Sample preparations
Animals
Wild-type mice were purchased from Jackson laboratories (Bar Harbor, ME, C57BL6/J; stock #000664). All mice were housed in a temperature-controlled environment with a 12:12 light/dark cycle and ad libitum access to standard laboratory chow and water. Euthanasia was performed by rapid cervical dislocation. All experimental animal procedures were approved by the JHU Institutional Animal Care and Use Committee.
RNAscope single molecule fluorescent in situ hybridization (smFISH) in cultured neurons
Mouse cortical neurons were cultured on a 96-well ibidi optical bottom culture plate as previously described [33, 34]. Briefly, following rapid cervical dislocation of timed-pregnant dams, embryos were removed and cortices dissected and incubated with papain (Worthington, catalog number LK003150). The tissue was triturated to obtain single cells, which were plated in FBS-containing media for 24 h. Following complete media replacement on day in vitro (DIV) 1, partial media exchanges were performed every 3 days. In situ hybridization assays were performed with RNAscope technology using the RNAscope Fluorescent Multiplex kit V2 and 4-plex Ancillary Kit (catalog numbers 323100, 323120 ACD, Hayward, CA) as above, with the exception of the protease step, where cultured cells were treated with a 1:15 dilution of protease III for 30 min. Cells were incubated with probes for Arc, Bdnf exon 1, Bdnf exon 4, and Fos (catalog number 316911, 457321-c2, 482981-c3 and 316921-c4, ACD, Hayward, CA) and stored overnight in a 4× saline sodium citrate (SSC) buffer. After amplification, probes were fluorescently labeled with Opal dyes (PerkinElmer; Opal520 was diluted 1:500 and assigned to Fos, Opal570 was diluted 1:500 and assigned to Bdnf exon 1, Opal620 was diluted 1:500 and assigned to Bdnf exon4 and Opal690 was diluted 1:500 and assigned to Arc) and stained with DAPI (4′,6-diamidino-2-phenylindole) to label nuclei, then stored in phosphate buffered saline (PBS) at 4 °C.
RNAscope smFISH in mouse brain tissue
Mice were sacrificed by rapid cervical dislocation as previously described, to ensure that transcription of activity-regulated gene programs were not triggered by anesthesia [17]. Mouse brain was extracted and rapidly frozen in 2-methylbutane (ThermoFisher), and stored at − 80 °C until slicing. Sixteen micrometer coronal sections were prepared using a Leica CM 1520 Cryostat (Leica Biosystems, Buffalo Grove, IL) and mounted onto glass slides (VWR, SuperFrost Plus). RNAscope was performed using the Fluorescent Multiplex Kit V2 (Cat # 323100, 323120 ACD, Hayward, California) according to manufacturer’s instructions as previously described [17]. Sections were incubated with specific probes targeting Gal, Th, Bdnf exon 9, and Npy (Cat # 400961, 317621-C2, 482981-C3, 313321-C4, ACD, Hayward, California) and were incubated at 40 °C with a series of fluorescent Opal Dyes (Perkin Elmer; Opal690 diluted at 1:500 and assigned to Npy; Opal570 diluted at 1:500 and assigned to Th; Opal620 diluted at 1:500 and assigned to Bdnf; Opal520 diluted at 1:500 and assigned to Gal). DAPI was used to label nuclei and slides were coverslipped with FluoroGold (SouthernBiotech).
Immunofluorescence staining in post-mortem human Alzheimer’s disease (AD) brain
Post-mortem human brain tissue was obtained by autopsy from the Offices of the Chief Medical Examiner of Maryland, all with informed consent from the legal next of kin collected under State of Maryland Department of Health and Mental Hygiene Protocol 12–24. Clinical characterization, diagnoses, and macro- and microscopic neuropathological examinations were performed on all samples using a standardized paradigm. Details of tissue acquisition, handling, processing, dissection, clinical characterization, diagnoses, neuropathological examinations, and quality control measures have been described previously [35]. Alzheimer’s disease diagnosis comprise standard neuropathology ratings of Braak staging schema [36] evaluating neurofibrillary tangle burden, and the CERAD scoring measure of senile plaque burden [37]. An Alzheimer’s likelihood diagnosis was then performed based on the published consensus recommendations for post-mortem diagnosis of Alzheimer’s disease [2] as with prior publications [38, 39].
