Alcnc-enhanced fluorescence meets machine learning classification

Abstract

Method and apparatus for detecting and differentiating neurotransmitters using ultraviolet plasmonic-engineered native fluorescence. In one example, the method includes determining a photobleaching rate constant of a neurotransmitter-containing analyte loaded onto a plasmonic-engineered biosensor and subjected to illumination by ultraviolet light. The method further includes submitting a query containing the determined rate constant to a database including calibration data representing a plurality of different neurotransmitters and a plurality of different biosensors. In at least some examples, the queried database returns a response indicating one or more of a predicted identity of the neurotransmitter together with a corresponding confidence score, an estimated amount of the neurotransmitter in the analyte, and an estimated percentage of the neurotransmitter relative to another neurotransmitter in the analyte.

Description

[0001] 026389-0063-W001

[0002] ALCNC-ENHANCED FLUORESCENCE MEETS MACHINE LEARNING CLASSIFICATION

[0003] CROSS REFERENCE TO RELATED APPLICATIONS

[0004] This application claims priority to U.S. Provisional Patent Application No. 63 / 748,162, filed on January 22, 2025, which is incorporated by reference herein in its entirety.

[0005] This application is related to U.S. Pat. App. No. 18 / 774,407, filed on July 16, 2024, which is incorporated by reference herein in its entirety.

[0006] BACKGROUND

[0007] The precise quantification and identification of monoamine neurotransmitters (MANTs) play a pivotal role in understanding neurological processes and in the early diagnosis of neurodegenerative diseases. Conventional analytical methods such as chromatography and mass spectrometry require complex sample preparation and expensive reagents, unsuitable for frequent assessments of MANT levels. Electrochemical approaches such as the Fast-Scan Cyclic Voltammetry (FSCV) method are cost-effective alternatives; however, they cannot distinguish MANTs with similar oxidation potentials. Genetically encoded fluorescence probes are able to detect MANTs with excellent sensitivity and specificity; however, they require the use of transgenic animals. Antibody or aptamer-based assays have demonstrated real-time sensing of MANTs. However, the long-term stability of these probes in biological fluids remains an issue.

[0008] MANTs have an aromatic ring structure that emits autofluorescence (AF) when excited by ultraviolet (UV) light. The AF absorption cross-sections are orders of magnitude higher than that of Raman or infrared absorption. Therefore, AF spectroscopy is a promising technique for sensitive, label-free, and probe-free quantification and identification of MANTs. However, the classification of similar MANTs based on their AF profile is challenging due to their overlapping spectrum. Prior work has shown that the AF of MANTs drop cast on a solid substrate (e.g., a silicon wafer or a plasmonic nano hole array, etc.) decays exponentially over time when continuously exposed to UV light. The decay rate constants were found to be distinct among similarly structured MANTs, and their differences were enlarged by a UV plasmonic nano holearray. Our prior work focused solely on the decay rate constants and has not realized the full potential of using UV plasmonic-engineered Auto Fluorescence Time Decay Series (AFTDS) in classifying MANTs. In this paper, we demonstrated excellent classification accuracy by combining artificial intelligence with the AFTDS of MANTs deposited on aluminum plasmonic concave nanocubes (AICNC). The assembly of concave cubes was reproducibly obtained by026389-0063-W001

[0009] drop casting a droplet of nanoparticle solution containing AICNCs in ambient conditions, offering a cost-effective and nanofabrication-free way to form a large area of plasmonic substrates.

[0010] Machine learning has increasingly become a powerful tool in nanoscale science, surface chemistry, and biosensors, enabling advances in fabrication, characterization, and property prediction by integrating experimental data with physics-informed models.

[0011] What is needed are methods for detecting and differentiating neurotransmitters using ultraviolet plasmonic-engineered and machine learning methods.

[0012] SUMMARY

[0013] One embodiment described herein is a system for detecting monoamine neurotransmitters, comprising: a plasmonic substrate comprising aluminum concave nanocubes; an ultraviolet light source configured to illuminate the plasmonic substrate; a spectrometer configured to collect fluorescence data from the surface of the plasmonic substrate, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired over a time period; and a processor configured to execute a machine learning model trained to classify the monoamine neurotransmitters based on the fluorescence data. In one aspect, the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers. In another aspect, the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 200-300 nanometers. In another aspect, the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 266 nanometers. In another aspect, the wherein the machine learning model comprises a Long Short-Term Memory network. In another aspect, the Long Short-Term Memory network comprises: three sequential Long Short-Term Memory layers; batch normalization layers following each of the three sequential Long Short-Term Memory layers; a fully connected dense layer; and a softmax output layer configured to classify the monoamine neurotransmitters into a plurality of classes. In another aspect, the plurality of classes comprises dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In another aspect, the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers.

[0014] Another embodiment described herein is a method for classifying monoamine neurotransmitters, comprising: depositing a sample containing monoamine neurotransmitters onto a plasmonic substrate comprising aluminum concave nanocubes; illuminating the plasmonic substrate with ultraviolet light; collecting fluorescence data comprising a plurality of fluorescence spectra over a time period during continuous illumination; preprocessing the fluorescence data to generate standardized input data; and classifying the monoamine neurotransmitters by applying026389-0063-W001

[0015] a trained machine learning model to the standardized input data. In one aspect, the sample comprises a solution and depositing the sample comprises drop casting the solution onto the plasmonic substrate to form a coffee ring pattern. In another aspect, illuminating the plasmonic substrate comprises directing the ultraviolet light toward an outer edge of the coffee ring pattern where a concentration of the monoamine neurotransmitters is higher. In another aspect, collecting the fluorescence data comprises acquiring the plurality of fluorescence spectra at intervals of approximately 0.5 seconds over a time period of 2 to 3 minutes. In another aspect, each fluorescence spectrum of the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers. In another aspect, preprocessing the autofluorescence time decay series data comprises: shifting time series data to eliminate leading missing values; truncating or padding the time series data to a standardized sequence length; and normalizing feature values using MinMax scaling. In another aspect, the standardized sequence length comprises 10-25 time steps. In another aspect, the wherein the trained machine learning model comprises a Long Short-Term Memory network.

[0016] Another embodiment described herein is a non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to: receive fluorescence data collected from a sample comprising one or more monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired during continuous ultraviolet illumination; preprocess the fluorescence data by aligning time series traces and standardizing sequence lengths to generate input tensors; apply a trained neural network to the input tensors to capture temporal dependencies in the autofluorescence time decay series data; and output a classification of the monoamine neurotransmitters based on analysis by the trained neural network. In one aspect, the trained neural network is a trained Long Short Term Memory network. In another aspect, the instructions further cause the processor to normalize feature values of the autofluorescence time decay series data to a range of 0 to 1 using MinMax scaling prior to applying the trained Long Short-Term Memory network. In another aspect, preprocessing the autofluorescence time decay series data comprises shifting time series traces to eliminate leading missing values by aligning valid entries to a common starting point. In another aspect, the standardized sequence lengths comprise 17 time steps, and wherein the instructions further cause the processor to truncate longer sequences to 17 regularly spaced sequences or pad shorter sequences to achieve the standardized sequence lengths. In another aspect, the trained Long Short-Term Memory network comprises: three sequential Long Short-Term Memory layers having 200, 150, and 100 units respectively; batch normalization layers following each of the three026389-0063-W001

[0017] sequential Long Short-Term Memory layers; a fully connected dense layer with 50 neurons; and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In another aspect, the instructions further cause the processor to process a single input tensor in under 1 millisecond to enable real-time classification of the monoamine neurotransmitters.

[0018] Another embodiment described herein is a biosensing device for label-free detection and classification of monoamine neurotransmitters, comprising: a substrate having a surface with aluminum concave nanocubes deposited thereon; an excitation source configured to direct ultraviolet light onto the substrate to excite native autofluorescence from monoamine neurotransmitters in contact with the aluminum concave nanocubes; a detector configured to acquire time-resolved fluorescence spectra from the monoamine neurotransmitters over a measurement period to generate autofluorescence time decay series data; and a classification module comprising a processor and a memory storing a trained recurrent neural network, wherein the classification module is configured to receive the autofluorescence time decay series data and output an identification of the monoamine neurotransmitters based on temporal decay characteristics captured by the trained recurrent neural network. In one aspect, the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers with a standard deviation of approximately 9 nanometers. In another aspect, the trained recurrent neural network comprises a Long Short-Term Memory network having three sequential Long Short-Term Memory layers with batch normalization layers following each of the three sequential Long Short-Term Memory layers. In another aspect, the Long Short-Term Memory network further comprises a fully connected dense layer and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In another aspect, the excitation source comprises a continuous wave solid state ultraviolet laser configured to emit light at a wavelength of approximately 266 nanometers, and wherein the detector is configured to acquire the time-resolved fluorescence spectra at intervals of approximately 0.5 seconds over a measurement period of 2 to 3 minutes.

[0019] DESCRIPTION OF THE DRAWINGS FIG. 1 shows a schematic for the approach for classifying monoamine neurotransmitters by applying machine learning on UV plasmonic-engineered auto fluorescence time decay series (AFTDS).

[0020] FIG. 2 shows an illustration of data indexing adjustment during preprocessing.026389-0063-W001

[0021] FIG. 3 shows a workflow for classifying DA, DOPAC, and NE after preprocessing UV autofluorescence raw data.

[0022] FIG. 4 shows a plot illustrating the training and validation curves of the LSTM model, illustrating model accuracy (blue) and model loss (red) over 50 training epochs.

[0023] FIG. 5A-F show the characterization of aluminum concave nanocubes. FIG. 5A shows a coffee ring formed by evaporating a drop of deionized water containing molecules on AICNCs. FIG. 5B shows extinction spectrum of AICNC and FIG. 5C shows an EDS line scan profile over a selected region of an AICNC, illustrating the distribution of Al and O, with the inset displaying the scan area superimposed on an AICNC. FIG. 5D shows STEM images highlight the morphology and concave geometry of the AICNCs, with oxide layer thickness estimated to be 6 ± 2 nm. FIG.

[0024] 5E-F show EDS elemental mapping images showing the spatial distribution of Al (cyan) and the oxide layer (red) on the AICNC surfaces.

[0025] FIG. 6A shows a SEM image of the coffee ring pattern formed by the molecular solution. FIG. 6B shows a borderline of the coffee ring pattern. FIG. 6C shows STEM image of an individual AICNC particle.

[0026] FIG. 7 shows a graph of absorption vs. concentration and data fits for DA, DOPAC, and NE in deionized water.

[0027] FIG. 8A-C shows fluorescence spectra of DA (FIG. 8A), DOPAC (FIG. 8B), and NE (FIG.

[0028] 8C) in five different concentrations in DI water.

[0029] FIG. 9A-C shows absorption spectra of DA (FIG. 9A), DOPAC (FIG. 9B), and NE (FIG.

[0030] 9C) in five different concentrations in DI water.

[0031] FIG. 10A-C show AFTDS for three neurotransmitters DA (FIG. 10A), DOPAC (FIG. 10B), and NE (FIG. 10C) drop cast and dried on AICNC substrates (1 pL of 500 pM). FIG. 5D-E show fluorescence intensity of DA, DOPAC, and NE (1 pL of 500 pM) collected at 0 to 0.5 seconds on a silicon wafer (FIG. 10D) and AICNC (FIG. 10E) substrates. FIG. 10F shows fluorescence spectra of DA, DOPAC, and NE solution dissolved in water at a concentration of 50 pM measured by a fluorometer.

[0032] FIG. 11A shows integrated fluorescence intensity and FIG. 11 B shows net enhancement factors from three neurotransmitters: DA, DOPAC, and NE, comparing signal levels between AF data on silicon wafers and AICNCs.

[0033] FIG. 12A-C shows AFTDS spectra of DA (FIG. 12A), DOPAC (FIG. 12B), and NE (FIG.

[0034] 12C) on ALCNC in three different concentrations.

[0035] FIG. 13A-E shows confusion matrices and Table 4 shows both percentage and sample count per classification outcome.026389-0063-W001

[0036] FIG. 14A-E show confusion matrices for LSTM, KNN and RF classifiers distinguishing DA, DOPAC and NE from AFTDS and AF in-solution data. FIG. 14A-C (AFTDS) retain high diagonal accuracies, whereas the accuracies in FIG. 14D-E (AF in-solution) drop sharply, especially for DA and NE, signaling greater inter-class confusion. Cell values are percentages; diagonals denote correct predictions, and off-diagonals indicate misclassifications.

[0037] FIG. 15A-C shows bar charts of Precision (FIG. 15A), Recall (FIG. 15A), and F1 Score (FIG. 15A) for DA, NE, DOPAC, Weighted Average and Macro Average.

[0038] FIG. 16A-B shows schematic diagrams illustrating an optical system with which various embodiments can be practiced.

[0039] FIG. 17 shows a flowchart illustrating a method of detecting and differentiating neurotransmitters according to some examples.

[0040] FIG. 18 shows a block diagram illustrating a computing device used according to some examples.

[0041] DETAILED DESCRIPTION

[0042] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. For example, any nomenclatures used in connection with, and techniques of biochemistry and computer science described herein are well known and commonly used in the art. In case of conflict, the present disclosure, including definitions, will control the interpretation of terms. Exemplary methods and materials are described below, although methods and materials similar or equivalent to those described herein can be used in practice or testing of the embodiments and aspects described herein.

[0043] As used herein, terms such as “include,” “including,” “contain,” “containing,” “having,” and the like mean “comprising.” The present disclosure also contemplates other embodiments “comprising,” “consisting essentially of,” and “consisting of” the embodiments or elements presented herein, whether explicitly set forth or not. As used herein, “comprising,” is an “open-ended” term that does not exclude additional, unrecited elements or method steps. As used herein, “consisting essentially of” limits the scope of a claim to the specified materials or steps and those that do not materially affect the basic and novel characteristics of the claimed invention. As used herein, “consisting of” excludes any element, step, or ingredient not specified in the claim.

[0044] As used herein, the term “a,” “an,” “the” and similar terms used in the context of the disclosure (especially in the context of the claims) are to be construed to cover both the singular026389-0063-W001

[0045] and plural unless otherwise indicated herein or clearly contradicted by the context. In addition, “a,” “an,” or “the” means “one or more” unless otherwise specified.

[0046] As used herein, the term “or” can be conjunctive or disjunctive.

[0047] As used herein, the term “and / or” refers to both the conjunctive and disjunctive.

[0048] As used herein, the term “substantially” means to a great or significant extent, but not completely.

[0049] As used herein, the term “about” or “approximately” as applied to one or more values of interest, refers to a value that is similar to a stated reference value, or within an acceptable error range for the particular value as determined by one of ordinary skill in the art, which will depend in part on how the value is measured or determined, such as the limitations of the measurement system. In one aspect, the term “about” refers to any values, including both integers and fractional components that are within a variation of up to ± 10% of the value modified by the term “about.” Alternatively, “about” can mean within 3 or more standard deviations, per the practice in the art. Alternatively, such as with respect to biological systems or processes, the term “about” can mean within an order of magnitude, in some embodiments within 5-fold, and in some embodiments within 2-fold, of a value. As used herein, the symbol means “about” or “approximately.” All ranges disclosed herein include both end points as discrete values as well as all integers and fractions specified within the range. For example, a range of 0.1-2.0 includes 0.1, 0.2, 0.3, 0.4 . . . 2.0. If the end points are modified by the term “about,” the range specified is expanded by a variation of up to ±10% of any value within the range or within 3 or more standard deviations, including the end points, or as described above in the definition of “about.”

[0050] As used herein, the terms “room temperature,” “RT,” or “ambient temperature” refer to the typical temperature in an indoor laboratory setting. In one aspect, the laboratory setting is climate controlled to maintain the temperature at a substantially uniform temperature or with a specific range of temperatures. In one aspect, “room temperature” refers a temperature of about 15-30 °C, including all integers and endpoints within the specified range. In another aspect, “room temperature” refers a temperature of about 15-30 °C; about 20-30 °C; about 22-30 °C; about 25-30 °C; about 27-30 °C; about 15-22 °C; about 15-25 °C; about 15-27 °C; about 20-22 °C; about 20-25 °C; about 20-27 °C; about 22-25 °C; about 22-27 °C; about 25-27 °C; about 15 °C ± 10%; about 20 °C ± 10%; about 22 °C ± 10%; about 25 °C ± 10%; about 27 °C ± 10%; ~20 °C, ~22 °C, ~25 °C, or ~27 °C, at standard atmospheric pressure.

[0051] As used herein, the terms “control,” or “reference” are used herein interchangeably. A “reference” or “control” level may be a predetermined value or range, which is employed as a026389-0063-W001

[0052] baseline or benchmark against which to assess a measured result “Control” also refers to control experiments.

[0053] As used herein, the use of the ordinal adjectives “first,” “second,” “third,” etc., to refer to an object of a plurality of like objects merely indicates that different instances of such like objects are being referred to, and is not intended to imply that the like objects so referred-to have to be in a corresponding order or sequence, either temporally, spatially, in ranking, or in any other manner.

[0054] As used herein, in addition to its plain meaning, the conjunction “if” may also or alternatively be construed to mean “when” or “upon” or “in response to determining” or “in response to detecting,” which construal may depend on the corresponding specific context. For example, the phrase “if it is determined” or “if [a stated condition] is detected” may be construed to mean “upon determining” or “in response to determining” or “upon detecting [the stated condition or event]” or “in response to detecting [the stated condition or event].”

[0055] As used herein, the terms “couple,” “coupling,” “coupled,” “connect,” “connecting,” or “connected” refer to any manner known in the art or later developed in which energy is allowed to be transferred between two or more elements, and the interposition of one or more additional elements is contemplated, although not required. Conversely, the terms “directly coupled,” “directly connected,” etc., imply the absence of such additional elements.

