LSPR-susceptible metasurface platform for spectrometer-less and AI-empowered diagnostic biomolecule detection
| dc.contributor.author | Li, Jinke | en |
| dc.contributor.author | Kim, Jin Tae | en |
| dc.contributor.author | Li, Hongliang | en |
| dc.contributor.author | Cho, Hyo Young | en |
| dc.contributor.author | Kim, Jin Soo | en |
| dc.contributor.author | Choi, Duk Yong | en |
| dc.contributor.author | Wang, Chenxi | en |
| dc.contributor.author | Lee, Sang Shin | en |
| dc.date.accessioned | 2025-05-23T07:25:42Z | |
| dc.date.available | 2025-05-23T07:25:42Z | |
| dc.date.issued | 2024-10-16 | en |
| dc.description.abstract | In response to the growing demand for biomolecular diagnostics, metasurface (MS) platforms based on high-Q resonators have demonstrated their capability to detect analytes with smart data processing and image analysis technologies. However, high-Q resonator meta-atom arrays are highly sensitive to the fabrication process and chemical surface functionalization. Thus, spectrum scanning systems are required to monitor the resonant wavelength changes at every step, from fabrication to practical sensing. In this study, we propose an innovative dielectric resonator-independent MS platform that enables spectrometer-less biomolecule detection using artificial intelligence (AI) at a visible wavelength. Functionalizing the focused vortex MS to capture gold nanoparticle (AuNP)-based sandwich immunoassays causes the resulting vortex beam profiles to be significantly affected by the localized surface plasmon resonance (LSPR) occurring between AuNPs and meta-atoms. The convolutional neural network algorithm was carefully trained to accurately classify the AuNP concentration-dependent focused vortex beam, facilitating the determination of the concentration of the targeted diagnostic biomolecule. Successful in situ identification of various biomolecule concentrations was achieved with over 99 % accuracy, indicating the potential of combining an LSPR-susceptible MS platform and AI for continuously tracking various chemical and biological compounds. | en |
| dc.description.sponsorship | The FVM was constructed with eight carefully chosen nanopillars crafted from hydrogenated amorphous silicon (a-Si:H) (Section S1, Supporting Information). This selection was driven by the higher bandgap energy of the material compared with silicon in the visible spectrum, effectively enhancing the efficiency of silicon-type dielectric MSs [40,41]. Furthermore, the material exhibits a high RI and is cost-effective [42]. The FVMs were fabricated through electron beam lithography (EBL) (Methods section). The desired phase profile of the FVM is described by [Formula presented], where (x,y) represents the coordinates corresponding to the center of each unit cell in the middle of the FVM, l0 is the topological charge of the vortex, and f is the focal length. Moreover, adjustments were made by setting the values of f and l0 to 200 \u03BCm and +5, respectively, to address focusing distortion [43] and facilitate a sufficiently large vortex beam spot for vision intelligence-based imaging identification [44]. The fabricated device surface underwent chemical functionalization by capturing the antibody through liquid-phase amination using (3-Aminopropyl)triethoxysilane (APTES) and glutaraldehyde (GA), followed by immobilization of the capture anti-CRP (see Methods Section) [45]. Subsequently, sandwich immunoassays were constructed by incubating the surface with different concentrations of CRP, followed by incubation with 10-nm AuNPs conjugated with anti-CRP (see Methods Section and Section S2, Supporting Information). Finally, the anti-CRP/CRP/anti-CRP/AuNP sandwich immunoassays were formed on the device surface. Fig. 2(a) shows the attachment of anti-CRP/CRP/anti-CRP/AuNP sandwich immunoassays (depicted by white dots) on the silica substrate, arranged around FVM, at four distinct concentrations (0, 1, 2, and 4 \u03BCg/mL). Given that each sandwich immunoassay complex incorporates an AuNP, the examination of targeted diagnostic molecules could be facilitated by quantifying the number of AuNPs present. In Fig. 2(b), the successful adherence of anti-CRP/CRP/anti-CRP/AuNP sandwich immunoassays on the fabricated FVMs is confirmed.The original dataset was passed through the preprocessing steps, including segmentation, data scrubbing, data conversion, and outlier removal, to enhance the predictability of ResNet. A total of 120 images were captured for each concentration, with 80% of the images allocated for training and the remaining 20% for validation. The criterion for optimizing data classification was the cross-entropy loss. In Fig. 5(b), the cross-entropy losses and precision rates are plotted in relation to training epochs. The Adam Optimizer was employed to improve the optimization