Fresh frozen inferior temporal cortex from a donor with clinically confirmed Alzheimer’s disease (AD) was sectioned at 10 μm and stored at − 80 °C. Immunofluorescence staining was performed following a demonstrated protocol provided by 10× Genomics available online (CG000312, 10X Genomics, Pleasanton, California). Briefly, slides were thawed for 1 min at 37 °C and fixed with pre-chilled methanol (Cat #34860, Sigma-Aldrich, St. Louis, Missouri) for 30 min at − 20 °C. Sections were blocked with Human TruStain FcX (Cat #422301, Biolegend, San Diego, California) and 2% BSA (Cat #130-091-376, Miltenyi Biotec, Auburn, California) diluted in Blocking Buffer for 5 min at room temperature (RT). Primary antibodies were added in Antibody Diluent (3× SSC, 2% BSA and 0.1% TritonX-100 in nuclease free water) and incubated for 30 min at RT. All primary antibodies were diluted from each stock solution at a concentration of 1:100: mouse anti-beta-amyloid (Cat #803001, Biolegend, San Diego, California), rabbit anti-pTau Ser202/Thr205 (Cat # SMC-601, StressMarq Biosciences, Cadboro Bay, Victory, Canada), and chicken anti-MAP2 (Cat #ab92434, Abcam, Cambridge, Massachusetts). The slides were subjected to 5 subsequent washes, each of which takes 30 s with Wash Buffer (3× SSC, 2% BSA and 0.1% TritonX-100 in nuclease free water). The tissue sections were then incubated with corresponding fluorescently labeled secondary antibodies diluted from each stock solution at a concentration of 1:500 for 30 min at RT. All secondary antibodies were purchased from Thermo Fisher Scientific (Waltham, Massachusetts): goat anti-mouse IgG (H + L) conjugated to Alexa Fluor 488 (Cat #A-11001), donkey anti-rabbit IgG (H + L) conjugated to Alexa Fluor 555 (Cat #A-31572), and goat anti-chicken IgY (H + L) conjugated to Alexa Fluor 633. DAPI was added to visualize the nuclei. After 5 washes with Wash Buffer, which takes 30 s for every round, and subsequent 20 quick immersions in 3× SSC (Millipore-Sigma, S6639L, St. Louis, Missouri), slides were coverslipped in 85% glycerol and stored at 4 °C.
RNAscope smFISH in post-mortem human dorsolateral prefrontal cortex (DLPFC)
Fresh frozen dorsolateral prefrontal cortex (DLPFC) samples from 2 healthy individuals were sectioned as previously described [17]. Single molecule fluorescent in situ hybridization assays were performed with RNAscope Fluorescent Multiplex Kit V2 and 4-plex Ancillary Kit (Cat # 323100, 323120 ACD, Hayward, California) according to manufacturer’s instructions. Tissue sections were incubated with probes for SNAP25, SLC17A7, GAD1, and MBP (Cat #518851, 415611-C2, 573061-C3, 573051-C4, ACD, Hayward, California) and labeled with Opal Dyes (Perkin Elmer, Waltham, MA; Opal690 at 1:1000 for SNAP25; Opal570 at 1:1500 for SLC17A7; Opal620 at 1:500 for GAD1; Opal520 at 1:1500 for MBP) and stained with DAPI to label the nucleus [40].
Fluorescent imaging
Using a Zeiss LSM780 confocal microscope equipped with 20× (0.8 NA) and 63× (1.4NA) objectives, a GaAsP spectral detector, and 405, 488, 561, and 633 lasers, lambda stacks were acquired in z-series with the same settings and laser power intensities. Stacks were linearly unmixed in ZEN software using previously created reference emission spectral profiles [17] and saved as Carl Zeiss Image “.czi” files to retain image metadata. Raw lambda stacks were unmixed with SUFI and compared to ZEN unmixed results. Single-fluorophore positive fingerprints were generated from samples prepared as above.