[0056] The functions of the various elements shown in the figures, including any functional blocks labeled as “processors” and / or “controllers,” may be provided through the use of dedicated hardware as well as hardware capable of executing software in association with appropriate software. When provided by a processor, the functions may be provided by a single dedicated processor, by a single shared processor, or by a plurality of individual processors, some of which may be shared. Moreover, explicit use of the term “processor” or “controller” should not be construed to refer exclusively to hardware capable of executing software, and may implicitly include, without limitation, digital signal processor (DSP) hardware, network processor, application specific integrated circuit (ASIC), field programmable gate array (FPGA), read only memory (ROM) for storing software, random access memory (RAM), and non volatile storage. Other hardware, conventional and / or custom, may also be included. Similarly, any switches shown in the figures are conceptual only. Their function may be carried out through the operation of program logic, through dedicated logic, through the interaction of program control and dedicated logic, or even manually, the particular technique being selectable by the implementer as more specifically understood from the context.026389-0063-W001

[0057] As used herein, the terms “circuit,” “circuitry” may refer to one or more or all of the following: (a) hardware-only circuit implementations (such as implementations in only analog and / or digital circuitry); (b) combinations of hardware circuits and software, such as (as applicable): (i) a combination of analog and / or digital hardware circuit(s) with software / firmware and (ii) any portions of hardware processor(s) with software (including digital signal processor(s)), software, and memory(ies) that work together to cause an apparatus, such as a mobile phone or server, to perform various functions); and (c) hardware circuit(s) and or processor(s), such as a microprocessor(s) or a portion of a microprocessor(s), that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation.” This definition of circuitry applies to all uses of this term in this application, including in any claims. As a further example, as used in this application, the term circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and / or firmware. The term circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in server, a cellular network device, or other computing or network device.

[0058] It should be appreciated by those of ordinary skill in the art that any block diagrams herein represent conceptual views of illustrative circuitry embodying the principles of the disclosure. Similarly, it will be appreciated that any flow charts, flow diagrams, state transition diagrams, pseudo code, and the like represent various processes which may be substantially represented in computer readable medium and so executed by a computer or processor, whether or not such computer or processor is explicitly shown.

[0059] Described herein is a hybrid approach integrating advanced plasmonic nanomaterials and machine learning (ML) for high-precision biomolecule detection. Aluminum concave nanocubes (AICNCs) are used as an innovative plasmonic substrate to enhance the native fluorescence of neurotransmitters, including dopamine (DA), norepinephrine (NE), and 3,4-dihydroxyphenylacetic acid (DOPAC). AICNCs amplify weak fluorescence signals, enabling probe-free, label-free detection and differentiation of these molecules with great sensitivity and specificity. To further improve classification accuracy, ML algorithms were empolyed, with Long Short-Term Memory (LSTM) networks playing a central role in analyzing time-dependent fluorescence data. Comparative evaluations with / r-nearest neighbors (KNN) and Random Forest (RF) demonstrate the superior performance of LSTM in distinguishing neurotransmitters. The results reveal that AICNC substrates provide up to a 12-fold enhancement in fluorescence intensity for DA, 9-fold for NE, and 7-fold for DOPAC compared to silicon substrates. At the same time, ML algorithms026389-0063-W001

[0060] achieve classification accuracy exceeding 89%. This methodology bridges the gap between nanotechnology and ML, showcasing the synergistic potential of AICNC-enhanced native fluorescence and ML in biosensing. The framework paves the way for probe-free, label-free biomolecule profiling, offering transformative implications for biomedical diagnostics and neuroscience research. Similar methods without machine learning or aluminum concave nanocubes have been described in U.S. Pat. App. Pub. No. US 2025 / 0027876 A1 , which is incorporated by reference herein for such teachings

[0061] While ML has been applied to Raman and fluorescence spectroscopy for biochemical sensing, this study is the first application of ML for the classification of neurotransmitters based on AFTDs. Comparative analysis were performed using three ML techniques - Long Short-Term Memory (LSTM), -nearest neighbors (KNN), and Random Forest (RF). LSTM on AFTDS collected on AICNCs achieved the highest classification accuracy, followed by a slightly poorer performance by KNN and RF on AFTDS. KNN and RF on AF collected in the solution phase without plasmonic nanoparticles achieved poorer performances. These results emphasize the importance of plasmonic-engineered AFTDS in achieving high classification accuracy among similarly structured MANTs. In addition, the superiority of LSTM over KNN and RF in analyzing time-dependent AF data was demonstrated.

[0062] Monoamine neurotransmitters (MANTs) may play a role in neurological processes and in the diagnosis of neurodegenerative diseases. Conventional analytical methods such as chromatography and mass spectrometry may involve complex sample preparation and expensive reagents. Electrochemical approaches such as Fast-Scan Cyclic Voltammetry (FSCV) may provide cost-effective alternatives; however, such approaches may have difficulty distinguishing MANTs with similar oxidation potentials. Genetically encoded fluorescence probes may detect MANTs with sensitivity and specificity; however, such probes may involve the use of transgenic animals. Antibody or aptamer-based assays may demonstrate real-time sensing of MANTs; however, the long-term stability of such probes in biological fluids may present challenges.

[0063] MANTs have an aromatic ring structure that emits autofluorescence (AF) when excited by ultraviolet (UV) light. The AF absorption cross-sections may be orders of magnitude higher than that of Raman or infrared absorption. Therefore, AF spectroscopy may provide a technique for sensitive, label-free, and probe-free quantification and identification of MANTs. However, classification of similar MANTs based on AF profiles may be challenging due to overlapping spectra among structurally similar molecules.

[0064] The present disclosure describes systems and methods that may address these challenges by combining plasmonic nanomaterials with machine learning (ML) for biomolecule026389-0063-W001

[0065] detection. In some cases, aluminum concave nanocubes (AICNCs) may serve as a plasmonic substrate to enhance the native fluorescence of neurotransmitters. The AICNCs may amplify weak fluorescence signals, enabling probe-free, label-free detection and differentiation of molecules such as DA, NE, and DOPAC.

[0066] In some cases, the autofluorescence of MANTs deposited on a solid substrate may decay exponentially over time when continuously exposed to UV light. The decay rate constants may be distinct among similarly structured MANTs, and differences in decay characteristics may be enlarged by UV plasmonic nanostructures. The present disclosure describes approaches that may leverage autofluorescence time decay series (AFTDS) data in classifying MANTs.

[0067] To improve classification accuracy, ML algorithms may be employed, with Long Short-Term Memory (LSTM) networks playing a role in analyzing time-dependent fluorescence data. In some cases, comparative evaluations with k-nearest neighbors (KNN) and Random Forest (RF) may demonstrate performance characteristics of LSTM in distinguishing neurotransmitters. The AICNC substrates may provide fluorescence enhancement compared to non-plasmonic substrates such as silicon substrates, while ML algorithms may achieve classification accuracy for distinguishing between DA, NE, and DOPAC.

[0068] The assembly of concave nanocubes may be reproducibly obtained by drop casting a droplet of nanoparticle solution containing AICNCs in ambient conditions, offering a cost-effective approach to form plasmonic substrates without complex nanofabrication techniques. The combination of plasmonic-engineered AFTDS and ML may provide a framework for probe-free, label-free biomolecule profiling with applications in biomedical diagnostics and neuroscience research.

[0069] Systems

[0070] A system for classifying monoamine neurotransmitters is described herein. The system may include a plasmonic substrate comprising aluminum concave nanocubes configured to enhance autofluorescence signals of monoamine neurotransmitters deposited thereon. In some cases, the aluminum concave nanocubes may provide localized surface plasmon resonance (LSPR) effects in the ultraviolet range, which may amplify both excitation and fluorescent emission near the nanocube surface. The plasmonic substrate may receive samples containing monoamine neurotransmitters such as dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0071] The system may include an ultraviolet light source configured to illuminate the plasmonic substrate to excite autofluorescence from the monoamine neurotransmitters. In some cases, the026389-0063-W001

[0072] ultraviolet light source may direct ultraviolet light onto the plasmonic substrate, causing the aromatic ring structures of the monoamine neurotransmitters to emit native autofluorescence. The ultraviolet light source may provide continuous illumination to the plasmonic substrate during data acquisition.

[0073] The system may further include a spectrometer configured to collect autofluorescence time decay series data from the monoamine neurotransmitters under continuous illumination by the ultraviolet light source. The autofluorescence time decay series data may comprise a plurality of fluorescence spectra acquired over a time period. In some cases, the spectrometer may be coupled with a detector such as a charge-coupled device (CCD) camera to capture the fluorescence spectra. The plurality of fluorescence spectra may capture the temporal evolution of the autofluorescence signal as the fluorescence intensity decays over time during continuous ultraviolet illumination.

[0074] The system may include a processor configured to execute a machine learning model trained to classify the monoamine neurotransmitters based on the autofluorescence time decay series data. The machine learning model may comprise a Long Short-Term Memory (LSTM) network configured to capture temporal dependencies in the autofluorescence time decay series data. In some cases, the LSTM network may analyze the time-dependent characteristics of the autofluorescence signals to distinguish between structurally similar monoamine neurotransmitters. The gated recurrent structure of the LSTM network may retain information from both early and later portions of the decay curve, which may occur at different times during the measurement period.

[0075] In some cases, the system may operate by depositing monoamine neurotransmitters onto the plasmonic substrate, illuminating the substrate with ultraviolet light, collecting the autofluorescence time decay series data comprising the plurality of fluorescence spectra, and processing the collected data through the machine learning model to output a classification of the monoamine neurotransmitters. The combination of the plasmonic substrate comprising aluminum concave nanocubes and the LSTM network may enable label-free and probe-free classification of monoamine neurotransmitters based on temporal patterns in the enhanced autofluorescence signals.

[0076] The plasmonic substrate may comprise aluminum concave nanocubes (AICNCs) configured to enhance autofluorescence signals of monoamine neurotransmitters. In some cases, the aluminum concave nanocubes may have a nominal diameter of approximately 80 nanometers. The aluminum concave nanocubes may have a standard deviation of approximately 9 nanometers in diameter. In some cases, the aluminum concave nanocubes may comprise an026389-0063-W001

[0077] oxide layer on a surface thereof having a thickness of between 4 and 8 nanometers. The oxide layer may form on the surface of the aluminum concave nanocubes after exposure to air.

[0078] The concave geometry of the aluminum concave nanocubes may feature sharp corners and edges similar to aluminum bowtie nano antennas. In some cases, the sharp corners and edges may enable more effective light trapping and concentration compared to smoother geometries or spherical shapes such as aluminum hole arrays or aluminum nanotriangles. The concave geometry may contribute to stronger and more consistent fluorescence enhancement compared to other nanoparticle configurations.

[0079] In some cases, the aluminum concave nanocubes may have an extinction dip near 300 nm that closely aligns with the emission wavelengths of the neurotransmitters. The alignment between the extinction dip and the emission wavelengths may support plasmonic enhancement of the autofluorescence signals from the monoamine neurotransmitters.

[0080] The plasmonic substrate may provide a fluorescence enhancement factor of at least 7-fold compared to a non-plasmonic silicon substrate. In some cases, the plasmonic substrate may provide a fluorescence enhancement factor of up to 12-fold for dopamine compared to the non-plasmonic silicon substrate. The plasmonic substrate may provide a fluorescence enhancement factor of up to 9-fold for norepinephrine and up to 7-fold for 3,4-dihydroxyphenylacetic acid compared to the non-plasmonic silicon substrate. The fluorescence enhancement may be attributed to localized surface plasmon resonance effects of the aluminum concave nanocubes in the ultraviolet range, which may amplify both excitation and fluorescent emission near the nanocube surface.

[0081] In the context of a system for classifying monoamine neurotransmitters, the aluminum concave nanocubes having a nominal diameter of approximately 80 nanometers and comprising an oxide layer on a surface thereof having a thickness of between 4 and 8 nanometers may provide the fluorescence enhancement characteristics described above. In the context of a biosensing device for label-free detection and classification of monoamine neurotransmitters, the aluminum concave nanocubes deposited on a substrate surface may be configured to provide localized surface plasmon resonance enhancement in an ultraviolet wavelength range.

[0082] Methods

[0083] Example analytical techniques for detecting, differentiating, and / or quantifying MANTs include liquid chromatography / mass spectroscopy (LC-MS), fast-scan cyclic voltammetry (FSCV), and nanomaterial-based biosensors. LC-MS provides relatively high sensitivity and selectivity but typically involves extensive sample preparation. In some cases, such sample026389-0063-W001

[0084] preparation may disadvantageously result in an unacceptably high percentage of sample loss. FSCV can be carried out with relatively little sample preparation and can detect MANTs in vivo. However, in some cases, the FSCV techniques may not be sufficiently selective because different MANT molecules may have overlapping redox potentials. In many examples, nanomaterialbased biosensors rely on aptamers, antibodies, enzymatic reactions, or chemical reactions for selective MANT detection. However, the corresponding selection of highly specific aptamers or antibodies can be challenging and / or time-consuming in its own right, and nonspecific binding to interfering molecules may still occur even with carefully selected aptamers. Enzymatic or chemical reactions, although specific, may suffer from relatively low sensitivity.

[0085] Intrinsic-fluorescence-based sensors can be used in environmental monitoring, cell imaging, biomedical applications, and analytical chemistry. For example, protein arrays can be detected using time-resolved ultraviolet (UV) native fluorescence. In some examples, multiphoton excitation of native protein fluorescence can be used to discriminate ligands with different binding affinities. However, the relatively low quantum yield and the unstable nature of intrinsic fluorescence of biomolecules may typically constrain the usage of at least some intrinsic-fluorescence-based biosensors.

[0086] At least some of the above-indicated problems in the state of the art can beneficially be addressed using the analytical techniques for detecting and differentiating neurotransmitters with relatively high sensitivity and specificity disclosed herein. For example, we realized that the use of UV plasmonic-enhanced native fluorescence could improve the sensitivity and lower the limits of detection for at least some UV biosensors. Accordingly, some examples disclosed herein use UV plasmonic-engineered native fluorescence of neurotransmitters as a sensing mechanism to achieve highly sensitive and specific detection of at least some neurotransmitters. In some examples, an aluminum (Al) hole array achieves a net enhancement of 50x for tryptophan. In some other examples, similar enhancement factors are achieved for dopamine (DA), norepinephrine (NE), and 3,4-dihydroxyphenylacetic acid (DOPAC) on engineered Al hole arrays. We also observe that DA, NE, and DOPAC have distinct photobleaching rates under UV illumination, and the pairwise differences in their photobleaching rates can be enlarged on at least some plasmonic-engineered substrates, such as an Al thin film and an Al hole array. Some embodiments therefore use the photobleaching rates of UV plasmonic-engineered native fluorescence as a mechanism for differentiating neurotransmitters without the need for an aptamer, antibody, chemical reaction, or enzymatic reaction. At least some embodiments disclosed herein are expected to be of significant practical interest in various applications involving detection, differentiation, and / or quantification of neurotransmitters.026389-0063-W001

[0087] The method for classifying monoamine neurotransmitters may include depositing a sample containing monoamine neurotransmitters onto a plasmonic substrate comprising aluminum concave nanocubes. In some cases, depositing the sample may comprise drop casting a solution containing the monoamine neurotransmitters onto the plasmonic substrate and allowing the solution to dry to form a coffee ring pattern. The monoamine neurotransmitters may include dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In some cases, a volume of approximately 1 pL of the solution may be drop cast onto the plasmonic substrate. The solution may contain the monoamine neurotransmitters dissolved in deionized water at varying concentrations.

[0088] The method may include illuminating the plasmonic substrate with ultraviolet light to excite autofluorescence from the monoamine neurotransmitters. In some cases, illuminating the plasmonic substrate may comprise directing the ultraviolet light toward an outer edge of the coffee ring pattern where a concentration of the monoamine neurotransmitters is higher. The ultraviolet light may be provided by a continuous wave solid state laser configured to emit light at a wavelength of approximately 266 nanometers. In some cases, the ultraviolet light may be directed at an incident angle of approximately 60 degrees with an output power of approximately 5 mW and a circular spot size of approximately 100 pm in diameter.

[0089] The method may further include collecting autofluorescence time decay series data comprising a plurality of fluorescence spectra acquired over a time period during continuous illumination. In some cases, collecting the autofluorescence time decay series data may comprise acquiring the plurality of fluorescence spectra at intervals of approximately 0.5 seconds over a time period of 2 to 3 minutes. Each fluorescence spectrum of the plurality of fluorescence spectra may span a wavelength range of 280 to 360 nanometers. The autofluorescence time decay series data may capture the temporal evolution of the autofluorescence signal as the fluorescence intensity decays over time during continuous ultraviolet illumination. In some cases, each fluorescence time series may contain approximately 240 to 360 spectra collected at 0.5 second intervals over the 2 to 3 minute measurement timeframe.

[0090] The method may include preprocessing the autofluorescence time decay series data to generate standardized input data. In some cases, preprocessing the autofluorescence time decay series data may comprise shifting time series data to eliminate leading missing values. The shifting may align valid entries to a common starting point by moving valid data entries to the left and eliminating leading Not a Number (NaN) values. Preprocessing may further comprise truncating or padding the time series data to a standardized sequence length. In some cases, the standardized sequence length may comprise 17 time steps. Longer sequences may be026389-0063-W001

[0091] truncated to 17 regularly spaced sequences, and shorter sequences may be padded if necessary to achieve the standardized sequence length. Preprocessing may also comprise normalizing feature values using MinMax scaling. In some cases, the MinMax scaling may normalize feature values to a range of 0 to 1.

[0092] The method may include classifying the monoamine neurotransmitters by applying a trained machine learning model to the standardized input data. The trained machine learning model may comprise a Long Short-Term Memory (LSTM) network configured to analyze temporal patterns in the autofluorescence time decay series data. In some cases, the LSTM network may capture temporal dependencies in the autofluorescence time decay series data by retaining information from both early portions and later portions of the decay curve. The gated recurrent structure of the LSTM network may enable the network to learn from the fast initial quench and the slower tail of the decay curve, which may occur at different times during the measurement period.

[0093] In some cases, the LSTM network may achieve a classification accuracy of at least 89 percent for distinguishing between dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. The LSTM network may process the standardized input data and output a classification of the monoamine neurotransmitters based on the temporal decay characteristics captured from the autofluorescence time decay series data. In some cases, the trained LSTM network may process a single input tensor in under 1 millisecond to enable real-time classification of the monoamine neurotransmitters.

