Photonic biosensors based on nanoparticle superstructures: from data analysis to artificial intelligence (AI) detection

Abstract

The superstructures formed by the self-assembly of nanoparticles (NPs) can exhibit unique photonic collective properties (structural color, localized surface plasmon resonance [LSPR]), enhance the interaction between light and matter, and open up new possibilities for photonic sensing. Many photonic biosensors have addressed the limitations of current bioanalytical methods with their non-invasive nature, real-time monitoring, and high sensitivity. In recent years, the construction of photonic biosensors using super-structured materials could further enhance the sensors in terms of sensitivity, processing capacity, ease of use, and miniaturization. Superstructure-based photonic biosensors can analyze complex samples, but their development still needs to overcome limitations related to target binding specificity, long-term stability, and signal decoding efficiency. The development of artificial intelligence (AI) provides new opportunities to solve these problems. Deep learning (DL) algorithms can independently extract multi-dimensional data features such as spectra and images, distinguish weak biological signals from noise, optimize detection parameters, and achieve real-time dynamic calibration. In this review, we provide the photonic collective characteristics of superstructures and the applications of biosensors in intelligent diagnosis. The applications of superstructured photonic sensors in disease diagnosis, drug delivery, and cell imaging are summarized. The colorimetric, fluorescence-based sensor technologies assisted by DL are discussed along with challenges faced in integrating AI with superstructure-based photonic biosensors. As this field continues to evolve, the integration of AI and superstructure-based photonic biosensors will undoubtedly play a pivotal role in shaping the future of medical diagnostics and therapeutic interventions.

1. Introduction

Monitoring human health is critical for ensuring social stability, promoting overall well-being, and achieving long-term sustainability. Early diagnosis is particularly essential in this context, as it significantly enhances the effectiveness of medical treatments and interventions. Therefore, the development of biosensors has become increasingly important for disease diagnosis and evaluation of treatment plans. Current biosensors, such as electrochemical sensors, have the advantages of low cost and easy integration, but are susceptible to electrolyte interference [[1], [2], [3]]. Piezoelectric biosensors can detect mass changes based on the variation of quartz crystal vibration frequency, but they have relatively low sensitivity [4]. Thermal biosensors measure temperature changes through enzymatic reactions or microbial metabolic heat production, but are easily influenced by the ambient temperature [5]. Photonic biosensors have become advanced devices in health monitoring technology. They use the interaction between biomolecular receptors and specific analytes to convert photonic information into measurable output signals, especially their non-invasive analysis capabilities, real-time monitoring functions and enhanced detection sensitivity [[6], [7], [8]]. Photonic biosensors can be classified into colorimetric biosensors [[9], [10], [11], [12]], fluorescent biosensors [13], Raman spectroscopy sensors [14,15], surface plasmon resonance (SPR) sensors [16,17], etc., according to their working principles. These sensors can operate reliably in a wide range of complex biological samples, including blood, saliva, and urine, without the need for extensive sample preparation. However, there are still some challenges in enhancing the binding specificity, sensitivity, long-term stability of target biomarkers and the detection efficiency of the biosensing system, and new nanophotonic biosensors need to be developed.

Numerous nanophotonic biosensors have been developed to overcome the shortcomings of existing bioanalytical techniques in sensitivity, processing capacity, portability, and design for miniaturization [8]. NPs provide substantial advantages due to their small size, large surface area, and facile functionalization. The self-assembly process leads NPs to form stable, low-energy configurations, giving rise to superstructures with distinctive collective photonic properties [18,19]. By manipulating the self-assembly architecture, it becomes feasible to obstruct electromagnetic wave transmission or confine light to subwavelength dimensions, thereby enhancing light-matter interactions. Moreover, the photonic signal of the superstructure depends not only on single-component NPs but also on the symmetry, orientation, and dimension of the entire structure, demonstrating collective effects. These collective effects further promote the enhancement of photonic signals, such as photonic crystal-enhanced fluorescence [20,21], photonic crystal-enhanced Raman scattering [22,23], and superfluorescence (SF) [24], enabling highly sensitive environmental monitoring and single-molecule detection at ultralow concentrations, paving the way for advanced biosensing applications. Additionally, nanophotonic biosensors generate vast amounts of multi-dimensional data (e.g., spectra and images) during operation. To leverage these capabilities, advanced machine learning (ML) algorithms are urgently needed to streamline data analysis and optimize detection efficiency.

The advent of AI has introduced a transformative opportunity in the field of photonic biosensing, addressing challenges related to sensor design and data analysis [25,26]. ML techniques are capable of autonomously extracting signal characteristics, detecting faint biomarker signals, and managing the high-dimensional data challenges posed by intricate biological signals [27]. Through model training, these techniques can effectively differentiate target biological signals from background noise, thus enhancing the precision, dependability, and consistency of detection processes. AI algorithms possess the ability to evaluate data streams instantaneously, adjust detection parameters dynamically, and facilitate real-time monitoring alongside adaptive modifications [28,29]. Moreover, AI employs sophisticated data management and analytical methodologies. By synthesizing multi-faceted information such as biological imagery and spectral data, a more holistic biological information framework can be established. An active learning algorithm can also automatically calibrate sensor sensitivity according to environmental changes (such as temperature and pH value) to ensure long-term stability. Integrating nanophotonic devices with AI chips enables low-power consumption computing at the micro-nano scale while building portable detectors and conducting multi-scenario application expansions.

In this review, we provide the applications of superstructure-based photonic biosensors in intelligent diagnosis. Firstly, it introduces collective photonic phenomena such as structural color, chiral optics, LSPR and photoluminescence (PL) generated by superstructures. Then, it summarizes the achievements of devices based on superstructures in disease diagnosis, drug delivery, cellular imaging and biochips. Various photonic biosensors assisted by ML, including colorimetric sensors, fluorescence sensors and photonic biosensors based on Raman spectroscopy, are discussed. Finally, we provide insights into the future developments at the intersection of DL and superstructure-based photonic biosensors. As shown in Fig. 1, the synergy between AI and superstructure-based photonic biosensors can enable extensive data analysis and portable point-of-care diagnostics, thus shaping the future of medical diagnosis and treatment.

Fig. 1.

The process of disease diagnosis based on nanoparticle superstructure-based photonic biosensors combined with AI.

2. Photon collective phenomena of superstructures

The self-assembled upper structure exhibits distinctive collective photonic properties that markedly differ from those of the original monodisperse particles and bulk materials. Due to its adjustable collective photonic properties, research in this area is highly prevalent in nanoscience, encompassing studies on structural color, LSPR, chiral optics, and PL.

2.1. Structure color

In nature, light scatters and diffracts when it hits butterfly wings, beetle elytra, peacock feathers, pearls, and natural opal, etc., to produce structural color. This is due to the periodic structure of natural materials. Colloidal NPs with sizes comparable to the wavelength of visible light can also be artificially assembled into colloidal crystals with periodic arrangement through self-assembly, becoming photonic crystals (PCs) [[30], [31], [32]]. According to their structural periodicity, PCs can be classified into one-dimensional (1D), two-dimensional (2D) and three-dimensional (3D) PCs. One special property of PCs is that they have a photonic bandgap, in which electromagnetic waves with certain energy are prohibited from passing through the PC, which can produce better photonic effects. When light is incident on a periodic structure, it is reflected from each interface. Under appropriate conditions, these reflected waves undergo phase interference according to Bragg's law as shown in Eq. (1) [33]:

$$m\lambda =2d\cos \theta$$ (1)

where d is the interplanar distance, θ is the angle of incident light, m is the order of diffraction, and λ is the wavelength of the reflected light. Combining Bragg's law with Snell's law of refraction leads to Eq. (2):

$$m\lambda =2d{\left({n_{eff}}^2-{\sin }^2\theta \right)}^{1/2}$$ (2)

where n_eff is the mean effective refractive index. When the incident light is perpendicular to the sample, Eq. (2) can be simplified as:

$$m\lambda =2d{n}_{eff}$$ (3)

Furthermore, controlling the position of the photonic bandgap to align with the fluorescence emission peak or Raman spectrum can enhance the fluorescence intensity or Raman signal, which holds potential applications in areas such as optics, sensing, display, biomedicine, environment, and safety.

2.1.1. Self-assembled photonic crystals: materials

In general, PCs are typically prepared through ``top-down'' lithography techniques and ``bottom-up'' self-assembly strategies [30,34,35]. Lithography processing technology provides precise control over periodic structures, but the process is complicated and expensive, with limited types of available photoresist. Building PCs through self-assembly is a simple and cost-effective method. With the development of colloidal particles, the types of building blocks for constructing PCs are becoming increasingly diverse, including emerging metal-organic frameworks (MOFs), covalent organic frameworks (COFs), cellulose nanocrystals (CNCs), and other materials [36]. This undoubtedly increases the potential applications of PCs in sensing, photonic waveguides, coding anti-counterfeiting, displays, biomedicine, etc.

The advancement in the synthesis of monodisperse colloidal particles has led to the emergence of numerous novel polyhedral building units for the fabrication of well-ordered superstructures, as well as endeavors to explore uncharted collective properties and applications [36,37]. Maspoch et al. demonstrated the ability of zeolitic imidazolate framework-8 (ZIF-8) particles to spontaneously arrange themselves into large-scale structures with a 3D rhombohedral lattice, exhibiting properties of PCs [38] (Fig. 2a). The photonic bandgaps can be modified by adjusting the size of the ZIF-8 particles and changing the adsorption of guest molecules within their micropores (Fig. 2b). Wang et al. presented a technique to induce the self-assembly of ZIF-8 particles by evaporating a mixture of methanol and water [39]. The difference in volatility between these two components results in Marangoni flow towards the center, which prevents the formation of coffee rings (Fig. 2c). The addition of methanol not only speeds up the assembly process but also improves the orientation and gloss of the superstructure. Vogel et al. successfully constructed supraparticles using MOF particles, which exhibited uniform coloring due to the presence of ordered layers resembling onions on their outermost regions [40] (Fig. 2d). The supraparticles' structural color arises from the interaction between light and their constituent building blocks, resulting in interference effects [41]. Consequently, the coloration observed in these MOF supraparticles varied depending on the viewing angle (Fig. 2e). All dispersions containing these supraparticles displayed visible macroscopic coloration, indicating that local ordering alone was adequate for generating interference effects. Furthermore, apart from MOFs, the Maspoch group also utilized COFs as novel porous materials for constructing colloidal building particles capable of producing face-centered cubic (fcc) arrangements within porous PCs [42]. Anticipate that such innovative porous materials will facilitate the development of PCs with unparalleled functionality. Additionally, important organic compounds that compose the cell walls of plants, such as cellulose, hemicellulose, and lignin, can also self-assemble into macroscopic materials that exhibit PC properties. The helicoidal arrangement of CNC is responsible for the structural color observed in cellulose-based optics [43,44]. Tang et al. introduced chiral photonic crystals (CPCs) synthesized from CNCs and photosensitive inorganic polyoxometalates (POMs) [Eu(SiW₁₀MoO₃₉)₂]^13‒ using an evaporation-induced self-assembly method [45] (Fig. 2f). By varying the concentration of doped POMs in the films, tunable photonic band gaps (PBGs) are achieved. The presence of vibrant iridescence suggests that the helical pitch is comparable to the wavelength of visible light (Fig. 2g, h).

Fig. 2.

(a) Field emission scanning electron microscope (FE-SEM) image of a truncated rhombic dodecahedral (TRD) ZIF-8 superstructure. (b) Photographs of the PCs composed of TRD ZIF-8 particles of different sizes. Copyright 2017 Springer Nature [38]. (c) Optical microscopy images of the evaporative deposition margins of ZIF-8 particle droplets with pure water as the solvent (upper part) and 80% methanol concentration as the solvent (bottom part) on different substrates. Copyright 2023 Royal Society of Chemistry [39]. (d) FE-SEM image and optical photograph of a TRD ZIF-8 supraparticle. (e) Angle-dependent photographs of MOF supraparticles. Copyright 2022 Wiley-VCH GmbH [40]. (f) TEM image of CNC nanorods. (g) Photograph of CPC film demonstrating its transparency and iridescence. (h) Image of CPC film captured under natural light. Copyright 2023 Wiley-VCH GmbH [45].

