Contributed by Xiaowei Zhuang, October 25, 2016 (sent for review October 10, 2016; reviewed by Gaudenz Danuser and Taekjip Ha)
Multiplexed single-molecule FISH allows spatially resolved gene-expression profiling in single cells. However, because of off-target binding of FISH probes and cellular autofluorescence, background can become limiting in multiplexed single-molecule FISH measurements, especially when tissue samples are imaged or when the degree of multiplexing is increased. Here we report a sample clearing approach for FISH that substantially reduced these background sources by anchoring RNAs to a polymer matrix and then removing proteins and lipids. This approach allows measurements with higher detection efficiency and sensitivity across more color channels in both cell culture and tissue with no detectable loss in RNA. We anticipate that this clearing approach will greatly facilitate applications of multiplexed FISH measurements in a wide variety of biological systems.
Highly multiplexed single-molecule FISH has emerged as a promising approach to spatially resolved single-cell transcriptomics because of its ability to directly image and profile numerous RNA species in their native cellular context. However, background—from off-target binding of FISH probes and cellular autofluorescence—can become limiting in a number of important applications, such as increasing the degree of multiplexing, imaging shorter RNAs, and imaging tissue samples. Here, we developed a sample clearing approach for FISH measurements. We identified off-target binding of FISH probes to cellular components other than RNA, such as proteins, as a major source of background. To remove this source of background, we embedded samples in polyacrylamide, anchored RNAs to this polyacrylamide matrix, and cleared cellular proteins and lipids, which are also sources of autofluorescence. To demonstrate the efficacy of this approach, we measured the copy number of 130 RNA species in cleared samples using multiplexed error-robust FISH (MERFISH). We observed a reduction both in the background because of off-target probe binding and in the cellular autofluorescence without detectable loss in RNA. This process led to an improved detection efficiency and detection limit of MERFISH, and an increased measurement throughput via extension of MERFISH into four color channels. We further demonstrated MERFISH measurements of complex tissue samples from the mouse brain using this matrix-imprinting and -clearing approach. We envision that this method will improve the performance of a wide range of in situ hybridization-based techniques in both cell culture and tissues.
Single-molecule FISH (smFISH) is a powerful technique that allows the direct imaging of individual RNA molecules within single cells (1, 2). In this approach, RNAs are labeled via the hybridization of fluorescently labeled oligonucleotide probes, producing bright fluorescent spots for single RNA molecules, which reveal both the abundance and the spatial distribution of these RNAs inside cells (1, 2). The ability of smFISH to image gene expression at the single-cell level in both cell culture and tissue has led to exciting advances in our understanding of the natural noise in gene expression and its role in cellular response (3, 4), the intracellular spatial organization of RNAs and its role in posttranscriptional regulation (5, 6), and the spatial variation in gene expression within complex tissues and its role in the molecular definition of cell types and tissue functions (6, 7).
To extend the benefits of this technique to systems-level questions and high-throughput gene-expression profiling, approaches to increase the multiplexing of smFISH (i.e., the number of different RNA species that can be simultaneously quantified within the same cell) have been developed (8⇓⇓⇓⇓–13). Most of these approaches take advantage of color multiplexing, which has allowed a few tens of RNA species to be imaged simultaneously. We have recently introduced multiplexed error-robust FISH (MERFISH), a massively multiplexed form of smFISH that allows RNA imaging and profiling at the transcriptomic scale (13, 14). MERFISH achieves this level of multiplexing by assigning error-robust barcodes to individual RNA species, labeling RNAs combinatorically with oligonucleotide probes that contain a representation of these barcodes, and reading out these barcodes through sequential rounds of single-color (13) or multicolor (14) smFISH imaging (Fig. S1). Using this approach, we have imaged 140 and 1,000 RNA species in individual cells with two different encoding schemes, one of which allows both error detection and correction and the other that allows error detection (13), and the number of addressable RNA species can be further changed by using different encoding schemes. Recently, we have increased the measurement throughput of MERFISH and demonstrated the ability to profile gene expression in tens of thousands of cells in a single day-long measurement (14). A different multiplexed smFISH method using color-based barcodes and sequential imaging (seqFISH) has been independently proposed and initially demonstrated with measurements of 12 RNA species in individual cells (12). While this current paper was in review, a paper reporting an extension of seqFISH that adds the error-correction capability and demonstrates the capability of imaging 125 or 250 RNA species was published (15).