[0094] Biosensing Device

[0095] The biosensing device for label-free detection and classification of monoamine neurotransmitters may include a substrate having a surface with aluminum concave nanocubes deposited thereon. In some cases, the aluminum concave nanocubes may be configured to provide localized surface plasmon resonance enhancement in an ultraviolet wavelength range. The localized surface plasmon resonance enhancement may amplify both excitation efficiency and fluorescent emission from monoamine neurotransmitters in contact with the aluminum concave nanocubes. In some cases, the substrate may comprise a silicon wafer onto which the aluminum concave nanocubes are deposited by drop casting a nanoparticle solution and allowing the solution to dry in ambient conditions.

[0096] The biosensing device may include an excitation source configured to direct ultraviolet light onto the substrate to excite native autofluorescence from monoamine neurotransmitters in contact with the aluminum concave nanocubes. In some cases, the excitation source may026389-0063-W001

[0097] comprise a continuous wave solid state ultraviolet laser configured to emit light at a wavelength of approximately 266 nanometers. The excitation source may direct ultraviolet light at an incident angle onto the substrate surface where the monoamine neurotransmitters are deposited. In some cases, the ultraviolet light may be directed toward regions of the substrate where concentrations of the monoamine neurotransmitters are higher, such as at an outer edge of a coffee ring pattern formed during sample deposition.

[0098] The biosensing device may further include a detector configured to acquire time-resolved fluorescence spectra from the monoamine neurotransmitters over a measurement period to generate autofluorescence time decay series data. In some cases, the detector may comprise a spectrometer coupled with a charge-coupled device camera. The detector may acquire the time-resolved fluorescence spectra at intervals of approximately 0.5 seconds over a measurement period of 2 to 3 minutes. In some cases, each of the time-resolved fluorescence spectra may span a wavelength range of 280 to 360 nanometers. The autofluorescence time decay series data generated by the detector may capture the temporal evolution of the autofluorescence signal as the fluorescence intensity decays during continuous ultraviolet illumination by the excitation source.

[0099] The biosensing device may include a classification module comprising a processor and a memory storing a trained recurrent neural network. In some cases, the classification module may be configured to receive the autofluorescence time decay series data and output an identification of the monoamine neurotransmitters based on temporal decay characteristics captured by the trained recurrent neural network. The trained recurrent neural network may comprise a Long Short-Term Memory network configured to analyze temporal patterns in the autofluorescence time decay series data. In some cases, the Long Short-Term Memory network may have three sequential Long Short-Term Memory layers with batch normalization layers following each of the three sequential Long Short-Term Memory layers. The Long Short-Term Memory network may further comprise a fully connected dense layer and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0100] The classification module may preprocess the autofluorescence time decay series data prior to applying the trained recurrent neural network. In some cases, preprocessing may comprise aligning time series traces by shifting valid entries to eliminate leading missing values and standardizing sequence lengths to generate input tensors. The standardized sequence lengths may comprise 17 time steps. In some cases, the classification module may normalize feature values of the autofluorescence time decay series data to a range of 0 to 1 using MinMax scaling prior to applying the trained recurrent neural network. The processor of the classification026389-0063-W001

[0101] module may process a single input tensor in under 1 millisecond to enable real-time classification of the monoamine neurotransmitters.

[0102] The aluminum concave nanocubes of the biosensing device may have a nominal diameter of approximately 80 nanometers with a standard deviation of approximately 9 nanometers. In some cases, the aluminum concave nanocubes may comprise an aluminum oxide layer on surfaces thereof having a thickness of between 4 and 8 nanometers. The aluminum oxide layer may form on the surfaces of the aluminum concave nanocubes after exposure to air. In some cases, the concave geometry of the aluminum concave nanocubes may feature sharp corners and edges that enable effective light trapping and concentration, contributing to the localized surface plasmon resonance enhancement in the ultraviolet wavelength range. The aluminum concave nanocubes having the nominal diameter of approximately 80 nanometers and the aluminum oxide layer thickness of between 4 and 8 nanometers may provide fluorescence enhancement factors of up to 12-fold for dopamine, up to 9-fold for norepinephrine, and up to 7-fold for 3,4-dihydroxyphenylacetic acid compared to non-plasmonic substrates.

[0103] In the context of a system for classifying monoamine neurotransmitters, the ultraviolet light source may comprise a continuous wave solid state laser configured to emit light at a wavelength of approximately 266 nanometers. In some cases, the continuous wave solid state laser may provide stable and consistent ultraviolet illumination for exciting autofluorescence from monoamine neurotransmitters deposited on the plasmonic substrate. The wavelength of approximately 266 nanometers may be selected to excite the aromatic ring structures of the monoamine neurotransmitters, causing the monoamine neurotransmitters to emit native autofluorescence.

[0104] The ultraviolet laser may be positioned at an incident angle of 60 degrees relative to a substrate surface. In some cases, the incident angle of 60 degrees may facilitate illumination of the monoamine neurotransmitters deposited on the plasmonic substrate while allowing collection of the emitted autofluorescence by a detector positioned to receive the fluorescence signals. The ultraviolet laser may have an output power of approximately 5 mW. In some cases, the output power of approximately 5 mW may provide sufficient excitation energy to induce autofluorescence from the monoamine neurotransmitters while avoiding excessive photobleaching or sample degradation during the measurement period.

[0105] The ultraviolet laser may have a circular spot size of approximately 100 micrometers in diameter. In some cases, the circular spot size of approximately 100 micrometers in diameter may define an illumination area on the plasmonic substrate where the monoamine neurotransmitters are excited to emit autofluorescence. The spot size may be selected to026389-0063-W001

[0106] illuminate a region of the substrate where concentrations of the monoamine neurotransmitters are sufficient to produce detectable autofluorescence signals, such as at an outer edge of a coffee ring pattern formed during sample deposition.

[0107] In the context of a biosensing device for label-free detection and classification of monoamine neurotransmitters, the excitation source may comprise a continuous wave solid state ultraviolet laser configured to emit light at a wavelength of approximately 266 nanometers. In some cases, the detector may be configured to acquire time-resolved fluorescence spectra at intervals of approximately 0.5 seconds over a measurement period of 2 to 3 minutes. The acquisition interval of approximately 0.5 seconds may provide temporal resolution sufficient to capture the decay dynamics of the autofluorescence signals from the monoamine neurotransmitters during continuous ultraviolet illumination. In some cases, the measurement period of 2 to 3 minutes may allow collection of approximately 240 to 360 spectra that capture the temporal evolution of the autofluorescence signal as the fluorescence intensity decays over time.

[0108] The Long Short-Term Memory (LSTM) network may comprise three sequential Long Short-Term Memory layers configured to process the autofluorescence time decay series data. In some cases, the three sequential Long Short-Term Memory layers may have 200, 150, and 100 units respectively. The first Long Short-Term Memory layer may have 200 units, the second Long Short-Term Memory layer may have 150 units, and the third Long Short-Term Memory layer may have 100 units. The sequential arrangement of the three Long Short-Term Memory layers may enable the network to learn hierarchical temporal representations from the autofluorescence time decay series data.

[0109] The Long Short-Term Memory network may further comprise batch normalization layers following each of the three sequential Long Short-Term Memory layers. In some cases, a first batch normalization layer may follow the first Long Short-Term Memory layer having 200 units, a second batch normalization layer may follow the second Long Short-Term Memory layer having 150 units, and a third batch normalization layer may follow the third Long Short-Term Memory layer having 100 units. The batch normalization layers may stabilize learning during training of the Long Short-Term Memory network by normalizing activations between the sequential Long Short-Term Memory layers.

[0110] The Long Short-Term Memory network may comprise a fully connected dense layer following the three sequential Long Short-Term Memory layers and batch normalization layers. In some cases, the fully connected dense layer may have 50 neurons. The fully connected dense layer having 50 neurons may receive output from the third Long Short-Term Memory layer and026389-0063-W001

[0111] may transform the learned temporal representations into features suitable for classification. In some cases, a batch normalization layer may follow the fully connected dense layer.

[0112] The Long Short-Term Memory network may comprise a softmax output layer configured to classify the monoamine neurotransmitters into a plurality of classes. In some cases, the plurality of classes may comprise dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. The softmax output layer may output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In some cases, the softmax output layer may receive input from the fully connected dense layer and may generate probability values indicating the likelihood that an input autofluorescence time decay series corresponds to each of the three monoamine neurotransmitter classes.

[0113] The autofluorescence time decay series data may exhibit exponential decay over time when the monoamine neurotransmitters are continuously exposed to ultraviolet light. In some cases, the autofluorescence signals from the monoamine neurotransmitters may decrease in intensity following an exponential decay pattern during continuous ultraviolet illumination. The decay rate constants may be distinct among similarly structured monoamine neurotransmitters. In some cases, dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid may each exhibit different decay rate constants despite having similar aromatic ring structures. The distinct decay rate constants among the similarly structured monoamine neurotransmitters may provide temporal features that the Long Short-Term Memory network may capture for classification purposes.

[0114] The gated recurrent structure of the Long Short-Term Memory network may retain information from both early portions and later portions of the decay curve. In some cases, the Long Short-Term Memory network may learn from a fast initial quench region and a slower tail region of the decay curve, which may occur at different times during the measurement period. The three sequential Long Short-Term Memory layers having 200, 150, and 100 units respectively may enable the network to capture temporal dependencies across the autofluorescence time decay series data that distinguish the monoamine neurotransmitters based on the distinct decay rate constants.

[0115] In the context of a system for classifying monoamine neurotransmitters, the Long Short-Term Memory network may comprise three sequential Long Short-Term Memory layers, batch normalization layers following each of the three sequential Long Short-Term Memory layers, a fully connected dense layer, and a softmax output layer configured to classify the monoamine neurotransmitters into a plurality of classes comprising dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.026389-0063-W001

[0116] In the context of a biosensing device for label-free detection and classification of monoamine neurotransmitters, the trained recurrent neural network may comprise a Long Short-Term Memory network having three sequential Long Short-Term Memory layers with batch normalization layers following each of the three sequential Long Short-Term Memory layers. In some cases, the Long Short-Term Memory network of the biosensing device may further comprise a fully connected dense layer and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0117] Preprocessing of autofluorescence time decay series data may comprise multiple operations to prepare the data for machine learning model input. In some cases, the preprocessing operations may include shifting time series data to eliminate leading missing values, truncating or padding the time series data to a standardized sequence length, and normalizing feature values using MinMax scaling.

[0118] The autofluorescence time decay series data collected during experiments may contain time series traces that do not begin at a common starting point. In some cases, collection of actual spectra data may start at a later time after a spectrometer shutter opens due to adjustment of sample position to find a spot with high fluorescence signal. The delayed collection may result in leading missing values, represented as Not a Number (NaN) values, at the beginning of certain time series traces. Preprocessing the autofluorescence time decay series data may comprise shifting time series traces to eliminate leading missing values by aligning valid entries to a common starting point. In some cases, each row of time series data may be shifted left to eliminate the leading NaN values, effectively aligning all rows to start at the same initial time point. The shifting operation may move valid data entries to the left while removing the leading NaN values, ensuring that all time series traces begin from a common starting point for consistent processing by machine learning models.

[0119] The number of time-dependent spectra for each illumination spot may vary across experiments. In some cases, the sequence length may range from 17 to 240-360 spectra after preprocessing due to variability in experimental conditions. To address the varying lengths of the time series data, preprocessing may comprise truncating or padding the time series data to a standardized sequence length. In some cases, the standardized sequence length may comprise 17 time steps. Longer sequences may be truncated to 17 regularly spaced sequences to achieve the standardized sequence length. In some cases, the truncation may select 17 regularly spaced time points from longer sequences to preserve the overall decay dynamics while achieving consistent input dimensions. Shorter sequences may be padded if necessary to achieve the standardized sequence length of 17 time steps. The standardization to 17 time steps may ensure026389-0063-W001

[0120] consistent input dimensions for machine learning models while preserving the temporal decay characteristics of the autofluorescence signals.

[0121] Preprocessing may further comprise normalizing feature values using MinMax scaling. In some cases, the MinMax scaling may normalize feature values of the autofluorescence time decay series data to a range of 0 to 1. The normalization to the range of 0 to 1 may be performed prior to applying a trained Long Short-Term Memory network or other machine learning models. In some cases, the MinMax scaling may fit on a training partition within each fold of cross-validation and then be applied to validation data to prevent data leakage during model evaluation.

[0122] For sequence-based models such as Long Short-Term Memory networks, the preprocessed autofluorescence time decay series data may be converted into structured input tensors of shape corresponding to time steps and wavelength bins. In some cases, the input tensors may have a shape of 17 time steps by a number of wavelength bins spanning the 280 to 360 nanometer wavelength range. The structured input tensors may preserve the sequential nature of the autofluorescence time decay series data for analysis by the Long Short-Term Memory network.

[0123] For non-sequential models such as k-nearest neighbors (KNN) and Random Forest (RF), the autofluorescence time decay series data may be flattened into feature vectors rather than structured as input tensors. In some cases, the flattened feature vectors may be right-aligned intensity vectors derived from the preprocessed time series data. The flattening operation may convert the time-dependent spectral data into a one-dimensional feature representation suitable for processing by non-sequential machine learning algorithms. In some cases, the flattened feature vectors may enable KNN and RF classifiers to analyze the autofluorescence time decay series data without requiring the sequential processing capabilities of recurrent neural networks.

[0124] In the context of a method for classifying monoamine neurotransmitters, preprocessing the autofluorescence time decay series data may comprise shifting time series data to eliminate leading missing values, truncating or padding the time series data to a standardized sequence length, and normalizing feature values using MinMax scaling. In some cases, the standardized sequence length may comprise 17 time steps.

[0125] In the context of a non-transitory computer-readable medium storing instructions that when executed by a processor cause the processor to preprocess autofluorescence time decay series data, the instructions may cause the processor to normalize feature values of the autofluorescence time decay series data to a range of 0 to 1 using MinMax scaling prior to applying a trained Long Short-Term Memory network. In some cases, the instructions may cause the processor to shift time series traces to eliminate leading missing values by aligning valid026389-0063-W001

[0126] entries to a common starting point. The instructions may further cause the processor to truncate longer sequences to 17 regularly spaced sequences or pad shorter sequences to achieve standardized sequence lengths of 17 time steps.

[0127] A non-transitory computer-readable medium may store instructions that, when executed by a processor, cause the processor to perform operations for classifying monoamine neurotransmitters. In some cases, the instructions may cause the processor to receive autofluorescence time decay series data collected from monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes. The autofluorescence time decay series data may comprise a plurality of fluorescence spectra acquired during continuous ultraviolet illumination. In some cases, the plurality of fluorescence spectra may capture the temporal evolution of autofluorescence signals from the monoamine neurotransmitters as the fluorescence intensity decays overtime during the continuous ultraviolet illumination.

[0128] The instructions stored on the non-transitory computer-readable medium may cause the processor to preprocess the autofluorescence time decay series data by aligning time series traces and standardizing sequence lengths to generate input tensors. In some cases, aligning time series traces may comprise shifting valid entries to eliminate leading missing values, ensuring that all time series traces begin from a common starting point. Standardizing sequence lengths may comprise truncating longer sequences or padding shorter sequences to achieve a consistent number of time steps across all input data. In some cases, the standardized sequence lengths may comprise 17 time steps. The input tensors generated through preprocessing may have a shape corresponding to the standardized number of time steps and a number of wavelength bins spanning the wavelength range of the fluorescence spectra.

[0129] The instructions may further cause the processor to apply a trained Long Short-Term Memory network to the input tensors to capture temporal dependencies in the autofluorescence time decay series data. In some cases, the trained Long Short-Term Memory network may comprise three sequential Long Short-Term Memory layers having 200, 150, and 100 units respectively. The trained Long Short-Term Memory network may further comprise batch normalization layers following each of the three sequential Long Short-Term Memory layers. In some cases, the trained Long Short-Term Memory network may comprise a fully connected dense layer with 50 neurons. The trained Long Short-Term Memory network may comprise a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0130] The instructions stored on the non-transitory computer-readable medium may cause the processor to output a classification of the monoamine neurotransmitters based on analysis by the026389-0063-W001

[0131] trained Long Short-Term Memory network. In some cases, the classification output may correspond to one of dopamine, norepinephrine, or 3,4-dihydroxyphenylacetic acid based on the classification probabilities generated by the softmax output layer. The trained Long Short-Term Memory network may analyze the temporal patterns in the autofluorescence time decay series data to distinguish between the structurally similar monoamine neurotransmitters based on distinct decay characteristics captured in the input tensors.

[0132] The instructions stored on the non-transitory computer-readable medium may be implemented in various software. In some cases, the TensorFlow framework may provide functionality for implementing and executing the trained Long Short-Term Memory network architecture comprising the three sequential Long Short-Term Memory layers, batch normalization layers, fully connected dense layer, and softmax output layer. The implementation may include preprocessing operations for aligning time series traces, standardizing sequence lengths, and generating input tensors from the autofluorescence time decay series data.

[0133] Class labels for the monoamine neurotransmitters may be converted using different encoding methods depending on the type of classifier. In some cases, class labels may be converted using a LabelEncoder for traditional classifiers such as K-nearest neighbors and Random Forest classifiers. The LabelEncoder may convert categorical class labels such as dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid into numerical representations suitable for processing by traditional machine learning algorithms. For neural network models such as the Long Short-Term Memory network, class labels may be converted using one-hot encoding. In some cases, one-hot encoding may represent each class label as a binary vector where a single element corresponding to the class is set to one and all other elements are set to zero. The one-hot encoded class labels may be compatible with the softmax output layer of the Long Short-Term Memory network, which outputs classification probabilities across the plurality of classes comprising dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0134] The Long Short-Term Memory (LSTM) model may be trained for 50 epochs using a batch size of 64. In some cases, the training process may employ a learning rate scheduler configured to reduce the learning rate upon plateauing of validation loss. The learning rate scheduler may monitor the validation loss during training and may decrease the learning rate when the validation loss stops improving over a specified number of epochs. In some cases, the reduction in learning rate upon plateauing may help the model converge to a solution by allowing finer adjustments to model parameters when larger learning rate steps no longer produce improvements in validation performance.026389-0063-W001

[0135] Spectra from the same experimental replicate may be kept within a single fold to prevent data leakage. In some cases, ensuring that no replicate contributes to both training and validation sets may prevent the model from learning patterns specific to individual experimental replicates rather than generalizable features of the monoamine neurotransmitters. The partitioning of spectra from the same experimental replicate into a single fold may maintain independence between training and validation data, which may provide more accurate estimates of model performance on unseen data.

