By combining sample-expansion hydrogel chemistry with label-free chemical-specific stimulated Raman scattering microscopy, the protocol describes how to achieve label-free super-resolution volumetric imaging in biological samples. With an additional machine learning image segmentation algorithm, protein-specific multi-component images in tissues without antibody labeling were obtained.
The universal utilization of fluorescence microscopy, especially super-resolution microscopy, has greatly advanced knowledge about modern biology. Conversely, the requirement of fluorophore labeling in fluorescent techniques poses significant challenges, such as photobleaching and non-uniform labeling of fluorescent probes and prolonged sample processing. In this protocol, the detailed working procedures of vibrational imaging of swelled tissue and analysis (VISTA) are presented. VISTA circumvents obstacles associated with fluorophores and achieves label-free super-resolution volumetric imaging in biological samples with spatial resolution down to 78 nm. The procedure is established by embedding cells and tissues in hydrogel, isotropically expanding the hydrogel sample hybrid, and visualizing endogenous protein distributions by vibrational imaging with stimulated Raman scattering microscopy. The method is demonstrated on both cells and mouse brain tissues. Highly correlative VISTA and immunofluorescence images were observed, validating the protein origin of imaging specificities. Exploiting such correlation, a machine learning-based image-segmentation algorithm was trained to achieve multi-component prediction of nuclei, blood vessels, neuronal cells, and dendrites from label-free mouse brain images. The procedure was further adapted to investigate pathological poly-glutamine (polyQ) aggregates in cells and amyloid-beta (Aβ) plaques in brain tissues with high throughput, justifying its potential for large-scale clinical samples.
The development of optical imaging methods has revolutionized the understanding of modern biology because they provide unprecedented spatial and temporal information of targets across different scales, from subcellular proteins to whole organs1. Among them, fluorescence microscopy is the most well-established, with a large palette of organic dyes with high extinction coefficients and quantum yields2, easy-to-use genetic-encoded fluorescent proteins3, and super-resolution methods such as STED, PALM, and STORM for imaging nanometer-scale structures4,5. In addition, recent advancements in sample engineering and preservation chemistry, which expand specimens embedded in swellable polymer hydrogels6,7,8, enable sub-diffraction limited resolution on conventional fluorescence microscopes. For instance, typical expansion microscopy (ExM) effectively enhances the image resolution by four times with fourfold isotropic sample expansion7.
Despite its advantages, super-resolution fluorescence microscopy shares limitations that originate from fluorophore labeling. First, photobleaching and inactivation of fluorophores compromise the capacity for repetitive and quantitative fluorescence evaluations. Photobleaching is an inevitable event when light keeps pumping electrons into electronically excited states9. Second, labeling the fluorophores to the desired targets is not always a straightforward task. For instance, immunostaining demands a long and laborious sample preparation process and hinders imaging throughput10. It could also introduce artifacts due to inhomogeneous antibody-labeling, especially deep inside tissues11. Moreover, proper labeling strategies that target fluorophores for the desired proteins might be underdeveloped. For example, extensive screenings were required to find effective antibodies for Aβ plaques12. Smaller organic dyes, such as Congo red, often have limited specificity, only staining the core of the Aβ plaque. It is, therefore, highly desirable to develop a label-free super-resolution modality that circumvents the drawbacks of fluorophore-labeling and provides complementary high-resolution imaging from cells to tissues, and even to large-scale human samples.
Raman microscopy provides label-free contrast for chemical-specific structures and maps out the distribution of otherwise invisible chemical bonds by looking at the excited vibrational transitions13. In particular, stimulated Raman scattering (SRS) imaging on label-free or tiny-labeled samples has been demonstrated to have similar speed and resolution to fluorescence microscopy14,15. For example, healthy brain region has been readily differentiated from tumor-infiltrated region in human and mouse tissues16,17. Aβ plaques were also clearly imaged by targeting protein CH3 vibration (2940 cm−1) and amide I (1660 cm−1) on a fresh-frozen brain slice without any labeling18. Raman scattering, therefore, offers robust label-free contrast that overcomes the limitations of fluorophores. The question then became how one can accomplish super-resolution capacity using Raman scattering, which could reveal nanoscale structural details and functional implications in biological samples.
