the same fluidic channel could simultaneously monitor multiple biomarkers with dedicated sensing elements with corresponding antibodies from a single low-volume injection of one sample. By scaling up the microarray sensor design and the number of independently running microfluidic channels, a large set of different samples can be tested simultaneously, thereby increasing the analysis throughput and reducing the time and cost per analysis, which is not feasible with the existing IR-based detection methods. The scaling aspect presents challenges as it necessitates manufacturing sensor chips at low-cost and large areas. While, in this study, we used low-throughput electron beam lithography to pattern the nanorod antennas using research-grade fabrication protocols, the cost per chip is still one order of magnitude less than the commercially available bulky and expensive ATR crystals. There are recent advancements in wafer-scale manufacturing of nanophotonic substrates, such as deep ultraviolet lithography, interference lithography, and nanoimprint lithography, as well as leveraging complementary metal-oxide semiconductor (CMOS) compatible materials, including aluminum and low-loss dielectrics (77, 78). We foresee that translation of our fabrication to such wafer-scale methods could further reduce the manufacturing time and cost per chip, facilitating the commercial availability of the SEIRA substrates for low prices in the future. From a technical perspective, switching from bulky Fourier transform IR (FTIR) measurement tools to highly compact on-chip mid-IR biosensors using emerging IR technologies on Quantum cascade lasers (79), quantum cascade detectors (80), and flat meta-optics is another open challenge. But this also opens up new opportunities to take newly developed IR sensors into biological, medical, and pharmaceutical research and clinical applications.

MATERIALS AND METHODS

Numerical simulation of the plasmonic metasurfaces

A finite integration Maxwell solver CST Studio 2018 is used for the numerical design and characterization of the far-field and near-field properties of the nanostructures. A unit cell consisting of a nanorod with a height of 100 nm of Au (gold) over 5 nm of Cr (chromium) is created on calcium difluoride (CaF₂) substrate with a water layer on top of the antennas with optical constants retrieved from Olmon et al. (81), Rakić et al. (82), Li (83), and Hale et al. (84) for Au, Cr, CaF₂, and water respectively. The structures are modeled with a tetrahedral mesh of 40-nm size and simulated with periodic boundary conditions. The incident plane wave is set to have an inclination angle of 16.7° as an average of the incident angle spread (9.8° to 23.6°) with the Cassegrain objective used in the FTIR measurements. The far-field spectrum is calculated as the average of the transverse magnetic (TM) excitation mode perpendicular to the long edges of the nanorod and the transverse electric (TE) excitation mode perpendicular to the short edges.

Fabrication of the plasmonic metasurfaces

CaF₂ chips of 13-mm diameter and 0.5-mm thickness are used as the substrates (Crystran, UK). After RCA1 (NH₄OH:H₂O₂:H₂O = 1:1:5) cleaning protocol and subsequent washing with acetone and isopropyl alcohol (IPA), chips are spin-coated with low– and high–molecular weight PMMA [poly(methyl methacrylate)]. This is sputtered with 10 nm of Au as a conductive layer for electronic beam lithography. The nanorod arrays of 250 × 250 μm² are patterned with 5-nm resolution and Au mirrors of the same size with 100-nm resolution using a 100-keV beam. Afterward, the Au layer is removed by wet etching with KI + I2 solution and subsequently developed with a PMMA developer MiBK:IPA = 1:3. Then, electron beam evaporation is done to deposit 5 nm of Cr with 100 nm Au. Liftoff is done with acetone and followed by heated sonication in MICROPOSIT REMOVER 1165 and water to remove the resist residues completely. Last, structures are analyzed using scanning electron microscopy to verify the successful fabrication. The nanorods used in the experiments are 1500 nm in length, 100 nm in width and height in an array with a y-axis periodicity of 3200 nm and a gap of 80 nm between the nanorods in the x direction.

Fabrication of microfluidic micro-flowcell and chipcell

The upper layer of the micro-flowcell with small channels of width (300 μm) and depth (30 μm) is fabricated using standard soft lithography. The Si wafer is coated with a positive resist (AZ1512) and exposed via direct laser writing. After development, the wafer is dry-etched (Bosch process), followed by resist stripping to form the mold. The mold for the chipcell and the big channel layer of the micro-flowcell are created using a vinyl-based cutting plotter (GRAPHTEC). All the molds are silanized for PDMS fabrication by incubating them with TMCS (trimethylchlorosilane). Then, the mixture of PDMS and curing agent in the ratio of 10:1 is poured over the mold surfaces and baked at 80°C for over 2 hours. The upper and big channels for the micro-flowcell are then oxygen plasma–bonded with alignment to form the complete micro-flowcell part.