[0136] The larger dataset size for models trained on aluminum concave nanocube substrates may result from the time-variant spectra collection approach under continuous light illumination. In some cases, acquiring autofluorescence time decay series data with an acquisition time of 0.5 seconds may yield approximately 240 to 360 spectra from a single molecular concentration in 2 to 3 minutes. The time-variant spectra collection approach may provide a larger number of training samples compared to solution-based spectra collection, which may involve separate measurements at different concentrations to obtain comparable numbers of spectra.

[0137] In the context of a non-transitory computer-readable medium storing instructions that when executed by a processor cause the processor to classify monoamine neurotransmitters, the instructions may further cause the processor to process a single input tensor in under 1 millisecond to enable real-time classification of the monoamine neurotransmitters. In some cases, the input tensor may correspond to a standardized 17-time-step sequence derived from autofluorescence time decay series data collected from monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes. The sub-millisecond processing time may confirm suitability of the framework for real-time neurotransmitter classification applications.

[0138] Machine learning classification of monoamine neurotransmitters may alternatively be performed using a K-nearest neighbors (KNN) algorithm. In some cases, the KNN algorithm may classify data points based on the closest training examples in a feature space. The KNN algorithm may be an instance-based learning algorithm that does not require explicit model training in the same manner as neural network architectures. The grid search optimization may evaluate multiple values of k and select the value that produces the highest classification accuracy across the cross-validation folds.

[0139] The KNN model may achieve classification accuracy when applied to autofluorescence time decay series data collected from monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes. The classification accuracy values for the KNN model may indicate that the KNN algorithm can distinguish between the structurally similar026389-0063-W001

[0140] monoamine neurotransmitters based on features extracted from the autofluorescence time decay series data.

[0141] For application to autofluorescence time decay series data, the KNN model may receive flattened, right-aligned intensity vectors derived from the preprocessed time series data. In some cases, the flattening operation may convert the time-dependent spectral data into a onedimensional feature representation suitable for processing by the KNN algorithm. The right-aligned intensity vectors may preserve the temporal information from the autofluorescence time decay series while providing a format compatible with the distance-based classification approach of the KNN algorithm.

[0142] Machine learning classification of monoamine neurotransmitters may alternatively be performed using a Random Forest (RF) algorithm. In some cases, the Random Forest algorithm may comprise an ensemble learning method based on decision trees. The Random Forest algorithm may classify spectral data by aggregating predictions from multiple decision trees trained on different subsets of the training data.

[0143] The Random Forest model may achieve classification accuracy when applied to autofluorescence time decay series data collected from monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes. The classification accuracy values for the Random Forest model may indicate that the ensemble-based structure of the Random Forest algorithm can capture features from the autofluorescence time decay series data for distinguishing between the monoamine neurotransmitters.

[0144] For application to autofluorescence time decay series data, the Random Forest model may receive flattened, right-aligned intensity vectors derived from the preprocessed time series data. In some cases, the flattened feature vectors may enable the Random Forest classifier to analyze the autofluorescence time decay series data without requiring the sequential processing capabilities of recurrent neural networks. The Random Forest algorithm may provide built-in feature importance scores that may aid interpretation of which spectral and temporal features contribute to classification decisions.

[0145] Classification of monoamine neurotransmitters may be performed on static autofluorescence spectra recorded in solution without plasmonic nanoparticles. In some cases, the in-solution autofluorescence dataset may comprise static snapshots of paired wavelengthintensity values for each neurotransmitter. The static autofluorescence spectra may be restructured into wavelength-intensity pairs for each monoamine neurotransmitter to form a feature matrix amenable to machine learning approaches. In some cases, the in-solution autofluorescence spectra may not capture the temporal decay dynamics that are present in026389-0063-W001

[0146] autofluorescence time decay series data collected from monoamine neurotransmitters deposited on plasmonic substrates.

[0147] The KNN and Random Forest models may be applied to the in-solution autofluorescence spectra for classification of monoamine neurotransmitters. In some cases, the KNN model applied to in-solution autofluorescence spectra may achieve reduced classification accuracy compared to the KNN model applied to autofluorescence time decay series data on aluminum concave nanocubes. The Random Forest model applied to in-solution autofluorescence spectra may achieve reduced classification accuracy compared to the Random Forest model applied to autofluorescence time decay series data on aluminum concave nanocubes. In some cases, the reduced classification accuracy for in-solution measurements may result from the absence of temporal decay information and the absence of plasmonic enhancement provided by the aluminum concave nanocubes.

[0148] The reduced classification accuracy for in-solution autofluorescence spectra compared to autofluorescence time decay series on aluminum concave nanocubes may highlight the contribution of plasmonic-engineered autofluorescence time decay series data in achieving classification accuracy among structurally similar monoamine neurotransmitters. In some cases, the temporal decay characteristics captured in the autofluorescence time decay series data may provide distinguishing features that are not present in static autofluorescence spectra collected in solution. The fluorescence enhancement provided by the aluminum concave nanocubes may amplify the autofluorescence signals, providing stronger input data for the machine learning models compared to non-enhanced in-solution measurements.

[0149] Confusion matrices may display class-specific prediction distributions for the machine learning classification models. In some cases, the confusion matrices may present diagonal elements representing correct classifications (true positives) and off-diagonal elements indicating misclassifications.

[0150] The off-diagonal elements in the confusion matrices may indicate misclassification patterns among the monoamine neurotransmitter classes. In some cases, the misclassification patterns may reveal which pairs of monoamine neurotransmitters are more likely to be confused by the classification models. The confusion matrices may provide visualization of model effectiveness and misclassification tendencies that complement the precision, recall, and F1 score metrics.

[0151] The system for classifying monoamine neurotransmitters may operate through coordinated interaction of the plasmonic substrate, ultraviolet light source, spectrometer, and machine learning processor. In some cases, the interaction of these system elements may enable026389-0063-W001

[0152] label-free and probe-free classification of monoamine neurotransmitters based on temporal patterns in enhanced autofluorescence signals.

[0153] The workflow for classifying monoamine neurotransmitters may begin with sample deposition on the plasmonic substrate comprising aluminum concave nanocubes. In some cases, a solution containing monoamine neurotransmitters such as dopamine, norepinephrine, or 3,4-dihydroxyphenylacetic acid may be drop cast onto the plasmonic substrate. The solution may be allowed to dry in ambient conditions to form a coffee ring pattern on the substrate surface. In some cases, the evaporation dynamics during drying may cause the monoamine neurotransmitters to migrate toward an outer edge of the droplet, resulting in higher concentrations of the monoamine neurotransmitters at the outer edge of the coffee ring pattern compared to other regions of the dried sample.

[0154] Following sample deposition, the aluminum concave nanocubes of the plasmonic substrate may interact with the deposited monoamine neurotransmitters to provide localized surface plasmon resonance enhancement of autofluorescence signals. In some cases, the concave geometry of the aluminum concave nanocubes may feature sharp corners and edges that enable effective light trapping and concentration. The localized surface plasmon resonance effects of the aluminum concave nanocubes in the ultraviolet range may amplify both excitation efficiency and fluorescent emission from the monoamine neurotransmitters in contact with or in proximity to the nanocube surfaces. In some cases, the localized surface plasmon resonance enhancement may provide fluorescence enhancement factors of up to 12-fold for dopamine, up to 9-fold for norepinephrine, and up to 7-fold for3,4-dihydroxyphenylacetic acid compared to non-plasmonic substrates such as silicon wafers.

[0155] The ultraviolet light source may interact with the plasmonic substrate by directing ultraviolet light onto the substrate surface where the monoamine neurotransmitters are deposited. In some cases, the ultraviolet light source may comprise a continuous wave solid state laser configured to emit light at a wavelength of approximately 266 nanometers. The ultraviolet light may be directed toward the outer edge of the coffee ring pattern where concentrations of the monoamine neurotransmitters are higher. In some cases, the ultraviolet light may excite the aromatic ring structures of the monoamine neurotransmitters, causing the monoamine neurotransmitters to emit native autofluorescence. The continuous illumination provided by the ultraviolet light source may induce autofluorescence from the monoamine neurotransmitters throughout a measurement period during which the spectrometer collects time-resolved spectra.

[0156] The spectrometer may interact with the plasmonic substrate and ultraviolet light source by collecting autofluorescence signals emitted from the monoamine neurotransmitters during026389-0063-W001

[0157] continuous ultraviolet illumination. In some cases, the spectrometer may be coupled with a detector such as a charge-coupled device camera to capture fluorescence spectra across a wavelength range of 280 to 360 nanometers. The spectrometer may acquire time-resolved fluorescence spectra at intervals of approximately 0.5 seconds over a measurement period of 2 to 3 minutes. In some cases, the time-resolved spectra may capture the temporal evolution of the autofluorescence signal as the fluorescence intensity decays over time during the continuous ultraviolet illumination.

[0158] The autofluorescence signals from the monoamine neurotransmitters may exhibit exponential decay over time when the monoamine neurotransmitters are continuously exposed to ultraviolet light. In some cases, the autofluorescence intensity may decrease following an exponential decay pattern during the measurement period. The decay rate constants may be distinct among similarly structured monoamine neurotransmitters such as dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In some cases, the distinct decay rate constants may provide temporal features that distinguish the monoamine neurotransmitters despite overlapping spectral profiles. The localized surface plasmon resonance enhancement provided by the aluminum concave nanocubes may enlarge differences in decay characteristics among the monoamine neurotransmitters compared to non-plasmonic substrates.

[0159] The spectrometer may generate autofluorescence time decay series data comprising the plurality of time-resolved fluorescence spectra acquired during the measurement period. In some cases, each fluorescence time series may contain approximately 240 to 360 spectra collected at 0.5 second intervals over the 2 to 3 minute measurement timeframe. The autofluorescence time decay series data may capture both spectral information across the 280 to 360 nanometer wavelength range and temporal information reflecting the decay dynamics of the autofluorescence signals from the monoamine neurotransmitters.

[0160] The machine learning processor may receive the autofluorescence time decay series data from the spectrometer and may preprocess the data to generate standardized input tensors. In some cases, preprocessing may comprise shifting time series traces to eliminate leading missing values by aligning valid entries to a common starting point. Preprocessing may further comprise truncating or padding the time series data to a standardized sequence length of 17 time steps. In some cases, longer sequences may be truncated to 17 regularly spaced sequences to preserve the overall decay dynamics while achieving consistent input dimensions. Preprocessing may also comprise normalizing feature values using MinMax scaling to normalize feature values to a range of 0 to 1. In some cases, the standardized input tensors may have a shape corresponding to 17026389-0063-W001

[0161] time steps and a number of wavelength bins spanning the 280 to 360 nanometer wavelength range.

[0162] The machine learning processor may apply a trained Long Short-Term Memory network to the standardized input tensors to capture temporal dependencies in the autofluorescence time decay series data. In some cases, the Long Short-Term Memory network may comprise three sequential Long Short-Term Memory layers having 200, 150, and 100 units respectively, with batch normalization layers following each of the three sequential Long Short-Term Memory layers. The Long Short-Term Memory network may further comprise a fully connected dense layer with 50 neurons and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0163] The gated recurrent structure of the Long Short-Term Memory network may retain information from both early portions and later portions of the decay curve captured in the autofluorescence time decay series data. In some cases, the Long Short-Term Memory network may learn from a fast initial quench region and a slower tail region of the decay curve, which may occur at different times during the measurement period. The temporal dependencies captured by the Long Short-Term Memory network may enable the network to distinguish between the structurally similar monoamine neurotransmitters based on the distinct decay rate constants reflected in the autofluorescence time decay series data.

[0164] The machine learning processor may output a classification of the monoamine neurotransmitters based on analysis by the trained Long Short-Term Memory network. In some cases, the softmax output layer may generate probability values indicating the likelihood that an input autofluorescence time decay series corresponds to each of the three monoamine neurotransmitter classes comprising dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. The classification output may correspond to the monoamine neurotransmitter class having the highest probability value generated by the softmax output layer. In some cases, the trained Long Short-Term Memory network may process a single standardized input tensor in under 1 millisecond, enabling real-time classification of the monoamine neurotransmitters following data acquisition and preprocessing.

[0165] The interaction of the plasmonic substrate, ultraviolet light source, spectrometer, and machine learning processor may provide a coordinated workflow for label-free and probe-free classification of monoamine neurotransmitters. In some cases, the aluminum concave nanocubes of the plasmonic substrate may enhance the autofluorescence signals from the monoamine neurotransmitters through localized surface plasmon resonance effects. The ultraviolet light source may provide continuous illumination that excites autofluorescence and026389-0063-W001

[0166] induces temporal decay dynamics in the autofluorescence signals. In some cases, the spectrometer may capture the temporal evolution of the autofluorescence signals as time-resolved spectra exhibiting exponential decay. The machine learning processor may preprocess the time-resolved spectra to generate standardized input tensors and may apply the trained Long Short-Term Memory network to analyze temporal dependencies and output identification of the monoamine neurotransmitters based on the distinct decay characteristics captured in the autofluorescence time decay series data.

[0167] FIGS. 16A-B are diagrams illustrating an optical system 100 with which various embodiments can be practiced. More specifically, FIG. 16A is a block diagram illustrating the optical system 100 according to one example. FIG. 16B is a diagram illustrating a cross-sectional side view of a biosensor 130 that is interrogated with the optical system 100 according to one example.

[0168] In the example shown, the optical system 100 includes a light source 110, an excitation filter 120, lenses 124, 140, and 156, mirrors 126, 154, and 158, an emission filter 150, a spectrometer 160, a pixelated optical detector (e.g., a CCD) 170, and a computing device 180. In operation, the light source 110 generates an output light beam 112, which is directed through the excitation filter 120 and the focusing lens 124 toward the mirror 126. The mirror 126 redirects a received filtered excitation light beam 128 toward the biosensor 130. In response to the excitation light beam 128, the biosensor 130 generates an emission (including fluorescence) light beam 138, which is collected by the lens 140, passes through the emission filter 150 and is relayed by the mirrors 154 and 158 and the lens 156 to the spectrometer 160. The spectrometer 160 disperses the received light in wavelength, and the spectrally dispersed light is detected with the pixelated optical detector 170. A readout signal 172 from the pixelated optical detector 170 is directed to the computing device 180 for processing.

[0169] In various examples, the excitation filter 120 is a fixed low-pass filter, a fixed bandpass filter, or a variable bandpass filter. In some examples, one or both of the light source 110 and the excitation filter 120 are tunable, which enables the optical system 100 to change the spectral composition of the excitation light beam 128. In some examples, the light source 110 and the excitation filter 120 are configured to make the excitation light beam 128 substantially monochromatic or substantially dichromatic. In some examples, the excitation light beam 128 has a wavelength in the UV range, such as the wavelength of about 266 nm.

[0170] In various examples, the emission filter 150 is a fixed long-pass filter or a variable long-pass filter. In some examples, the emission filter 150 is absent, and the spectrometer 160 performs the rejection of unwanted wavelengths instead of the emission filter 150.026389-0063-W001

[0171] Referring to FIG. 16B, in the example shown, the biosensor 130 includes a silicon substrate layer 132 with a thin layer of aluminum concave nanocubes (AICNC) 134. The AICNC layer 134 is created by drop casting on the silicon layer 132 in solution (3.9 x12particles per mL) and allowed to dry at room temperature. This creates a random array of AICNC features of varying thickness. The biosensor 130 has a relatively thin sample layer 136 deposited onto the exterior surface of the AICNC layer 134 and comprising neurotransmitter molecules. In one example, the sample layer 136 is produced by dissolving the neurotransmitter molecules in water solution and then the drop casting the solution (1 pL) onto exterior surface of the AICNC layer 134.

[0172] In one example, the optical system 100 can be used to measure the native fluorescence of neurotransmitters on a substrate as follows. A 12-mW optical power output beam 112 generated by the high-intensity 266-nm UV continuous-wave (CW) laser 110 is focused by the plano-convex lens 124 (with the focal length of 100 mm) to a spot size of 60 pm by 60 pm on the sample layer 136 at a 60-degree incident angle. The native fluorescence emission 132 of neurotransmitters is excited in the illuminated spot and is collected by the UV objective 140 (with the focal length of 25 mm, F / 2.8). The fluorescence emission 132 passes through the long-pass filter 150 and the cylindrical lens 156 (with the focal length of 100 mm) and enters into the imaging spectrometer 160 coupled to the UV-enhanced CCD camera 170. A photobleaching series of fluorescence spectra is collected by the computing device 180 by operating the CCD camera 180 using a 0.5-second integration time for a total time duration of 90 seconds.

[0173] FIG. 17 is a flowchart illustrating a method 1000 of detecting and differentiating neurotransmitters according to some examples. The method 1000 can be implemented using the optical system 100 and further using appropriate software installed on the computing device 180 and / or a remote network-connected server. The method 1000 is described below.

[0174] A block 1002 of the method 1000 includes loading an sample onto a biosensor 130. In some examples, operations of the block 1002 include selecting the biosensor 130 from a plurality of different biosensors 130. Different biosensors 130 may differ in one or more of the following characteristics: (i) the material used in the AICNC layer 134; (ii) the amplitude of the feature height variation across the AICNC layer 134; (iii) the transverse (in-layer) size of the holes in the AICNC layer 134; (iv) the average distance between the concave nanocubes in the AICNC layer 134; (v) the geometric shapes of the holes in the AICNC layer 134, etc. One of the available biosensors 130 may be selected in the block 1002 based on auxiliary information about the sample. For example, when the neurotransmitter concentration(s) in the sample is (are) expected to be relatively low, a biosensor 130 providing a relatively high level of signal enhancement may be026389-0063-W001

[0175] selected. In another example, when the relative concentrations of two known neurotransmitters in the sample need to be determined, a biosensor 130 providing an approximately largest difference in one or more pertinent signal characteristics for the two neurotransmitters may be selected. In yet some other examples, other considerations (such as various trial and error approaches) for selecting the biosensor 130 in the block 1002 can also be employed.

