2.1.

NADH Fluorescence and Metabolism

Imaging endogenous fluorescence of NAD(P)H and FAD has shown enormous promise in providing metabolic readouts in vivo.^3,20,21 Taking advantage of the fact that NAD(P)H and FAD are autofluorescent, their oxidized or reduced counterparts [NAD(P)⁺ or ${\mathrm{FADH}}_2$] are not, allowing for the calculation of optical redox ratios, which represent the state of cellular metabolism at subcellular scales. Although varying definitions of redox ratio have been used, largely based on the pioneering work of Chance et al.,²² the optical redox ratio has been widely defined as the intensity of NADH divided by the intensity of FAD.²³ With the development of fluorescence lifetime imaging microscopy (FLIM),²⁴ the utility of NADH endogenous fluorescence has been further extended with FLIM able to measure glycolysis via the discrimination of bound versus free NADH.^25–27 Whether based on intrinsic fluorescence alone or via FLIM measurements, such capability of detecting spatial and temporal metabolic contrast has contributed to many areas of biology, including stem cell differentiation,²⁸ cancer,³ and its interplay with the immune system,²⁹ including microglia metabolism.³⁰ However, this technique has its limitations, as the NAD(P)H autofluorescence signal is indistinguishable between the relative contributions of NAD(P)H and NADH, both of which are metabolically important, thus obstructing definitive interpretation of these readouts. Traditional methods to determine NADH and NAD(P)H concentrations require the excision and disassociation of tissue, which destroy spatial information and leave cells potentially prone to oxidation. Oxidation is problematic because it modifies the innate molecular states within the cell, distorting the interpretation of the data. Therefore, autofluorescence imaging would greatly benefit from well-designed biosensors that could alleviate these concerns and serve as very useful validation.

Luckily, several biosensors have recently been developed that could address many of these issues. Traditional metabolic sensors based on ribonucleic acid (RNA) and transcription factors can generate readouts to different metabolic processes, oxidative phosphorylation, and lipid biogenesis. However, these genetic readouts rely on mechanisms of transcription and translation and often lag minutes to hours behind the initial metabolic changes. Other techniques based on “protein activity,” such as ones that utilize fluorescence resonance energy transfer³¹ or circularly permuted fluorescent proteins,³² have readout times closer to milliseconds to seconds, providing more informative temporal data. There are a few recently reported biosensors that are attempting to differentiate NADH and NAD(P)H: sensor of NAD(H) redox (SoNar) biosensors have been designed to detect NADH/NAD⁺-specific ratios and are based on circularly permuted yellow fluorescent protein.³³ Moreover, NADP⁺ levels have been characterized with anisotropy Apollo-NADP⁺ sensor, and a new NAD(P)H-estimating ratiometric non-destructive sensing tool has been reported to show NAD(P)H/NADP⁺ redox ratio through fluorescent readout, which are exciting validation tools to help advance NADH endogenous fluorescence.³⁴ However, limitations still exist because these sensors are sensitive to pH and temperature changes and often contain fluorophores with significant spectral overlap with NAD(P)H emission. One must consider the potential challenges in separating signals from autofluorescence and fluorescent dyes/probes, particularly when their spectral characteristics overlap. To address this issue, a combination of quantitative approaches can be employed, including FLIM and spectral deconvolution techniques. FLIM utilizes the distinct fluorescence decay profiles of different fluorophores and autofluorescent species, providing an additional layer of specificity. Spectral deconvolution can further enhance signal separation by analyzing the contributions of individual components across the emission spectrum. Together, these approaches enable robust discrimination of overlapping signals.

2.2.