Immunoassay protocol

The sensor chip is washed with acetone, ethanol, and water, followed by oxygen plasma cleaning. Afterward, within 30 min, it is incubated with a 2 mM thiol mixture of activated ester (HSC11EG4OCH2COONHS, ProChimia Surfaces) and spacer thiols (HS-C6-EG3OH, ProChimia Surfaces) in the ratio of 1:9 overnight. After washing with ethanol, the chip is mounted with the microfluidic parts to start the in situ immunoassay steps. With the continuous 1× PBS [10 mM phosphate buffer, 2.7 mM potassium chloride, and 137 mM sodium chloride (pH 7.4)] buffer flow started, 200-μl SYN-211 antibody (40 μg/ml) (#ab206675, Abcam; sc-12767, Santa Cruz Biotechnology) in sodium acetate buffer is injected. Once the antibody signal is stabilized after washing, 10 μM of 200-μl aSyn fibrils (in-house prepared) is injected. In the aSyn oligomer capture (fig. S10) and the monomer capture (fig. S11) experiments, 200 μl of protein solutions of 5 μM oligomers and 10 μM monomers are used, respectively.

Infrared in situ measurements

The in situ measurements are carried out using an FTIR spectrometer (Bruker Vertex 80v) coupled to a microscope (HYPERION 3000 IR microscope) with a reflective Cassegrain objective (15×, numerical aperture = 0.4), with high-power globar as the light source and liquid nitrogen–cooled mercury cadmium telluride as the detector. An external polarizer is used to apply incident light polarization parallel to the long axis of the nanorods. A knife-edge aperture limits the light collection to slightly less than 250 × 250 μm², which is the size of each sensing element and mirror. The measurements are done in reflection mode, illuminating the sensor chip from the backside of the CaF₂ substrate to avoid light propagation through buffer/water and in a purged dry air environment. The inlet of the microfluidic parts is connected to an external pump coupled with a flow rate sensor to precisely control the flow rate of buffer and analyte injection. The buffer is run continuously at a flow rate of 50 μl/min, and analytes are injected onto the sensor surface at 10 μl/min. For collecting data, one of the sensing elements in the middle fluidic channel and its neighboring mirror is chosen for sensing and referencing, respectively. The in-flow experiments are done by measuring the mirror once, followed by 20 times the sensing element in a continuous loop, with 32 scans per measurement. For the experiments described in Fig. 4 for the secondary structure analysis, measurements are done with 512 scans with repetitive loops of one mirror followed by four times the sensing element measurements.

FTIR data analysis and absorbance calculation

The continuous extraction of the protein absorbance fingerprints in the amide range (1500 to 1700 cm⁻¹) is carried out by normalizing the spectrally averaged reflectance spectra obtained following the protein injection (R) to that of the spectra of the reference sensor in the buffer in the previous functionalization state (R₀). Then, the differential absorbance spectrum in mOD is retrieved using the formula −1000*log10(R/R₀). We perform a baseline correction procedure on the obtained data before displaying the final absorbance spectra curves in the main and supplementary figures (see fig. S9). This is done by subtracting a second-order polynomial fitted to the absorption spectra in the range of 1450 to 1700 cm⁻¹. This extracted amide band absorbance is then integrated over the range 1525 to 1650 cm⁻¹ to output the integrated time absorbance plots used to monitor the binding kinetics.

Secondary structure analysis

Secondary structure analysis of the aSyn species shown in Fig. 4 follows the protocol from Yang et al. (56). The reflectance spectra are collected with 512 scans per measurement and 4-cm⁻¹ spectral resolution. Two hundred microliters of 5 μM aSyn fibrils, 5 μM aSyn oligomers, and 10 μM aSyn monomers are used in separate experiments for this analysis. The retrieved absorbance spectra after the stabilized protein binding are used to execute the secondary structure analysis. First, we performed a second-derivative analysis on the absorbance spectra to separate the overlapping bands without any execution bias. This is done by applying a second-order Savitzy-Golay filter with a seven-point calculation window. The number of subbands and their peak position data obtained from this analysis are then used to resolve the secondary structure information using FSD. We fitted the absorbance spectra as a linear combination of Lorentzian/Gaussian curves with the peak positions the same as the frequencies obtained from the second-derivative analysis through multiple iterations until convergence is reached. Last, the area under each curve is integrated to calculate each subband’s relative contribution, corresponding to the percentage presence of a particular secondary structure motif.