Schematic diagram of MERFISH. (A) Illustration of a barcoding process used by MERFISH to identify RNAs. In this implementation of MERFISH, each individual RNA species is assigned a unique binary barcode. This barcode is then read out through a series of smFISH staining and imaging rounds. Each smFISH image is associated with a specific bit in the binary barcode, and only a subset of the targeted RNAs is labeled such that it will fluoresce in each image. If an RNA fluoresces in a given image, then it is assigned a “1” in the corresponding bit. If it does not, then it is assigned a “0.” In this fashion, the specific on/off pattern of fluorescence across n smFISH images is used to construct a binary barcode for each measured RNA in the sample, and this measured binary barcode is then used to decode the identity of that RNA: for example, A, B, C. (B) Schematic depiction of the design of MERFISH probes used in this work. Individual RNAs are labeled with multiple “encoding” oligonucleotide probes. These encoding probes contain a central target region that has a sequence complementary to a portion of the RNA to which the encoding probes are targeted. This sequence is flanked by multiple readout sequences. These readout sequences are custom-designed, 20-nt sequences, and there is one unique readout sequence associated with each bit in the barcodes. If an RNA species is assigned a barcode with a “1” in a given bit, then the readout sequence associated with that bit will be contained within the encoding probes that target that RNA; thus, the set of readout sequences associated with each RNA define its barcode. In the MHD4 code used in this work to encode RNAs, each valid 16-bit barcode contains only four “1” bits and, hence, the set of encoding probes targeting each RNA together contain four different readout sequences. To limit the length of the encoding probes, we randomly chose three of the four readout sequences to be associated with each encoding probe. After staining the RNAs with encoding probes, the barcodes associated with the RNAs are then measured by a series of hybridizations with fluorescently labeled readout probes, each complementary to a readout sequence. (C) Schematic depiction of the MERFISH readout process used here. During each round of readout hybridization, one or more readout probes are added to the sample. Multiple different readout probes can be hybridized to the sample simultaneously if each is conjugated to a spectrally distinct dye, allowing multiple bits to be read out simultaneously. Illustrated here is a two-color readout scheme with the two distinct dyes marked as red and blue circles. The sample is imaged in two color channels and the presence or absence of a fluorescent spot determines if the corresponding readout sequence is present and, thus, if the barcode associated with each RNA copy has a “1” or a “0” in the corresponding bit. To remove the fluorescent signal before the next round of hybridization and imaging, a disulfide bond linking the fluorophores to the readout probes is reductively cleaved and the free fluorophores are washed away. The sample is then restained with a different set of readout probes and the process repeated to read out the remaining bits in the barcodes.
smFISH measurements typically benefit from high signal-to-background ratios, resulting in the detection of individual RNA molecules with high accuracy and detection efficiency (1, 2): In many cases, the bright fluorescent signals that arise from the tens of fluorescently labeled probes bound to each copy of an RNA substantially exceed the background that arises from probes binding off target or from cellular autofluorescence. However, we have observed that as the degree of multiplexing is increased, the background level also tends to increase. The resulting decrease in the signal-to-background ratio makes a number of important applications and extensions of multiplexed smFISH challenging. For example, efforts to further increase the degree of multiplexing, to thousands or potentially tens of thousands of RNAs, will likely be limited by increased background. In addition, many RNAs are not long enough to accommodate tens of oligonucleotide probes, limiting the ability to measure relatively short RNAs and to discriminate many different RNA isoforms. Finally, background is typically more pronounced in complex tissues, making multiplexed smFISH measurements in tissues more challenging.
Here we report a sample clearing approach aimed at improving the signal-to-background ratio in RNA FISH measurements by substantially reducing background fluorescence signal. Many modern tissue-clearing approaches are designed to preserve the protein content of the sample while reducing scattering and autofluorescence background by extracting lipids and matching refractive index (16⇓⇓⇓⇓⇓⇓–23). For example, embedding and cross-linking tissues to hydrogels provides a powerful approach to tissue clearing, minimizing sample distortion during lipid removal and index matching while maintaining the protein content of the sample (19, 20). These approaches have also been made compatible with RNA FISH by stabilizing RNA molecules, for example through cross-linking of RNAs to proteins, without removing the protein content of the cell (20, 24). Here we show that a major source of background in RNA FISH measurements is the nonspecific binding of FISH probes to cellular components other than RNAs, such as proteins. For this reason, a clearing method that preserves RNAs while removing proteins and lipids is desired for RNA FISH imaging. In the recently developed expansion microscopy method, proteins (25) and, more recently, RNAs (26, 27) are physically anchored to a solvent-expandable and clearable poly-electrolyte matrix, effectively imprinting signals of these components on this matrix and allowing these molecular signals to be expanded along with the matrix for increasing image resolution. Inspired by this approach, we anchored RNA molecules to a nonswellable polyacrylamide (PA) matrix and then removed unwanted, non-RNA components, such as proteins and lipids, with the aim to remove their contribution to background fluorescence. We demonstrated that this matrix-imprinting and -clearing approach substantially reduced the background because of off-target binding of FISH probes and cellular autofluorescence. By comparing the copy number of 130 RNAs measured via MERFISH in uncleared and cleared cultures of human cells, we demonstrated that this matrix-imprinting–based clearing approach improves the detection efficiency and detection limit of MERFISH with no detectable loss in RNAs. Moreover, the reduction in autofluorescence, in particular in the blue–green spectral range, facilitated extension of MERFISH imaging from two to four distinct color channels with no reduction in performance. This improvement reduced the number of hybridization rounds needed for MERFISH measurements, which should further increase the MERFISH measurement speed and throughput. Finally, we demonstrated that this clearing approach substantially reduces the background in tissue, facilitating high-performance MERFISH measurements in cryosections of adult mouse brain tissues. Given the simplicity and efficacy of this matrix-imprinting–based clearing method, we envision that this approach could be used to substantially improve the performance of a wide range of in situ hybridization methods for both RNA and DNA in cell cultures and tissues.
Our first step in the development of a sample clearing method for smFISH was to determine the physical origin of off-target binding of oligonucleotide probes: are these probes binding to the incorrect RNA or other cellular components, such as proteins or lipids? To address this question, we stained human lung fibroblast (IMR-90) cells using FISH probes targeting the Filamin A (FLNA) mRNA. As expected, we observed both bright fluorescence spots marking individual molecules of FLNA mRNA (Fig. 1A, Left) and a diffuse background because of off-target probe binding (Fig. 1A, Center) that was not present in samples not stained with FISH probes (Fig. S2). We then measured the RNase sensitivity of both the foreground RNA spots and the diffuse background, reasoning that if the background arose from off-target binding to incorrect RNAs, both the foreground spots and the background should be RNase-sensitive. We found that a brief RNase A treatment completely removed the bright foreground spots, but produced little if any reduction in the background (Fig. 1A, Right). Thus, we conclude that the vast majority of off-targeting binding of smFISH probes arose from binding to cellular components other than RNA, such as proteins and lipids.