Autofluorescence for Cell Type Identification

To fully harness the potential of autofluorescence, particularly in complex biological environments, there is a critical need for highly specific label-based methods. These methods are essential for advancing, defining, and validating label-free approaches by providing a reliable reference for accurate cell identification and metabolic profiling. There is a great promise that NAD(P)H and FAD⁺ autofluorescence could be used to provide cell-specific metabolic contrast in a variety of different contexts. Changes in³⁵ NADH and FAD⁺ autofluorescence have been used to monitor stem cell differentiation and macrophage and microglia activation in vivo. In fact, the metabolic characterization of cells, such as microglia and macrophages, in their three-dimensional local cellular environment, is perhaps one of the most exciting applications of endogenous fluorescence imaging. For example, the ability to identify and metabolically profile microglia in the brain during neurodegenerative pathology, such as Alzheimer’s disease, has been an aspiration in the biophotonics field.³⁶ The use of label-free fluorescent microscopy can facilitate a more granular dissection of the function of plaque-adjacent microglia, elucidating their previously unknown metabolic profiles and role in disease progression. In breast cancer, NADH and FAD⁺ autofluorescence can be used to distinguish macrophages from cancer cells in the breast tumor microenvironment.²¹ Moreover, the optical redox ratio can reveal the effectiveness of chemotherapeutic treatment³⁷ and add context to the importance of tumor-associated macrophage metabolism. Although these techniques show great promise in providing label-free cellular contrast in vivo for both microglia and macrophages, there is still a significant need for new probes that can be used in concert with these techniques to further enhance the diagnostic value of these observations. For example, current means for labeling cell types require the staining of fixed tissues, which loses dynamic context and destroys some metabolic information. Other alternatives are the genetic expression of fluorescent proteins based on cell type-specific promotors along with the creation of mouse models, which are often cost-prohibitive and can be complicated in model systems with limited genetic manipulability. In addition, dye-labeled antibodies specific to cell surface markers have been used,²¹ but these have limitations in depth of diffusion in in vivo applications, restricting accurate identification to just the most superficial contexts. Notably, due to their decreased size over traditional monoclonal antibodies, dye-labeled nanobodies specific to cell type receptors could aid in the identification of deeper cells.³⁸ Ultimately, the development of probes that easily diffuse through tissues and only get fluorescently activated when bound to specific cell surface receptors would push this area of exploration further. These probes would provide an invaluable biological ground truth to inform endogenous fluorescence and aid in its ability to answer specific biological questions by identifying specific cell types and their activation states.

2.3.

Label-Free Raman and Infrared Vibrational Spectroscopies

One of the key goals of optical microscopy in biology is to observe how biomolecules behave and interact in real time. Vibrational spectroscopy offers an exciting, label-free way to image the chemical makeup of cells. Although it is widely used in chemistry to identify substances, applying it to live-cell imaging is still a challenge. So far, it is mainly been used to detect broad categories of molecules such as lipids, nucleic acids, proteins, carotenoids, and porphyrin-containing proteins.^39–41

The difficulty lies in the nature of vibrational spectroscopy, which includes techniques such as Raman and mid-infrared (MIR) spectroscopy.⁴¹ Raman spectroscopy is limited by its extremely weak signal in which only 20 photons would be collected after a 10-s exposure. In contrast, fluorescent molecules absorb and emit light much more efficiently $\sim 10$ trillion times stronger than Raman scattering. Under the same conditions, fluorescence can collect around $>150$ photons in just $1\mu \mathrm{s}$, easily detecting individual molecules in real time. Such a weak signal limits both imaging speed and sensitivity to detecting changes in molecular constituents. Non-linear optical interactions that stimulate coherent Raman scattering (CRS) have been exploited to boost Raman imaging speed by $>{10}^6$ times—allowing video-rate CRS imaging of cells.^42,43 Despite these challenges, Raman microscopy has been shown to provide valuable information for stem cells and embryogenesis.⁴⁴

Although Raman spectroscopy is an impactful tool for specific applications, and highly specific, it is not yet widespread as a biomedical research tool, given its limited analytical performance in complex biosamples. Raman microscopy is an area where a continued community-driven effort to validate constituent signals amongst complex biological contexts could unlock significant gains in the application of unlabeled technologies.

MIR vibrational spectroscopy exploits dipole-allowed optical transitions, resulting in an optical response only 100 times weaker fluorescence.^45,46 However, MIR absorption spectroscopy comes with restraining challenges. First, the wavelength of infrared radiation in the MIR spectral range is $>10\times$ longer than visible wavelengths, which leads to poor spatial resolution, and is generally incapable of sub-cellular scale. Second, water is a strong and broad absorber in the MIR spectrum, which interferes with the molecular absorption signatures, hindering the ability for live cell analysis and deep tissue imaging.