AI-DNN analysis

The spectra for the entire dataset used in AI-DNN analysis are obtained as mentioned in the “Infrared in situ measurements” and “FTIR data analysis and absorbance calculation” sections above. The stock concentration of the in-house prepared oligomers and fibrils is measured using bicinchoninic acid (BCA) assay and NanoDrop measurements. The mixtures of the aggregates of different percentage combinations used in the analysis are carefully titrated from these stock protein solutions with a fixed final mixture concentration of 5 μM. In total, we prepared seven mixtures with oligomers and fibrils in the following relative percentages of each (oligomers-fibrils): 100-0, 75-25, 60-40, 50-50, 40-60, 25-75, and 0-100%. We followed the same protocol as in the “Immunoassay protocol” section. After stabilizing the integrated sensorgram signal from antibody binding [200 μl, SYN-211 (40 μg/ml)] followed by washing, 200 μl of the aggregate mixture is injected. We take the stable absorbance of the antibody as a baseline and start measuring the spectra of the injected mixture and the corresponding binding on the surface. The absorbance spectra in the wave number range of 1450 to 1700 cm⁻¹ are continuously monitored until the mixture flows over the surface and is washed. The collected raw absorbance spectra are used without any spectral averaging (i.e., one spectral line shown in Fig. 5 for each mixture corresponds to one 32-scan measurement). For uniformity, we chose 350 such spectra for each mixture that starts from the time of injection and up to the point of stabilization. This gives us a dataset of each mixture with 350 spectra of different time points with 133 data points of absorbance over the wave number range of 1450 to 1700 cm⁻¹. For each mixture, every spectrum involved is labeled with the value that corresponds to the percentage presence of fibrils (Table 1). The absorbance plots retrieved and their labels (shown in Fig. 5) for each mixture are used as the entire training/testing dataset for the AI model. We created the AI-DNN model in Python language using the scikit-learn library. We created a DNN based on the MLP model as we treated this as a regression problem.

To match with the input data points, the DNN model has 133 input nodes corresponding to each wave number point that will take in/fed by the absorbance value for that particular wave number at a given time point and 1 output node to feed/output the corresponding label of the spectra. To optimize and tune the parameters of the model, like the number of hidden layers and the total number of nodes per hidden layer, we used the K-fold CV method. For this purpose, the entire dataset was split randomly into 80 and 20% (i.e., the total spectra of each mixture). The 80% of the thus collected dataset is then used for the training and K-fold CV. We executed a fivefold CV on the training dataset using the solver Adam and the activation function as logistic. We chose the loss function as the mean squared error for the optimization step to yield the appropriate parameters leading to the least mean squared error possible. The parameters of 4 hidden layers with 26 nodes each are obtained from this training step. The rest of the 20% of the dataset is then fed to this optimized MLP-DNN model [depth (h) = 5, width (d_m) = 26, and input dimension (n) = 133], and the predicted output labels from the network for each spectrum of every mixture ratios are plotted against the actual label (percentage concentration of fibrils) in the plot shown in Fig. 5D.