Although extensive efforts have been made to achieve super-resolution for Raman microscopy with elegant optic instrumentations, the resolution enhancement on biological samples has been rather limited19,20,21. Here, based on the recent works22,23, we present a protocol that combines a sample-expansion strategy with stimulated Raman scattering for super-resolution label-free vibrational imaging, named Vibrational Imaging of Swelled Tissues and Analysis (VISTA). First, cells and tissues were embedded in hydrogel matrixes through an optimized protein-hydrogel hybridization protocol. The hydrogel tissue hybrids were then incubated in detergent-rich solutions for delipidation, followed by expansion in water. The expanded samples were then imaged by a regular SRS microscope by targeting CH3 vibrations from retained endogenous proteins. VISTA, owing to its label-free imaging feature, bypasses photobleaching and inhomogeneous labeling arising from fluorophore labeling, with much higher sample processing throughput. This is also the first sub-100 nm (down to 78 nm) label-free imaging reported. No additional optical instrumentation besides typical SRS setup22,24 is required, making it readily applicable. With correlative VISTA and immunofluorescence images, an established machine-learning image-segmentation algorithm was trained25,26 to generate protein-specific multiplex images from single-channel images. The method was further applied to investigate Aβ plaques in mouse brain tissues, providing a holistic image suited for sub-phenotyping based on the fine views of the plaque core and peripheral filaments surrounded by cell nuclei and blood vessels.
All animal procedures performed in this study were approved by the California Institute of Technology Institutional Animal Care and Use Committee (IACUC), and the protocol procedures complied with all relevant ethical regulations.
1. Preparation of stock solutions for fixation and sample expansion
2. Preparation of mammalian cell samples
3. Preparation of mouse brain samples
4. Hydrogel embedding, denaturation, and expansion of cell and tissue samples
5. Label-free imaging of endogenous protein distribution in expanded cell and tissue samples
6. Correlative VISTA and fluorescent imaging of immuno-labeled and expanded tissue samples
7. Construction, training, and validation of U-Net architecture
NOTE: Installation on Linux is recommended. A graphics card with >10 GB of RAM is required.
8. VISTA combined with U-Net predictions for protein-specific multiplexity in label-free images
After establishing the working principle of the imaging and analysis method, image registration was done to evaluate the expansion ratio and to ensure isotropic expansion during sample processing (Figure 1A,B). Both untreated and VISTA samples were imaged while targeting the bond vibration at 2940 cm−1, which originates from CH3 of endogenous proteins. In untreated samples, the protein-rich structures like nuclei were dark due to the overwhelming lipid content from surrounding tissues22 (Figure 1A). After sample processing that includes the delipidation treatment, the resulting image showed the same feature with an inversed contrast (Figure 1B). The shapes and the relative positions of nuclei and vessels were completely unaltered (Figure 1A,B; numbered structures), confirming that the treatment is an isotropic process. By comparing the sizes of the corresponding nuclei, it was concluded that the method achieves 3.4 times expansion in brain tissue samples as compared to untreated samples22,23.
Knowing the expansion ratio in brain tissue, VISTA can now resolve new features in the label-free SRS images that were previously unresolvable. Although actin and tubulin structures have been the gold standard for super-resolution demonstrations, the resolution improvements on actin and tubulin structures have been well-characterized by fluorescence-based sample-expansion strategies using similar hybridization chemistry28. Moreover, imaging specific actin/tubulin structures is less feasible with this technique because the signal comes from the total ensemble of endogenous proteins, where cytoskeleton structures like tubulins would not have sufficient contrast (signal-to-background ratio) to be clearly distinguished. Hence, we decided to pursue imaging other nanoscale structures. We showed that features can be captured from mouse cortexes down to 150 nm (Figure 1C,D). Based on the dispersive patterns around neuronal dendrites, the observed small structures are likely dendritic spine heads7, which have a size of 146 nm (Figure 1D). In addition, the method was used to image the fibrillar structures in Aβ plaques, which are believed to have a thickness around 100 nm29,30. Indeed, it was demonstrated that ~130 nm fibrillar structures can be resolved in a representative diffusive Aβ plaque using this method (Figure 1E,F).