2.1.2. Photonic crystal enhanced fluorescence

The fluorescence emission in the structures was either intensified or diminished, depending on the location of the photonic bandgap. This occurs because resonant structures can alter the density of optical states (DOS), which is linked to the electromagnetic modes of propagation, the rate of fluorescence decay, and the lifetime of the excited state [46,47]. On the other hand, adjusting the photonic bandgap of the PC can yield maximum enhancement in luminescence. Xu et al. presented the formation of 3D superstructures through tilted angle sedimentation in the self-assembly of colloidal zeolite LTA superballs (ZAS), resulting in ZAS superstructures that exhibit a PBG in the visible light range [48]. The ZAS superstructures exhibit a reversible response to a variety of chemical vapors, such as water, acetone, n-hexane, n-butanol, and cyclohexane. As temperature increases, the number of particles involved in forming R-helices also increases, leading to chiral photonic superstructures with net right-handedness. A film containing Ag nanocluster encapsulated ZAS exhibits a negative circular dichroism (CD) signal that overlaps with corresponding reflection bands and displays white luminescence under 360 nm irradiation as well as an unsymmetrical PL band across the visible regime. Additionally, it shows strong PBG-based LCP luminescence. Song et al. developed a microchip for enhancing fluorescence using PC technology [49]. The microchip was fabricated through inkjet printing and featured a micropattern. By exploiting the difference in wettability between the hydrophilic PC dot and the hydrophobic polydimethylsiloxane (PDMS) substrate, analytes were observed to concentrate in the hydrophilic PC region during drop evaporation. To detect cocaine, a DNA-aptamer-modified PC microchip was utilized, achieving a detection limit of 1.4 × 10¹⁷ mol L^-1. Additionally, they have developed a PC chip that can detect a variety of metal ions [50]. The fluorescence of metal quinolines, which is influenced by the presence of metal ions, was used as the detection signal. By incorporating five PCs with different photonic bandgaps into the microchip design, the entire range of metal quinoline fluorescence spectra could be covered. Through linear discriminant analysis (LDA), the PC microchip achieved accurate classification of 12 different metal ions and one control sample. This efficient multi-stopband PC microchip enables effective pattern recognition for multiple analytes through rational design strategies. Mihi et al. utilized perovskite nanocrystal (NC) to create photonic structures that effectively couple light to the NC layer, resulting in an increased electric field intensity within the perovskite film [51]. By employing PDMS templates with patterned holes to cover the droplets deposited on the glass substrate, colloidal droplets were dispersed into these template holes and self-assembled through solvent evaporation, forming a PC composed of CsPbBr₃ NPs. Due to the enhanced multi-photon absorption caused by the light confinement provided by the PC structure, the 2D PCs made from CsPbBr₃ exhibit amplified spontaneous emission (ASE) at lower photonic excitation fluences in the near-IR range.

2.1.3. Photonic crystal enhanced Raman

The PCs formed by colloidal particle self-assembly can serve as suitable substrates for metal NPs. When the Raman shift wavelength of the probe molecule overlaps with the PBG, Raman scattering may be enhanced, further improving the Raman signal. Klimonsky et al. fabricated inverse opal structured PC thin films through the polymerization of a photocurable resin [52]. Raman spectra were collected at varying light incidence angles to analyze the position of the photonic band gap in different samples. By adding a small amount of Au NPs to the anti-quartz structure PC, the enhancement factor of the Raman signal can be further increased by 1–2 × 10⁵ [53]. Additionally, it is also feasible to utilize self-assembled monolayer PS spheres for constructing 2D PCs by depositing alternating layers of amorphous silicon and silver (Ag). This process allows the fabrication of a metal-dielectric PC [54]. The enhancement of the Raman signal through PCs still originates from the “hot spots” located between sharp points and gaps of metal particles. However, only a small portion of the analytes are trapped in these areas, necessitating new approaches to enrich the target analytes. Zhao et al. have developed a hydrophobic microchip with PCs (HPCM) that serves dual purposes: it enables hydrophobic enrichment and enhances surface-enhanced Raman spectroscopy (SERS) activity by applying a thin Ag film on the surface of the PCs [55]. By placing liquid droplets containing the analyte onto HPCM, coffee ring formation can be prevented due to its hydrophobic surfaces. This allows for effective enrichment and improves both uniformity and reproducibility of Raman signals. The SERS signal of organic dyes on HPCM experiences a maximum enhancement of approximately fivefold. This improvement is attributed to achieving an overlap between PBG and the excitation wavelength used in Raman spectroscopy, thereby maximizing SPR effects and enhancing light utilization efficiency.

2.2. Chiral optics

Chiral structures refer to objects that possess a mirror image that cannot be superimposed on the original [56]. In recent years, the growing attention and extensive investigation into self-assembled chiral superstructures have primarily been motivated by their unique optical characteristics, such as CD and circularly polarized luminescence (CPL) [57,58]. These traits arise from the assembly of individual building blocks, which are highly sensitive to the structural arrangement of the materials and are typically not observed in other types of self-assembled superstructures. CD refers to the difference in absorption of LCP and RCP circularly polarized light by chiral materials, quantified as the absorption asymmetry factor (g_abs). Consequently, optical CD is defined by Eq. (4) [59]:

$$\text{CD}=\Delta A=A_L-A_R$$ (4)

where A_L and A_R denote the absorbance of LCP and RCP light, respectively. In this context, CD takes on dimensionless values ranging from −1 to +1. The g_abs is usually defined by Eq. (5):

$${\mathrm{g}}_{abs}=\frac{2\left(A_L-A_R\right)}{A_L+A_R}$$ (5)

CPL refers to the intensity difference between LCP and RCP light emitted by the material after excitation, quantified as the luminescence asymmetry factor (g_lum). The g_lum is expressed as Eq. (6):

$${\mathrm{g}}_{lum}=\frac{I_L-I_R}{(I_L+I_R)/2}$$ (6)

where I_L and I_R represent the intensities of left-handed and right-handed circularly polarized emissions, respectively. Unlike individual chiral metal clusters in solution, self-assembled structures exhibit intense PL and a significantly higher g_lum. Within these assemblies, ligands are immobilized within crystalline lattices, restricting intramolecular rotation and causing excitation energy to dissipate primarily via radiative pathways. Consequently, PL QYs are enhanced due to aggregation-induced emission effects. Typically, two distinct design principles are employed to create these chiral superstructures: structural chirality and media chirality. The former involves arranging achiral building blocks in a helical manner while relying on inherent chirality for media chirality [60,61].

2.2.1. Structural chirality

Inorganic nanomaterials can be developed into complex 3D chiral components through a bottom-up self-assembly process. The bottom-up strategy based on structural units that utilizes external fields or chiral templates to construct inorganic chiral upper structures has attracted widespread attention. Tang et al. described the production of chiral inorganic films using a general technique that involves Langmuir-Schaefer assembly of colloidal inorganic nanowires [[62], [63], [64]] (Fig. 3a). They first transferred an aligned nanowire film onto a quartz plate as the initial layer using the Langmuir-Schaefer method. Then, by horizontally rotating the quartz plate at a specified angle either clockwise or anticlockwise before adding subsequent layers of nanowire film, they were able to generate left- or right-handed chiral inorganic films accordingly. By precisely controlling the rotation angle, it is possible to finely adjust the optical activity and obtain various structures for these chiral inorganic films. The ultrathin films fabricated by Au nanowires exhibit significant optical activity, with an anisotropy factor as high as 0.285 [62] (Fig. 3b, c). Using NiMoO₄·xH₂O nanowires as building blocks, a biomimetic chiral photonic crystal structure is self-assembled [63] (Fig. 3d). This structure can be observed using either right- or left-circular polarizers (Fig. 3e, f). They also introduced a method to achieve intense and tunable CPL from inorganic CPCs doped with semiconductor quantum dots (QDs) [64] (Fig. 3g). The chiral membrane, comprised of NiMoO₄·xH₂O nanowires and CdSSe@ZnS QDs, exhibits a significant absolute dissymmetry factor (|g_lum|) reaching up to 0.25 (Fig. 3h). 3D chiral superstructures in quantum sensing have become an important research direction in the intersection of photonics and condensed matter physics in recent years. Jung et al. presented a method to build multi-dimensional photonic nanostructures using colloidal QDs [65]. The structures were fabricated via high-resolution transfer printing (TP) with multi-layer quantum dot patterns. This approach allows flexible stacking angles and precise control over structural parameters, making the nanostructures adaptable for various photonic applications. Compared to QD films, the 2D QD nanonets showed an 8-fold increase in PL. Furthermore, asymmetric stacking of 1D QD layers enabled effective polarization control, resulting in strong chiral optical activity with a CD intensity of up to 20.5° This offers a promising route for designing polarization-sensitive 3D chiral structures.

Fig. 3.

(a) Fabrication procedure of inorganic nanowire chiral ultrathin films. (b) SEM images of one layer of Au nanowire assembly. (c) Magnified view of (b). Copyright 2017 Wiley-VCH GmbH [62]. (d) AFM image of aligned nanowires. (e) Chiral PCs viewed through a right- or left-circular polarizer. (f) CD and absorption spectra of seven samples. Copyright 2019 Wiley-VCH GmbH [63]. (g) Structural characterization of inorganic CPCs doped with QDs. (h) CD spectra and corresponding CPL spectra of CPCs. The |g_lum| value as a function of pitch number. Copyright 2022 Wiley-VCH GmbH [64].

2.2.2. Media chirality

The primary characteristic of chiral media is their ability to induce a modification in the orientation of the polarization plane when polarized light traverses through them, commonly referred to as optical activity. The pivotal attribute of chiral molecular and nanoscale structures lies in their capability to cause rotation in the polarization direction. Tang et al. utilized a template-based approach to create 3D chiral polymer inverse opal PCs (3D CPIOPCs) [66] (Fig. 4b). They employed a chiral polymer precursor solution (poly(N-propargyl acrylamide)-random-poly(N-propargyl-(R/S)-camphanamide) (PM1-r-PM2)) to fabricate these 3D CPIOPCs (Fig. 4a). By introducing chiral polymer, the refractive indexes of LCP and RCP lights differ within the chiral media, resulting in distinct PBG structures for CPIOPCs (Fig. 4c). They also utilized atomically precise Au₃[(R)-Tol-BINAP]₃Cl and Au₃[(S)-Tol-BINAP]₃Cl clusters to form well-organized assemblies consisting of nanocubes with a uniform body-centered cubic (BCC) packing arrangement [67] (Fig. 4d, e). The initially non-luminescent Au clusters gradually exhibited high luminescence during the self-assembly process. Once the ordered structure was established, chiral Au displayed significantly enhanced CD intensity and an impressive CPL response (Fig. 4f). The nanocubes' BCC packing structure ensured that each inward p-tolyl ring on the surfaces of the Au clusters was securely held in place through intermolecular CH/π interactions, resulting in maximum PL intensity and a g factor value of 7 × 10^–3. Gao et al. formed chiral superstructures by incorporating small chiral molecules into preexisting achiral superstructures [68] (Fig. 4g). Upon introducing R, R-diaminocyclohexane (R, R-DACH), a specific type of small chiral molecule, the CdSe nanoplatelet (NPL) superlattice underwent a remarkable transformation from its initial achiral state to exhibit chirality (Fig. 4h, i). This transformation resulted in the generation of highly extensive and adjustable chiroptical activities, with the maximum g-factor reaching 3.09 × 10^–2. Liu et al. utilized gold nanorods (NRs), which have a low-g chiral nature, to bind with human islet amyloid polypeptides (hIAPPs) [69]. Through the coassembly of NRs and hIAPPs, the g-factors were significantly enhanced, achieving a value of 0.12 at 661 nm. By inducing strong polarization-dependent spectral shifts and reducing scattering of energy states associated with dipoles aligned in opposite directions within assembled helices, the optical asymmetry g-factors were amplified by over 4600 times. Duan et al. created superlattices called chiral molecular intercalation superlattices (CMIS) by inserting specific chiral molecules, such as R-α-methylbenzylamine and S-α-methylbenzylamine, into layered 2D atomic crystals (2D ACs) like TaS₂ and TiS₂ [70]. The results of CD spectroscopy analysis indicated that the CMIS exhibited distinct opposite CD absorption characteristics within the wavelength range of 255–275 nm. Kotov et al.'s study revealed that when aqueous solutions containing Cd²⁺ are mixed with either l- or d-cystine, it results in the formation of bowtie-shaped structures known as cadmium cystinate [71]. The factors responsible for varying chirality are embedded within tiny components at a nanoscale level. Initially, adaptable hydrogen bonds enable flexibility in bond angles. Additionally, the ionization capabilities possessed by nanoribbons themselves create long-distance repulsive interactions among these minuscule constituents, which can be adjusted extensively by modifying pH levels and ionic strength; thus, further strengthening their assembly's handedness.

Fig. 4.

(a) Protocol for preparation of CPIOPC. (b) SEM images of CPIOPCs. (c) Diffused transmittance CD (DTCD) (upper half) and ultraviolet visible (UV–vis) absorption spectra (bottom half) of R-CPIOPCs and S-CPIOPCs. Copyright 2018 American Chemical Society [66]. (d) Structural schematic diagrams of Au₃[(R)-Tol-BINAP]₃Cl and Au₃[(S)-Tol-BINAP]₃Cl clusters. (e) SEM image of nanocube assemblies. (f) CD spectra of Au clusters. Copyright 2017 Wiley-VCH GmbH [67]. (g) Schematic diagram of the process of the chiral molecules-induced deformation of the NPL superlattice. (h) Comparison of the CD intensity between NPL superlattices and the separated ones with R, R-DACH as the chiral inducer. (i) CD spectra of the NPL superlattice without chiral DACH (olive curve), with R, R-DACH (blue curve), and with S, S-DACH (orange curve). Copyright 2024 American Chemical Society [68].