Matrix imprinting and clearing reduces background in smFISH measurements. (A) A human fibroblast cell (IMR-90) stained with smFISH probes targeting the FLNA mRNA before (Left and Center) and after (Right) treatment with RNase A. The contrast of the Center and Right panels has been increased fivefold from that of the Left panel to better visualize the background from probes bound off-target. (Scale bars, 10 µm.) (B) Schematic diagram of a matrix-imprinting and -clearing approach to reduce background in smFISH measurements. Cells are stained with smFISH probes or encoding probes for MERFISH measurements, and a poly-dT anchor probe, which targets the polyA tail of mRNAs. Cells are then embedded in a PA matrix, to which the poly-dT anchor probes are covalently linked via a terminal acrydite moiety. Proteins and lipids are then digested and extracted, freeing off-target bound smFISH probes to diffuse out of the PA matrix and removing cellular components that contribute to autofluorescence. (C) U-2 OS cells labeled with MERFISH-encoding probes targeting 130 RNAs followed by staining with a readout probe conjugated to Cy5 that binds to the encoding probes in an uncleared sample (Upper) and a sample treated with the matrix-imprinting and -clearing protocol (Lower). (Scale bars, 20 µm.)
Off-target binding of FISH probes is largely insensitive to RNase treatment. Images of different background sources in IMR-90 cells: cells stained with encoding probes but no fluorescently labeled readout probe (Left), cells stained with a fluorescently labeled readout probe but no encoding probes (Center), and cells stained with encoding probes, a fluorescently labeled readout probe that can bind to a readout sequence on these encoding probes, and then treated with RNase A to remove all specific RNA signals (Right). All three images are displayed at the same contrast to illustrate the relative intensity of the signal from the autofluorescence background of the cell (Left), the very low level (if any) of nonspecific binding of readout probes, and the signal from the off-target (RNase-insensitive) binding of encoding probes followed by binding of the readout probes to the encoding probes. The encoding probes used here target the FLNA mRNA only, and the readout probe used here is the Bit-1 readout probe conjugated to Cy5 (Table S1). (Scale bars, 5 µm.)
Hence, we reasoned that one way to reduce the background would be to remove the non-RNA components, such as proteins and lipids, from the sample. Moreover, because these components are also a major source of autofluorescence, the autofluorescence background might be reduced by such an approach as well. To this end, we fixed the sample and hybridized it with oligonucleotide probes as in standard smFISH (1, 2) or MERFISH measurements (13, 14), and then embedded the sample in an inert, nonfluorescence matrix to which RNA molecules were anchored, effectively imprinting the desired RNA signal onto this matrix (Fig. 1B). Once RNAs were anchored, cellular proteins and lipids were removed (for example, by digestion and extraction) without, in principle, affecting the number and localization of RNAs within the sample. smFISH probes bound off-target to these components should then be free to diffuse from the matrix. We used PA as the inert matrix and a 15-nt poly-dT oligonucleotide to bind and anchor polyadenylated (polyA) RNAs to the PA matrix, although other methods may also be used to link RNAs to the matrix. This anchor probe was comprised of 50% locked-nucleic acid bases to stabilize the hybridization to polyA tails of the RNAs (28) and additionally contained a terminal acrydite moeity, which can be covalently incorporated into the PA matrix as it polymerizes.
To test whether this clearing approach led to a reduction in off-target binding, we first measured the efficacy of protein and lipid removal and observed that this protocol efficiently removed cellular proteins and lipids from embedded cultured human osteosarcoma (U-2 OS) cells (Fig. S3). Next, we performed labeling as in MERFISH experiments and tested whether off-target probe binding was indeed reduced by clearing. In a MERFISH measurement, we typically stain cells first with a complex library of “encoding” oligonucleotide probes (13, 14). These encoding probes are not themselves fluorescently labeled. Instead, each encoding probe contains a targeting sequence that directs its binding to a cellular RNA and multiple readout sequences. Multiple encoding probes are targeted to each RNA, and the set of readout sequences contained within these encoding probes form a specific barcode that is unique to that RNA species (Fig. S1 B and C). These barcodes are then measured in a series of hybridizations, each round of hybridization using either one fluorescently labeled “readout” probe complementary to a specific readout sequence, reading out one bit per round in the single-color imaging mode (13), or multiple readout probes labeled with spectrally distinct dyes, reading out multiple bits simultaneously in the multicolor imaging mode (14) (Fig. S1 B and C). One advantage of this two-step labeling approach with encoding hybridization followed by readout hybridization is that it substantially reduces the time required for each hybridization round, because hybridization of the readout probes to encoding probes (including all fluid handling and sample washing) requires <30 min (13, 14) compared with the overnight hybridization typically required for direct hybridization of FISH probes to cellular RNAs, because the readout sequences on the encoding probes do not form a secondary structure and are not occluded by cellular proteins.