criterion. The training and validation error rates were 5.35% and 1.02%, respectively, with corresponding accuracies of 99.61% and 100%. Fig. 5(c) showcases the classification matrix depicting the evaluation patterns. The true and CNN-predicted biomolecular concentrations are performed in the vertical and horizontal axes, respectively. A value of 1 in a diagonal cell indicates 100% classification accuracy for the specific concentration from 0 to 4 \u03BCg/mL. The confusion matrix demonstrates precise recognition across all concentrations, achieving an average of near-complete accuracy while actively identifying in real time. Furthermore, the dataset, centered around incident light at \u03BB = 650 nm, was used for training, and the trained algorithm demonstrated an average accuracy of 99.74% in identifying biomolecule concentrations (Section S6, Supporting Information). For comparison with a CNN-based image analysis, a principal component analysis (PCA) (Section S7, Supporting Information), a dimensionality reduction technique often used in image processing to reduce the number of features while retaining the most important information [58,59], was performed. First, each image was converted into a one-dimensional vector and the dataset was standardized. The covariance matrix of the standardized data, representing the variance between different features, was subsequently calculated. Finally, the eigenvalues and eigenvectors of the covariance matrix were computed. The eigenvectors indicate the directions of the maximum variance (principal components), while the eigenvalues correspond to the magnitude of variance in the corresponding directions. This procedure revealed that in the spectral regime at \u03BB = 550 nm, the vortex beams were successfully classified based on the PCA to precisely observe the CRP concentration. Meanwhile, for \u03BB = 650 nm, the donut-shaped vortex beam patterns showed too much correlation to allow for the desired classification.This work was supported by the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. 2018R1A6A1A03025242), Ministry of Science and ICT (2020R1A2C3007007), the research grant of Kwangwoon University in 2024, and Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean Government (21YR2710, Nanophotonic vision intelligence for airborne virus detection). This work was supported by the National Research Foundation of Korea (NRF) funded by the Ministry of Education (No. 2018R1A6A1A03025242), Ministry of Science and ICT (2020R1A2C3007007), the research grant of Kwangwoon University in 2024, and Electronics and Telecommunications Research Institute (ETRI) grant funded by the Korean Government (21YR2710, Nanophotonic vision intelligence for airborne virus detection). | en |
| dc.description.status | Peer-reviewed | en |
| dc.identifier.issn | 0003-2670 | en |
| dc.identifier.other | PubMed:39260911 | en |
| dc.identifier.other | ORCID:/0000-0002-5339-3085/work/184100568 | en |
| dc.identifier.scopus | 85202059378 | en |
| dc.identifier.uri | http://www.scopus.com/inward/record.url?scp=85202059378&partnerID=8YFLogxK | en |
| dc.identifier.uri | https://hdl.handle.net/1885/733751738 | |
| dc.language.iso | en | en |
| dc.rights | Publisher Copyright: © 2024 Elsevier B.V. | en |
| dc.source | Analytica Chimica Acta | en |
| dc.subject | Artificial intelligence | en |
| dc.subject | Diagnostic biomolecule detection | en |
| dc.subject | Gold nanoparticle | en |
| dc.subject | Localized surface plasmon resonance | en |
| dc.subject | Metasurface | en |
| dc.title | LSPR-susceptible metasurface platform for spectrometer-less and AI-empowered diagnostic biomolecule detection | en |
| dc.type | Journal article | en |
| dspace.entity.type | Publication | en |
| local.contributor.affiliation | Li, Jinke; Kwangwoon University | en |
| local.contributor.affiliation | Kim, Jin Tae; Electronics and Telecommunications Research Institute | en |
| local.contributor.affiliation | Li, Hongliang; Kwangwoon University | en |
| local.contributor.affiliation | Cho, Hyo Young; Electronics and Telecommunications Research Institute | en |
| local.contributor.affiliation | Kim, Jin Soo; Korea University | en |
| local.contributor.affiliation | Choi, Duk Yong; Department of Quantum Science & Technology, Research School of Physics, ANU College of Science and Medicine, The Australian National University | en |
| local.contributor.affiliation | Wang, Chenxi; Kwangwoon University | en |
| local.contributor.affiliation | Lee, Sang Shin; Kwangwoon University | en |
| local.identifier.citationvolume | 1326 | en |
| local.identifier.doi | 10.1016/j.aca.2024.343094 | en |
| local.identifier.pure | fd078db3-017f-4d26-be5a-25c77dfa8708 | en |
| local.identifier.url | https://www.scopus.com/pages/publications/85202059378 | en |
| local.type.status | Published | en |