2.3. Localized surface plasmon resonance

Plasmon polaritons can be observed in noble metals such as Au, Ag, Cu, and Pd, with the energy wave being distributed between the oscillations of the electromagnetic field and the internal excitations within the medium [72,73]. Noble metals exhibit remarkable optical properties and plasmon resonances. Plasmonic superstructures made from these noble metals demonstrate LSPR as an optical property, which represents the collective resonance of electrons when exposed to incident light. Based on Mie theory, the LSPR of a metal NP can be described using the extinction cross section (σ_ext) [74]:

$${\sigma }_{ext}=9\frac{\omega }{c}{{\epsilon }_m}^{3/2}V\left[\frac{{\epsilon }_i}{{\left({\epsilon }_r+2{\epsilon }_m\right)}^2+{{\epsilon }_i}^2}\right]$$ (7)

where V represents the volume of the NPs, ω denotes the plasma frequency, ε_m is the relative dielectric constant of the surrounding medium, and ε_r and ε_i are the real and imaginary parts of the sample’s dielectric function, respectively. The value of σ_ext is strongly influenced by the dielectric characteristics of the sample. When the denominator approaches zero (ε_r approaches −2ε_m and ε_i approaches 0), a strong resonance occurs at:

$${\omega }_{LSPR}=\frac{{\omega }_p}{{\left(1+2{\epsilon }_m\right)}^{1/2}}$$ (8)

ω_p is influenced by the electron density (n) and the electron effective mass (m_e), and consequently, it varies with the size and shape of the NP. It can be calculated according to the following equation:

$${\omega }_p={\left(\frac{n{e}^2}{{\epsilon }_0{m}_e}\right)}^{1/2}$$ (9)

where ε₀ is the vacuum permittivity and e is the unit charge, respectively. Besides size, the anisotropic shape of NPs can also influence their plasmonic resonance. The coefficient 2 preceding ε_m adjusts according to the nanostructure geometry, reflecting the variation in restoring force between negative and positive charges. Thoughtfully engineered noble metal superstructures can enhance the coupling and hybridization of various LSPR modes [75].

2.3.1. Surface-enhanced Raman scattering

For noble-metal NCs, the plasmon resonance in their superstructures, which refers to the collective oscillation of free electrons, offers a viable approach for achieving light concentration and manipulation on a small scale. This plasmon resonance enables a wide range of potential applications for noble-metal NC superstructures in diverse areas, including optical waveguides, superlensing, photon detection, and SERS. When irradiated with wavelengths that couple with the plasmon resonance of the inner NCs, the junction regions between adjacent NCs act as ``hotspots,'' leading to amplified local electromagnetic fields within the superstructure [76]. Consequently, this amplification significantly enhances the Raman scattering signals from species detected at these junctions. Wang et al. utilized an excessive number of nonvolatile ligands to create highly ordered superlattices of oleylamine Au (OA-Au) NPs [77] (Fig. 5a, b). They employed in situ Raman spectroscopy to observe the assembly dynamics at multiple NP levels in real time (Fig. 5c). This was achieved by utilizing the SPR coupling of Au NPs and the SERS sensitivity to the distance between adjacent Au NPs. The collective properties exhibited by superlattice nanostructures allowed for convenient monitoring of the assembly process. They also utilized a modified method of assembling oil-in-water microemulsion to synthesize gold superparticles (GSPs), which were then coated with a layer of ZIF-8 [78] (Fig. 5d, e). The Raman scattering intensity of R6G molecules on GSP@ZIF-8 core-shell structures demonstrated an approximate 1.5-fold enhancement compared to the signal obtained on uncoated GSP substrates. The SERS sensitivity of GSPs was improved due to the porous structure of the MOF shell, leading to enhanced gas adsorption capacity (Fig. 5f). Note that within these superstructures, the strength of SERS is determined not only by characteristics such as the type, form, and dimensions of individual NC units but also by factors like the distance between NCs and arrangement pattern. The excitation of surface plasmons on metal surfaces leads to an amplification of the electromagnetic field, which significantly contributes to Raman enhancement in SERS. However, it is important to note that the EM field surrounding individual particles primarily exists within a near-field region referred to as the hot zone, typically spanning less than 3 nm. Tang et al. created a core-shell structure by incorporating a hydrogen-bonded organic framework (HOF) shell onto a gold NP core ([Au@HOF]−1) [79] (Fig. 5h). This resulted in the concentration and preferential accumulation of probe molecules on the highly active region of the plasmonic gold core, resulting in significantly improved SERS performance compared to bare Au NPs and MOF-based core-shell structures (Fig. 5g, i). In practical applications, [Au@HOF]−1 was fabricated as a sensor film for real-time SERS detection of thiram, demonstrating exceptional sensitivity down to 0.12 ng mL^-1 and establishing its reliability as a field-ready detection tool. To better comprehend and optimize the SERS effect, it is highly desirable to develop large-scale NC superstructures with controllable shapes. Tang et al. demonstrate the utilization of three different shapes (octahedral, cubic and rhombic dodecahedral) of Au NCs with identical sizes as fundamental units for constructing superstructures [80] (Fig. 5j). The enhancement factors for rhombic dodecahedral (RD), octahedral, and cubic NC superstructures having similar layer numbers are 1.0 × 10⁷, 0.8 × 10⁷, and 0.7 × 10⁷, respectively. Furthermore, chemical reactions can be facilitated by hot electrons generated through Landau damping of localized SPR, which convert light energy and enhance catalysis in practical applications. Tang et al. fabricated ordered Pd nanoarrays on the surfaces of Au NR [81]. Incorporating transition metal Pd into hybridized Au NPs may eliminate energy barriers and enhance the catalytic activity of resulting hetero-nanostructures. The extensive field distribution observed in Au@Pd superstructures is attributed to the well-organized open structure of Pd nanoarrays along with their strong plasmonic antenna effect. Consequently, the light-enhanced catalytic activity exhibited by Au@Pd superstructures can be significantly enhanced. Goda et al. utilized a metal-free nanostructure with a specific topology, comprising porous carbon nanowires arranged in an array [82]. This nanostructure exhibited exceptional sensitivity, biocompatibility, and reproducibility as a substrate for SERS, achieving an impressive current enhancement factor of 10⁶. Carbon material possesses excellent charge-transfer efficiency. Additionally, the presence of residual H, N, and S atoms following carbonization further facilitates charge transfer between the substrate and molecules, thereby contributing to the enhancement observed in SERS signals. Xu et al. present a universal method for creating Pickering emulsions without the use of modifiers. In this process, colloidal Au NPs are assembled from the solution onto the emulsion surfaces, leading to a significant enhancement in the SERS signal intensity of the 3-mercaptopropionic acid (3-MPA) ligands that are adsorbed on the Au NPs [83]. Pickering emulsions are characterized by the encapsulation of fine liquid droplets within a layer of solid particles dispersed in another immiscible fluid. These unique emulsions hold great promise for various applications in sustainability and healthcare.

Fig. 5.

(a) Schematic illustration of the nonvolatile ligand-regulated Au NP assembly process. (b) SEM image of assembled OA-Au. (c) Variation of R6G Raman intensity over time. Copyright 2023 Elsevier [77]. (d, e) TEM images of GSP and GSPs@ZIF-8. (f) Schematic illustration of GSPs and GSPs@ZIF-8 with gas collisions. Copyright 2018 Wiley-VCH GmbH [78]. (g) Schematic diagram illustrating the adsorption between analytes and different substrates. (h) TEM image of the shell of [Au@HOF]−1. (i) SERS spectra of 1,2-bis (4-pyridyl) ethylene (BPE) in the presence of different substrates. Copyright 2023 American Chemical Society [79]. (j) SEM and schematic images of superstructures self-assembled by octahedral and cubic Au NCs. Copyright 2011 Wiley-VCH GmbH [80].

2.3.2. Surface-enhanced fluorescence

Surface-enhanced fluorescence (SEF) relies on plasmonic nanostructures that create intense local electromagnetic fields, which impact the excitation and emission of fluorophores through near-field coupling [75,84,85]. The presence of these nanostructures in close proximity to a fluorophore can lead to modifications in its spectral properties due to changes in both excitation and emission. This change is caused by the increased density of photonic states in the vicinity of metal NPs. Some studies also refer to this phenomenon as plasma-enhanced fluorescence (PEF) or metal-enhanced fluorescence (MEF). Rinaldi et al. fabricated well-organized arrays of gold nanostructures that are combined with CdSe/ZnS NCs [86]. The close proximity between the metallic nanostructures' surface plasmon (SP) resonance band and the NCs' excitation/emission bands may lead to a significant enhancement in fluorescence, estimated to be up to 30 times higher. The proximity between the plasma NPs and the fluorescent group plays a critical role in adjusting the fluorescence enhancement intensity. Duan et al. utilized plasmonic substrates consisting of self-assembled polydopamine (PDA)-coated NCs [87]. They achieved a maximum enhancement factor of 62.1 for the IR-800CW dye by optimizing the spectral overlap between IR-800CW fluorophores and PDA substrates with a thickness of 20 nm. It is important to note that as the distance from the metal surface decreases, both the strength of the electric field and nonradiative energy transfer to the metal increase, leading to potential fluorescent quenching. Therefore, finding an optimal distance between fluorophore and substrate is crucial for achieving fluorescence enhancement while maintaining this delicate balance between these two factors. Therefore, it is essential to devise methods that allow precise control over the placement of Au NPs and fluorescent emitters on a plasmonic substrate to fully utilize the potential of plasmon-enhanced fluorescence systems.

2.4. Photoluminescence

Metal halide perovskite (MHP) NPs possess various advantages, including multifield coupled responses, photoelectric conversion, electrooptical conversion, and all-optical conversion [[88], [89], [90]]. These advantages have significantly contributed to the advancement of optoelectronic devices. Ensembles of uniform MHP NCs can easily form superlattices, which may protect the NCs against detrimental environmental factors. As a result, these superlattices exhibit slightly improved PL stability compared to dispersed NCs. The unique optical characteristics displayed by these superlattices include SF, ASE, and spectral redshift. When comparing individual NCs to their superlattices, it is observed that fluorescence exhibits a common exponential damping time τ_FL, whereas SF demonstrates a short collective decay period τ_SF [88]. Regarding exciton behavior, ASE resembles SF, with the key distinction being that the dipole phase of excitons in the emitting state is random in ASE but coherent in SF. ASE can be characterized as a series of optical transitions in gain media, where all forms of ASE within the gain region can be amplified through stimulated emission. The emission experiencing the greatest enhancement gradually becomes dominant, ultimately leading to cooperative output.

Consequently, using such superlattice structures can enhance the performance of optoelectronic devices while enabling the development of high-efficiency directional quantum light sources. Wang et al. showed that orderly arranged CsPbBr₃ perovskite nanocubes can spontaneously form superlattices [91] (Fig. 6a-d). By aligning along the z-axis, these nanocube-based superlattices exhibit unique optical characteristics, creating a unidirectional crystal-like structure with a directed transition dipole moment. These hierarchical assemblies comprise well-organized crystalline lattices filled with anisotropic NPs, leading to improved performance in micro-scale light-emitting devices. In contrast, randomly aggregated NPs emit photons indiscriminately and lack sufficient brightness due to their lack of convergence. The resultant crystalline superlattice arrays are approximately 5.2 times brighter than randomly aggregated NP arrangements (Fig. 6e). This breakthrough holds promise for meeting demands for highly luminous and multi-photon quantum light sources while also facilitating future advancements in miniaturized light-emitting displays featuring diverse surface patterns. Wang et al. also introduced a novel technique called NP self-assembly assisted low-temperature sintering (NSALS) to fabricate CsPbBr₃ single-crystal microstructures directly on different substrates [92] (Fig. 6f, g). They achieved this by organizing CsPbBr₃ NPs into an ordered superlattice precursor, followed by controlled sintering at an optimal temperature to promote atomic migration fusion among the particles. The resulting CsPbBr₃ single-crystal microstructure exhibited a significantly enhanced fluorescence intensity when sintered at 150 °C (Fig. 6h). Furthermore, the CsPbBr₃ single-crystal photodetectors fabricated by this method exhibit excellent switching performance and long-term stability. Hpye et al. showed the direct emission of linearly polarized light from light-emitting diodes (LEDs) fabricated using CsPbI₃ perovskite nanoplatelet superlattices [93] (Fig. 6i). The self-assembly of these highly confined and aligned nanoplatelets resulted in strongly linearly polarized electroluminescence from the LEDs, enabling the creation of pure red CsPbI₃ LEDs with a peak polarization degree (DOP) in electroluminescence of 74.4% (Fig. 6j). Ionic surfactants, both anionic and cationic, can bind to halide NC surfaces in an ionic manner and can also undergo unfavorable solubilization equilibria with the ions present in the inorganic NC. To overcome this issue, Kovalenko et al. developed zwitterionic phospholipids as capping ligands for inorganic NCs to prevent adverse reactions between ionic surfactants and NC cores [94] (Fig. 6k). This approach resulted in the production of highly emissive MAPbX₃ and FAPbX₃ NCs with robust surfaces (Fig. 6l). Notably, films composed of C₈C₁₂-PEA-capped FAPbBr₃ QDs demonstrated a high room-temperature PL quantum yield (QY) of 96%–97%, highlighting the exceptional optical properties exhibited by FAPbBr₃ NCs as a whole (Fig. 6m).

Fig. 6.

(a) Schematic of a long-range order across the multiscale. (b) TEM image of the CsPbBr₃ NPs. (c) High-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) of superlattices projected along the z direction. (d) High-resolution (HR) TEM image of adjacent local NPs in superlattices. (e) Schematic of the formation of crystalline superlattice array and aggregate array, respectively. The insets are the corresponding fluorescence micrographs. Copyright 2022 American Association for the Advancement of Science [91]. (f) Schematic of the preparation of the CsPbBr₃ single-crystal microstructure. (g) TEM images of the CsPbBr₃ single-crystal. (h) PL spectra of the CsPbBr₃ single-crystal microstructure sintered at different temperatures. Copyright 2024 American Chemical Society [92]. (i) Schematic illustrating how the orientation of the NPLs could be tuned. The insets show TEM images of the face-down (top), mixed (middle) and edge-up (bottom) NPLs. (j) Polarization dependence of the EL. Copyright 2024 Springer Nature [93]. (k) Zwitterionic molecules. (l) Typical HAADF-STEM image of FAPbBr₃ NCs. (m) Highlighted optical properties of FAPbBr₃ ensemble. Copyright 2024 Springer Nature [94].