Protease digestion and detergent treatment efficiently remove protein and lipid from PA-embedded cells. (A) Images of U-2 OS cells stained with Krypton, a nonspecific protein dye, before (Uncleared) or after matrix imprinting and clearing (Cleared). The contrast at which the Right image is displayed has been increased 10× relative to the Center image to better illustrate the reduction in fluorescence signal. (B) The average fluorescence signal observed from the samples in A. The average fluorescence has been normalized to the fluorescence observed in the uncleared sample. The error bar represents SEM (n = 3 replicates). (C) As in A but for DiD, a nonspecific lipid stain. (D) As in B but for the samples stained with DiD. (Scale bars, 20 µm.)
To demonstrate the clearing efficacy, we stained U-2 OS cells with encoding probes used for a MERFISH measurement of 130 RNAs at a total concentration of 300 µM, which is threefold higher than typically used in our MERFISH experiments, to generate high background. We then embedded and cleared the sample in the PA matrix as described above, and stained the RNA-imprinted matrix with a readout probe labeled with a Cy5 dye. Fig. 1C shows that the cleared samples contained visible smFISH spots but substantially lower background than uncleared samples, demonstrating that this approach indeed reduced the background as a result of off-target probe binding.
As an aside, MERFISH measurements require repeated sample staining with a series of readout probes and, in cases where the FISH signal is removed by chemical cleavage of the fluorophores (14), the efficient removal of cleaved fluorophores. To facilitate the rapid penetration of readout probes as well as the rapid removal of cleaved dyes, we embedded samples in 50- to 100-µm-thick PA films. These films were thick enough to cover cultured cells or moderately sized tissue slices, yet thin enough that the rate of readout probe hybridization and the rate of dye cleavage/removal were not substantially changed from those observed in uncleared samples (Fig. S4).
Matrix imprinting and clearing in PA films does not reduce the rate of readout probe binding or reductive cleavage of fluorescent dyes. (A) The average brightness of individual RNA spots as a function of time exposed to a readout probe conjugated to Cy5 in uncleared samples (orange) or matrix imprinted and cleared samples (blue). The average brightness was normalized to the average of the brightness observed in the final two time points. (B) The average brightness of individual RNA spots as a function of time exposed to cleavage buffer (SI Materials and Methods). The average brightness has been normalized to that observed before exposure to cleavage buffer. Both measurements were conducted on IMR-90 cells stained with encoding probes targeting the FLNA mRNA and the first readout probe (Bit 1) (Table S1). The readout hybridization buffer used in A differed slightly from that described previously (14) in that it contained 3 nM of the readout probe and no dextran sulfate. All error bars represent SEM (n = 3 replicates).
To determine if any RNAs were lost during matrix imprinting and clearing, we used MERFISH to determine the copy number of 130 RNAs in a cleared sample of U-2 OS cells and compared these numbers to that derived previously from an uncleared sample (14). We used our previously published 16-bit, modified Hamming-distance-4 (MHD4) code to encode RNAs (13). In this encoding scheme, all valid binary barcodes used to encode RNAs are separated by a Hamming distance of at least 4, which means that at least four bits must be read incorrectly to change one valid barcode to another, drastically reducing the probability of misidentifying RNAs. Furthermore, this scheme also allows us to correct single-bit errors because every single-bit error produces a barcode uniquely close to a single valid barcode. This specific MHD4 code contains 140 valid barcodes (13); we only used 130 of them to encode RNAs, leaving the remaining 10 barcodes to serve as “blank” controls to determine the rate of spurious RNA detection and estimate misidentification rates.
We performed this MERFISH measurement of these 130 RNA species, as described previously (14), using two-color imaging to read out 16 bits in 8 rounds of hybridization and imaging (two bits per round), as well as reductive cleavage of disulfide bonds to remove the fluorophores linked to the readout probes between consecutive rounds of smFISH imaging (Fig. S1C). Fig. 2A shows that individual RNA molecules could be clearly detected in each of the eight hybridization and imaging rounds, allowing their identity to be decoded. As described previously (14), we used the depletion of RNAs near cell boundaries to perform cell segmentation. Fig. 2B shows that the copy number per cell observed for these RNAs measured in the cleared sample correlated strongly with those measured in an uncleared sample with a Pearson correlation coefficient of 0.94 between the log10 copy numbers (ρ10 = 0.94 for the 116 RNA species whose measured copy numbers were larger than that observed for the largest blank barcode count). On average, the ratio between the copy numbers measured in the cleared sample to those measured in the uncleared sample was 1.12 ± 0.04 (SEM, n = 116 RNAs), and this ratio was largely independent of the length of the RNAs (Fig. 2C). Although we conservatively used only those RNAs with copy numbers greater than the largest blank count for analysis here, the results were similar when all 130 RNAs were used.
Matrix imprinting and clearing improves MERFISH performance with no loss in RNA. (A, Left) Two-color smFISH images from each of the eight rounds of hybridization and imaging in a MERFISH measurement of 130 RNA species in matrix-imprinted and -cleared U-2 OS cells using readout probes labeled with Cy5 (green) or Alexa750 (red). Yellow represents the overlay between the two dyes. Only a small portion of the MERFISH imaging FOV is shown. (Scale bars, 2 µm.) (Right) All identified RNAs (colored markers) detected in a single FOV with the barcodes of the RNAs represented by the colors of the markers. The white box represents the portion of this FOV displayed on the Left. (Scale bar, 25 µm.) (B) The average copy numbers per cell observed for these RNA species in matrix-imprinted and -cleared U-2 OS cells versus the copy numbers obtained from previously published measurements in an uncleared sample (14). Copy numbers were corrected by subtracting the average copy number observed for the blank barcodes. Uncorrected copy numbers are displayed in Fig. S5B. The log10 counts correlate with a Pearson correlation coefficient of 0.94 (P value: 10−54). The dashed line represents equality. (C) The average ratio of the copy number per cell for a sample that was matrix imprinted and cleared to that observed for an uncleared sample for RNAs within the specified RNA length range. Error bars represent SEM (n = 26 genes for each bin). (D) Average copy number per cell of the blank barcodes (i.e., barcodes not assigned to an RNA) in an uncleared sample and in a matrix-imprinted and -cleared sample. Error bars represent SEM (n = 10 blank barcodes).