As VISTA enables effective protein retention and protein imaging22, one can clearly distinguish protein-rich nucleoli in nuclei and the ribbon-like cytoskeleton structures in the cytosols of cultured HeLa cells (Figure 2A, arrowhead). The method was further applied to study poly-glutamine (polyQ) aggregates that are transiently expressed in mammalian cells (Figure 2B,C). The results confirmed that the aggregates, as an expectedly densely packed structure, were expanded isotropically by comparing the same aggregate structures before and after expansion across multiple replicate samples23. High-resolution structures that are absent/blurred in the normal-resolution SRS images were obtained using this method. The VISTA-aggregate images revealed fibril-like protrusions on the periphery of the polyQ aggregates and a hollow structure in the center (Figure 2B, arrowhead). The observation that protrusions seamlessly attach to cytosolic contents might suggest that aggregates engage with functional proteins in cytosol. In hindsight, the capacity to expand dense aggregates also becomes plausible because the fixation reagent formaldehyde and the hydrogel monomers acryl amide and sodium acrylates are all small molecules that can diffuse in and out of protein aggregates. Once the aggregate co-polymerizes with monomers into hydrogel, the expansion process should proceed as normal.
We then applied this method to mouse brain tissue to further extend its scope. Although tissue samples pose challenges like reduced permeability, increased thickness, and heterogeneous mechanical strength, the mouse brain samples were successfully imaged using this method (Figure 2D). Similar to cell samples, protein-rich structures including cell nuclei, blood vessels, and neuronal processes were observed (Figure 2D, arrowhead). The limitation of brain tissues is that only 3.4 times expansion was achieved, which makes the effective resolution in brain samples 99 nm22. We validated the structural origin of SRS signals by correlative dye and antibody staining, in which DAPI stains for nuclei and lectin stains for blood vessels (Figure 2E,F). Neuronal cell bodies and processes were also delineated by immunofluorescence from NeuN and MAP222. With the trained convolutional neural network (CNN) algorithm utilizing the correlative fluorescence images as ground truth, the single-channel images were then segmented into specific protein-structure channels for multiplex images22.
Finally, we aimed to interrogate pathologic Aβ plaques in the brains of 5xFAD mice, a well-known animal model for Alzheimer's disease31. After following the procedures, a 3-dimensional SRS image of amyloid plaques deposition in brain tissues was acquired (Figure 2G). Puncta with high protein concentrations were observed (Figure 2G, orange arrow), representing the core of the Aβ plaque. Such image also revealed the peripheral Aβ plaques (Figure 2G, magenta arrowhead), which are often neglected by conventional Congo red staining that only targets the Aβ core. When combined with the trained segmentation algorithm, the label-free image could be transformed into a target-specific multiplex image (Figure 2F) and can be performed jointly with immunofluorescence23 to study plaque-astrocyte and plaque-microglia microenvironment interactions32 in a comprehensive and high-throughput manner.
Figure 1: Sample expansion strategy enables super-resolution label-free imaging in mouse brain tissues. (A) An SRS image at CH3 frequency of mouse hippocampus. (B) A VISTA image in the same field of view of mouse hippocampus. Labeled area shows the corresponding features before and after treatments. (C) A VISTA image of a normal mouse cortex that shows finer features. Inset shows the region of interest. (D) Resolution quantification for the fine structure observed in the expanded sample. FWHM of 497 nm corresponds to an effective resolution of 146 nm with 3.4 times expansion. (E) A VISTA image of an amyloid-beta (Aβ) plaque in mouse brain tissues. Inset shows the enlarged region of interest. (F) Resolution quantification for the extrusion fiber structure of the expanded amyloid-beta plaque. FWHM of 442 nm corresponds to an effective resolution of 130 nm with 3.4 times expansion. Scale bars = 20 µm. Please click here to view a larger version of this figure.