SF is an extremely demanding collective quantum phenomenon that has been witnessed merely in a restricted number of condensed matter systems. For the first time, Stöferle et al. have successfully observed SF in a superlattice composed of nanocubes made from CsPbBr₃ [24]. They organized the perovskite QDs of CsPbBr₃ into a 3D superlattice structure, facilitating the coherent emission and collection of photons, ultimately resulting in the phenomenon known as SF. Maksym V. et al. successfully synthesized superlattices (SLs) of perovskite-type (ABO₃) binary and ternary NC through co-assembly techniques [95]. In the binary SLs, larger spherical Fe₃O₄ or NaGdF₄ NCs occupy the A-sites, while smaller cubic CsPbBr₃ NCs are present on both B- and O-sites. Truncated-cuboid PbS NCs are found to occupy the B-sites in ternary SLs. Remarkably, when using 8.6 nm CsPbBr₃ and 16.5 nm NaGdF₄ perovskite mesostructures as precursors for ABO₃-type binary SL formation, they observed SF under high excitation density conditions.

3. Advances in photonic biomedicine based on superstructures

Prompt prevention and detection of diseases are of particular significance for early treatment. Nevertheless, the current reliable diagnostic approaches all demand professional medical expertise. There is a necessity for the innovative development of solutions that empower patients to self-monitor, are inexpensive, user-friendly, and effective. Assemblies with structural color, SERS, or fluorescent properties can be used for the assessment and diagnosis of disease states. Next, we present the most recent advancements in optically-enabled superstructures for biological applications, encompassing disease diagnosis, drug delivery and release monitoring, cell imaging and tracking, and biochips.

3.1. Disease diagnosis

Blood tests and imaging studies, commonly employed in medical facilities, carry inherent risks that can potentially harm patients' well-being. Conversely, a promising approach to diagnosing diseases lies in non-invasive exhaled breath detection technology [96,97]. The production of distinct volatile organic compounds (VOCs) by tumor cells presents an opportunity for early-stage cancer identification through analyzing exhaled breath samples [98]. Timely recognition of cancer holds immense significance in achieving favorable treatment outcomes. Wang et al. utilized hollow structures of ZIF-8 coated on gold superparticles (GSPs@H-ZIF-8) [99]. As gas molecules traverse the porous ZIF material, their diffusion rate decreases, leading to an accumulation within the cavities and a higher concentration of molecules on the surface of the Raman-enhanced superparticle (Fig. 7a). In a complex environment, the hollow structure demonstrates a detection limit for aldehyde molecules that is more than one order of magnitude lower compared to solid structures. There is a significant difference in the intensity of characteristic peaks in the Raman spectra of healthy individuals and lung cancer patients (Fig. 7b, c). Furthermore, the GSPs@H-ZIF-8 material has excellent stability. Raman spectra can be obtained from human subjects at different time periods in their daily lives for three consecutive days. Wang et al. also prepared a hollow Co-Ni layered double hydroxide (LDH) nanocage on Ag nanowires (Ag@LDH) to enhance the adsorption and capture ability of gaseous molecules, resulting in more efficient reaction or absorption of targeted analytes on the SERS active site [100] (Fig. 7d-f). The detection limit of SERS sensors for aldehydes is 1.9 × 10⁹ v/v (1.9 ppb). The SERS microarray sensors were fabricated using a templating method, and it is easy to distinguish exhaled breath samples as either negative or positive for simulated lung cancer biomarkers (Fig. 7g, h). The Ag@LDH sensor was still capable of collecting well-performing Raman spectra after being placed for 180 days. Raman spectral changes can merely be observed with the aid of instruments, while the detection of VOCs by colorimetry is more straightforward. Wang et al. utilized ice template-assisted techniques to facilitate the self-assembly of ZIF-8 particles into layered porous structures [101]. The 3D framework with large pores enables efficient gas flow, while the inherent micropores of the MOF enhance gas molecule enrichment capability, achieving a detection limit (LOD) of 4.6 ppm for formic acid.

Fig. 7.

(a) Schematic diagram of adsorption between exhaled gas molecules and different substrates. (b) Raman spectra from tests in healthy people and lung cancer patients. (c) Principal component-linear discriminant analysis (PC-LDA) classifier based on SERS spectrum. Copyright 2022 Wiley-VCH GmbH [99]. (d) Illustration of the strategy conformational transition of the analyte with the substrate. (e) Diagram of a SERS sensor for VOCs detection. (f) TEM image of Ag@LDH. (g) Principal component analysis (PCA) of the dataset of the detection system of mixed mimetic exhalation. Copyright 2019 Wiley-VCH GmbH [100].

Tears consist of various components such as proteins, peptides, lipids, metabolites, and electrolytes. These components can be used as indicators for ocular diseases like dry eye syndrome (DES), keratitis, night blindness, acute conjunctivitis, etc. [102]. Moreover, tear analysis is a patient-friendly alternative to blood testing since it is convenient and noninvasive. Wang et al. developed a strategy using Au NPs and a sputtering Au layer on SiO₂ nanosphere arrays to create highly intense and dense SERS hotspots [103]. The periodic large pores in the SiO₂ nanosphere array allow impurities from biological sources to enter the lower part while separating them from the target analyte (Fig. 8a, b). This ensures that impurities from the SERS substrate surface are eliminated so that trace target biomolecules in complex tear fluid can be detected accurately by the tear assay kit (Fig. 8c). It can effectively screen for patients with jaundice (Fig. 8d, e). They also utilized a single-layer SiO₂ array as a foundation to facilitate the self-assembly of Au NPs into the spaces between SiO₂ microspheres [104] (Fig. 8f). This procedure resulted in the formation of an Au NP-bridge array characterized by vertical nanochannels and microcavities. This specific configuration exhibits vertical permeability and displays a microvortex effect, highlighting its capability for efficient mass transfer at the solid-liquid interface. The application of this array involves detecting the concentrations of microRNA-21 in HeLa and MCF-7 cells, effectively demonstrating swift DNA hybridization kinetics (Fig. 8g, h).

Fig. 8.

(a) A schematic diagram of SiO₂ NPs sieving substances in tear fluid. (b) Structural unit simplified model of SiO₂@Au@Au NPs. (c) The optical photos of the tear assay kit. (d) Representative Raman spectra of tear collected from the jaundice-positive patients and jaundice-negative patients. (e) PC-LDA analysis based on Raman spectra of tear for discriminating jaundice-positive patients. Copyright 2022 Wiley-VCH GmbH [103]. (f) Schematic illustrating the principle of microRNA-21 detection on the Au NP-bridge array. (g) The anti-interference test. (h) MicroRNA-21 concentration in different numbers of cancer cells. Copyright 2022 American Chemical Society [104].

Optical biosensors enable the simultaneous detection of multiple analytes, representing a significant advancement in modern analytical detection. This approach leverages the high sensitivity and non-invasive sensing features of photonics, along with its ability to perform real-time multi-parameter analysis, integrated with high-throughput multiplexing techniques to allow fast and precise identification of various biomarkers within complex biological samples. The process relies on functional materials that offer selective recognition, photonic architectures that distinguish multidimensional signals, and advanced computational algorithms that interpret intricate datasets.

Detecting cardiac troponin I (cTnI), B-type natriuretic peptide (BNP), and myoglobin (Myo) is essential for the early treatment of acute myocardial infarction (AMI) and heart failure (HF) [105]. Shang et al. utilized magnetic assistance during self-assembly within droplet templates to create structural color spindles (SCSPs) [106]. These SCSPs possess a highly organized lattice structure, which results in the observation of structural colors under light. By incorporating specific DNA aptamers onto differently structured colored SCSPs, they were able to encode information for detecting various types of tumor cell-derived exosomes through coding techniques. Furthermore, due to their distinctive spindle shape and magnetic responsiveness, these SCSPs can be effectively utilized as rotary stirrers to enhance detection efficiency compared to static modes by inducing controlled turbulence during spinning experiments conducted. The authors also investigated the LOD for two types of tumor cell-derived exosomes (A549 and HepG2), which were found to have LOD values at 1.67 × 10⁴ particles/mL and 1.38 × 10⁴ particles/mL, respectively. A hydrogel composed of polyethylene glycol diacrylate (PEG-DA), polyethylene glycol (PEG), and AA effectively encapsulated self-assembled SiO₂ NPs [107]. By etching the SiO₂ NPs, You et al. obtained a PC hydrogel anti-opal film (PC-GELIO). This PC-GELIO membrane exhibited a reflection wavelength that matched the emission peak wavelength of fluorescent molecules, resulting in enhanced fluorescence for improved detection of the tumor marker alpha-fetoprotein (AFP). The researchers achieved a limit of quantification (LOQ) for AFP, reaching as low as 23 pg mL^-1. The fluorescence intensity collected over 7 days showed an RSD of 3.35%, indicating good reproducibility of the method. Su et al. developed a colorimetric biosensor based on the template method for self-assembly of colloidal particle chains [108]. First, the protein A, anti-CEA antibodies, and anti-Tau antibodies are modified on the PS chains in different regions. Then, when mixed with the sample, the surface of Se NPs effectively adsorbs the target proteins (IgG, CEA, and Tau proteins). Finally, through antibody-antigen recognition, PS chains and Se NPs assemble into PS@Se heterochains. By analyzing the imaging color of these heterochains, multiple protein biomarkers can be quantitatively detected simultaneously. The colorimetric biosensor with heterochains was evaluated for stability over a two-week duration, showcasing its capability for reliable and prolonged protein level detection. This approach enables the rapid and efficient measurement of protein expression levels in various clinical samples, including buffer solutions, urine, and serum, by utilizing a neural network algorithm. The mean accuracy attained is 97.3%. Furthermore, nanochains modified with specific SARS-CoV-2 spike antibodies can be used to detect SARS-CoV-2 pseudoviruses in biological fluids directly [109].

3.2. Drug delivery and release monitoring

PCs exhibit superior photonic characteristics and distinct diffraction wavelengths. Upon stimulation from physical, chemical, or biological factors, their diffraction peaks can shift as a result of alterations in lattice spacing, arrangement, or the refractive index of the surrounding environment [110]. By monitoring the drift amount (Δλ) of the reflected/transmitted spectrum in real time, the changes in the measured parameters can be retrieved. Therefore, assembling PCs with materials that have good biocompatibility can construct photonic biosensors for monitoring drug release and detecting diseases. Zhao et al. proposed a method for promoting wound healing by utilizing self-bonded hydrogel inverse opal particles as a patch [111]. They fabricated inverse opal scaffolds particles through the reverse replication of PC templates, followed by the incorporation of drug-loaded pre-gel into these particles to create an adhesive patch specifically designed for wound healing. The scaffolds were fabricated using biocompatible materials, specifically hyaluronic acid methacryloyl (HAMA) and gelatin methacryloyl (GelMA), and were incorporated with graphene oxide quantum dots (GO QDs). The release of the drug caused changes in the refractive index of the particles, resulting in observable color shifts that could be used to monitor the delivery process. Zhao et al. introduced innovative microneedle patches with encoded structural colors (EMNs) that integrate PC and microneedle arrays (MNs) [112]. PC composed of SiO₂ NPs was loaded with biocompatible polymer compositions, including polyethylene glycol diacrylamide (PEGDA) and acrylamide (AM). PEGDA finds extensive applications in biomedical domains, such as in drug delivery systems, scaffolds for tissue engineering, and hydrogel-based wound dressings. These EMNs can detect small molecules such as pH, glucose, and histamine directly at the site of the wound. Upon interaction with a target substance, the volume of the EMN responds to the change, causing alterations in both the color structure and characteristic peaks shifts of the PC. EMNs, when used for treating severe bacterial infection wounds, exhibit noticeable variations in color that can be easily observed without the aid of any magnifying device. Shang et al. proposed a self-healing hydrogel dressing based on microspheres with structural color [113]. These microparticles possess an inverse opal framework that exhibits photothermal responsiveness. Their composition involves good biocompatibility hyaluronic acid methacryloyl, naturally metabolizable silk fibroin methacryloyl, and low doses of black phosphorus quantum dots (BPQDs). Within this framework lies a dynamic hydrogel created through the interaction between cyanoacetate and benzaldehyde-functionalized dextran (DEX-CA and DEX-BA). Additionally, eumenitin and vascular endothelial growth factor were simultaneously loaded into these microspheres. By leveraging alterations in the coloration of the microsphere structure, it becomes possible to effectively monitor drug release progress towards achieving optimal wound management. Chai et al. introduced a novel approach for delivering drugs to the inner ear and treating noise-induced hearing loss (NIHL) using composite microcarriers made of silk fibroin-polydopamine (PDA@SFMCs) [114]. These microcarriers were designed using SiO₂ NPs that self-assembled through an emulsion template method. By filling the gaps between the superparticles with a silk fibroin pre-gel, SFMCs were obtained after removing the template microspheres. The incorporation of PDA allowed for localized drug administration, enabling precise and targeted delivery of loaded drugs. Silk fibroin provides mechanical support, and PDA enhances surface functionalization. Both have good biocompatibility and can be degraded in vivo. The richly porous nanostructures and interpenetrating nanochannels within PDA@SFMCs facilitated rapid absorption of therapeutic molecules as a drug vehicle. In experiments conducted on guinea pigs exposed to noise, N-acetylcysteine (NAC) loaded PDA@SFMCs demonstrated partial restoration of hearing levels. Lei et al. fabricated hydrogels with PC properties (PCHs) that incorporated molecules of (4-((2-acrylamidoethyl)carbamoyl)-3-fluorophenyl)boronic acid (AFPBA) [115]. Hydrogels are composed of acrylic acid (AA) and AM cross-linked. They are easily hydrolyzed or enzymatically degraded in vivo, with degradation products being non-toxic carboxylic acids and amides, and thus have good biocompatibility. The presence of AFPBA enables the hydrogel to undergo expansion or contraction by competitively binding with glucose. This deformation process results in a visible change in structural color, allowing for the monitoring of glucose levels within the physiological range.