Reference spectral profile creation in ZEN software for validation
Reference emission spectral profiles, referred to as ‘fingerprints’ or ‘endmembers’, were created for each Opal dye in ZEN software and validated for specificity as previously described [17]. Briefly, a control probe against the housekeeping gene Polr2a was used to generate four single positive slides in mouse brain tissue according to manufacturer’s instructions. Mouse tissue was used for the absence of confounding lipofuscin signals and therefore lower tissue autofluorescence. Polr2a was labeled with either Opal520, Opal570, Opal620, or Opal690 dye to generate single positive slides. For DAPI, a single positive slide was generated using identical pre-treatment conditions without probe hybridization. To create a fingerprint for lipofuscin autofluorescence, a negative control slide was generated using a 4-plex negative control probe against four bacterial genes (Cat #321831, ACD, Hayward, CA) in DLPFC tissue. All Opal dyes were applied to the slide, but no probe signal was amplified due to the absence of bacterial gene expression. Within a field of view, a single pure region of interest (i.e. any isolated strong puncta for Polr2a slides) was manually selected with the crosshair tool in ZEN software to generate a spectral reference profile.
In a similar approach, reference emission spectral profiles were generated for immunofluorescent staining in post-mortem human AD brain tissue. For amyloid plaques and tau tangles, each single positive slide was prepared by labeling β-amyloid (Abeta) or phospho-tau (pTau) with appropriate primary and secondary antibodies conjugated with Alexa fluor (AF) 488 and AF555, respectively. A lipofuscin fingerprint was created for human AD brain tissue using a negative control slide treated only with fluorescently labeled secondary antibodies in the absence of primary antibodies. For DAPI-stained nuclei and MAP2-positive neurites (labeled with AF633), single positive slides were generated using mouse brain tissue to avoid lipofuscin autofluorescence.
Spectral unmixing
Spectral unmixing is the process of decomposing composite multichannel images into spectral profiles and abundances of each endmember in each pixel [2, 41, 42]:
$$\left[\begin{array}{ccc}{F}_{(\mathrm{1,1})}& \cdots & {F}_{(1,n)}\\ \vdots & \ddots & \vdots \\ {F}_{(C,1)}& \cdots & {F}_{(C,n)}\end{array}\right]= \left[\begin{array}{ccc}{S}_{(\mathrm{1,1})}& \cdots & {S}_{(1,k)}\\ \vdots & \ddots & \vdots \\ {S}_{(C,1)}& \cdots & {S}_{(C,k)}\end{array}\right]\left[\begin{array}{ccc}{A}_{(\mathrm{1,1})}& \cdots & {A}_{(1,n)}\\ \vdots & \ddots & \vdots \\ {A}_{(k,1)}& \cdots & {A}_{(k,n)}\end{array}\right]$$
(1)
Which can be denoted as F = SA.
In Eq. (1), F denotes the fluorescence intensities of n pixels recorded in C different spectral channels. S is the spectral signatures of k fluorophores, and A is the abundance of each fluorophore in each pixel. To this end, the unmixing process is usually divided into three different steps: (i) estimation of the number of endmembers, (ii) extraction of endmembers, (iii) estimation of abundance. In fluorescence microscopy, the number of endmembers is known in advance. We discuss the latter two steps below.
Automated extraction of spectral signatures
An essential part of the proposed pipeline is the automated extraction of the spectral signatures, or ‘endmembers,’ from the observed multispectral image. To achieve this, we use the Vector Component Analysis (VCA)—an Endmember Extraction Algorithm [32] that can be used to extract fingerprints (i.e. spectral signatures) from multiplex lambda stacks. We approach the extraction of fingerprints in two different ways, (i) Using lambda stacks acquired from the single positive slides discussed above to extract fingerprints for individual fluorophores. (ii) Using a multiplex lambda stack and extracting fingerprints for all fluorophores in one go. We discuss the pros and cons of each method and provide additional details in the Results section.
Estimation of abundance
This step involves the estimation of the proportion of different fluorophores in each pixel. Here we implement and compare three different methods derived from remote sensing and adapt them for unmixing in fluorescence microscopy: (i) fully constrained least square unmixing (FCLSU) algorithm [43] tries to minimize the squared error in the linear approximation of multispectral image, imposing the non-negative constraint and the sum-to-one constraint for the abundance calculations. (ii) extended linear mixing model (ELMM) algorithm [30] extends the idea of FCLSU unmixing by taking into account the spectral variability—particularly, scaling of reference spectra. (iii) generalized extended linear mixing model (GELMM) algorithm [31] extends ELMM to account for complex spectral distortions where different wavelength recordings are affected unevenly.