[0176] The operations of the block 1002 further include forming the sample layer 136 on the selected biosensor 130 using the sample. In some examples, the spin-coating technique described above in reference to FIG. 1 B may be used for this purpose. In some other examples, other suitable techniques for forming the sample-containing sample layer 137 on the selected biosensor 130 can similarly be employed in the block 1002.

[0177] A block 1004 of the method 1000 includes the computing device 180 (serving as an electronic controller of the optical system 100) controlling various system components to measure a photobleaching series of native fluorescence spectra using the loaded biosensor 130 prepared in the block 1002. In some examples, the measurements of the block 1004 are performed as described above. In such examples, the excitation light at 266 nm serves a dual purpose of both exciting the native fluorescence in the sample and inducing the photochemical reactions that alter the neurotransmitter molecules in the sample thereby making them unable to fluoresce and quenching the observed native fluorescence. In some other examples, the fluorescence measurements in the block 1004 are implemented such that a first wavelength is used to excite the native fluorescence whereas a different second wavelength is used to induce the photochemical reactions. In such examples, the computing device 180 configures the light source 110 and / or the excitation filter 120 to switch the excitation light beam 128 between the first and second wavelengths during the acquisition of the photobleaching series as needed.

[0178] A block 1006 of the method 1000 includes the computing device 180 processing the photobleaching series of native fluorescence spectra measured in the block 1004 to determine a corresponding set of photobleaching parameters. In some examples, the processing operations include (i) integrating the fluorescence intensity in accordance with Eq. 1 and ii fitting the obtained values of S(f) with the two-term exponential function of Eq. (2).

[0179]

[0180] The photobleaching rates are obtained by fitting the integrated fluorescence intensity S(t) with the following two-term exponential function:

[0181] S(t) = a x exp(— k t) + b x exp(— k2t) (2)026389-0063-W001

[0182] where a and b are amplitudes. The first exponential term in Eq. 2 represents the fast decay with the rate constant i, and the second exponential term in Eq. 2 represents the slow decay with the rate constant fe.

[0183] Herein, the term “photobleaching” refers to photochemical alteration of a fluorophore molecule that makes it permanently unable to fluoresce because some portions thereof are broken down by photochemical reactions. Irreversible photobleaching of neurotransmitters is observed on biosensors 130. The photobleaching rate(s) can be obtained as described above and then used to characterize the corresponding analytes according to various embodiments described herein.

[0184] The corresponding set of parameters obtained form the fit include the rate constants i, and fe and the amplitudes a and b. In some other examples, other suitable mathematical models and / or parameterization of the photobleaching kinetics can also be used.

[0185] A block 1008 of the method 1000 includes the computing device 180 submitting a query to a database with the set of photobleaching parameters determined in the block 1006. In one example, the database includes a plurality of LUTs having stored therein various photobleaching calibration data. The plurality of LUTs includes a first set of calibration LUTs having stored therein the photobleaching calibration data exemplified by Table 1 and Table 2. The plurality of LUTs also includes a second set of calibration LUTs having stored therein the photobleaching calibration data. In some examples, the plurality of LUTs may also include one or more additional sets of calibration LUTs having stored therein additional photobleaching calibration data. In different examples, the database may be stored locally in the computing device 180 or be stored in a remote server that is network-connected to the computing device 180.

[0186] In some examples, the query submitted in the block 1008 is configured based on the user input and contains a numerical field specifying the query type. The query type determines which set of calibration LUTs the query will be directed to in the database. For example, when the query type is set to “type=0,” the query will be routed in the database to the above-described first set of calibration LUTs. When the query type is set to “type=1 ,” the query will be routed in the database to the above-described second set of calibration LUTs, and so on. The query submitted in the block 1008 also contains at least a subset of the set of photobleaching parameters determined in the block 1006 and the identifier of the biosensor 130 selected in the block 1002. In various examples, the query submitted in the block 1008 also typically contains any applicable auxiliary information about the sample and other pertinent metadata.

[0187] A block 1010 of the method 1000 includes the computing device 180 receiving a response to the query submitted in the block 1008 and optionally displaying for the user the received026389-0063-W001

[0188] response on a display device. The contents of the received response typically depend on the query type. For example, a response to the query “type=O” may contain a predicted identity of the neurotransmitter in the sample and its estimated concentration or amount in the sample layer 136 of the biosensor 130. The neurotransmitter identity can be predicted by matching one or both of the rate constants , and k2to the corresponding data in the first set of calibration LUTs. The concentration or amount of the neurotransmitter in the sample layer 137 can be estimated based on one or both of the amplitudes a and b. In some examples, the received response contains a ranked list of predicted identities along with their confidence scores computed based on the “distances” between the set of photobleaching parameters determined in the block 1006 and the corresponding calibration sets of photobleaching parameters. As another example, a response to the query “type=1” may contain a percentage ratio of the two neurotransmitters in the sample determined based on the photobleaching calibration data stored in the second set of calibration LUTs. Upon completion of the operations of the block 1010, the processing of the method 1000 is terminated.

[0189] FIG. 18 is a block diagram illustrating a computing device 1100 used in the optical system 100 according to some examples. In various examples, the optical system 100 may include a single computing device 1100 or multiple computing devices 1100. In some examples, the computing device 1100 implements the computing device 180 (see FIG. 16A). In various examples, an instance of the computing device 1100 can be used to implement the method 1000.

[0190] The computing device 1100 of FIG. 18 is illustrated as having a number of components, but any one or more of these components may be omitted or duplicated, as suitable for the application and setting. In some embodiments, some or all of the components included in the computing device 1100 may be attached to one or more motherboards and enclosed in a housing. In some embodiments, some of those components may be fabricated onto a single system-on-a-chip (SoC) (e.g., the SoC may include one or more electronic processing devices 1102 and one or more storage devices 1104). Additionally, in various embodiments, the computing device 1100 may not include one or more of the components illustrated in FIG. 18, but may include interface circuitry for coupling to the one or more components using any suitable interface (e.g., a Universal Serial Bus (USB) interface, a High-Definition Multimedia Interface (HDMI) interface, a Controller Area Network (CAN) interface, a Serial Peripheral Interface (SPI) interface, an Ethernet interface, a wireless interface, or any other appropriate interface). For example, the computing device 1100 may not include a display device 1110, but may include display device interface circuitry (e.g., a connector and driver circuitry) to which an external display device 1110 may be coupled.026389-0063-W001

[0191] The computing device 1100 includes a processing device 1102 (e.g., one or more processing devices). As used herein, the terms “electronic processor device” and “processing device” interchangeably refer to any device or portion of a device that processes electronic data from registers and / or memory to transform that electronic data into other electronic data that may be stored in registers and / or memory. In various embodiments, the processing device 1102 may include one or more digital signal processors (DSPs), application-specific integrated circuits (ASICs), central processing units (CPUs), graphics processing units (GPUs), server processors, field programmable gate arrays (FPGA), or any other suitable processing devices.

[0192] The computing device 1100 also includes a storage device 1104 (e.g., one or more storage devices). In various embodiments, the storage device 1104 may include one or more memory devices, such as random-access memory (RAM) devices (e.g., static RAM (SRAM) devices, magnetic RAM (MRAM) devices, dynamic RAM (DRAM) devices, resistive RAM (RRAM) devices, or conductive-bridging RAM (CBRAM) devices), hard drive-based memory devices, solid-state memory devices, networked drives, cloud drives, or any combination of memory devices. In some embodiments, the storage device 1104 may include memory that shares a die with the processing device 1102. In such an embodiment, the memory may be used as cache memory and include embedded dynamic random-access memory (eDRAM) or spin transfer torque magnetic random-access memory (STT-MRAM), for example. In some embodiments, the storage device 1104 may include non-transitory computer readable media having instructions thereon that, when executed by one or more processing devices (e.g., the processing device 1102), cause the computing device 1100 to perform any appropriate ones of the methods disclosed herein below or portions of such methods.

[0193] The computing device 1100 further includes an interface device 1106 (e.g., one or more interface devices 1106). In various embodiments, the interface device 1106 may include one or more communication chips, connectors, and / or other hardware and software to govern communications between the computing device 1100 and other computing devices. For example, the interface device 1106 may include circuitry for managing wireless communications for the transfer of data to and from the computing device 1100. The term “wireless” and its derivatives may be used to describe circuits, devices, systems, methods, techniques, communications channels, etc., that may communicate data via modulated electromagnetic radiation through a nonsolid medium. The term does not imply that the associated devices do not contain any wires, although in some embodiments they might not. Circuitry included in the interface device 1106 for managing wireless communications may implement any of a number of wireless standards or protocols, including but not limited to Institute for Electrical and Electronic Engineers (IEEE)026389-0063-W001

[0194] standards including Wi-Fi (IEEE 802.11 family), IEEE 802.16 standards, Long-Term Evolution (LTE) project along with any amendments, updates, and / or revisions (e.g., advanced LTE project, ultramobile broadband (UMB) project (also referred to as “3GPP2”), etc.). In some embodiments, circuitry included in the interface device 1106 for managing wireless communications may operate in accordance with a Global System for Mobile Communication (GSM), General Packet Radio Service (GPRS), Universal Mobile Telecommunications System (UMTS), High Speed Packet Access (HSPA), Evolved HSPA (E-HSPA), or LTE network. In some embodiments, circuitry included in the interface device 1106 for managing wireless communications may operate in accordance with Enhanced Data for GSM Evolution (EDGE), GSM EDGE Radio Access Network (GERAN), Universal Terrestrial Radio Access Network (UTRAN), or Evolved UTRAN (E-UTRAN). In some embodiments, circuitry included in the interface device 1106 for managing wireless communications may operate in accordance with Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Digital Enhanced Cordless Telecommunications (DECT), Evolution-Data Optimized (EV-DO), and derivatives thereof, as well as any other wireless protocols that are designated as 3G, 4G, 5G, and beyond. In some embodiments, the interface device 1106 may include one or more antennas (e.g., one or more antenna arrays) configured to receive and / or transmit wireless signals.

[0195] In some embodiments, the interface device 1106 may include circuitry for managing wired communications, such as electrical, optical, or any other suitable communication protocols. For example, the interface device 1106 may include circuitry to support communications in accordance with Ethernet technologies. In some embodiments, the interface device 1106 may support both wireless and wired communication, and / or may support multiple wired communication protocols and / or multiple wireless communication protocols. For example, a first set of circuitry of the interface device 1106 may be dedicated to shorter-range wireless communications such as Wi-Fi or Bluetooth, and a second set of circuitry of the interface device 1106 may be dedicated to longer-range wireless communications such as global positioning system (GPS), EDGE, GPRS, CDMA, WiMAX, LTE, EV-DO, or others. In some other embodiments, a first set of circuitry of the interface device 1106 may be dedicated to wireless communications, and a second set of circuitry of the interface device 1106 may be dedicated to wired communications.

[0196] The computing device 1100 also includes battery / power circuitry 1108. In various embodiments, the battery / power circuitry 1108 may include one or more energy storage devices (e.g., batteries or capacitors) and / or circuitry for coupling components of the computing device 1100 to an energy source separate from the computing device 1100 (e.g., to AC line power).026389-0063-W001

[0197] The computing device 1100 also includes a display device 1110 (e.g., one or multiple individual display devices). In various embodiments, the display device 1110 may include any visual indicators, such as a heads-up display, a computer monitor, a projector, a touchscreen display, a liquid crystal display (LCD), a light-emitting diode display, or a flat panel display.

[0198] The computing device 1100 also includes additional input / output (I / O) devices 1112. In various embodiments, the I / O devices 1112 may include one or more data / signal transfer interfaces, audio I / O devices (e.g., microphones or microphone arrays, speakers, headsets, earbuds, alarms, etc.), audio codecs, video codecs, printers, sensors (e.g., thermocouples or other temperature sensors, humidity sensors, pressure sensors, vibration sensors, etc.), image capture devices (e.g., one or more cameras), human interface devices (e.g., keyboards, cursor control devices, such as a mouse, a stylus, a trackball, or a touchpad), etc.

[0199] Depending on the specific embodiment of the optical system 100, various components of the interface devices 1106 and / or I / O devices 1112 can be configured to send and receive suitable control messages, suitable control / telemetry signals, and streams of data. In some examples, the interface devices 1106 and / or I / O devices 1112 include one or more analog-to-digital converters (ADCs) for transforming received analog signals into a digital form suitable for operations performed by the processing device 1102 and / or the storage device 1104. In some additional examples, the interface devices 1106 and / or I / O devices 1112 include one or more digital-to-analog converters (DACs) for transforming digital signals provided by the processing device 1102 and / or the storage device 1104 into an analog form suitable for being communicated to the corresponding components of the optical system 100.

[0200] According to an example embodiment disclosed above, provided is an apparatus comprising: a fluorimeter configured to measure a time series of fluorescence spectra of a biosensor having loaded thereon an analyte and subjected to illumination by ultraviolet light, the biosensor including an engineered layer of plasmonic material, the analyte including a neurotransmitter, different ones of the fluorescence spectra in the time series corresponding to different respective illumination times; and a computing device configured to: determine a first rate constant based on the time series, the first rate constant corresponding to a photochemical reaction of the neurotransmitter caused by the ultraviolet light; submit to a database a query containing at least the determined first rate constant, the database including photobleaching calibration data representing a plurality of different neurotransmitters and a plurality of different biosensors; and display on a display device a response to the submitted query received from the database.026389-0063-W001

[0201] Machine readable storage including machine-readable instructions, when executed, to implement a method or realize an apparatus in any of the examples of the present application.

[0202] Various techniques, or certain aspects or portions thereof, may take the form of program code (i.e., instructions) embodied in tangible media, such as floppy diskettes, CD-ROMs, hard drives, a non-transitory computer readable storage medium, or any other machine-readable storage medium wherein, when the program code is loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the various techniques. In the case of program code execution on programmable computers, the computing device may include a processor, a storage medium readable by the processor (including volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device. The volatile and non-volatile memory and / or storage elements may be a RAM, an EPROM, a flash drive, an optical drive, a magnetic hard drive, or another medium for storing electronic data. The eNB (or other base station) and UE (or other mobile station) may also include a transceiver component, a counter component, a processing component, and / or a clock component or timer component. One or more programs that may implement or utilize the various techniques described herein may use an application programming interface (API), reusable controls, and the like. Such programs may be implemented in a high-level procedural or an object-oriented programming language to communicate with a computer system. However, the program(s) may be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or an interpreted language, and combined with hardware implementations.

[0203] It should be understood that many of the functional units described in this specification may be implemented as one or more components, which is a term used to more particularly emphasize their implementation independence. For example, a component may be implemented as a hardware circuit comprising custom very large scale integration (VLSI) circuits or gate arrays, off-the-shelf semiconductors such as logic chips, transistors, or other discrete components. A component may also be implemented in programmable hardware devices such as field programmable gate arrays, programmable array logic, programmable logic devices, or the like.

[0204] Components may also be implemented in software for execution by various types of processors. An identified component of executable code may, for instance, comprise one or more physical or logical blocks of computer instructions, which may, for instance, be organized as an object, a procedure, or a function. Nevertheless, the executables of an identified component need not be physically located together, but may comprise disparate instructions stored in different locations that, when joined logically together, comprise the component and achieve the stated purpose for the component.026389-0063-W001

[0205] Indeed, a component of executable code may be a single instruction, or many instructions, and may even be distributed over several different code segments, among different programs, and across several memory devices. Similarly, operational data may be identified and illustrated herein within components, and may be embodied in any suitable form and organized within any suitable type of data structure. The operational data may be collected as a single data set, or may be distributed over different locations including over different storage devices, and may exist, at least partially, merely as electronic signals on a system or network. The components may be passive or active, including agents operable to perform desired functions.

[0206] Reference throughout this specification to “an example” means that a particular feature, structure, or characteristic described in connection with the example is included in at least one embodiment of the present invention. Thus, appearances of the phrase “in an example” in various places throughout this specification are not necessarily all referring to the same embodiment.

[0207] As used herein, a plurality of items, structural elements, compositional elements, and / or materials may be presented in a common list for convenience. However, these lists should be construed as though each member of the list is individually identified as a separate and unique member. Thus, no individual member of such list should be construed as a de facto equivalent of any other member of the same list solely based on its presentation in a common group without indications to the contrary. In addition, various embodiments and examples of the present invention may be referred to herein along with alternatives for the various components thereof. It is understood that such embodiments, examples, and alternatives are not to be construed as de facto equivalents of one another, but are to be considered as separate and autonomous representations of the present invention.

[0208] Many different arrangements of the various methods and components depicted, as well as steps or components not shown, are possible without departing from the spirit and scope of the present disclosure. Embodiments of the present disclosure have been described with the intent to be illustrative rather than restrictive. Alternative embodiments will become apparent to those skilled in the art that do not depart from its scope. A skilled artisan may develop alternative means of implementing improvements without departing from the scope of the present disclosure. It should thus be noted that the matter contained in the above description or shown in the accompanying drawings is to be interpreted as illustrative and not in a limiting sense.

[0209] The functionality described herein as being performed by one component may be performed by multiple components in a distributed manner. Likewise, functionality performed by multiple components may be consolidated and performed by a single component. Similarly, a component described as performing particular functionality may also perform additional026389-0063-W001

[0210] functionality not described herein. For example, a device or structure that is “configured” in a certain way is configured in at least that way but may also be configured in ways that are not listed.

[0211] With regard to the processes, systems, methods, heuristics, etc. described herein, it should be understood that, although the steps of such processes, etc. have been described as occurring according to a certain ordered sequence, such processes could be practiced with the described steps performed in an order other than the order described herein. It further should be understood that certain steps could be performed simultaneously, that other steps could be added, or that certain steps described herein could be omitted. In other words, the descriptions of processes herein are provided for the purpose of illustrating certain embodiments and should in no way be construed so as to limit the claims.

[0212] Accordingly, it is to be understood that the above description is intended to be illustrative and not restrictive. Many embodiments and applications other than the examples provided would be apparent upon reading the above description. The scope should be determined, not with reference to the above description, but should instead be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled. It is anticipated and intended that future developments will occur in the technologies discussed herein, and that the disclosed systems and methods will be incorporated into such future embodiments. In sum, it should be understood that the application is capable of modification and variation.