3.3. Cell imaging and tracking

By incorporating polymers with responsive properties, the stimulation-induced alteration of lattice parameters can result in a modification of structural color. This facilitates the investigation of cell behavior and state from multiple perspectives, while also enhancing our comprehension of cellular movement and mechanical characteristics [116,117]. Gu et al. developed a system called PC cellular force microscopy (PCCFM) for visualizing cellular forces [118]. By utilizing a PC hydrogel substrate (PCS), they were able to convert small-scale deformations into visible color changes. This substrate consists of self-assembled PCs made up of SiO₂ NPs and polyacrylamide (PAA) hydrogels that fill in the spaces between these particles. By analyzing spectra obtained from a single-color image, it is possible to quantitatively analyze both deformations and cellular forces. The PCCFM system was successfully used for monitoring real-time beating patterns of cardiac cell monolayers as well as variations in force within cell aggregates, suggesting its potential application in studying pathophysiological effects on cardiac tissue caused by static stress. They expanded the application of PCCFM to enable real-time and high-capacity examination for identifying the fluctuating cellular force in MDCK cells and the process of osteogenic differentiation in bone marrow mesenchymal stem cells (BMSCs) [119]. Zhang et al. demonstrated the micro-groove structures of hydrogels with anisotropic structural colors using oriented tunicate NCs (TCNCs) into the hydrogel networks [120]. These hydrogel networks are composed of methacrylated gelatin, which exhibits excellent biocompatibility. The incorporation of micro-groove structures in these hydrogels promotes alignment of cardiomyocytes, and their autonomous beating leads to deformation of the hydrogels, resulting in a shift in interference color. This approach can be utilized to measure heart rate in cardiac tissue. Zhao et al. introduced a hydrogel with both electroconductive and anisotropic structural color properties [121]. The self-assembled carbon nanotubes form superarranged carbon nanotube sheets (SACNTs) on a glass substrate. These SACNTs were then pre-polymerized using GelMA solution, followed by coating them with AAm pre-gel solution containing dispersed SiO₂ NPs. Finally, the mixture was cured to form a conductive color hydrogel. The unique anisotropic morphology of sacnt effectively influenced the alignment of cardiomyocytes, while their conductivity facilitated synchronized beating among these cells. As a result, the hydrogel base underwent deformation and exhibited dynamic changes in its structural color and reflection spectrum. Consequently, this innovative approach led to the development of a visual heart-on-a-chip system capable of dynamically sensing cardiomyocytes and screening drugs. Ren et al. proposed a decoding method that utilizes a well-defined local gradient field and a separable polarizability derivatives-based algorithm with a rotational coarse-grained model (SPARC) [122]. This approach allows for the rapid computation of SERS spectra for proteins through the assistance of full-atom molecular dynamics (MD) simulation. By comparing the experimental spectra with the simulated ones, it becomes possible to determine the protein structure within seconds.

3.4. Biochip

Combining NP self-assembly technology and techniques such as inkjet printing, lithography, and custom functional devices, it is possible to construct and then fabricate chips for metabolite detection, biological protein and cell recognition, etc. [[123], [124], [125], [126]]. Wang et al. employed template printing techniques to fabricate a film composed of patterned Au NPs [127] (Fig. 9a, b). This particular film made from the arrangement of Au NPs was utilized for the development of a durable wearable sweat sensor based on SERS (Fig. 9c). The sensor chip enabled real-time collection of sweat and simultaneous detection of multiple parameters, including volume, pH, and lactate concentration (Fig. 9d-f). Hao et al. introduced a SERS chip that can be worn and is composed of a hydrogel film containing plasmonic Ag-Ag-Ag trimers [128]. Chitosan hydrogels have the ability to capture uric acid and lactic acid found in sweat. The wearable SERS chip was utilized to investigate the impact of regular exercise on levels of uric acid. Song et al. presented a novel approach for creating a biochip with divisional photonic [129]. Biochips are composed of printed nanochains, and the assembly of these nanochain arrays is achieved through the utilization of carboxyl-functionalized PS NPs (Fig. 9g). By taking advantage of resonance-induced near-field enhancement, the resulting nanochains exhibit noticeable alterations in color upon capturing target exosomes (Fig. 9h, i). These captured exosomes can be easily distinguished by visual inspection using an optical microscope, enabling the simultaneous identification of multiple target exosomes within just 30 min and with a sensitivity level of 6 × 10⁷ particles mL^-1 (Fig. 9j-l). Song et al. have developed a biosensor that utilizes optical colorimetry to directly detect pathogenic bacteria in water, urine, and serum samples [130]. The researchers employed the template method to self-assemble PS NPs into well-organized linear nanoarrays. By leveraging capillary action within evaporated droplets, the nanoarray efficiently captures bacterial pathogens within minutes. The near-field positioning of the nanoarray greatly enhances the scattered light signal emitted by the bacteria, enabling visual analysis of their growth, reproduction, and cellular activity. Zhao et al. created a 3D microphysiological lung-on-a-chip system using structural color materials, allowing for visualization of breathing [131]. The PDMS films were specifically designed to resemble alveolar arrays on a physiological scale. Through solvent evaporation, SiO₂ NPs self-assemble onto the surface of the PDMS film, forming ordered colloidal crystals that exhibit distinct structural colors. Human lung cells are co-cultured on this SiO₂ layer while cyclic airflow is introduced into the pulmonary array, mimicking the expansion and contraction processes experienced by alveoli in the human body. As these breathing movements cause deformations in the PDMS, corresponding shifts in structural colors occur simultaneously, providing valuable insights into cell mechanics and enabling real-time monitoring of cultivation progress. Zhao et al. introduced an approach to encode multichannel bioinformation using a PhC array integrated with Mxene [132]. The PC, formed through self-assembly of SiO₂, exhibits the property of enhancing fluorescence. MXene nanosheets were utilized as a platform for immobilizing QD-labeled oligonucleotides onto the surface of PhC arrays. Through fluorescence resonance energy transfer (FRET), the presence of MXene nanosheets resulted in quenching of QD fluorescence. However, upon introduction of the corresponding target molecule, the QDs kept away from the MXene nanosheets and restored their fluorescence signal. By combining different fluorescent signals in this manner, we can achieve multichannel bioinformation coding.

Fig. 9.

(a) Photograph of large-area patterned Au NP superlattice film. (b) SEM images of a patterned Au NP superlattice film. (c) Schematic of the wearable SERS-based sweat sensor. (d) Correlation between the sweat volume collected by the microfluidic channel and total bodyweight loss. (e) Plot showing continuous long-term monitoring of sweat pH. (f) Representative SERS spectra of sweat under high- and moderate-intensity activity. Copyright 2024 American Chemical Society [127]. (g) Distinguish glioblastoma patients from healthy people using a four-channel biochip. (h, i) The corresponding optical images and the intensity changes of biochips after detecting exosome samples from healthy people and glioblastoma patients. (j, k) Detecting exosomes in PBS, serum, and urine, respectively. (l) Semiquantitative testing of exosome samples. Copyright 2024 Wiley-VCH GmbH [129].

The biomedical field benefits from various collective photonic effects derived from different superstructures, each with specific advantages and complementary functions. For instance, structural color, which relies on the diffraction or interference of light by periodic nanostructures to produce color changes, demonstrates unique value in visual monitoring applications. The LSPR of metal NPs (such as Au and Ag) enhances light-matter interactions, enabling highly sensitive detection capabilities. Meanwhile, the excitation-emission characteristics of fluorescent molecules excel in precise quantification and biomolecular labeling. To illustrate these differences, we have compiled and classified relevant case studies in Table 1, comparing three key technologies: structural color based on PCs, plasmon resonance, and fluorescence.

Table 1.

The working principles, structural designs, key performance indicators and application fields of different types of photonic biosensors.

Case	Working principle	Working principle	Application Areas	Ref.
Ice-template ZIF-8	Porous structure promotes gas diffusion; structural color changes with molecule enrichment	Detection limit: 4.6 ppm (formic acid)	Disease diagnosis (VOC detection)	[101]
Colloidal Particle Chain Sensor	Colorimetric detection via protein adsorption-induced structural color changes	Accuracy: 97.3 %	Multiplexed protein detection	[108]
PC Hydrogel Patch	Structural color shift monitors drug release	Visual color changes	Wound healing monitoring	[112]
3D Lung-on-a-Chip	Structural color changes visualize lung breathing movements	Real-time alveolar mechanics monitoring	Pulmonary biomechanics research	[131]
PC-MXene Chip	Photonic crystal-enhanced fluorescence for multichannel bioinformatics	Multichannel fluorescence encoding	Nucleic acid detection	[132]
GSPs@H-ZIF-8	SERS enhancement via hollow ZIF-8 structure for VOC detection	Detection limit: 21.2 ppb (4-ethylbenzaldehyde)	Early cancer diagnosis	[99]
Ag@LDH	SERS for aldehyde detection	Detection limit: 1.9 ppb (aldehyde)	Cancer biomarker detection	[100]
Au NP Bridge Array	LSPR for microRNA detection	Detect microRNA-21 concentration	Cancer biomarker analysis	[104]
Nanochain SARS-CoV-2 Sensor	Virus detection via plasmonic enhancement	Rapid pseudovirus detection	Viral diagnostics	[109]
PC-GELIO	Fluorescence enhancement via inverse opal template for AFP detection	LOQ: 23 pg/mL	Tumor marker detection	[107]
PDA@SFMCs	Fluorescent polymer labeling for drug delivery	Restore hearing loss	Hearing loss therapy	[114]
Nanochain Biochip	Resonance-induced near-field enhancement for exosome detection	Sensitivity (6 × 10⁷ particles/mL)	Exosome detection	[129]

4. Application of AI-integrated photonic biosensing

Photonic biosensors exhibit fast response times and demonstrate greater resistance to external disturbances. For example, Raman spectroscopy identifies biochemical changes via molecular-specific spectral peaks [78,103,104,133], fluorescence microscopy tracks labeled biomolecules in cells [50,134], and structural color shifts monitor real-time parameters like pH and temperature [63,107]. However, traditional analysis methods often depend on manual expertise, which can limit both efficiency and accuracy, particularly when interpreting complex spectra that require extensive physicochemical knowledge. By integrating AI with photonic biosensors, the precision and reliability of diagnostic outcomes are significantly improved, thereby overcoming these limitations. DL models, including convolutional neural networks (CNNs) for image classification [135,136], long short-term memory (LSTM) networks for dynamic fluorescence analysis [137,138], generative adversarial networks (GANs) for synthetic image generation [139,140], automate feature extraction, enhance data processing speed, and minimize human bias. A pre-trained classifier can efficiently perform predictions or diagnoses on new, unseen patient samples and transmit results to portable devices via cloud computing servers. This process provides clinicians with valuable information to support their decision-making [141] (Fig. 12a). Additionally, this integration enables real-time monitoring, personalized therapeutic approaches, and prompt responses to medical emergencies. Advanced algorithms can process large volumes of intricate biological data, enabling sensors to accurately detect and interpret subtle changes in biomarkers. Consequently, this reduces the likelihood of false positives and false negatives, thereby minimizing misdiagnosis. These AI-driven advancements facilitate rapid screening of disease markers, multi-modal data fusion, and reliable analytical insights, accelerating progress in precision medicine (Fig. 10).

Fig. 12.

(a) The feature extraction of the response data for sweat biomarker classification. Copyright 2022 American Chemical Society [178]. (b) Schematic illustration of detecting unsaturated fatty acids utilizing a smartphone-aided nanozyme colorimetric sensor array in conjunction with DL techniques. Copyright 2024 American Chemical Society [181].

Fig. 10.

ML can effectively consolidate a wide and abundant range of information to aid in clinical decision-making. Copyright 2019 Springer Nature [141].

4.1. Brief overview of deep learning

DL, as a branch of ML, is inspired by the mechanisms through which the human brain interprets and processes information (Fig. 11a). ML itself represents a central approach to realizing AI, with its performance improving automatically through data-driven model training [142]. Over time, DL has become the dominant computational technique within ML, demonstrating exceptional effectiveness across a wide range of complex cognitive tasks—frequently achieving results on par with, or even exceeding, human-level accuracy[143]. One of the major strengths of DL is its ability to extract knowledge from large-scale datasets. In recent years, this field has seen rapid advancement and has been successfully applied to numerous conventional tasks, particularly in the domains of spectroscopy and biospectral imaging [[144], [145], [146], [147]].

Fig. 11.

(a) The interconnection among AI, ML, and DL. (b) Categorization of ML techniques and associated algorithms. Copyright 2024 American Chemical Society [155].(c) DL frameworks, including unsupervised learning, supervised learning, semi-supervised learning, and reinforcement learning, along with their typical use cases. Copyright 2024 Wiley-VCH GmbH [156].