Figure 2: Label-free super-resolution volumetric imaging in cells and tissues enabled by VISTA. (A) Volumetric image of a normal HeLa cell. Arrowhead: cytoskeleton-like structure. (B) Single z-slice image of a polyQ aggregate expressed in HeLa cells. Arrowhead: hollow structure and fibril extrusions. (C) Maximum intensity projections of x-y, x-z, and y-z directions show the volumetric view of the polyQ aggregate-containing cell. (D) Volumetric image of a coronal section of mouse brain. Arrowhead: neuronal processes. (E) Fluorescence image of nuclei (DAPI staining) at the same sample region shows 1 to 1 correlation with nuclei in the VISTA image. (F) Fluorescence image of blood vessels (anti-lectin) at the same sample region shows a 1 to 1 correlation with vessel structures in the VISTA image. (G) Volumetric image of Aβ (orange arrow) containing brain tissue. Pink arrowhead: peripheral Aβ plaque. (H) Multiplex image from (G), predicted by the trained image segmentation algorithm. v-congo red represents the core of the Aβ plaque; v-GLUT1 represents blood vessels; v-DAPI represents nuclei; v-peripheral plaque represents the Aβ plaque not stained by the Congo red dye. Scale bars = 10 µm. The length scale is in terms of distance before expansion (adjusted for different expansion ratios). Please click here to view a larger version of this figure.
In summary, we present the protocol for VISTA, which is a label-free modality to image protein-rich cellular and subcellular structures of cells and tissues. By targeting endogenous CH3 from proteins in hydrogel-embedded cell and tissues, the method achieves an effective imaging resolution down to 78 nm in biological samples and resolves minor extrusion in Huntingtin aggregates and fibrils in Aβ plaques. This technique is the first instance to report sub-100 nm resolution for label-free imaging modalities22. Compared to existing expansion methods6,7,8,28,33,34, the technique inherits the merit of label-free SRS imaging and, hence, is free from photobleaching, inactivation, or quenching caused by laser illuminations. In addition, as a label-free method, it circumvents the demanding, inefficient, and potentially artifact-causing antibody-labeling that is always involved in methods such as DISCO12,35 and ExM33,34 and, thus, offers high-throughput sample preparation and uniform imaging throughout tissues. To address the lack of multiplexity in the label-free approach, VISTA, implemented with a CNN-based image segmentation algorithm25, provides protein-specific multi-component images without any labels in brain tissues22. The method was further applied on 5xFAD mouse brains and enabled a holistic volumetric view of aggregates core and periphery fibrils, nuclei, and blood vessels23. We envision that VISTA would scale up well for larger samples such as primate or human brain slices and could, ultimately, be useful for clinical investigations.
There are three essential steps that ensure the successful implementation of this method. First, maximum protein retention in the hydrogel sample hybrid is crucial and required22. To achieve this goal, the fixation condition was modified to contain a high concentration of acrylamide28 and to replace the protein digestion procedure with high-concentration detergent delipidation that saves significant protein loss from protein digestion. The addition of AA quenches the intermolecular crosslinking of proteins and enables the isotropic expansion without protein digestions28. In a previous study, deuterated monomers were used to prove that the aliphatic CH bonds in acrylamide hydrogels only give rise to constant background22. Second, proper correlations between SRS and immuno-labeling and distinctions between different protein targets need to be established. As the method relies on image-segmentation algorithms to add multiplexity to monochromatic SRS images, crosstalk between different protein targets in immunofluorescence will significantly compromise the quality of images. We meticulously selected protein-rich structures that are obvious in SRS images and validated their corresponding immunofluorescence features. Third, before using the model to predict fluorescence patterns from new SRS data sets, the validity and reliability of the trained machine-learning model should be testified. Distinct features that are not included in the training sets will likely cause issues in prediction. If the prediction results are not satisfying, the user should try to include more data for training and avoid predicting patterns that are not included in the training sets. Pearson's correlations of the testing sets and validation sets should also be monitored to ensure the quality of the prediction22,23. It is suggested to have at least 100 corresponding image sets for training.
While the method has immense potential for biological studies, there are certain limitations awaiting creative solutions. First and foremost, the sensitivity needs further improvement. The detection limit of label-free simulated Raman scattering is in the low millimolar range, and the isotropic expansion of samples in three dimensions significantly dilutes chemical bonds and weakens the signal. We, hence, are limited to imaging the total ensemble of endogenous proteins, which lacks specificity and multiplexity. Combining VISTA with ultrasensitive SRS36 could possibly extend this to image low-abundance proteins and study aggregate structures and compositions at super-resolution level by targeting orthogonal chemical bonds37. Second, the current 3.4 times expansion ratio in brain tissues only gives moderate resolution improvement. Although we have already resolved minor extrusions in Aβ plaques that were previously indistinguishable, higher resolution is always desirable. In this case, innovations in protein-anchoring and hydrogel chemistry would greatly benefit. For example, different gel formulation could enable larger expansion ratios for even higher image resolution38,39,40. New procedures in sample processing would allow it to be applied with widely available FFPE histology samples38,41, making it even better suited for large-scale clinical studies.