[0213] Embodiments described herein, including devices, control systems, or methods described herein may be implemented using specifically designed hardware, configurable hardware, programmable data processors configured by the provision of software (which may optionally comprise “firmware”) capable of executing on the data processors, special purpose computers or data processors that are specifically programmed, configured, or constructed to perform one or more steps in a method as explained in detail herein and / or combinations of two or more of these. Examples of specifically designed hardware are: logic circuits, application-specific integrated circuits (“ASICs”), large scale integrated circuits (“LSIs”), very large scale integrated circuits (“VLSIs”), and the like. Examples of configurable hardware are: one or more programmable logic devices such as programmable array logic (“PALs”), programmable logic arrays (“PL s”), and field programmable gate arrays (“FPGAs”)). Examples of programmable data processors are: microprocessors, digital signal processors (“DSPs”), embedded processors, graphics processors, math co-processors, general purpose computers, server computers, cloud computers, mainframe computers, computer workstations, and the like. For example, one or more data processors in a026389-0063-W001

[0214] computer system for a device may implement methods as described herein by executing software instructions in a program memory accessible to the processors.

[0215] Processing may be centralized or distributed. Where processing is distributed, information including software and / or data may be kept centrally or distributed. Such information may be exchanged between different functional units by way of a communications network, such as a Local Area Network (LAN), Wide Area Network (WAN), or the Internet, wired or wireless data links, electromagnetic signals, or other data communication channels.

[0216] While processes or blocks are presented in a given order, alternative examples may perform routines having steps, or employ systems having blocks, in a different order, and some processes or blocks may be deleted, moved, added, subdivided, combined, and / or modified to provide alternative or subcombinations. Each of these processes or blocks may be implemented in a variety of different ways. Also, while processes or blocks are at times shown as being performed in series, these processes or blocks may instead be performed in parallel or may be performed at different times.

[0217] In addition, while elements are at times shown as being performed sequentially, they may instead be performed simultaneously or in different sequences. It is therefore intended that the following claims are interpreted to include all such variations as are within their intended scope.

[0218] Embodiments described herein may also be provided in the form of a program product. The program product may comprise any non-transitory medium which carries a set of computer-readable instructions which, when executed by a data processor, cause the data processor to execute a method of the invention. Program products according to the invention may be in any of a wide variety of forms. The program product may comprise, for example, non-transitory media such as magnetic data storage media including floppy diskettes, hard disk drives, optical data storage media including CD ROMs, DVDs, electronic data storage media including ROMs, flash RAM, EPROMs, hardwired or preprogrammed chips (e.g., EEPROM semiconductor chips), nanotechnology memory, or the like. The computer-readable signals on the program product may optionally be compressed or encrypted.

[0219] The methods described herein may be implemented in software. “Software” includes any instructions executed on a processor and may include (but is not limited to) firmware, resident software, microcode, and the like. Both processing hardware and software may be centralized or distributed (or a combination thereof), in whole or in part, as known to those skilled in the art. For example, software and other modules may be accessible via local memory, via a network, via a browser or other application in a distributed computing context, or via other means suitable for the purposes described above.026389-0063-W001

[0220] Where a component (e.g., a controller, software module, processor, server, client, device, etc.) is referred to, unless otherwise indicated, reference to that component (including a reference to a “means”) should be interpreted as including as equivalents of that component any component which performs the function of the described component (i.e., that is functionally equivalent), including components which are not structurally equivalent to the disclosed structure which performs the function in the illustrated exemplary embodiments of the invention.

[0221] Furthermore, some embodiments described herein may include one or more electronic processors configured to perform the described functionality by executing instructions stored in non-transitory, computer-readable medium. Similarly, embodiments described herein may be implemented as non-transitory, computer-readable medium storing instructions executable by one or more electronic processors to perform the described functionality. As used in the present application, “non-transitory computer-readable medium” comprises all computer-readable media but does not consist of a transitory, propagating signal. Accordingly, non-transitory computer-readable medium may include, for example, a hard disk, a CD-ROM, an optical storage device, a magnetic storage device, a ROM (Read Only Memory), a RAM (Random Access Memory), register memory, a processor cache, or any combination thereof.

[0222] Many of the modules and logical structures described are capable of being implemented in software executed by a microprocessor or a similar device or of being implemented in hardware using a variety of components including, for example, application specific integrated circuits (“ASICs”). Terms like “controller” and “module” may include or refer to both hardware and / or software. The claims should not be limited to specific examples or terminology or to any specific hardware or software implementation or combination of software or hardware. Also, if an apparatus, method, or system is claimed, for example, as including a controller, module, logic, electronic processor, or other element configured in a certain manner, for example, to perform multiple functions, the claim or claim element should be interpreted as meaning one or more controllers, modules, logic elements, electronic processors other elements where any one of the one or more elements is configured as claimed, for example, to perform any one or more of the recited multiple functions.

[0223] One embodiment described herein is a system for detecting monoamine neurotransmitters, comprising: a plasmonic substrate comprising aluminum concave nanocubes; an ultraviolet light source configured to illuminate the plasmonic substrate; a spectrometer configured to collect fluorescence data from the surface of the plasmonic substrate, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired over a time period; and a processor configured to execute a machine learning model trained to classify the monoamine026389-0063-W001

[0224] neurotransmitters based on the fluorescence data. In one aspect, the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers. In another aspect, the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 200-300 nanometers. In another aspect, the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 266 nanometers. In another aspect, the wherein the machine learning model comprises a Long Short-Term Memory network. In another aspect, the Long Short-Term Memory network comprises: three sequential Long Short-Term Memory layers; batch normalization layers following each of the three sequential Long Short-Term Memory layers; a fully connected dense layer; and a softmax output layer configured to classify the monoamine neurotransmitters into a plurality of classes. In another aspect, the plurality of classes comprises dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In another aspect, the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers.

[0225] Another embodiment described herein is a method for classifying monoamine neurotransmitters, comprising: depositing a sample containing monoamine neurotransmitters onto a plasmonic substrate comprising aluminum concave nanocubes; illuminating the plasmonic substrate with ultraviolet light; collecting fluorescence data comprising a plurality of fluorescence spectra over a time period during continuous illumination; preprocessing the fluorescence data to generate standardized input data; and classifying the monoamine neurotransmitters by applying a trained machine learning model to the standardized input data. In one aspect, the sample comprises a solution and depositing the sample comprises drop casting the solution onto the plasmonic substrate to form a coffee ring pattern. In another aspect, illuminating the plasmonic substrate comprises directing the ultraviolet light toward an outer edge of the coffee ring pattern where a concentration of the monoamine neurotransmitters is higher. In another aspect, collecting the fluorescence data comprises acquiring the plurality of fluorescence spectra at intervals of approximately 0.5 seconds over a time period of 2 to 3 minutes. In another aspect, each fluorescence spectrum of the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers. In another aspect, preprocessing the autofluorescence time decay series data comprises: shifting time series data to eliminate leading missing values; truncating or padding the time series data to a standardized sequence length; and normalizing feature values using MinMax scaling. In another aspect, the standardized sequence length comprises 10-25 time steps. In another aspect, the wherein the trained machine learning model comprises a Long Short-Term Memory network.026389-0063-W001

[0226] Another embodiment described herein is a non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to: receive fluorescence data collected from a sample comprising one or more monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired during continuous ultraviolet illumination; preprocess the fluorescence data by aligning time series traces and standardizing sequence lengths to generate input tensors; apply a trained neural network to the input tensors to capture temporal dependencies in the autofluorescence time decay series data; and output a classification of the monoamine neurotransmitters based on analysis by the trained neural network. In one aspect, the trained neural network is a trained Long Short Term Memory network. In another aspect, the instructions further cause the processor to normalize feature values of the autofluorescence time decay series data to a range of 0 to 1 using MinMax scaling prior to applying the trained Long Short-Term Memory network. In another aspect, preprocessing the autofluorescence time decay series data comprises shifting time series traces to eliminate leading missing values by aligning valid entries to a common starting point. In another aspect, the standardized sequence lengths comprise 17 time steps, and wherein the instructions further cause the processor to truncate longer sequences to 17 regularly spaced sequences or pad shorter sequences to achieve the standardized sequence lengths. In another aspect, the trained Long Short-Term Memory network comprises: three sequential Long Short-Term Memory layers having 200, 150, and 100 units respectively; batch normalization layers following each of the three sequential Long Short-Term Memory layers; a fully connected dense layer with 50 neurons; and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In another aspect, the instructions further cause the processor to process a single input tensor in under 1 millisecond to enable real-time classification of the monoamine neurotransmitters.

[0227] Another embodiment described herein is a biosensing device for label-free detection and classification of monoamine neurotransmitters, comprising: a substrate having a surface with aluminum concave nanocubes deposited thereon; an excitation source configured to direct ultraviolet light onto the substrate to excite native autofluorescence from monoamine neurotransmitters in contact with the aluminum concave nanocubes; a detector configured to acquire time-resolved fluorescence spectra from the monoamine neurotransmitters over a measurement period to generate autofluorescence time decay series data; and a classification module comprising a processor and a memory storing a trained recurrent neural network, wherein the classification module is configured to receive the autofluorescence time decay series data026389-0063-W001

[0228] and output an identification of the monoamine neurotransmitters based on temporal decay characteristics captured by the trained recurrent neural network. In one aspect, the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers with a standard deviation of approximately 9 nanometers. In another aspect, the trained recurrent neural network comprises a Long Short-Term Memory network having three sequential Long Short-Term Memory layers with batch normalization layers following each of the three sequential Long Short-Term Memory layers. In another aspect, the Long Short-Term Memory network further comprises a fully connected dense layer and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid. In another aspect, the excitation source comprises a continuous wave solid state ultraviolet laser configured to emit light at a wavelength of approximately 266 nanometers, and wherein the detector is configured to acquire the time-resolved fluorescence spectra at intervals of approximately 0.5 seconds over a measurement period of 2 to 3 minutes.

[0229] It will be apparent to one of ordinary skill in the relevant art that suitable modifications and adaptations to the compositions, formulations, methods, processes, and applications described herein can be made without departing from the scope of any embodiments or aspects thereof. The compositions and methods provided are exemplary and are not intended to limit the scope of any of the specified embodiments. All of the various embodiments, aspects, and options disclosed herein can be combined in any variations or iterations. The scope of the compositions, formulations, methods, and processes described herein include all actual or potential combinations of embodiments, aspects, options, examples, and preferences herein described. The exemplary compositions and formulations described herein may omit any component, substitute any component disclosed herein, or include any component disclosed elsewhere herein. The ratios of the mass of any component of any of the compositions or formulations disclosed herein to the mass of any other component in the formulation or to the total mass of the other components in the formulation are hereby disclosed as if they were expressly disclosed. Should the meaning of any terms in any of the patents or publications incorporated by reference conflict with the meaning of the terms used in this disclosure, the meanings of the terms or phrases in this disclosure are controlling. Furthermore, the foregoing discussion discloses and describes merely exemplary embodiments. All patents and publications cited herein are incorporated by reference herein for the specific teachings thereof.

[0230] Various embodiments and aspects of the inventions described herein are summarized by the following clauses:

[0231] Clause 1. A system for detecting monoamine neurotransmitters, comprising:026389-0063-W001

[0232] a plasmonic substrate comprising aluminum concave nanocubes;

[0233] an ultraviolet light source configured to illuminate the plasmonic substrate;

[0234] a spectrometer configured to collect fluorescence data from the surface of the plasmonic substrate, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired over a time period; and

[0235] a processor configured to execute a machine learning model trained to classify the monoamine neurotransmitters based on the fluorescence data.

[0236] Clause 2. The system of clause 1, wherein the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers.

[0237] Clause 3. The system of clause 1 or 2, wherein the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 200-300 nanometers.

[0238] Clause 4. The system of any one of clauses 1-2, wherein the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 266 nanometers.

[0239] Clause 5. The system of any one of clauses 1-3, wherein the wherein the machine learning model comprises a Long Short-Term Memory network.

[0240] Clause 6. The system of any one of clauses 1—4, wherein the Long Short-Term Memory network comprises:

[0241] three sequential Long Short-Term Memory layers;

[0242] batch normalization layers following each of the three sequential Long Short-Term Memory layers;

[0243] a fully connected dense layer; and

[0244] a softmax output layer configured to classify the monoamine neurotransmitters into a plurality of classes.

[0245] Clause 7. The system of any one of clauses 1-7, wherein the plurality of classes comprises dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0246] Clause 8. The system of any one of clauses 1-8, wherein the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers.

[0247] Clause 9. A method for classifying monoamine neurotransmitters, comprising:

[0248] depositing a sample containing monoamine neurotransmitters onto a plasmonic substrate comprising aluminum concave nanocubes;

[0249] illuminating the plasmonic substrate with ultraviolet light;026389-0063-W001

[0250] collecting fluorescence data comprising a plurality of fluorescence spectra over a time period during continuous illumination;

[0251] preprocessing the fluorescence data to generate standardized input data; and classifying the monoamine neurotransmitters by applying a trained machine learning model to the standardized input data.

[0252] Clause 10. The method of clause 9, wherein the sample comprises a solution and depositing the sample comprises drop casting the solution onto the plasmonic substrate to form a coffee ring pattern.

[0253] Clause H. The method of clause 9 or 10, wherein illuminating the plasmonic substrate comprises directing the ultraviolet light toward an outer edge of the coffee ring pattern where a concentration of the monoamine neurotransmitters is higher.

[0254] Clause 12. The method of any one of clauses 9-11, wherein collecting the fluorescence data comprises acquiring the plurality of fluorescence spectra at intervals of approximately 0.5 seconds over a time period of 2 to 3 minutes.

[0255] Clause 13. The method of any one of clauses 9-12, wherein each fluorescence spectrum of the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers. Clause 14. The method of any one of clauses 9-13, wherein preprocessing the autofluorescence time decay series data comprises:

[0256] shifting time series data to eliminate leading missing values;

[0257] truncating or padding the time series data to a standardized sequence length; and normalizing feature values using MinMax scaling.

[0258] Clause 15. The method of any one of clauses 9-14, wherein the standardized sequence length comprises 10-25 time steps.

[0259] Clause 16. The method of any one of clauses 9-15, wherein the wherein the trained machine learning model comprises a Long Short-Term Memory network.

[0260] Clause 17. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to:

[0261] receive fluorescence data collected from a sample comprising one or more monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired during continuous ultraviolet illumination; preprocess the fluorescence data by aligning time series traces and standardizing sequence lengths to generate input tensors;026389-0063-W001

[0262] apply a trained neural network to the input tensors to capture temporal dependencies in the autofluorescence time decay series data; and

[0263] output a classification of the monoamine neurotransmitters based on analysis by the trained neural network.

[0264] Clause 18. The non-transitory computer-readable medium of clause 17, wherein the trained neural network is a trained Long Short Term Memory network.

[0265] Clause 19. The non-transitory computer-readable medium of clause 17 or 18, wherein the instructions further cause the processor to normalize feature values of the autofluorescence time decay series data to a range of 0 to 1 using MinMax scaling prior to applying the trained Long Short-Term Memory network.

[0266] Clause 20. The non-transitory computer-readable medium of any one of clauses 17-19, wherein preprocessing the autofluorescence time decay series data comprises shifting time series traces to eliminate leading missing values by aligning valid entries to a common starting point.

[0267] Clause 21. The non-transitory computer-readable medium of any one of clauses 17-20, wherein the standardized sequence lengths comprise 17 time steps, and wherein the instructions further cause the processor to truncate longer sequences to 17 regularly spaced sequences or pad shorter sequences to achieve the standardized sequence lengths.

[0268] Clause 22. The non-transitory computer-readable medium of any one of clauses 17-21, wherein the trained Long Short-Term Memory network comprises:

[0269] three sequential Long Short-Term Memory layers having 200, 150, and 100 units respectively;

[0270] batch normalization layers following each of the three sequential Long Short-Term Memory layers;

[0271] a fully connected dense layer with 50 neurons; and

[0272] a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0273] Clause 23. The non-transitory computer-readable medium of any one of clauses 17-22, wherein the instructions further cause the processor to process a single input tensor in under 1 millisecond to enable real-time classification of the monoamine neurotransmitters. Clause 24. A biosensing device for label-free detection and classification of monoamine neurotransmitters, comprising:

[0274] a substrate having a surface with aluminum concave nanocubes deposited thereon;026389-0063-W001

[0275] an excitation source configured to direct ultraviolet light onto the substrate to excite native autofluorescence from monoamine neurotransmitters in contact with the aluminum concave nanocubes;

[0276] a detector configured to acquire time-resolved fluorescence spectra from the monoamine neurotransmitters over a measurement period to generate autofluorescence time decay series data; and

[0277] a classification module comprising a processor and a memory storing a trained recurrent neural network, wherein the classification module is configured to receive the autofluorescence time decay series data and output an identification of the monoamine neurotransmitters based on temporal decay characteristics captured by the trained recurrent neural network.

[0278] Clause 25. The biosensing device of clause 24, wherein the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers with a standard deviation of approximately 9 nanometers.

[0279] Clause 26. The biosensing device of clause 24 or 25, wherein the trained recurrent neural network comprises a Long Short-Term Memory network having three sequential Long Short-Term Memory layers with batch normalization layers following each of the three sequential Long Short-Term Memory layers.

[0280] Clause 27. The biosensing device of any one of clauses 24-26, wherein the Long Short-Term Memory network further comprises a fully connected dense layer and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

[0281] Clause 28. The biosensing device of any one of clauses 24-7, wherein the excitation source comprises a continuous wave solid state ultraviolet laser configured to emit light at a wavelength of approximately 266 nanometers, and wherein the detector is configured to acquire the time-resolved fluorescence spectra at intervals of approximately 0.5 seconds over a measurement period of 2 to 3 minutes.026389-0063-W001

[0282] EXAMPLES

[0283] Aluminum concave nanocubes solution with nominal diameters of 80 ± 9 nm was purchased from NanoComposix (particle concentration: 3.9 * 1012particles per mL, mass concentration: 2.8 mg mL-1, with a surface area of 27.3 m2g-1, and were dissolved in 1-propanol). Dopamine, 3,4-dihydroxyphenylacetic acid, and norepinephrine (>99.9%) were obtained from Sigma-Aldrich.