DL utilizes extensive datasets to automatically establish the relationship between input data and their associated output labels. The methodology is built upon hierarchical representation-learning techniques, where multiple abstraction layers are formed through the combination of basic non-linear components. Each component transforms the data from one level—beginning with the original input—into a more abstract and refined representation at the next level [143]. Representation learning, in general, refers to a collection of approaches that empower machines to process raw data and independently identify the relevant features necessary for tasks such as classification or recognition [148]. Using the backpropagation algorithm, DL systems detect intricate patterns within large-scale datasets and determine how internal parameters should be modified accordingly. These parameter updates allow the model to enhance its feature representations at each stage by leveraging knowledge derived from earlier layers. As an example, conventional ML relies on human-engineered ``feature extractors'' [149,150]. In image processing, experts must manually develop algorithms to extract visual attributes like position and color—essentially translating raw pixel values into meaningful numerical vectors—a task demanding substantial domain expertise and technical effort [147,151,152]. By comparison, DL enables the automatic extraction of multi-level feature structures directly from data, without the need for manual intervention.

ML continues to evolve, offering enhanced capabilities across a range of ML tasks. It has also facilitated progress in various application domains, including image high-resolution, sample detection, and image recognition [146,147,151] (Fig. 11b). DL is generally divided into three primary categories: supervised learning, semi-supervised learning (sometimes termed partially supervised learning), and unsupervised learning [153,154]. In addition to these, reinforcement learning (RL) constitutes another distinct type of learning framework [143] (Fig. 11c). While it is frequently grouped within the semi-supervised paradigm, it can also be linked to unsupervised learning strategies in certain contexts.

4.1.1. Supervised learning

DL incorporates a range of supervised learning techniques, such as recurrent neural networks (RNNs), CNNs, and deep neural networks (DNNs). Within the RNN architecture, specialized models like gated recurrent units (GRUs) and LSTM networks are frequently employed [154,157]. A major advantage of these methods is their ability to utilize existing knowledge in data acquisition or prediction generation. However, a notable limitation occurs when the decision boundary becomes too rigid due to missing samples from specific classes in the training set [143,158]. Even with this constraint, supervised learning approaches are typically viewed as more direct and efficient than other learning paradigms, providing robust results with comparatively easier implementation.

Many ML algorithms have been used to analyze spectra, artificial neural network (ANN) based models such as the multilayer perceptron (MLP), CNN, or support vector machine (SVM). MLP is a fundamental form of ANN, and subsequently, more complex DL models (such as CNN and ResNet) have all been developed based on this architecture [159]. MLP has been demonstrated to effectively address nonlinear problems [152,160]. It consists of more than three layers, specifically the input layer, output layer, and at least one intervening hidden layer. Each neuron in the MLP gathers signals from the preceding layer and propagates its output to the neurons in the subsequent layer. Through the use of activation functions, MLP is capable of converting a spectrum into a variable-length vector, demonstrating strong nonlinear mapping abilities. Consequently, MLP offers great versatility and can be applied to both supervised and unsupervised learning tasks, including regression, classification, and feature extraction [147]. MLP is often considered limited due to its computationally intensive characteristics and reduced effectiveness in processing grid-structured data, particularly in applications involving image and spectral analysis tasks.

A SVM is an optimization-based classification model designed to find optimal hyperplanes, or decision boundaries, that effectively distinguish between different classes in a high-dimensional space [153,161,162]. Unlike ANNs, SVMs are deterministic, yielding a unique and globally optimal solution[162]. The SVM training model was converted into a program designed to operate on the ATMega microprocessor. This model was optimized to correctly classify the corresponding glucose levels of artificial respiration samples with an accuracy rate of 97.1% [163]. However, the scale of the solution is not predetermined. Since the support vectors are selected from the training dataset, a sufficient amount of data is required to prevent the number of vectors needed for separation from becoming equivalent to the size of the entire dataset.

Compared with other supervised learning methods, CNNs have become the leading methodology for nearly all recognition and object detection tasks [[148], [149], [150],154,164]. A standard CNN structure typically comprises convolutional (Conv) layers, an activation layer, pooling (POOL) layers, and fully connected (FC) layers [142,165]. In a trained neural network, the early layers are mainly responsible for extracting the ``low-level features'' of the data; for example, basic elements such as edges, lines, shapes, and textures in image data [148,149,159,166]. Since these low-level features are universal, the feature extraction capabilities of the early layers can be ``transferred'' to the analysis of other image data. Following a series of alternating Conv and POOL layers that progressively extract and refine hierarchical features, the FC layer functions as a classifier [142]. It is usually applied at the final stage of the model to categorize input images into pre-defined classes by synthesizing high-level features and contextual information from earlier layers. CNNs have overcome the constraints of conventional visual algorithms through their distinctive architectural design, establishing themselves as a cornerstone technology in DL. As their applications evolve from identifying handwritten digits to interpreting intricate visual data, the potential uses of CNNs continue to grow extensively.

DL development marked a significant breakthrough in algorithm design and also facilitated the adoption of two core concepts—data augmentation and transfer learning—by showcasing their effectiveness in visual processing tasks. The training of neural networks usually requires a large amount of labeled data. Sometimes, due to limited conditions, it may be difficult to obtain sufficient data, and a small-scale dataset may lead to overfitting problems. To address this issue, various data augmentation strategies have been introduced. For instance, rotation, flipping, translation, shearing, adjusting color saturation, adding blur, and random cropping are utilized to expand the scale of the dataset [149,163]. These methods alter the visual appearance of input data while preserving its essential semantic content, effectively expanding the training set and improving the model’s generalization performance. In addition, transfer learning approaches enable the adaptation of pre-trained algorithm models by leveraging their relevant components for application to novel yet related tasks [145]. Managing multiple tasks using the same image dataset can be approached by formulating each task's detection as an individual classification problem. Feature transfer offers a solution in such cases, where features applicable to several tasks are learned, thus effectively expanding the dataset size and/or imposing regularization on the classifier. Handling similar tasks across distinct image groups indicates that the source and target tasks are not identical, a challenge that can be mitigated via instance transfer. This approach involves assigning weights to the source training data such that only the most relevant samples receive higher importance [167].

4.1.2. Semi‑supervised learning

Semi-supervised algorithms operate based on a limited set of labeled samples, which are sufficient to guide the training process [167]. These algorithms develop an autonomous learning mechanism, enabling them to produce their own labels during model refinement. In some cases, GANs and deep reinforcement learning (DRL) are also applied in a similar manner [168,169]. Moreover, RNNs, including variants such as GRUs and LSTMs, are utilized for partially supervised learning tasks [144,150]. Conventional data analysis methods frequently face challenges due to insufficient event data and a high prevalence of censored observations. Incorporating a semi-supervised approach enables the model to leverage the scarce labeled event data while also extracting valuable insights from the unlabeled censored data, thus effectively mitigating the risk of overfitting associated with limited training samples [170]. A key benefit of this method is its ability to significantly reduce the need for large amounts of labeled data; at the same time, the generalization ability of the model has been improved. However, a notable drawback is that irrelevant features in the training data may lead to inaccurate decision-making.

4.1.3. Unsupervised learning

Geoffrey Hinton, the winner of the 2024 Nobel Prize in Physics, once mentioned in his published article that human and animal learning primarily occurs without supervision—we understand the structure of the world through observation rather than being explicitly informed about each object [142]. Unsupervised learning allows the training process to proceed even in the absence of labeled data, as no predefined categories are required [167]. In such scenarios, the learning system detects significant features or inherent patterns that help reveal concealed structures and relationships within the input. Widely used methods in this domain include clustering, dimensionality reduction techniques, and generative modeling, all of which fall under the umbrella of unsupervised learning [171]. Several DL architectures have shown promising results in tackling complex tasks related to nonlinear dimensionality reduction and grouping similar data points. Prominent among these are restricted autoencoders (AEs) and GANs, which reflect recent progress in the field [149,150]. Despite their advantages, unsupervised approaches often struggle to provide accurate categorization and typically require substantial computational resources.

Extracting essential information from complex spectral data without supervision necessitates the development of customized network architectures and loss functions that align with specific objectives [171]. Feature extraction approaches facilitate the conversion of elaborate spectral inputs into more straightforward and meaningful representations. At present, widely used techniques for feature extraction in spectral analysis encompass various AE frameworks as well as contrastive learning approaches. AEs, for instance, aim to reduce the discrepancy—such as mean squared error—between the original input spectrum and the reconstructed output by fine-tuning model parameters like weights and biases [149]. Advanced versions of AEs, including variational autoencoders (VAEs) and β-VAEs [172], surpass conventional linear methods such as PCA by capturing complex nonlinear dependencies. These techniques are capable of uncovering latent structures within spectral data, thereby supporting further quantitative interpretation and identification of underlying patterns.

To decrease dimensionality while preserving the most significant characteristics of the data, PCA is a method that transforms high-dimensional datasets into a lower-dimensional space using orthogonal projection [173]. PCA serves as a tool for exploratory multivariate analysis, where a reduced set of uncorrelated components—known as principal components or loadings—is derived to capture the essential spectral information [160]. This is achieved by eliminating redundant, highly correlated original variables that have minimal impact on decision-making processes. As an unsupervised technique, PCA does not require labeled data and instead identifies key components based on the inherent covariance structure of the dataset.

Moreover, to efficiently suppress noise and improve resolution in the preprocessing of spectral data, it is beneficial to employ GANs for constructing models that can produce realistic data instances starting from random noise. GANs are fundamentally composed of two key modules: a generator and a discriminator [149,154,161,169]. The discriminator is responsible for identifying differences between genuine and artificially generated samples, while the generator is trained simultaneously to refine the authenticity of its outputs so as to deceive the discriminator. Such techniques enable faster acquisition of spectral data and contribute to the advancement of real-time spectroscopic measurement. Despite the solid theoretical foundation of GANs, practical implementation still encounters issues like mode collapse and instability during training [149,154].

4.1.4. Reinforcement learning

RL functions by interacting with an external environment, as opposed to supervised learning, which relies on predefined sample data. The feedback provided by RL typically takes the form of a noisy reward signal. This approach is occasionally categorized under the umbrella of semi-supervised learning. Based on this principle, various hybrid methods combining supervised and unsupervised strategies have been developed [159]. Compared to conventional supervised learning techniques, RL presents greater challenges due to the absence of a direct loss function to guide the learning process. Additionally, two key distinctions exist between supervised and RL: first, the objective function in RL is not fully accessible and must be explored through interaction rather than being explicitly provided; second, the state transitions are determined by the environment, where each new input depends on the previous actions taken within that environment [174]. When addressing a specific task, the choice of RL approach depends on the problem's scope or domain. RL finds applications in areas such as business strategy formulation and industrial automation through robotics [175]. One major limitation of this learning method is that certain parameters can significantly affect the rate at which learning occurs. During real-time monitoring, RL can adjust sensor parameters (such as exposure time and light source intensity) in real time based on feedback, thereby enhancing the recognition accuracy and response speed of target features [143,154]. RL has significantly enhanced the performance boundaries of photonic sensors through dynamic optimization and adaptive decision-making, particularly demonstrating potential in multi-objective optimization under complex environments. However, its application is still constrained by issues such as computational cost, data quality, and interpretability.

4.2. Deep learning-assisted photonic biosensors

Fueled by the convergence of AI with photonic and biosensing technologies, AI-enhanced photonic biosensors represent a cross-disciplinary innovation that underscores the connection between the digital world and the real world. The growing abundance of data and improvements in computing power have greatly accelerated the application of DL techniques in both image and spectral processing. Particularly in quantitative and classification tasks, advanced methods like transfer learning, data enhancement strategies, and GANs are being utilized to tackle specific challenges in image and spectrum analysis, thereby meeting the demand for large-scale training data in neural network development [147,152,154,159,164]. In addition, research has focused on creating more interpretable techniques for spectral information, with the goal of explaining how models make decisions, pinpointing key spectral features, and visually emphasizing critical distinguishing areas within images.

A major challenge in the field of image computing lies in managing the vast volume of data that must be handled. The typical workflow involves several stages: Initially, raw visual data is captured using equipment like cameras and sensors. Following this, preprocessing techniques—such as noise reduction, color calibration, and correction of geometric distortions—are applied to enhance image quality. Subsequently, important features are identified using methods like edge detection and texture analysis, which serve as critical inputs for further processing. Based on these features, ML algorithms (including models like CNN, GNN and RNN) are employed to perform tasks such as object identification, semantic segmentation, or categorization. In the final stage, post-processing strategies are implemented to refine the model’s output, with results interpreted through visualization or statistical analysis, leading to actionable and structured insights applicable in domains like medical diagnostics [176,177].

In this subsection, we will focus on the detection research cases of colorimetric, fluorescent and Raman sensors, and introduce the breakthroughs achieved in the deep integration of photonic biosensing and intelligent algorithms.

4.2.1. Deep learning-assisted colorimetric sensor

Colorimetric methods have garnered significant attention for their rapid detection, high sensitivity, and low resource consumption [10,12]. However, issues such as non-uniform colorimetric responses and variations in ambient light conditions can introduce substantial errors in chromaticity detection, thereby compromising the overall accuracy and limiting the reliability and professional applicability of colorimetric sensing. DL algorithms, characterized by their extensive and comprehensive parameter networks, excel in complex pattern recognition and data analysis. By inputting multi-channel spectral data from colorimetric sensors into CNNs, hidden features can be effectively extracted, thereby enhancing detection sensitivity [178]. Additionally, employing GANs to simulate environmental noise can significantly reduce detection errors [179]. The integration of colorimetric sensors with AI holds considerable potential for improving sensitivity, intelligence, and expanding application scenarios.