The authors have nothing to disclose.
We acknowledge the Caltech Biological Imaging Facility for software support. L.W. acknowledges the support of the National Institutes of Health (NIH Director's New Innovator Award, DP2 GM140919-01), Amgen (Amgen Early Innovation Award), and the start-up funds from the California Institute of Technology.
1.0 M Tris pH 8 | Sigma-Aldrich | 648314 | |
16% Paraformaldehyde | Electron microscopy science | 15710 | diluted to 4% in PBS |
25x water immersion objective | Olympus | XLPLN25XWMP2 | NA 1.05 |
5XFAD Mice | Mutant Mouse Resource and Research Centers and the Jackson Laboratory | B6SJL-Tg (APPSwFlLon, PSEN1*M146L*L286 V) 6799Vas/Mmjax | Alzheimer brain |
60x water immersion objective | Olympus | UPLSAPO60XWIR | NA 1.2 |
Acrylamide | Sigma-Aldrich | A9099 | |
ammonium persulfate | Sigma-Aldrich | A3678 | |
anti-MAP2 | Cell Signaling Technology | 8707 | |
anti-NeuN | Cell Signaling Technology | 24307 | |
borosilicate coverslip #1.5 | Fisher Scientific | 1254581 | |
C57BL/6J Mice | Jackson Laboratory (JAX) | 664 | Normal mice |
D2O | Sigma-Aldrich | 151882 | for SRS calibration |
DAPI | Thermo Fisher | D1306 | |
DMEM | GIBCO | 10566-016 | |
FBS | GIBCO | A4766 | |
glass slide 3" x 1" x 1 mm | VWR | 16004-430 | |
goat anti-chicken IgY, Alexa Fluor 647 | Invitrogen | A-21449 | |
goat anti-mouse IgG, Alexa Fluor 647 | Invitrogen | A-21236 | |
goat anti-rabbit IgG, Alexa Fluor 488 | Invitrogen | A-11034 | |
goat anti-rat IgG, Alexa Fluor 568 | Invitrogen | A-11077 | |
Grace Bio-Labs Press-To-Seal silicone isolators | Sigma-Aldrich | GBL664108 | microscope spacer |
Htt-97Q-GFP Plasmid | Gift from Prof. R. Kopito and Prof. F.-U.Hartl. | ||
Laser scanning microscope | Olympus | FV3000 | laser scanning confocal microscope |
lipofectamine 3000 | Thermo Fisher | L3000001 | transfection agent |
Lycopersicon Esculentum Lectin DyLight®594 (lectin) | Vector Laboratories | DL-1177-1 | |
Microscope spacer | Grace Bio-Labs | 621502 | |
N,N′-methylenebisacrylamide (BIS) | Sigma-Aldrich | M1533 | bought as 2% solution in water |
Nuclease free water | Thermo Fisher | 10977-015 | |
Penicillin-Streptomycin | GIBCO | 15140-122 | |
poly-strene beads | Sigma-Aldrich | 43302 | for resolution characterization |
Sodium Acrylate | Sigma-Aldrich | 408220 | |
sodium dodecyl sulfate | Sigma-Aldrich | 71725 | |
soft-wool paint brush #3 | TANIS | 000333 | |
SRS Laser | A.P.E | picoEmerald | 2ps pulse width |
tetramethylethylenediamine | Sigma-Aldrich | T9281 | |
Tissue culture flask 25 cm2 | Corning | 430639 | |
Triton X-100 | Sigma-Aldrich | T8787 | |
Tween-20 | Sigma-Aldrich | P9416 | |
tweezer | Fine Science Tool | 11295-51 | |
Vibrotome | Leica | VT1200S | the vibratome |
.