[0284] Fluorescence measurements of the neurotransmitters were conducted on two substrates: AICNC and plain silicon wafers as a reference. The AICNC structure was selected based on our hypothesis that its concave geometry, featuring sharp corners and edges similar to those in aluminum bowtie nano antennas, enables it to more effectively trap and concentrate light, leading to stronger and more consistent fluorescence enhancement compared to smoother geometries or spherical shapes, such as aluminum hole arrays, and aluminum nanotriangles. A standard 2-inch silicon wafer was cut into four pieces and treated with a plasma cleaner for 90 seconds at a base pressure of 0.4 torr to render the surface hydrophilic. Subsequently, 5 pL of the AICNC nanoparticle (NP) solution was drop-cast onto the silicon substrate and allowed to dry in ambient conditions; offering a simple and time-efficient fabrication route without requiring complex methods or equipment such as e-beam lithography or sputtering. After drying, a multi-layer pattern of the nanoparticles remained on the substrate surfaces. The structural and compositional analysis of AICNCs was performed using scanning electron microscopy (SEM) and scanning transmission electron microscopy (STEM). Following substrate preparation, solutions of three MANTs in deionized water at varying concentrations were prepared, and 1 pL of each solution was drop-cast onto two different substrates. The outer edge of the coffee ring pattern of the AICNC where high concentration of AICNCs were found and a Si wafer to measure the fluorescence spectrum. The dried MANTs were then exposed to a 266 nm CW solid state UV laser CryLas FQCW266-10 (incident angle of 60°) with an output power of ~5 mW and a circular spot size of ~100 pm in diameter, and AFTDS were collected using a Horiba I HR 550 spectrometer coupled with a CCD camera. The AFTDS were taken over 2-3 minutes and an exposure time of 0.5 seconds was used for each spectrum within the AFTDS.

[0285] A comparative analysis was performed using three ML techniques: first, LSTM was used; a type of recurrent neural network well-suited for sequence prediction, which can capture temporal dependencies in the fluorescence data, potentially improving the accuracy for distinguishing closely related biomolecules. While LSTM has been applied for capturing sequential patterns and forecasting time-series tasks such as molecular generation, property prediction, and dynamic process modeling, this study is the first application of LSTM in classifying molecules based on026389-0063-W001

[0286] AFTDS. Secondly, KNN was used; a simple, instance-based learning algorithm that classifies data points based on the closest training examples in the feature space. Third, we used RF, an ensemble learning method based on decision trees, to classify the spectral data and assess its performance compared to other algorithms. The AFTDS data was recorded as a time-dependent spectrum spanning 280-360 nm. Each fluorescence time series originally contained up to 240-360 spectra (0.5 s interval over 2-3 minutes). However, because the number of time-dependent spectra for each illumination spot (sequence length) varied across experiments (ranging from 0-17 to 240-360 spectra after preprocessing), all AFTDS inputs were standardized to the shortest complete sequence length of 17. Longer sequences were truncated to 17 regularly spaced sequences, and shorter ones were padded if necessary. This variability was due to the adjustment of sample position to find a high fluorescence signal spot that leads to delayed collection of fluorescent data after shutter opens at time = 0. This length adjustment ensured consistent input dimensions for the LSTM while preserving the overall decay dynamics.

[0287] Adjustment details are shown in FIG. 2, Table 1, and Table 2. These were converted into structured input tensors of shape (time_steps wavelength_bins) for sequence-based models (LSTM) or flattened into feature vectors for non-sequential models (KNN, RF). Processing the AFTDS with ML was accomplished in different steps; the fluorescence data gathered from multiple experiments was used to train the ML model. Each input sample was labeled with the corresponding neurotransmitter class (DA, DOPAC, NE) to train supervised classifiers. The training process involved pre-processing the fluorescence data, feature extraction, and applying data augmentation techniques to improve the model’s accuracy and generalizability.

[0288] Distinguishing neurotransmitters with similar structures solely based on their native fluorescence is challenging due to their overlapping fluorescence spectra. To address this challenge, we employed ML classification models and developed a reliable classification framework capable of differentiating neurotransmitters based on their unique fluorescence signatures. This approach allows for more precise detection and identification of neurotransmitters. The integration of ML not only improves classification accuracy but also provides a scalable method for analyzing large experimental datasets.

[0289] All fluorescence data was scaled with MinMaxScaler and evaluated model performance with 4-fold cross-validation (75% train, 25% test) as FIG. 3 shows the workflow. The dataset was first stratified by neurotransmitter class to maintain balanced representation in each fold. For every split, three folds (75%) were used exclusively for training, and the remaining fold (25%) was held out for validation. This process was repeated four times so that each fold served once as the validation set, and the reported metrics correspond to the mean across folds. To prevent data 026389-0063-W001

[0290] leakage, preprocessing steps such as MinMax scaling fit only on the training partition within each fold and then applied to the validation data. Importantly, spectra from the same experimental replicate were kept within a single fold, ensuring that no replicate contributed to both training and validation sets. For the in-solution emission spectra, each sample was a static snapshot of paired wavelength-intensity values fed directly to KNN and a 150-tree RF. For the AFTDS, we supplied KNN and RF with flattened, right-aligned intensity vectors, while the LSTM received the raw 17-step sequence, letting its three stacked LSTM layers learn temporal dependencies before batch normalization, a dense layer, and soft-max output. This unified preprocessing lets each model exploit either static spectral structure or time dynamics as appropriate, yielding consistent classification performance across modalities.

[0291] The AFTDS data was recorded as a series of spectra spanning 280-360 nm after sequential laser exposure over 2-3 minutes with 0.5 seconds as the exposure time. Each fluorescence time series originally contained 240-360 spectra (collected at 0.5 s intervals over 2-3 minutes). When saving autofluorescence data, the spectrometer software starts numbering each spectra from to when shutter opens. However, collection of actual spectra data might start at a later time with higher t number (e.g., tn) since adjustment of sample position might be necessary to find one spot with high fluorescence signal. The figure above illustrates this for clarification.

[0292] Before the data can be utilized effectively for machine learning models, it must first undergo curation to ensure it is appropriately structured. The goal of the task is to predict the molecule type based on the time series data of fluorescent intensities. Therefore, the dataset needs to be reshaped, where the time intervals are reflected in separate columns, representing the intensity values captured at different time points. Each row should correspond to the time dependent fluorescent intensity at a particular wavelength, with its name and spot number (Seg) serving as the label. This reshaped data structure is illustrated in Table 1.

[0293]

[0294] 026389-0063-W001

[0295]

[0296] As depicted in Table 1, it can be observed that some rows do not begin at time zero. This results in missing values (NaN) at the start of certain rows. To address this issue, each row needs to be shifted left to eliminate the leading Nan values, effectively aligning all rows to start at the same initial time point. The corrected data are displayed in Table 2.

[0297]

[0298] Once the rows are adjusted to start from time zero, the next step is to compute the minimum and maximum lengths of the non-Nan values in each row. This is essential for understanding the variation in the recorded intensity values across time and ensuring consistency in data processing.

[0299] It was assumed that the intensity values are dependent on each other through time, making the time series aspect crucial for accurate prediction of the molecule type. By aligning and trimming the time series data appropriately, the machine learning model will be better equipped to capture the underlying patterns in fluorescent intensity across different molecules.

[0300] To address the varying lengths of the time series data, it is necessary to feed fixed-length vectors into the LSTM model. Based on the minimum length of the time series, which is 17 (data from t0to fi6), any time series longer than this length will be sliced into smaller segments, each of length 17. This technique, referred to as slicing time series, augments the dataset by creating026389-0063-W001

[0301] multiple fixed-length inputs from longer sequences. By ensuring that each segment conforms to the fixed input size (17), the model can better generalize and handle varying input lengths during training.

[0302] Methods for Machine Learning Classification

[0303] Data Preprocessing

[0304] Two independent datasets were utilized in this study: (i) AFTDS data obtained from MANTs on an AICNC substrate, and (ii) static AF spectra recorded in solution. For the AFTDS dataset, temporal autofluorescence signals corresponding to DA, DOPAC, and NE were acquired under continuous illumination. These signals were arranged into a matrix with each row representing a single time-dependent trace. Missing values were handled by shifting Not a Number values, (NaNs) to the right, and shorter sequences were padded to match the length of the longest trace; additionally, traces with mixed labels were discarded and the remaining tags were standardized to “DA,” “DOPAC,” or “NE.” Conversely, in the AF-in-solution dataset, each sample consisted of a static autofluorescence intensity recorded at specific wavelengths. The values were restructured into [wavelength, intensity] pairs for each neurotransmitter, thereby forming a feature matrix amenable to conventional machine learning approaches. In both cases, feature values were normalized to the range [0, 1] using MinMax scaling. Class labels were converted using LabelEncoder for traditional classifiers and one-hot encoding for neural network models.

[0305] Model Design and Training

[0306] AF-in-Solution Models (KNN and RF):

[0307] The AF-in-solution dataset was analyzed using classical classifiers:

[0308] KNN: Following an extensive grid search, a value of k = 11 was selected to optimize crossvalidation performance using a 4-fold scheme.

[0309] RF: A forest comprising 150 decision trees (n_estimators = 150) was used, and model hyperparameters were tuned through grid search in a 4-fold cross-validation framework. Final performance was assessed on an independent 25% hold-out test set. Model performance was quantified using per-class precision, recall, Flscore, overall accuracy, and both macro and weighted averages.

[0310] Models on AFTDS (AICNC Substrate):

[0311] The AFTDS dataset was modeled using deep learning as well as classical approaches:026389-0063-W001

[0312] LSTM Network: The proposed LSTM architecture was implemented in Keras and comprised:

[0313] An input layer accommodating the time-series shape.

[0314] Three sequential LSTM layers with 200, 150, and 100 units, respectively, each followed by batch normalization to stabilize learning.

[0315] A fully connected (dense) layer with 50 neurons, followed by batch normalization. A final SoftMax output layer for classification into the three neurotransmitter classes.

[0316] The model was trained for 50 epochs using a batch size of 64, with a 75 / 25 trainingvalidation split. A learning rate scheduler was applied to reduce the learning rate upon plateauing of the validation loss.

[0317] KNN and RF: In addition to the LSTM model, both KNN and RF classifiers were also applied to the AFTDS dataset. The same preprocessing pipeline was used for feature extraction, with models evaluated through 4-fold cross-validation and subsequently on a stratified 25% hold-out test set. Performance metrics, including class-wise F1 -scores and confusion matrices, were reported.

[0318] Data Transformation for Time-Series Modeling

[0319] For sequence-based modeling, the AFTDS data were further preprocessed by rightaligning the time traces through shifting valid entries to the left and padding shorter sequences. Only traces exhibiting sufficient length and unambiguous class labels were retained to ensure that the dynamic spectral evolution was accurately captured by the LSTM network.

[0320] Performance Evaluation

[0321] Model performance was assessed using standard metrics: overall accuracy, per-class precision, recall, and F1-score, as well as macro and weighted averages. Confusion matrices displaying both raw counts and row-wise percentages were used to visualize classification performance and to elucidate error patterns, particularly highlighting class boundary confusion.

[0322] The AFTDS collected on AICNC substrates contain only seventeen regularly spaced intensity values per trace, so the data are sequential and short. A LSTM network suits this setting because its gated recurrent structure keeps information from both the fast initial quench and the slower tail of the decay curve, the two regions that together separate DA, DOPAC, and NE even though they occur at very different times.026389-0063-W001

[0323] The in-solution spectra, by contrast, are single wavelength-intensity snapshots. For these static measurements, we wanted algorithms that run quickly while still capturing subtle spectral patterns. An 11 -neighbor KNN classifier provides a simple non-parametric baseline that classifies by proximity in wavelength-intensity space, while a RF with 150 trees captures non-linear interactions among neighboring wavelengths and offers built-in feature importance scores that aid interpretation.

[0324] Simpler alternatives such as logistic regression and single decision trees were rejected before testing. Linear decision boundaries cannot recover the class separation that appears only after part of the fluorescence has decayed, and a single tree is sensitive to noise in individual wavelengths. Either option would likely underfit or overfit without reducing inference time, since RF and KNN already respond well under a millisecond on the intended devices. Therefore, we concentrated our experiments on the LSTM model for the sequential traces and on KNN together with RF as fast, interpretable baselines for the static spectra.

[0325] All analyses were performed in Python 3.10 using TensorFlow 2.18.0. The LSTM achieved 89% accuracy using 17-point standardized sequences derived from the original variable-length fluorescence time series, under 4-fold cross-validation with replicate-level partitioning to avoid data leakage. The training data was partitioned into four equal folds, with three folds (75%) used for training and one-fold (25%) reserved for validation in each iteration. This process was repeated four times, allowing each fold to serve as the validation set once. By exposing the model to diverse training and validation distributions, cross-validation effectively addressed class imbalance, enhanced generalization, and minimized bias. Real-time inference was benchmarked on Google Colab Pro with an NVIDIA A100 GPU and Intel Xeon CPU (2.20 GHz); the trained LSTM processed a single 17-point AFTDS trace in under 1 ms, confirming the framework’s suitability for real-time neurotransmitter classification.

[0326] FIG. 4 presents the training and validation curves of the LSTM model over 50 epochs, showing model accuracy and model loss. Both training and validation accuracy increase rapidly during the initial epochs and stabilize around 0.89, indicating good generalization and minimal overfitting. The k-fold cross-validation results further support the model’s consistency, with accuracy ranging from 88.2% to 90.9% and loss values stabilizing around 0.3 across folds.

[0327] The dataset used for model training and evaluation consisted of 146,432 samples for models trained on AICNC and 4113 samples for in-solution models. A 75-25% split was applied, allocating approximately 109,824 samples for training and 36,608 samples for testing in the AICNC dataset, while the in-solution dataset contained 2817 training samples and 939 test026389-0063-W001

[0328] samples the dataset captures both spectral and temporal fluorescence dynamics, providing a rich input for classification.

[0329] Results

[0330] FIG. 5A shows a SEM image of the coffee ring pattern (the dark circle) left by a drop of deionized water containing molecules after it evaporated on the AICNCs. The UV laser focuses near the dark circle where the concentration of molecules is higher than in other spots to maximize the fluorescence signal. FIG. 5B shows the extinction spectrum of AICNC measured by a UV-Vis spectrometer, and the extinction dip near 300 nm closely aligns with the emission wavelengths of the neurotransmitters, supporting the choice of AICNC for optimal plasmonic enhancement. FIG.

[0331] 5C-F present the STEM analysis of AICNCs, highlighting their composition and surface features. The Energy Dispersive Spectroscopy (EDS) in FIG. 5C confirms aluminum (Al) as the dominant component, with oxygen (O) mainly on the surface of the AICNCs. The average diameter of AICNCs were estimated to be 80 ± 9 nm by STEM tools, and the thickness of the oxide layer on the surface of AICNCs after exposure to air is estimated to be between 4 to 8 nm (FIG. 5D). The cyan-colored EDS map represents aluminum in FIG. 5E, while the red map corresponds to the oxide layer (FIG. 5F). These results provide a detailed characterization of the AICNCs and their surface properties. SEM images of AICNC nanoparticles, along with the coffee ring pattern formed by drop-casting 1 pL of molecule solution, are also presented in FIG. 6.

[0332] AF intensity is influenced by concentration, absorption efficiency (s), excitation light intensity, and experimental conditions. While a molecule with a higher quantum yield generally exhibits greater fluorescence intensity, the molecule with a higher extinction coefficient can absorb more light, leading to stronger fluorescence despite a lower quantum yield. For example, although NE has a higher quantum yield than DA, the higher extinction coefficient of DA (2110 L mo1-1cm-1compared to NE’s 1070 L mol-1cm-1) results in higher fluorescence intensity from DA. In contrast, although DOPAC exhibits an extinction coefficient comparable to DA, its much lower quantum yield (1.7%) leads to significantly reduced fluorescence intensity. The extinction coefficients of DA, NE, and DOPAC were determined from the slope of linear fit data of absorption vs. concentration plots in FIG. 7, using five data points employed in our previous work for quantum yield calculations. The calculated extinction coefficients were 2110, 1070, and 2210 L mol-1cm-1for DA, NE, and DOPAC, respectively, as shown in Table 3. The fluorescent data for DA, NE, and DOPAC in deionized water (DI) at 5 different concentrations are shown in FIG. 8 and the corresponding absorption data are shown in FIG. 9.026389-0063-W001

[0333]

[0334] We employed AICNCs that were drop cast and dried on a bare Si wafer (as reference) as plasmonic substrates to differentiate and analyze three similar neurotransmitters: DA, DOPAC, and NE. FIG. 10A-C presents AFTDS signals, decreasing over time for DA, DOPAC, and NE, and the topmost curve (highest intensity) corresponds to the first spectrum. The AF spectra, shown in FIG. 10D-E, compare the intensities of DA, DOPAC, and NE acquired at 0-0.5 seconds with an acquisition time of 0.5 seconds ( / o, the highest intensity spectrum in the AFTDS) on a silicon wafer (FIG. 10D) and AICNC substrates (FIG. 10E). DA exhibited the highest fluorescence intensity on both substrates, followed by NE and DOPAC. Notably, fluorescence intensities on the AICNC substrate were significantly higher compared to a silicon substrate, highlighting the fluorescence-enhancing properties of AICNCs. This enhancement is specifically attributed to the localized surface plasmon resonance (LSPR) effects of AICNCs in the UV range, which amplifes both the excitation and fluorescent emission (quantum yield) near the nanocube surface, unlike the non-plasmonic silicon substrate. This enhancement suggests strong plasmonic amplification or more efficient excitation / emission conditions on the AICNC substrate while maintaining consistent spectral profiles. Additionally, fluorescence spectra data acquired using a fluorometer for the neurotransmitters dissolved in water are presented in FIG. 10F.