Multiplexing experiment

The sensor chip is incubated with a 2 mM thiol mixture of activated ester (HSC11EG4OCH2COONHS, ProChimia Surfaces) and spacer thiols (HS-C6-EG3OH, ProChimia Surfaces) in the ratio of 1:9 overnight. Afterward, the chip is washed with ethanol to wash away unbound thiols and dried. Without any delay, the chip is put in a petri dish and placed inside the tool sciFLEXARRAYER S3 from Scienion (piezoelectric noncontact ultralow-volume dispensing system). The dispensing system is set to be on humidity control at dew point (66% relative humidity) and internal temperature of 16°C to prevent evaporation of the spotted solution. The bioprinting solutions consist of 50 μg/ml each of SYN-211 and HT7 antibodies separately in 10 mM sodium acetate buffer. The middle sensor linear array is focused using the camera, and then one-half of the array containing four sensing elements and two mirrors are dispensed with multiple microarray spots of 450 pl of SYN-211 solution and the other half with HT7 solution, respectively. Water droplets are pipetted inside the petri dish surrounding the chip, and the lid is closed and then covered with paraffin film to prevent evaporation. Then, the chip is incubated at room temperature for approximately 2 hours. Afterward, the chip is washed with 1× PBST (PBS containing 0.1% Tween 20) solution, followed by water, and dried. Afterward, the chip is placed on the microfluidic chipcell and incorporated with the micro-flowcell to start the in-flow SEIRA measurements. To eliminate nonspecific binding, the sensor is blocked using 300 μl of 2× milk buffer [Pierce Clear Milk Blocking Buffer (10×), Thermo Fisher Scientific]. For the cross-reactivity experiments shown in Fig. 6, 5 μM of 200-μl aSyn fibrils or 5 μM of 200-μl tau fibrils were used. For the combination injection step in the experiment shown in Fig. 7, the 200-μl sample contained an equal amount of aSyn fibrils and tau fibrils, each at a total concentration of 5 μM.

In vitro preparation of aSyn structural species

Monomers

Recombinant overexpression and purification of human wild-type (WT) aSyn monomers were purified as previously reported (53). pT7-7 plasmids encoding human WT aSyn were used for transformation in BL21 (DE3) Escherichia coli cells on an ampicillin agar plate. A single colony was transferred to 400 ml of LB medium comprising ampicillin (100 μg/ml; AppliChem, A0839), followed by overnight incubation at 37°C and 180 rpm. As a next step, the preculture was used to inoculate 3 to 6 liters of LB medium, including ampicillin (100 μg/ml). The induction of aSyn protein expression was further performed upon A600 forthcoming 0.4 to 0.6 via adding 1 mM 1-thio-β-d-galactopyranoside (AppliChem, A1008). Next, the cells were incubated at 37°C and 180 rpm for 4 hours, followed by centrifugation at 4000 rpm at 5°C for 30 min, using JLA 8.1000 rotor (Beckman Coulter, Bear, CA, USA). The resulting pellets were stored at −20°C until further steps. The cell lysis was conducted by dissolving the resulting pellet in 20 mM Tris-HCl (pH 7.5) containing protease inhibitors [1 mM EDTA (Sigma-Aldrich, 11,873,580,001] and 1 mM phenylmethylsulfonyl fluoride (Applichem, A0999) (buffer A), which was ultrasonicated (VibraCell VCX130, Sonics, Newtown, CT, USA) using the following conditions: time: 5 min; cycle: 30 s ON, 30 s OFF; amplitude: 70%. As a next step, the samples were centrifugated for 30 min at 12,000 rpm and 4°C for 30 min, and the supernatant was collected in a 50-ml Falcon tube and further located in boiling water (100°C) for approximately 15 min. The protein sample was subsequently centrifuged similarly to the abovementioned conditions, and the resultant supernatant was filtered through 0.45-μm filters and injected into a sample loop connected to HiPrep Q FF 16/10 (GE Healthcare). The supernatant was injected at 2 ml/min and eluted using 20 mM tris-HCl, 1 M NaCl (pH 7.5) (buffer B) from gradient 0 to 70% at 3 ml/min. All fractions were analyzed by SDS–polyacrylamide gel electrophoresis (PAGE), and the positive pure aSyn was pooled. For the aSyn monomers prepared for using monomers as such and for the preparation of fibrils, further purification is done by reverse-phase high-performance liquid chromatography (Jupiter 300 C4, 20 mm I.D. × 250 mm, 10-μm average bead diameter, Phenomenex) and lyophilized. Toward the oligomer preparation, after the HiPrep test, the aSyn-positive samples are pooled and dialyzed against deionized water at 4°C overnight to remove salts and subsequently snap-frozen and lyophilized.

Unmodified oligomers

Unmodified WT aSyn oligomers were prepared as previously described (53). Briefly, 60 mg of lyophilized protein was dissolved in 5 ml of PBS buffer (pH 7.4) and incubated at 900 rpm with constant shaking at 37°C for 5 hours. The sample was centrifuged at 12,000g at 4°C for 10 min, aiming to remove any insoluble species. As a next step, the supernatant was loaded onto a HiLoad 26/600 Superdex 200 preparation grade (GE Lifesciences) column pre-equilibrated with PBS buffer (pH 7.4). Protein was eluted as 2.5-ml fractions at a flow rate of 1 ml/min. All fractions were further analyzed by SDS-PAGE, and the positive oligomer fractions (i.e., void volume peak) were snap-frozen and stored at −20°C for further analyses.