These measurements showed that several aspects of MERFISH performance were improved with matrix imprinting and clearing. Previously, we observed a MERFISH detection efficiency of ∼90% (14); thus, a copy number ratio of ∼1.1 between the cleared and uncleared samples suggested that clearing increased the detection efficiency to near 100%. Second, we observed that the average frequency at which the blank barcodes were observed per cell in the cleared samples dropped substantially relative to that observed in the uncleared samples (Fig. 2D). The average level of blank barcode counts observed in the uncleared sample (Fig. 2D) was comparable to the observed copy number for the lowest abundance RNAs measured here, leading to the possibility that the copy number observed for these low abundance RNAs might have been biased by a background rate of spurious RNA counts in uncleared samples. Indeed, we observed an excess of these low-abundance RNAs in uncleared samples relative to that expected from bulk RNA-seq (Fig. S5A), whereas this bias was substantially reduced in cleared samples (Fig. S5). Thus, we conclude that the increased signal-to-background in cleared samples results in an improvement in both the detection efficiency and the detection limit in MERFISH measurements.
Matrix imprinting and clearing reduces bias in the detection of low abundance RNAs. (A) The ratio of the copy number per cell determined via MERFISH to the abundance determined via RNA-seq (46) as measured in FPKM for uncleared samples (blue) and for matrix-imprinted and -cleared samples (red). Error bars represent SEM (n = 26 RNAs in each abundance range). (B) The copy number per cell determined via MERFISH in a matrix-imprinted and -cleared sample compared with that determined for an uncleared sample. These copy numbers have not been corrected for the average rate of blank barcode detection as in Fig. 2B. The dashed line represents equality. The deviation from equality in B and the excess MERFISH counts relative to those estimated from bulk RNA-seq at the low abundance range are consistent with the increased rate of blank barcode detection observed for uncleared samples (Fig. 2D).
In addition to providing a substantial decrease in the background because of off-target binding of FISH probes, the removal of proteins and lipids from the sample may also reduce the level of autofluorescence. To quantify this decrease, we measured the fluorescence of unlabeled U-2 OS cells in uncleared and cleared samples with four excitation wavelengths: 750, 647, 561, and 488 nm. Consistent with the expectation that cell autofluorescence is substantially higher in the blue–green spectral range than in the red range, the clearing protocol had little effect on the already low autofluorescence background in the 750- and 647-nm channels, but produced a substantial reduction in the autofluorescence observed in the 561- and 488-nm channels (Fig. 3A).
Autofluorescence reduction by matrix imprinting and clearing facilitates four-color MERFISH. (A) The average autofluorescence observed for unstained U-2 OS cells before (blue) and after (red) matrix imprinting and clearing when excited with 750-, 647-, 561-, or 488-nm light. Error bars represent SEM (n = 3 replicates). (B) Images of cleared U-2 OS cells stained with MERFISH-encoding probes targeting 130 RNAs and the first four readout probes each conjugated to one of the following dyes: Alexa750, Cy5, ATTO565, or Alexa488. Samples were imaged with excitation light listed in A. (Scale bars, 10 µm.) (C) Average copy number per cell determined via four-color MERFISH to that determined with two-color MERFISH, both in cleared samples. The copy numbers have been corrected by subtracting the average rate of blank barcode detection as in Fig. 2B. The dashed line represents equality. The Pearson correlation coefficient between the log10 abundances is 0.99 (P value: 10−98). (D) The average rate of observing a “1” to “0” error (blue) or a “0” to “1” error (red) per bit for bits that are read out with each of the four color channels, as indicated by the excitation wavelength. Each error rate (“1” to “0” or “0” to “1”) was calculated for each individual bit using the frequency at which errors were corrected at that bit, as described previously (13), and then these per bit error rates were averaged over the bits that were detected in the same color channel (Table S1). Error bars represent SEM (n = 4 bits read out with each color channel).
With this reduction in the autofluorescence, we explored the possibility of using all four excitation channels to read out four different bits of the 16-bit code simultaneously in each round of imaging during MERFISH measurements. We again stained U-2 OS cells with the same MERFISH-encoding probe set for the 130 RNA species as described above and performed MERFISH measurements in which each round of hybridization used four different readout probes, conjugated respectively to Alexa750, Cy5, ATTO565, or Alexa488 via a disulfide bond (Fig. 3B and Table S1). With such four-color imaging, the 16-bit MERFISH measurement only required four rounds of hybridization and imaging.
Readout probe sequences
We then compared the measured copy numbers derived from this four-color measurement to those determined with a two-color (750 and 647 nm) measurement in the cleared sample. Fig. 3C demonstrates that these copy numbers correlated strongly with a ρ10 of 0.99 and had an average ratio of 1.01 ± 0.02 (SEM, n = 109 RNAs with copy numbers greater than that observed for the largest blank count). To confirm that imaging in the new color channels did not introduce additional error, we determined the “1” to “0” or “0” to “1” error rates per bit and found that these error rates did not vary substantially with the color channel (Fig. 3D).