SUFI toolbox
SUFI is a MATLAB-based command line toolbox for automated spectral unmixing of fluorescent images. Briefly, the analysis pipeline involves data normalization, automated extraction of spectral signatures using VCA algorithm, and application of spectral unmixing algorithms (Fig. 1). Bio-formats toolbox ‘bfmatlab’, which is compatible with images acquired on multiple microscope systems, is used to read the image data into a MATLAB structure with fields containing fluorescent channels, DAPI and lipofuscin. SUFI toolbox is publicly available at https://github.com/LieberInstitute/SUFI.
Performance metrics
The root mean squared error (RMSE) between a true image (yref), i.e. ZEN unmixed image and its estimate (yest) i.e. FCLSU (or ELMM or GELMM) unmixed image is defined as,
$$RMSE= \sqrt{\sum \frac{{({y}_{est}- {y}_{ref})}^{2}}{N}}$$
(2)
The structural similarity index (SSIM) is based on the computation of luminance, contrast and structure of true image vs. estimated image [44]. The range of values are between [0, 1] with a value of SSIM = 1 indicating 100 percent structural similarity.
$$SSIM= {\left[l({y}_{est}, {y}_{ref})\right]}^{\alpha }\cdot {\left[c({y}_{est}, {y}_{ref})\right]}^{\beta }\cdot {\left[s({y}_{est}, {y}_{ref})\right]}^{\gamma }$$
(3)
$$l\left({y}_{est}, {y}_{ref}\right)= \frac{2{\mu }_{{y}_{est}}{\mu }_{{y}_{ref}}+ {C}_{1}}{{\mu }_{{y}_{est}}^{2}+ {\mu }_{{y}_{ref}}^{2}+ {C}_{1}}$$
(4)
$$c\left({y}_{est}, {y}_{ref}\right)= \frac{2{\sigma }_{{y}_{est}}{\sigma }_{{y}_{ref}}+ {C}_{2}}{{\sigma }_{{y}_{est}}^{2}+ {\sigma }_{{y}_{ref}}^{2}+ {C}_{2}}$$
(5)
$$s\left({y}_{est}, {y}_{ref}\right)= \frac{{\sigma }_{{y}_{est}{y}_{ref}}+ {C}_{3}}{{\sigma }_{{y}_{est}}{\sigma }_{{y}_{ref}}+ {C}_{3}}$$
(6)
where \(\mu_{{y_{est} }} ,\mu_{{y_{ref} }} ,\sigma_{{y_{est} }} ,\sigma_{{y_{ref} }}\), and \(\sigma_{{y_{est} , y_{ref} }}\) are the local means, standard deviations, and cross-covariance for images\({y}_{est}, {y}_{ref}\). C1, C2, C3 are constants.
The Sørensen–Dice Similarity coefficient (DICE) ranges between [0, 1] where a value of DICE = 1 indicates a 100 percent match of segmentation between two images.
$$DICE= \frac{2* \left|intersection({y}_{est}, {y}_{ref})\right|}{\left|{y}_{est}\right|+\left|{y}_{ref}\right|}$$
(7)
where \(\left|{y}_{est}\right|\) represents the cardinal of \({y}_{est}\).
Data segmentation with dotdotdot
Using previously published MATLAB scripts [17], we automatically segment and quantify nuclei and RNA transcripts using SUFI generated unmixed outputs. Briefly, the dotdotdot processing pipeline involves smoothing, thresholding, watershed segmentation, autofluorescence masking, and dot metrics extraction. Specifically, adaptive 3D segmentation is performed on image stacks using the CellSegm MATLAB toolbox, and nuclei are further separated using the DAPI channel and 3D watershed function. Single dots are detected using histogram-based thresholding and assigned to nuclei based on their 3D location in the image stack. Lipofuscin signal is used as a mask to remove pixels confounded by autofluorescence.