Fibrils

WT aSyn fibrils were similarly prepared as previously stated (8, 53). Briefly, 4 mg of aSyn monomers was diluted in 600 μl of PBS (pH 7.4), and the pH was adjusted to ~7.2 to 7.4. After, the monomeric aSyn solution was filtered through 0.2-μm filters (Merck, SLGP033RS) before being incubated under constant orbital agitation (1000 rpm) at 37°C for 5 days. The extent of the formation of fibrils was assessed by SDS-PAGE and negative-staining TEM (53). aSyn WT fibrils were further sonicated on ice with a fine tip (Sonics Vibra cell) for 20 s, at 20% amplitude, a pulse of 1 s ON/1 s OFF (Sonic Vibra-Cell, Blanc-Labo, Switzerland). The number of monomers and oligomers released from sonicated fibrils was quantified using the filtration protocol that we developed (53). Sonicated aSyn fibrils were characterized by TEM and further aliquoted, snap-frozen, and stored at −80°C until the subsequent analyses.

SDS-PAGE and Western blot analysis

For SDS-PAGE analysis, Human WT aSyn monomers, oligomers, and fibrils were mixed with 5 × Laemmli buffer and subsequently loaded onto 15% polyacrylamide gels (28, 53). The gel was run at 180 V for 1 hour, and the protein bands were visualized by Coomassie’s brilliant blue staining. Human WT aSyn monomers, oligomers, and fibril samples were also analyzed and characterized through Western blot (28, 53). The aSyn species samples were loaded and separated on 15% polyacrylamide gels and run under similar conditions as the abovementioned. Subsequently, the proteins were transferred onto a nitrocellulose membrane (Thermo Fisher Scientific, Switzerland) using a semi-dry system (Bio-Rad, Switzerland) at 25 V, 0.5 A, and 45 min. The transferred proteins were further blocked for 1 hour using Odyssey blocking buffer [Li-Cor, Lincoln, NE, USA, (P/N: 927-40000)] and incubated overnight at 4°C with the primary antibody SYN-1 (#610787, BD Biosciences). On the next day, the membranes were washed three times in 1× PBST for 10 min at room temperature and incubated with IR dye–conjugated secondary antibodies for 1 hour protected from light also at room temperature. After, the membranes were washed three times in a similar way as aforesaid. Last, protein bands were visualized by fluorescence imaging using the Odyssey CLx System (Li-Cor, NE, USA).

aSyn depletion protocol from healthy human CSF

The immunoprecipitation assay was performed using a Dynabeads Antibody coupling kit (Invitrogen, USA) with superparamagnetic Dynabeads M-270 Epoxy beads following the manufacturer’s instructions. The scheme of this depletion process is shown in fig. S9. Briefly, 2 to 5 μg of mouse antibody (#848302, BioLegend, USA) was mixed with 1 mg of magnetic beads and incubated overnight at 37°C. On the next day, 1 ml of crude Human CSF samples was thawed on ice, followed by the addition of protease and phosphatase inhibitors. As a first step, a preclearing of the CSF sample was conducted to decrease/deplete the levels of endogenous immunoglobulin Gs (IgGs) by mixing the CSF with Pierce Protein G Magnetic Beads (Thermo Fisher Scientific, USA) for 2 hours at 4°C. The resulting antibody-conjugated epoxy beads were mixed with the IgG depleted CSF sample and incubated overnight on a rocking platform at 4°C. As a next step, the magnetic beads/sample solution was transferred to a magnetic particle processor, and the supernatant was collected (unbound fraction), which is the aSyn-depleted CSF used in the experiment shown in Fig. 8 as the complex biomatrix. The antibody-conjugated epoxy beads were washed twice with 1× PBST and once with PBS (pH 7.4). aSyn was eluted by adding a solution based on 50% acetonitrile/50%water/0.1% TFA (IP sample). The IP’ed CSF sample was dried in a SpeedVac and resuspended in PBS (pH 7.4).

Supplementary Materials

This PDF file includes:

Supplementary Text

Figs. S1 to S19