Finally, to confirm that the improved performance that we observed with cleared samples was reproducible, we performed additional two-color and four-color MERFISH measurements in cleared samples. Fig. S6 shows that the copy numbers derived from all of these measurements correlated strongly (all ρ10 are 0.94 or greater). By comparing each of these datasets to the previously determined detection efficiency of MERFISH measurements in uncleared samples (14), we estimated an average MERFISH detection efficiency of 96 ± 7% (SEM, n = 4 replicates) for cleared samples. Furthermore, we observed an ∼fourfold reduction in the average rate of blank barcode detection: 0.08 ± 0.03 counts per cell (SEM, n = 4 replicates) for cleared samples versus 0.30 ± 0.07 counts per cell (SEM, n = 7 replicates) (14) for uncleared samples.
Two- and four-color MERFISH measurements in matrix imprinted and cleared samples are reproducible. Comparison of the average copy number per cell measured in different two-color or four-color MERFISH measurements in matrix imprinted and cleared U-2 OS cells. ρ10 represents the Pearson correlation coefficient between the log10 copy numbers for all RNAs. The P values associated with all ρ10 are less than 10−44.
To explore whether clearing can overcome the increased background that we have observed in tissues, we performed MERFISH measurements of 130 RNA species on four cryosections taken from adult mouse hypothalamus, each ∼2 mm × 2 mm wide and 10 μm thick (Fig. 4 A and B). We performed 3D imaging with seven ∼1.5-µm-thick optical sections measured per field-of-view (FOV). These RNAs were again encoded with the 16-bit MHD4 code and read out with eight rounds of hybridization using two-color imaging per round. These samples were matrix-imprinted and cleared as described above but with the addition of a brief treatment with 4% (wt/vol) SDS before PA embedding, which further improved tissue clearing. Fig. 4 C and D illustrate that this matrix-imprinting and -clearing approach substantially reduced the background observed in these tissue slices. smFISH spots representing individual RNA molecules were clearly observable in the cleared sample in each round of imaging with low “1” to “0” or “0” to “1” error rates, similar to those detected in cultured cells, allowing individual RNAs to be decoded (Fig. 4 E–G).
MERFISH measurements of adult mouse brain tissue. (A) NissI-stained images of coronal and sagittal slices of an adult mouse brain taken from the Allen brain atlas (42). The black box and dashed line represent the region of the mouse hypothalamus studied. (Scale bar, 2 mm.) (B) Image of a single, 10-µm-thick cryosection of the mouse hypothalamus stained with DAPI. The complete volume of the central 2-mm × 2-mm region of this slice was imaged with MERFISH using seven 1.5-µm-thick optical sections per FOV. (Scale bar, 1 mm.) (C and D) Images of a small portion of a mouse hypothalamus slice stained with an encoding probe set used for a MERFISH measurement of 130 RNAs and a readout probe conjugated to Cy5. (C) Single optical section image of an uncleared sample. (D) Single optical section image of a matrix-imprinted and -cleared sample. (Scale bars, 50 µm.) (E) Zoom-in of the region of D marked with the white dashed box. (F) Decoded RNAs (different colors represent different barcodes) from all seven optical sections of the region shown in E. Not all RNA molecules shown in F are observed in E because E represents only one of the seven optical sections and one of the 16 bits. (Scale bars, 10 μm.) (G) The average rate of observing a “1” to “0” error (blue) or a “0” to “1” error (red) per bit for bits that are read out with each of the two color channels, as indicated by the excitation wavelength. Error rates were calculated as in Fig. 3D. Error bars represent SEM (n = 8 bits read out with each color channel). (H) The density of 130 RNA species as determined via MERFISH versus the abundance as determined via bulk RNA-seq for the region of the mouse hypothalamus shown in A. The Pearson correlation coefficient between the log10 abundances is 0.84 (P value: 10−35).
We compared the average RNA density determined via MERFISH from these four tissue slices with the abundance determined via bulk RNA-seq data derived from the same region of the hypothalamus (29), and observed a strong correlation (ρ10 = 0.84) between our MERFISH results and the bulk RNA-seq results (Fig. 4H). At the very low abundance range corresponding to those RNAs that are expressed poorly in the hypothalamus [<0.5 fragments per kilobase of transcript per million (FPKM)], the correlation between MERFISH and bulk RNA-seq results reduced substantially, suggesting that the abundance of these RNAs was near or below our current detection limit, a conclusion supported by the similarity between the copy numbers of these RNAs and the average copy number observed for the blank barcodes [6 × 106 ± 2 × 106/mm3 (SEM, n = 10 blank barcodes)].
Massively multiplexed smFISH allows spatially resolved gene-expression profiling within single cells. However, a number of important applications of and advances to this approach are limited by the fluorescence background encountered in these experiments. Here we described a clearing approach that substantially reduced several background sources in RNA FISH measurements by effectively imprinting the desired RNA signal onto an inert, nonfluorescent, PA matrix and then removing unwanted cellular components that give rise to background because of off-target probe binding and autofluorescence. The reduction in fluorescence background provided by this approach led to improvement in both the detection efficiency and the detection limit in MERFISH measurements. Moreover, this matrix-imprinting and -clearing approach produced a substantial reduction in the background observed for measurements in tissue samples, allowing high-performance MERFISH measurements in brain tissue sections, which should facilitate spatial mapping of distinct cell types in the brain.