The potential power of vibrational spectroscopic imaging motivates the development of new spectroscopic imaging methods to improve the speed and sensitivity of chemical imaging. Recent strategies were adopted to improve imaging speed for Raman spectroscopy using high-speed optical delay lines,^47,48 computational imaging,^49,50 and widefield imaging.⁵¹ Efforts to improve sensitivity in Raman detection have made use of novel detection strategies,^52–54 rear-resonant enhancement,^55,56 and photothermal detection^57,58 In addition, developments pushing technology for low-frequency Raman spectroscopy (with vibrational frequencies below $200{\mathrm{cm}}^{-1}$) offer the potential for studying mechanical soft materials such as biological tissues in a non-invasive manner.⁵⁴

Technological developments in infrared photonics have also advanced MIR spectrochemical detection. For example, efforts have focused on MIR signal enhancement with metasurfaces,⁵⁹ MIR photothermal imaging,^60,61 and combined MIR and fluorescent detection⁶² that allows for rapid imaging in live cells with visible or near-infrared detection—vastly improving spatial resolution. Future work with these new chemical imaging modalities will reveal the extent to which these recent advances expand our routine toolsets used for biological studies.

Vibrational spectroscopy has been applied to numerous biological systems, facilitating spatially resolved molecular fingerprint chemical imaging. However, the complexity of molecular composition in biological systems makes the interpretation of specific biomolecular interactions difficult to discern. This challenge lies in the fact that the molecular building blocks for biological macromolecules use a limited pallet for critical components such as nucleic acids, proteins, and lipids. Thus, it is critically important for the community to establish standards for chemical imaging through Raman or MIR spectroscopic measurements and validate the molecular information retrieved from the spectrochemical image datasets.

Here too is a place where fluorescent labels shine through with powerful capabilities because specificity in the labeling of molecules of interest facilitates detailed and specific studies. The combined use of fluorescent labels with Raman and MIR imaging could help establish the utility of chemical imaging in real-world biomedical applications. In addition, information from fluorescent labels could reveal additional ground truth information that, when coupled with spectroscopic imaging, could help test and validate unique molecular fingerprints in spectral signal analysis. Moreover, the fluorescent images can be used as ground-truth labels in supervised machine learning (ML) models, enabling the retrieval of quantitative and molecular-specific information from vibrational spectroscopic image datasets. Such capabilities will greatly benefit from sensitive spectral measurements capable of capturing accurate molecular information, as well as large datasets for reliable and unbiased training, testing, and validation of the ML models. As these ML models rely on large datasets, improved imaging speed is of great importance for various applications, such as high throughput screening (e.g., drug responses to patient-derived cells), whole-slide tissue imaging, flow cytometry, or studying the chemical processes of rapidly moving inter- and intra-cellular components such as exosomes and organelles.

In applications where fluorescent labels fall short in providing the specific molecular information to cross-correlate with the vibrational spectral data, multimodal imaging approaches can be implemented.^63–65 For example, in a recent work, second harmonic generation imaging was used in tandem with MIR chemical imaging to study tumor microenvironment in pancreatic tissue samples.⁶³ In this study, the unique specificity of SHG microscopy to fibrillar collagen was leveraged using the SHG images as ground-truth collagen labels in an ML model to identify spectral signatures of fibrous tissue regions. Such multimodal imaging investigations can help elucidate the chemical alterations that lead to the architectural changes in tissues associated with various pathologies. Similar label-free multimodal approaches can be used between the vibrational spectroscopies and multiphoton metabolic imaging techniques to decode the spectral cues associated with the metabolic states of cells.

The inherent differences in spatial resolution between fluorescence-based imaging techniques and vibrational spectroscopies, such as Raman or infrared, present unique challenges. For scenarios requiring higher spatial resolution or validation of subcellular signal sources, super-resolution microscopy or expansion microscopy offers complementary solutions. Expansion microscopy, in particular, enables physical magnification of the sample, facilitating the co-registration of spatially resolved fluorescence and vibrational data.