[0335] FIG. 11 compares the average integrated fluorescence intensities and net enhancement factors for DA, NE, and DOPAC; each droplet contains 1 pL of 500 pM of MANTs that drop cast and dried on a silicon wafer and AICNC substrates. The integrated fluorescence intensity and net enhancement factors were calculated using the methods published previously. To quantify the variability in enhancement measurements, the error was calculated as the standard deviation (STD) across multiple experimental replicates. For each neurotransmitter (DA, DOPAC, and NE), fluorescence spectra were collected at multiple distinct spots (n = 5-7) on both AICNC and silicon (Si) substrates. The integrated fluorescence intensity for each measurement was obtained by subtracting the baseline signal (dark spectrum) and computing the area under the spectrum using the trapezoidal rule. Enhancement factors were calculated by normalizing the integrated intensity on AICNC to the corresponding average intensity on Si. The reported enhancement values represent the mean of these normalized measurements, and the associated errors reflect the standard deviation across the set of replicates, thereby accounting for experimental variation due026389-0063-W001

[0336] to sample heterogeneity, measurement noise, and substrate uniformity. As shown in FIG. 11A, fluorescence intensities on the AICNC substrate were significantly higher than those on the silicon wafer, with DA exhibiting the strongest signal, followed by NE and DOPAC. FIG. 11 B highlights the net enhancement factors achieved on AICNC, with DA, NE, and DOPAC showing enhancements of 12, 9, and 7 respectively, compared to silicon wafers.

[0337] These results demonstrate that AICNCs drop cast and dried on a substrate are effective in enhancing the AF signal of MANTs. Although other plasmonic nanostructures have achieved higher fluorescence enhancement factors, the presented method achieves a plasmonic substrate without complicated, lengthy and expensive nanofabrication techniques.

[0338] The AFTDS data on different concentrations for model training and evaluation are shown in FIG. 12. The dataset contained three classes — DA, DOPAC, and NE — with slightly varying support: 54,784, 51,712, and 39,936 samples, respectively, in the AICNC dataset, and 1105, 1326, and 1326 samples in the in-solution dataset and more details about samples in each class can be found in FIG. 13. Each sample corresponds to a unique fluorescence data point defined as a pair of fluorescence intensity and wavelength values, represented as a function FL(Wv, / ). As shown in FIG. 12, each fluorescence time series was acquired over about 2 to 3 minutes measurement timeframe with a 0.5-second acquisition interval, yielding approximately 240-360 spectra per measurement. Each of these spectra spans a 60 nm range (280-360 nm), representing the AFTDS for a specific molecule at a defined concentration. The weighted and macro averages in performance metrics were used to ensure a balanced evaluation across all classes.

[0339] Classification Results

[0340] FIG. 14 presents the confusion matrices for the LSTM, KNN, and RF models used to classify three neurotransmitters: DA, DOPAC, and NE; based on AF data obtained from AICNC and in-solution environments. The matrices display class-specific prediction distributions, with diagonal elements representing correct or true positives and off-diagonal elements indicating misclassifications. The LSTM model on AFTDS achieved high classification accuracy, with 88.0% for DA, 90.4% for DOPAC, and 89.2% for NE, indicating strong temporal feature learning. The KNN model on AFTDS also demonstrated high performance, achieving 89.3% for DA, 86.1% for DOPAC, and 83.0% for NE. Similarly, the RF on AFTDS achieved 87.4% for DA, 85.6% for DOPAC, and 76.7% for NE, supporting its strong ensemble classification ability on temporal AFTDS data. Although RF outperformed KNN on the AF in-solution dataset (RF achieving 60.9% for DA, 82.1% for DOPAC, and 68.6% for NE versus KNN’s 44.8% for DA, 89.0% for DOPAC,026389-0063-W001

[0341] and 66.6% for NE), both models performed substantially worse compared to those applied to AFTDS data. This drop in accuracy highlights increased confusion among class boundaries in the solution-based measurements. The absolute number of samples per matrix cell was provided inside the brackets in FIG. 13 and in Table 4, revealing the distribution and support for each class. These matrices provide a clear visual assessment of model effectiveness and misclassification tendencies.

[0342] Table 4. Classification Models Performance Metrics Across Classes (DA, DOPAC, NE) .. . . Weighted Macro .

[0343] Mode C ass DA NE DOPACA. Accuracy Avg. Avg. * LSTM using AFTDS of MANTs on AICNC

[0344] Precision 0.9 0.88 0.89 0.89 0.89 - Recall 0.88 0.89 0.9 0.89 0.89

[0345] F1 Score 0.89 0.88 0.9 0.89 0.89 0.89 Support 54784 39936 51712 146432 146432

[0346] KNN using AFTDS of MANTs on AICNC

[0347] Precision 0.83 0.88 0.89 0.87 0.87

[0348] Recall 0.89 0.83 0.86 0.87 0.86

[0349] F1 Score 0.86 0.86 0.87 0.87 0.87 0.86 Support 54784 39936 51712 146432 146432

[0350] RF using AFTDS of MANTs on AICNC

[0351] Precision 0.81 0.89 0.84 0.84 0.85

[0352] Recall 0.87 0.77 0.86 0.84 0.83

[0353] F1 Score 0.84 0.82 0.85 0.84 0.84 0.84 Support 54784 39936 51712 146432 146432

[0354] KNN using AF of MANTs in solution

[0355] Precision 0.7 0.63 0.71 0.68 0.68 - Recall 0.45 0.67 0.83 0.67 0.65

[0356] F1 Score 0.55 0.65 0.79 0.66 0.67 0.68 Support 1105 1326 1326 3757 3757

[0357] RF using AF of MANTs in solution

[0358] Precision 0.67 0.67 0.78 0.71 0.71

[0359] Recall 0.45 0.63 0.82 0.66 0.63

[0360] F1 Score 0.55 0.65 0.79 0.66 0.67 0.71 Support 1105 1326 1326 3757 3757

[0361] The classification performance metrics, precision, recall, and F1 score, for LSTM, KNN, and RF models across three molecular classes: DA, DOPAC, and NE, are presented in Table 5. These metrics, derived from confusion matrices, provide insight into each model’s classification, accuracy, and efficiency. We included macro and weighted averages to summarize model performance across all classes. Macro average treats all classes equally, useful for balanced026389-0063-W001

[0362] datasets, while the weighted average reflects class distribution, favoring the majority classes. These metrics offer complementary views but can mislead if used alone in imbalanced settings. For better clarity, bar charts are provided in FIG. 15.

[0363] Table 5. Classification Models Performance Metrics Across Classes (DA, DOPAC, NE). ,, . . .ir- Weighted Macro .

[0364] Mode C ass DA NE DOPAC . . Accuracy Avg. Avg.7LSTM using AFTDS of MANTs on AICNC

[0365] Precision 0.90 0.88 0.89 0.89 0.89 - Recall 0.88 0.89 0.90 0.89 0.89

[0366] F1 score 0.89 0.88 0.90 0.89 0.89 0.89 KNN using AFTDS of MANTs on AICNC

[0367] Precision 0.83 0.88 0.89 0.87 0.87 - Recall 0.89 0.83 0.86 0.87 0.86 — F1 score 0.86 0.86 0.87 0.87 0.87 0.86 RF using AFTDS of MANTs on AICNC

[0368] Precision 0.81 0.89 0.84 0.84 0.85 - Recall 0.87 0.77 0.86 0.84 0.83

[0369] F1 score 0.84 0.82 0.85 0.84 0.84 0.84 KNN using AF of MANTs in the solution

[0370] Precision 0.70 0.63 0.71 0.68 0.68 - Recall 0.45 0.67 0.89 0.67 0.68

[0371] F1 score 0.55 0.65 0.79 0.66 0.67 0.68 RF using AF of MANTs in the solution

[0372] Precision 0.67 0.67 0.78 0.71 0.71

[0373] Recall 0.61 0.69 0.82 0.71 0.71

[0374] F1 score 0.64 0.68 0.80 0.71 0.71 0.71 The best-performing values for each metric and class within each dataset are bolded for emphasis. Values for recall are also shown in the confusion matrices

[0375] The results demonstrate that the LSTM model consistently outperforms the KNN model when applied to AFTDS data collected from MANTs on AICNCs. The LSTM model achieves a weighted and macro average F1 score of 0.89, indicating strong generalization and superior classification performance across all classes. In contrast, the KNN model yields a slightly lower weighted and macro average F1 score of 0.86, reflecting its comparatively reduced ability to capture dynamic spectral features.

[0376] For AF spectra data collected from MANTs in solution (without plasmonic nanostructures), the KNN model exhibits reduced classification performance compared to its performance on AFTDS data. It achieves a weighted and macro average F1 score of 0.68. Notably, KNN performs best on DOPAC but struggles with NE and exhibits the weakest performance on DA.026389-0063-W001

[0377] This discrepancy may be attributed to differences in molecular intensity within solution-based measurements. Conversely, classification accuracy using AFTDS data remains relatively consistent across different molecules, underscoring its robustness in molecular classification.

[0378] In addition to KNN and LSTM, an RF model was evaluated under the same conditions. On AFTDS data, the RF model achieved a weighted and macro average F1 score of 0.84, slightly lower than the KNN and LSTM models but still demonstrating strong performance. When applied to AF data collected in solution, the RF model outperformed KNN, with a weighted and macro average F1 score of 0.71. This indicates that while RF does not match LSTM performance on AFTDS data, it provides more reliable classification than KNN in solution-based environments, likely due to its ensemble-based structure and resilience to data noise.

[0379] The sample size for data collected on AICNC is much larger than that of solution-based spectra. Since ML benefits from large data sets, our methods of collecting time variant spectra under continuous light illumination are advantageous compared with solution-based spectra collection. For instance, using AICNC and acquiring AFTDS data with an acquisition time of 0.5 seconds, we obtained 300 spectra from a single molecular concentration in 2 minutes. In contrast, obtaining the same number of spectra with solution-based UV-Vis would require 300 different concentrations and separate measurements, which would be an impractical, time-consuming, and costly approach.

[0380] Overall, these findings highlight the LSTM model’s advantage in leveraging the timedependent characteristics of AFTDS signals, allowing for more accurate molecular differentiation. In contrast, the KNN model, which relies on static data, demonstrates lower classification performance. The confusion matrix results further reinforce the LSTM model’s effectiveness in distinguishing structurally similar MANT s based on AFTDS data while also showing the RF model as a better alternative than KNN when temporal data is limited or unavailable.

[0381] This work demonstrates a transformative approach to probe-free and label-free biosensing by merging autofluorescence using novel plasmonic nanoparticles with ML. Aluminum concave nanocubes, a UV plasmonic substrate, significantly enhanced the native fluorescence signals of neurotransmitters, achieving up to a 12-fold improvement compared to silicon substrates. This enhancement amplified the autofluorescence intensity signals, and AFTDS was the key factor that enabled the differentiation of similar neurotransmitter data in a label-free manner. When analyzed using ML techniques, the AFTDS on AICNCs provided high classification accuracy, highlighting the critical role of plasmonic-engineered fluorescence dynamics in molecular identification.026389-0063-W001

[0382] Integrating ML models, particularly LSTM networks, proved critical in analyzing and classifying complex fluorescence data. With a classification accuracy of 89%, the ML models effectively captured subtle variations in spectral data, enabling the reliable identification of neurotransmitters. The study also highlighted the comparative advantage of LSTM over KNN and RF in handling time-dependent fluorescence data (AFTDS). In addition, RF models demonstrated competitive performance, particularly in environments where temporal patterns are less pronounced. While not as effective as LSTM on AFTDS data, the RF model surpassed KNN in classifying solution-based spectra, emphasizing its robustness and versatility as a static-data classifier and suggesting RF as a better approach in scenarios where dynamic signal acquisition may be limited.

[0383] These findings underscore the potential of combining nanotechnology and ML to create smart biosensing systems that are both sensitive and selective. To the best of our knowledge, the proposed methodology may pave the way for developing non-invasive, real-time diagnostic tools for complex biological environments, offering new avenues for applications in biomedical diagnostics, neuroscience, and beyond.

Claims

026389-0063-W001CLAIMSWhat is claimed:

1. A system for detecting monoamine neurotransmitters, comprising:a plasmonic substrate comprising aluminum concave nanocubes;an ultraviolet light source configured to illuminate the plasmonic substrate;a spectrometer configured to collect fluorescence data from the surface of the plasmonic substrate, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired over a time period; anda processor configured to execute a machine learning model trained to classify the monoamine neurotransmitters based on the fluorescence data.

2. The system of claim 1, wherein the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers.

3. The system of claim 1 , wherein the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 200-300 nanometers.

4. The system of claim 1 , wherein the ultraviolet light source comprises a continuous wave solid state laser configured to emit light at a wavelength of approximately 266 nanometers.

5. The system of claim 1 , wherein the wherein the machine learning model comprises a Long Short-Term Memory network.

6. The system of claim 4, wherein the Long Short-Term Memory network comprises:three sequential Long Short-Term Memory layers;batch normalization layers following each of the three sequential Long Short-Term Memory layers;a fully connected dense layer; anda softmax output layer configured to classify the monoamine neurotransmitters into a plurality of classes.026389-0063-W0017. The system of claim 7, wherein the plurality of classes comprises dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

8. The system of claim 1, wherein the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers.

9. A method for classifying monoamine neurotransmitters, comprising:depositing a sample containing monoamine neurotransmitters onto a plasmonic substrate comprising aluminum concave nanocubes;illuminating the plasmonic substrate with ultraviolet light;collecting fluorescence data comprising a plurality of fluorescence spectra over a time period during continuous illumination;preprocessing the fluorescence data to generate standardized input data; and classifying the monoamine neurotransmitters by applying a trained machine learning model to the standardized input data.

10. The method of claim 9, wherein the sample comprises a solution and depositing the sample comprises drop casting the solution onto the plasmonic substrate to form a coffee ring pattern.

11. The method of claim 9, wherein illuminating the plasmonic substrate comprises directing the ultraviolet light toward an outer edge of the coffee ring pattern where a concentration of the monoamine neurotransmitters is higher.

12. The method of claim 9, wherein collecting the fluorescence data comprises acquiring the plurality of fluorescence spectra at intervals of approximately 0.5 seconds over a time period of 2 to 3 minutes.

13. The method of claim 12, wherein each fluorescence spectrum of the plurality of fluorescence spectra spans a wavelength range of 280 to 360 nanometers.

14. The method of claim 9, wherein preprocessing the autofluorescence time decay series data comprises:shifting time series data to eliminate leading missing values;026389-0063-W001truncating or padding the time series data to a standardized sequence length; and normalizing feature values using MinMax scaling.

15. The method of claim 14, wherein the standardized sequence length comprises 10-25 time steps.

16. The method of claim 9, wherein the wherein the trained machine learning model comprises a Long Short-Term Memory network.

17. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to:receive fluorescence data collected from a sample comprising one or more monoamine neurotransmitters deposited on a plasmonic substrate comprising aluminum concave nanocubes, wherein the fluorescence data comprises a plurality of fluorescence spectra acquired during continuous ultraviolet illumination; preprocess the fluorescence data by aligning time series traces and standardizing sequence lengths to generate input tensors;apply a trained neural network to the input tensors to capture temporal dependencies in the autofluorescence time decay series data; andoutput a classification of the monoamine neurotransmitters based on analysis by the trained neural network.

18. The non-transitory computer-readable medium of claim 17, wherein the trained neural network is a trained Long Short Term Memory network.

19. The non-transitory computer-readable medium of claim 17, wherein the instructions further cause the processor to normalize feature values of the autofluorescence time decay series data to a range of 0 to 1 using MinMax scaling prior to applying the trained Long Short- Term Memory network.

20. The non-transitory computer-readable medium of claim 17, wherein preprocessing the autofluorescence time decay series data comprises shifting time series traces to eliminate leading missing values by aligning valid entries to a common starting point.026389-0063-W00121. The non-transitory computer-readable medium of claim 17, wherein the standardized sequence lengths comprise 17 time steps, and wherein the instructions further cause the processor to truncate longer sequences to 17 regularly spaced sequences or pad shorter sequences to achieve the standardized sequence lengths.

22. The non-transitory computer-readable medium of claim 18, wherein the trained Long Short-Term Memory network comprises:three sequential Long Short-Term Memory layers having 200, 150, and 100 units respectively;batch normalization layers following each of the three sequential Long Short-Term Memory layers;a fully connected dense layer with 50 neurons; anda softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

23. The non-transitory computer-readable medium of claim 22, wherein the instructions further cause the processor to process a single input tensor in under 1 millisecond to enable realtime classification of the monoamine neurotransmitters.

24. A biosensing device for label-free detection and classification of monoamine neurotransmitters, comprising:a substrate having a surface with aluminum concave nanocubes deposited thereon; an excitation source configured to direct ultraviolet light onto the substrate to excite native autofluorescence from monoamine neurotransmitters in contact with the aluminum concave nanocubes;a detector configured to acquire time-resolved fluorescence spectra from the monoamine neurotransmitters over a measurement period to generate autofluorescence time decay series data; anda classification module comprising a processor and a memory storing a trained recurrent neural network, wherein the classification module is configured to receive the autofluorescence time decay series data and output an identification of the monoamine neurotransmitters based on temporal decay characteristics captured by the trained recurrent neural network.026389-0063-W00125. The biosensing device of claim 24, wherein the aluminum concave nanocubes have a nominal diameter of approximately 80 nanometers with a standard deviation of approximately 9 nanometers.

26. The biosensing device of claim 24, wherein the trained recurrent neural network comprises a Long Short-Term Memory network having three sequential Long Short-Term Memory layers with batch normalization layers following each of the three sequential Long Short- Term Memory layers.

27. The biosensing device of claim 26, wherein the Long Short-Term Memory network further comprises a fully connected dense layer and a softmax output layer configured to output classification probabilities for dopamine, norepinephrine, and 3,4-dihydroxyphenylacetic acid.

28. The biosensing device of claim 24, wherein the excitation source comprises a continuous wave solid state ultraviolet laser configured to emit light at a wavelength of approximately 266 nanometers, and wherein the detector is configured to acquire the time-resolved fluorescence spectra at intervals of approximately 0.5 seconds over a measurement period of 2 to 3 minutes.

Images