This matrix-imprinting and -clearing approach complements several existing methods that have been used to improve the signal or reduce nonspecific binding background for RNA detection in FISH experiments (15, 24, 26, 30, 31) or in situ sequencing experiments (32, 33). For example, signal amplification techniques, such as branched DNA (34), rolling-circle amplification (35, 36), and hybridization chain reaction (37) have been used to increase the signal associated with each RNA molecule (15, 24, 26, 30⇓⇓⇓–34), thereby increasing the ratio between RNA signals and autofluorescence background, although background as a result of off-target probe binding may also be amplified concurrently by these approaches. Alternatively, proximity-dependent approaches (38, 39), in which a fluorescent signal is only produced when two separate probes or two ends of the same probe are in close proximity, have been used to reduce the effect of nonspecific probe binding (30, 32, 36, 38, 39). Because matrix imprinting and clearing is compatible with each of these approaches, it may be combined with these techniques to further improve the performance of both conventional and highly multiplexed smFISH measurements.
In addition, the reduction in autofluorescence in the blue and green color channels provided by matrix imprinting and clearing allowed us to extend MERFISH measurements from two colors to four colors with no loss in performance, which should substantially increase the measurement speed and throughput of MERFISH. We previously reported a high-throughput MERFISH imaging platform and demonstrated the ability to profile 140 RNA species in 40,000 human cells in ∼18 h using a 16-bit MHD4 code, two-color imaging, and eight hybridization rounds (14). This platform also allowed us to profile a similar number of RNAs across an ∼16-mm2 × 10-µm tissue volume in ∼18 h here. This throughput was facilitated by our ability to rapidly bind FISH probes (through hybridization of readout probes to encoding probes, which requires <30 min) and then rapidly remove FISH signals (through the use of chemical cleavage to remove fluorophores from readout probes, which requires <20 min) in each round of hybridization. With the increase to four color channels, we can now read out the 16-bit code in just four rounds of hybridization with four bits detected per round, further reducing the measurement durations and allowing tens of thousands of cells to be profiled in <10 h with MERFISH. Similarly, we also anticipate that our previously demonstrated MERFISH measurement of 1,001 RNA species using a 14-bit MHD2 encoding scheme (13) can also be completed in just four hybridization rounds. While this current paper was in review, another paper demonstrated the profiling of >100 RNA species in tissue using a different multiplexed smFISH method (seqFISH) with a similar number of hybridization rounds: for example, 125 RNA species could be measured with a four-letter, five-color error-correcting barcoding scheme using four rounds of hybridization, but the slower approaches used there for probe hybridization (overnight hybridization of FISH probes directly to cellular RNA) and signal removal (4 h of DNase I digestion of FISH probes) in each round made the sample processing and measurement time per round longer (15).
Finally, we envision that the substantial reduction in background provided by this matrix-imprinting and -clearing approach will facilitate several additional extensions of MERFISH. First, an increase in the degree of multiplexing—to the simultaneous measurement of several thousand or tens of thousands of RNAs—would likely require substantially higher encoding probe concentrations than are currently used and, thus, would benefit from the much lower off-target probe binding achieved here in cleared samples. Second, we have performed MERFISH using tens of encoding probed per RNA [e.g., 92 encoding probes for RNAs that are 3 kb or longer (13, 14)]. With the dramatic decrease in background enabled by this clearing approach, it should be possible to detect RNAs that are much shorter, potentially with as few as <10 encoding probes per RNA. This advance would facilitate the detection of relatively short messenger and long-noncoding RNAs, and possibly some small RNAs. The ability to detect RNA molecules with relatively few FISH probes will also substantially improve the ability to distinguish RNA isoforms. Third, the combination of expansion microscopy (25) with MERFISH may be facilitated by a common matrix-imprinting approach, which may help RNA profiling in RNA-dense regions of cells and further increase in the degree of multiplexing. We also anticipate that the alternative RNA anchoring approach reported for expansion microscopy (26)—in which RNAs are alkylated with a cross-linker that is covalently incorporated into the PA gel—may be used for background reduction by matrix imprinting and clearing as well. Finally, whereas our current implementation of the matrix-imprinting and -clearing approach removes cellular proteins and, thus, information regarding the protein content of the sample, this information could be restored by labeling samples with antibodies conjugated to oligonucleotides before embedding. These oligonucleotides could then be anchored to the matrix, followed by digestion of the antibodies along with cellular proteins, allowing the original location of the antibodies to be determined via FISH imaging of these oligonucleotides (25). Such labeling approaches could be used to label cell boundary markers to facilitate cell segmentation in cases where a depletion of RNA at the edge of the cell is not sufficient to identify cell boundaries. Moreover, such approaches can convert protein identities into oligonucleotide signals, which may be used to perform highly multiplexed proteomic and transcriptomic measurements simultaneously in single cells.
Detailed descriptions of all protocols are provided in the SI Materials and Methods. All animal experiments were approved by the Institutional Animal Care and Use Committee of Harvard University.