DL-assisted image analysis has proven to be an effective approach for processing big data. Li et al. developed programmable colorimetric chips utilizing sodium alginate capsules containing various indicators [178] (Fig. 12a). They compiled a dataset of 4600 colorimetric response images, which are the responses of the programmable colorimetric chip to different concentrations of glucose, pH values, and lactic acid. An improved CNN model based on ResNet-18 was adopted that automatically extracts nonlinear features in images through its deep structure, including color intensity, spatial distribution, and response patterns between capsules. Class activation mapping (CAM) further shows that CNN mainly focuses on the regions with strong coloration in the chromatic capsules, verifying the consistency between feature extraction and chemical design. The network parameters (such as epochs, batch size, and learning rate) were optimized through transfer learning, and the fully connected layer was replaced by the global average pooling layer to reduce the number of parameters. The testing results obtained from real sweat samples using the DL-assisted colorimetric method were consistent with laboratory measurements. The interpretable DL process enabled validation of the rationality of the colorimetric chip and provided guidance for optimizing its design. They also developed an enzyme/indicator-immobilized colorimetric patch with Zn²⁺, glucose, and Ca²⁺ in human sweat that could be accurately classified and quantified via the CNN model [180]. Through DL of 5625 colorimetric images with a CNN model, feature extraction mainly relies on the CNN's automatic capture of the color change regions of the color sensor patches. CAM is used to generate heat maps to visually display the regions of interest of the CNN. The concentration of the analyte is directly correlated with the color change, and the volume differences of the samples are corrected by combining the adsorption-swell characteristics of the hydrogel. Su et al. introduced a smartphone-assisted colorimetric sensor array (CSA) utilizing MnO₂ nanozymes [181] (Fig. 12b). The CSAs were integrated with an ``image segmentation-feature extraction'' DL (ISFE-DL) approach for the analysis of unsaturated fatty acids (UFAs). The images of CSA were captured using the built-in camera of a smartphone and processed through the ``Quick Viewer'' application. MobileNetV3-small, a lightweight CNN, was utilized for feature extraction and regression modeling. The color images were converted to grayscale, and the OTSU algorithm was applied for binary segmentation to distinguish the foreground (sensor units) from the background. Small noise points were removed through morphological opening operations to denoise the image. After identifying the foreground contour, the radius and center position of the minimum enclosing circle were determined to select the appropriate region of interest (ROI). The red, green, blue (RGB) average values of the central area were extracted after sorting the ROIs by rows. The RGB differences between the ``before reaction'' and ``after reaction'' images were calculated to generate a color difference map, and these data were transformed into high-dimensional, nonlinear spectral features. Image quality was ensured by maintaining fixed shooting conditions (professional mode, dark box, uniform illumination) and standardizing ROI. The system successfully distinguished between oleic acid (OA), linoleic acid (LA), α-linolenic acid (ALA), and their mixtures. When dealing with small sample high-dimensional data efficiently, LDA and PLSR can be utilized to achieve classification and quantification. Nezhad et al. developed a colorimetric sensor array utilizing Au NPs [182]. In the presence of catecholamine neurotransmitters (CNs), Au NPs exhibit distinct fingerprint response patterns under different buffering conditions. Ten specific wavelengths were extracted from the UV–vis spectrum, which were based on the LSPR band of AuNPs and the spectral changes caused by their aggregation. Spectral profiles at different concentrations were generated for visual recognition, showing color tone changes and intuitively displaying the response patterns of each analyte at different concentrations. LDA was used for classification tasks, and PLSR was used for quantitative analysis, maximizing inter-class differences and minimizing intra-class differences through linear combinations. By integrating the colorimetric sensor array with these algorithms, the system can accurately differentiate between individual neurotransmitters and their mixtures.

4.2.2. Deep learning-assisted fluorescence sensor

Fluorescence-based biosensors offer remarkable sensitivity for analytical applications and have been effectively utilized to measure biomarkers in various biological samples. Researchers have focused on advancing multiplex detection methods within fluorescence sensing to achieve thorough analysis using minimal sample volumes, thereby substantially reducing both costs and processing time. Through the application of ML techniques, fluorescence signals can be swiftly categorized, enabling accurate and efficient multiplexed detection. Ozcan et al. introduced a paper-based fluorescence vertical flow assay (fxVFA) platform [183]. This platform can be used to quantify three key cardiac biomarkers (myoglobin, creatine kinase MB, and cardiac-type fatty acid-binding protein) from human serum samples. The fluorescence signal of the sensing membrane is focused onto the mobile phone camera through a filter and an external focusing lens. Fluorescence images are collected using a handheld fluorescence reader based on a smartphone. Customized automated segmentation code is used to segment 17 immune reaction spots from the image and calculate the average pixel intensity of each spot. The 17 signals are standardized; that is, the mean of the training set is subtracted and divided by the standard deviation, and they are used as input for the neural network. A fully connected neural network (FCNN) is used, and the hyperparameters, such as the number of hidden layers, number of units, regularization, and dropout, are optimized through grid search (Fig. 13a). The predicted concentrations from the fxVFA platform showed strong correlation with reference measurements across the entire clinical range. Tan et al. designed a liquid biopsy system for breast cancer detection by integrating a fluorescence sensor array with the DL tool AggMapNet [184] (Fig. 13b). A 12-unit sensor array was utilized, comprising conjugated polyelectrolytes, fluorophore-labeled peptides, and monosaccharides or glycans, for capturing fluorescence signals originating from cells and exosomes. To process the unordered fluorescence spectra, the CNN-based DL tool AggMapNet was employed to generate feature maps (Fmaps). The spectral data were reduced in dimension and classified using LDA, PCA, and HCA. By aggregating the relevant spectral signals to adjacent pixels, feature discrimination was enhanced. A 100% accuracy rate was achieved in the classification of breast cancer patients and healthy individuals, which was superior to traditional LDA. This approach successfully demonstrated the ability to differentiate between healthy and diseased samples using the transformed feature maps generated by AggMapNet. Additionally, multiplexed fluorescence sensing, which simultaneously detects RNA targets of respiratory syncytial virus (RSV), influenza A virus (IAV), and severe acute respiratory syndrome coronavirus (SARS-CoV-2), has been realized using cost-effective photonic components and pre-trained ML models [185]. A compact lensless fluorescence sensing device was used, combined with low-cost photonic components and a spectral sensor. Under three excitation light sources (blue, green, and red), fluorescence signals were captured through eight detection channels, generating 24 normalized fluorescence units (RFUs) as input features. The model automatically learned the nonlinear relationships among the spectra. The normalization of the original RFUs was performed to eliminate baseline differences in the measurements. Three models were used to process the spectral data. Firstly, MLR was used for preliminary analysis, but its effect was limited. SVR with an RBF kernel was adopted to handle the nonlinear relationships. NN, as a more complex model, was used to deal with higher degrees of nonlinearity. This approach enhances the efficiency and versatility of diagnostic tools in detecting multiple pathogens.

Fig. 13.

(a) Computationally refined neural network models for quantifying myoglobin (DNN_Myo). Copyright 2023 Wiley-VCH GmbH [183]. (b) Illustration of the AggMap workflow and accurate diagnosis of breast cancer by AggMapNet. Copyright 2022 American Chemical Society [184].

4.2.3. Deep learning-assisted Raman sensor

Raman spectroscopy enables the acquisition of unique fingerprint information from analytes, facilitating rapid and label-free detection. However, due to its inherently weak signals and sensitivity to sensor imperfections, it necessitates both signal enhancement and the removal of noise originating from instruments and backgrounds. While plasma nanomaterials can substantially boost Raman scattering intensity, challenges remain in capturing low signal-to-noise ratio (SNR) Raman or SERS signals, particularly in scenarios involving ultra-low concentration molecular biosensing or measurements that require brief acquisition times. Consequently, developing denoising algorithms to mitigate noise in collected Raman and SERS spectra is essential. Moreover, many analytes and matrix components exhibit similar or overlapping spectral profiles, complicating manual differentiation. The application of DL techniques for spectral preprocessing and feature selection can help extract pertinent information from spectral data, thereby enhancing dataset quality [159,160,173].

Raman spectroscopy has the potential to identify the species and antibiotic resistance of bacteria. Cui et al. introduced a novel approach that integrates open-set DL (OSDL) with single-cell Raman spectroscopy for pathogen identification [186]. To enhance this method, they emphasized the importance of establishing Raman datasets specifically for pathogens in aerosol form. The spectral baseline is automatically corrected by a Python script to eliminate background interference. The spectral intensity is normalized to the range of 0 to 1 within 600–1800 cm⁻¹, and only single-cell data is retained. An OSDL algorithm is developed based on the traditional CNN and improved by an attention neural network (Fig. 14a). The OSDL system optimizes feature extraction ability by adding an attention module, which can capture subtle differences in Raman spectra, reducing the false positive rate by approximately 36% and achieving an accuracy rate of 93% in identifying five target airborne pathogens. This advancement allows Raman-based pathogen detection to move beyond pure culture environments, significantly expanding its applicability in environmental monitoring. Dionne et al. developed a Raman-CNN to classify bacterial spectra based on isolate characteristics, empirical treatments, and antibiotic resistance profiles [187]. Single-layer bacterial samples were measured using a Raman spectrometer to ensure that most of the spectral data originated from individual cells. A low SNR = 4.1 was introduced by short measurement times. The dataset contained 60,000 reference spectra (from 30 types of bacteria and yeast) and 12,000 clinical patient sample spectra. A 1D-CNN was used to automatically extract features directly from the raw Raman spectra. Unique molecular compositions of different bacterial phenotypes lead to distinct variations in their Raman spectra. The CNN achieved an accuracy of 82.2% in the 30-class bacterial classification task, significantly higher than that of logistic regression (75.7%) and SVM (74.9%), along with antibiotic treatment identification accuracies at 97.0% ± 0.3%. This method enables precise and targeted treatment of bacterial infections within hours, significantly improving diagnostic efficiency and patient care. Wang et al. employed Raman spectroscopy combined with a VGG-16-based CNN to rapidly and non-invasively differentiate carcinoma tissues from adjacent non-tumor tissues [188]. Raman spectroscopy imaging facilitates the precise delineation of tumor margins and visualization of lesion regions. Extensive Raman datasets from liver tissues were used to train a DL model based on CNNs, enabling accurate classification of spectral data from different tissue types. This model effectively distinguished various pathological types of liver cancer tissues, highlighting its potential for pathological identification and intraoperative guidance (Fig. 14b). The spontaneous Raman spectra are affected by relatively low signal intensities and have a high SNR.

Fig. 14.

(a) Strategies for optimizing the OSDL algorithm. Copyright 2025 American Association for the Advancement of Science [186]. (b) Illustration of the workflow for histopathological diagnosis of liver cancer utilizing Raman spectroscopy in conjunction with an intelligent algorithm. Copyright 2023 Springer Nature [188].

DL is capable of extracting molecular characteristics from the intricate, high-dimensional spectral datasets generated by SERS, uncovering concealed chemical insights and addressing the classification challenges posed by peak overlapping in conventional approaches. It enables optimization of the analytical workflow for various types of interference, effectively eliminating background noise and fluctuating signals from SERS spectra, resolving standardization issues in large-scale datasets, delivering high-quality data for downstream analysis, and markedly improving the sensitivity for detecting low-concentration analytes. These benefits enhance the applicability of SERS in areas such as trace-level detection and biosensing.

4.3. AI-integrated photonic biosensing in practical applications

Combining AI with photonic biosensing technology to build instant and portable devices through intelligent, miniaturized, and multimodal integration is expected to realize the vision of universal healthcare that is accessible to everyone and available everywhere. However, before being widely applied in practice, the integrated photonic biosensing technology of AI must overcome some limitations and obstacles. Optical biosensors often need to be combined with devices such as cameras and spectrometers and integrate multi-source data (such as visible light, Raman, and fluorescence signals) through AI algorithms to enhance the accuracy of biological feature recognition. Factors such as sensor performance, photonic system parameters, and environmental conditions can lead to increased image noise and blurred features, which require improvement through active lighting and image enhancement algorithms (such as denoising and contrast enhancement) [149,164]. The image data streams generated by photonic biosensors have high-dimensional and nonlinear characteristics, and real-time analysis needs to be achieved through streaming processing frameworks and lightweight algorithms (such as random forests and CNNs) [146,147,151]. To reduce latency, AI models need to be deployed on local portable devices for edge computing [189]. Additionally, the collection of biological data may infringe on personal privacy and involve risks of privacy leakage, which necessitate the use of federated learning (FL) or blockchain technology for secure transmission and storage.

4.3.1. Image or spectral data acquisition

One key factor in obtaining high-quality images or data is the hardware structure, while also ensuring an economic, compact, and lightweight design. Expensive single-lens reflex or mirrorless cameras and various large spectrometers are not suitable for portable detection, although they can obtain high-quality data. Micro spectrometers, handheld Raman spectrometers, and smartphones—widely applicable and compact—have gradually become the core hardware structure of mobile medical platforms [164]. Reducing the size of sensors often comes at the expense of resolution or sensitivity. To enhance the image quality of portable imaging devices, it is necessary to develop specific algorithms to reduce noise caused by sensors or environmental factors, thereby improving the SNR. Bergholt et al. developed the DeepeR DL framework to improve the throughput of Raman spectroscopic imaging [190]. The method first denoises low SNR spectra using 1D ResUNet, then applies HyRISR for spatial super-resolution reconstruction of the denoised low-resolution images. The training utilized a dataset of 169 hyperspectral Raman images, expanded through data augmentation to better capture spatial-spectral correlations. Ren et al. used multi-layer convolution to extract features from low-resolution images and applied ReLU and the Adam optimizer to accelerate training [191]. They then built a mapping using synthetic low–high-resolution image pairs, enabling precise reconstruction of line-scan Raman images of live cells. A wide slit was used to retain full spectral information and prevent high-frequency loss, allowing the CNN to reconstruct high-resolution images more effectively.

Furthermore, developing generalized image processing algorithms is essential for achieving data standardization, restoration, and enhancement. The establishment of standardized benchmarking protocols holds critical significance in the domain of photonic biosensing, particularly due to the diversity of imaging techniques, sensor technologies, and variations in data acquisition criteria and hardware configurations [166,177]. For example, U-Net is a DL model that is particularly suitable for biomedical image analysis. Ronneberger et al. utilized the U-Net model through precise segmentation and detection techniques, which enabled them to effectively identify and count cells as well as conduct morphometric measurements [192]. Standardized protocols play a crucial role in connecting advancements in photonic biosensing with real-world application needs. Resolving technical inconsistencies, establishing unified metrics, and enhancing data normalization via algorithmic approaches can significantly improve interoperability and collaborative functionality across different systems.

Owing to issues like data privacy, security concerns, and the expense associated with data annotation, acquiring an adequate number of labeled training samples becomes challenging in many practical applications. To tackle these difficulties, methods such as few-shot learning and active learning can be employed to effectively train models using limited labeled data while minimizing annotation efforts. Few-shot learning (FSL) is designed to enable models to perform targeted tasks using only a limited set of labeled examples, effectively mitigating the risk of overfitting commonly seen in conventional supervised learning approaches under data-scarce conditions [193]. To offset the limitations caused by small sample sizes, FSL incorporates prior knowledge into the learning process. First, it involves augmenting or synthesizing training data to expand the available information base. Second, it focuses on constructing simpler model architectures or constrained hypothesis spaces to control model complexity. Finally, the learning framework can be adapted, such as through meta-learning mechanisms that teach the model to rapidly adjust to novel tasks or via metric learning techniques that assess sample similarity (e.g., cosine similarity) within an embedding space, allowing classification of query instances based on minimal support examples [[194], [195], [196]]. Transfer learning and domain-specific expertise play a crucial role in improving the model’s generalization performance. Active learning seeks to minimize the number of training samples required while maintaining the effectiveness of the classifier. Selecting informative examples strategically helps reduce the overall labeling burden. The approach involves prioritizing unlabeled images that are expected to contribute the most to model improvement, thereby balancing the trade-off between classification accuracy and annotation expenses [197]. Nikolaos et al. introduced a scalable active learning framework designed for multi-class image classification, which operates based only on binary feedback [198]. This type of feedback enables highly efficient annotation in complex classification tasks, leading to considerable savings in both time and effort during training. In the context of multi-label image classification, building an effective training set is crucial, and active learning enhances performance by iteratively identifying and querying the most valuable samples from human annotators [197].

4.3.2. Real-time data processing

Transforming photonic biosensors into portable, real-time monitoring tools involves combining various technologies such as hardware engineering, algorithm refinement, power efficiency strategies, and data handling. Patient monitoring through these sensors can be facilitated using connected devices that track multiple health indicators simultaneously and produce continuous data flows [145,168]. As a result, such interconnected systems serve as the backbone of IoT-based sensor networks, enabling real-time data acquisition. Efficient communication between these connected sensing nodes depends on overcoming issues related to transmission delay, network capacity, and prompt decision-making.

However, many image- and video-related applications require intricate processing steps, necessitating the use of hardware accelerators to meet real-time performance demands. Certain edge devices, such as NVIDIA’s Jetson Nano [199] and Intel’s Myriad2 processor, stand out due to their low power consumption and high performance. These devices offer advantages including energy efficiency, real-time processing capabilities, resistance to interference, and enhanced data privacy, all of which contribute significantly to the advancement of photonic biosensor technologies. Furthermore, the progress in Field-Programmable Gate Arrays (FPGAs), particularly in terms of expanded hardware capacity and improved programming tool efficiency, has had a major impact on real-time applications [200]. FPGA is a semiconductor device based on programmable logic units, which can dynamically configure its internal logic functions through hardware description languages. Tasks in image processing—especially those leveraging DL methods and CNNs—can achieve notable performance improvements through FPGA-based acceleration [165,201,202].

Edge computing has emerged as a promising solution to address these challenges. This distributed computing approach processes information near its source, reducing delays, enhancing responsiveness, and decreasing reliance on centralized cloud infrastructures [203]. When combined with AI, edge computing opens up new opportunities for intelligent systems capable of adapting and responding instantly to monitoring demands [204,205]. For example, in detection and surveillance tasks, image recognition models can be developed using extensive datasets and methods like CNNs or transfer learning frameworks to improve accuracy in complex settings. After training, lightweight versions of these models can be implemented on edge platforms (e.g., smart sensors) for immediate analysis, identifying visual changes, and delivering instant system feedback. Conducting AI inference directly on edge devices enables local data processing without transmitting large volumes of data to distant servers [206]. This leads to faster response times and more efficient use of network resources. However, despite its benefits, edge computing still encounters several limitations. Constraints in processing power and memory often necessitate balancing functionality with performance, while also safeguarding sensitive data.

4.3.3. Data privacy and ethics

Without robust security protocols and protective measures, the integration of IoT and edge computing could lead to unauthorized surveillance, breaches of privacy, and loss of public confidence. In response, researchers at Google introduced the FL framework in 2016, enabling ML models to be trained directly on user devices without the need to upload raw data to a centralized server [203,207,208]. FL is a distributed machine learning technique that allows multiple devices to work together to build a shared model, while ensuring that their original data remains stored locally. This approach effectively addresses concerns related to data privacy and security. The integration of hardware accelerators, edge computing, and FL offers privacy, security, low latency, and high energy efficiency solutions for intelligent sensing devices [209].

5. Conclusion and outlook

The self-assembly of NPs enables the creation of superstructures with unique collective properties (Fig. 15a), which hold significant potential in biosensing, cellular research, and related fields. Recent advancements in AI technology have driven substantial progress in ML applications for materials synthesis, process automation, and image analysis [155]. When strategically implemented, these technologies can significantly enhance the efficiency of data interpretation. This review summarizes recent advancements in photonic biosensors utilizing nanoparticle-based superstructures, with particular emphasis on the unique collective photonic responses arising from self-assembled nanostructures in optics and photonics. It systematically examines the emerging applications oaf superstructure photonic biosensors in biosensing and biomedical imaging, while highlighting the transformative potential of integrating ML to enhance data analysis accuracy and advance disease diagnostics through photonic biosensing platforms. The integration of self-assembly of NPs and AI-driven methods has promoted the development of materials science and nanotechnology. While notable breakthroughs have been achieved, this interdisciplinary field continues to face critical challenges requiring innovative solutions.

• (1)In the field of nanoparticle self-assembly, inter-batch variability and defects in assemblies pose a significant challenge for both industrial production and reproducibility in scientific research. It also affects the accuracy, reliability and validity of the data of the photonic biosensors based on the superstructure. The underlying issues stem from variations in reaction conditions, inconsistencies in raw material properties, fluctuations in process parameters, and the impact of human factors. However, optimizing the self-assembly process of NPs often necessitates fine-tuning multiple variables, which can be time-consuming and labor-intensive. At present, researchers have developed autonomous laboratories integrating robot and AI systems for applications such as the cultivation of Escherichia coli strains and the synthesis of photoactive inorganic perovskite NCs [210,211]. Automation in these laboratories ensures precise control over reagent addition, mixing, heating, and separation processes. AI algorithms are crucial for identifying correlations between reaction conditions and nanoparticle properties, enabling the recommendation of optimal experimental settings without exhaustively exploring the entire chemical space. In the future, they serve as an effective approach to mitigate batch-to-batch variability in self-assembled superstructures and have significantly advanced ML-assisted synthesis and the manufacturing of nanomaterials.
• (2)Self-assembled superstructures display distinctive collective properties that arise from the precise organization and interaction of NPs, and they can significantly enhance the interaction between light and biomarkers, achieving precise detection of biomolecules, cellular activities, and environmental changes (Fig. 15b). Therefore, the development of assemblies with new collective properties has become an important driving force for promoting the progress of superstructure-based photonic biosensing technology. In recent years, significant advances have been made in the computational prediction of colloidal particle self-assembly, especially regarding the role of ML technology in crystal structure classification and identification of suitable ordered parameters [207,[212], [213], [214]]. Through ML, the interaction parameters of colloidal particles that can self-assemble into the target structure can be reverse-designed from the desired material properties, and the rational design of materials can be realized [215,216]. However, it remains a challenge to predict with ML-assisted self-assembly whether the upper structures formed by colloidal particles with different shapes and interactions will exhibit new collective properties. For instance, the self-assembly process involves complex multi-body interactions, and the acquisition of experimental data is costly (such as cryo-electron microscopy and synchrotron radiation characterization), resulting in insufficient training data. Moreover, the interaction mechanisms among different self-assembled systems (such as Au NPs and MOF NPs) vary greatly, and a single model is difficult to be universally applicable.
• (3)The signals generated by superstructure-based photonic sensors (such as fluorescence, reflected light, and Raman signals) possess high-dimensional and nonlinear characteristics (Fig. 15c). The incorporation of AI presents novel opportunities for analyzing photonic spectral graphs and image data. AI not only streamlines the labor-intensive process of manual image analysis but also detects features invisible to the human eye. However, while AI models need to handle multi-source heterogeneous data, efficient data processing and analysis are required; issues such as sensor noise and background interference affect data quality. These issues primarily arise from the inherent limitations of photonic principles, the intricacies of instrument design, and environmental factors. Therefore, there is an urgent need for advanced algorithms and ML techniques to more accurately and reliably identify and diagnose biomarkers, avoiding or reducing the influence of external factors such as environmental changes, equipment drift, or sample preparation differences on superstructure photonic biosensors.

Fig. 15.

ML-assisted self-assembly of nanoparticles and optical biosensing. (a) SEM images of ZIF-8 NPs, cadmium cystinate nanoplatelets, Au NPs, and CsPbBr₃ nanocube self-assembled into different superstructures. Copyright Springer Nature [24,38,71]. Copyright Wiley-VCH GmbH [78]. (b) Optical images of structural color microneedle patches, chiral PCs, CsPbBr₃ superlattices, ZIF-8 PCs, and ZIF-8 supraparticles. Copyright Wiley-VCH GmbH [40,63,112]. Copyright Springer Nature [24,38]. (c) Optical images of the tear assay kit. Illustration showing the microRNA-21 detection using the AuNP-bridge array. Illustration showing the detection process of heterochains for highly sensitive and large-scale protein analysis. Copyright Wiley-VCH GmbH [103]. Copyright American Chemical Society [104,108].

The integration of nanoparticle superstructures with AI opens up new possibilities for biosensing and imaging, driving the advancement of personalized medicine and point-of-care diagnostics. Nevertheless, transforming these promising scientific breakthroughs into practical biosensor devices for daily use demands additional work. Upcoming research will concentrate on designing innovative nanoparticle superstructures, optimizing AI algorithms, and constructing miniaturized and low-cost photonic biosensors based on superstructures, thereby enhancing human health and quality of life.

CRediT authorship contribution statement

Jikun Yin: Writing – original draft, Conceptualization. Bo Wang: Supervision. Tie Wang: Supervision. Zhiyong Tang: Supervision.

Declaration of competing interest

The authors declare that they have no conflicts of interest in this work.

The author is an Editorial Board Member/Editor-in-Chief/Associate Editor/Guest Editor for this journal and was not involved in the editorial review or the decision to publish this article.

Biographies

Jikun Yin is currently a Ph.D. student of Prof. Tie Wang at the Life and Health Intelligent Research Institute, Tianjin University of Technology. He received his B.S. degree from Xiangtan University in 2020 and his M.S. degree from Xiangtan University in 2023. His research focuses on the assembly of MOF particles and the design and preparation of MOF composites.

Bo Wang is an assistant professor at the Life and Health Intelligent Research Institute, Tianjin University of Technology. He received his B.E. (2012) from Dalian Minzu University, M.S. (2016) from Tiangong University and Ph.D. (2020) from Tianjin University, China. His research focuses on the novel membrane materials design and scale up research for gas separation.

Tie Wang is a professor at the Life and Health Intelligent Research Institute, Tianjin University of Technology. He received his B.S. (2002) from Xi’an Jiaotong University and his Ph.D. (2008) from Changchun Institute of Applied Chemistry, China. He then worked as a postdoctoral fellow at Rensselaer Polytechnic Institute, Troy, NY, and the University of Florida, Gainesville, FL. His research focuses on nanoparticle assemblies and their applications.

Zhiyong Tang obtained his PhD degree in 1999 from the Chinese Academy of Sciences. Following that, he went to Swiss Federal Institute of Technology Zurich, Switzerland, and to the University of Michigan, USA, for his postdoctoral research. In November of 2006, he joined the National Center for Nanoscience and Technology in China and took up a full professor position. His current research interests focus on assembly and optical properties of functional nanomaterials, as well as their applications in energy, catalysis, separation, and sensing.