Human osteosarcoma cells (U-S OS, American Type Culture Collection) and Human fibroblasts (IMR-90, American Type Culture Collection) were cultured, fixed, permeabilized, and stained with smFISH probes or MERFISH-encoding probes, as described previously (13, 14, 40). Mouse hypothalamus tissue was freshly frozen, cryosectioned into 10-μm-thick slices, postfixed onto coverslips, cleared with 4% (wt/vol) SDS, permeabilized with 70% (vol/vol) ethanol, and then stained with encoding probes. Cells or tissue samples were embedded in a 4% (vol/vol) solution of a 19:1 ratio of acrylamide to bis-acrylamide containing 50 mM Tris HCl (pH 8), 300 mM NaCl, 0.03% (wt/vol) ammonium persulfate, and 0.15% (vol/vol) N,N,N′,N′-tetramethylethylenediamine (TEMED). Protein and lipids were removed with a >12 h, 37 °C digestion with proteinase K in 0.8 M guanidine HCl, 0.5% (vol/vol) Triton-X 100, 50 mM Tris pH 8, and 1 mM EDTA.
MERFISH measurements with U-2 OS cells were performed with a published encoding probe set (14). The encoding probe set for measurements in mouse brain tissue was designed as described previously (14). Readout probes were purchased from Biosynthesis and are described in Table S1. Encoding probes were constructed by modifying the Oligopaints approach (41) with a high-yield enzymatic amplification protocol and a high-speed probe design algorithm (13, 14, 40). Samples were imaged on custom setups with readout hybridization, buffer exchange, and reductive cleavage procedures detailed in SI Materials and Methods.
MERFISH measurements in human osteosarcoma cells (American Type Culture Collection, U-2 OS) were performed with the same MERFISH-encoding probe set as previously described (14). Briefly, this encoding scheme used a 16-bit MHD4 code to encode the RNAs (13). In this encoding scheme, each of the 140 possible barcodes required at least four errors to accumulate to be converted into another barcode. This property permitted the detection of errors at up to any two bits, and the correction of errors to any single bit. In addition, this encoding scheme used a constant Hamming weight (i.e., the number of “1” bits in each barcode) of 4, to avoid potential bias in the measurement of different barcodes because of a differential rate of “1” to “0” and “0” to “1” errors, as described previously (13). We used 130 of the 140 possible barcodes to encode cellular RNAs and the remaining 10 barcodes were left unassigned to serve as blank controls. The encoding probe set that we used contained 92 encoding probes per RNA, with each encoding probe containing three of the four readout sequences assigned to each RNA (Fig. S1B).
The MERFISH-encoding probes for measurements in the mouse hypothalamus were designed using the same 16-bit MHD4 code as above. Again, 130 of the 140 possible barcodes were assigned to RNAs that were selected to cover roughly three orders-of-magnitude in average expression in the hypothalamus with expression levels estimated from previously published RNA-seq (29). The remaining 10 barcodes were left unassigned to serve as blank controls. Encoding probes were designed using a previously published pipeline with the same stringency conditions and design criteria as described previously (14). Transcript sequences were derived from the mouse genome (mm9) downloaded from ensembl (useast.ensembl.org/Mus_musculus/Info/Index).
Construction of the encoding probe sets was conducted from complex oligonucleotide pools, as described previously (14, 40). Briefly, we amplified the oligopools (CustomArray) via limited-cycle PCR to make in vitro transcription templates, converted these templates into RNA via in vitro transcription, converted the RNA back to DNA via reverse transcription, and then purified the DNA via alkaline hydrolysis (to remove RNA templates), phenol-chloroform extraction (to remove proteins), and ethanol precipitation (to remove nucleotides and concentrate probes). To improve probe purity and reaction yield, we modified the previous protocol (14, 40) in the following ways. First, excess NTPs or dNTPs were removed via desalting columns (40K molecular weight cut-off Zeba; ThermoFisher, 89894) after both the in vitro transcription and the phenol-chloroform extraction. We found that removal of stray NTPs improved the performance of the reverse transcription and that removal of excess dNTPs aided in quantification of the final yield of the protocol. In addition, to further improve yield, we switched the salt in our ethanol purification from 2.5 M ammonium acetate (which allows nucleotides to be removed, but decreases DNA recovery in our hands) to 300 mM sodium acetate.
To stabilize the PA film, we coated our coverslips with a silane layer containing an allyl moiety, which could be actively incorporated into PA gels during polymerization, covalently cross-linking the PA film to the coverslip. Coverslips were silanized using a modified version of a published protocol (43). Briefly, 40-mm-diameter #1.5 coverslips (Bioptechs, 0420-0323-2) were washed for 30 min via immersion in a 1:1 mixture of 37% (vol/vol) HCl and methanol at room temperature. Coverslips were then rinsed three times in deionized water and once in 70% (vol/vol) ethanol. Coverslips were dried in a 70 °C oven and then immersed in 0.1% (vol/vol) triethylamine (Millipore, TX1200) and 0.2% (vol/vol) allyltrichlorosilane (Sigma, 107778) in chloroform for 30 min at room temperature. Coverslips were washed once each with chloroform and ethanol and then baked in a 70 °C oven for 1 h to dehydrate the silane layer. Silanized coverslips could then be stored at room temperature in a desiccated chamber for weeks with no obvious reduction in the quality of the silane layer.
To promote cell adhesion, silanized coverslips were coated with 0.1 mg/mL poly-d-lysine (PDL) (molecular weight 30,000–70,000 Da; Sigma, P7886) diluted in nuclease-free water for 1 h at room temperature. Coverslips were washed three times with nuclease-free water, incubated in water at room temperature overnight, and then dried and UV-sterilized before plating cells.
U-2 OS cells were cultured, fixed, and permeabilized using protocols previously described (14, 40), before staining with encoding probes. Briefly, cells cultured with EMEM (American Type Culture Collection, 30-2003) containing 10% (vol/vol) FBS (ThermoFisher, 10437) were plated on PDL-coated, silanized coverslips at a density of 300,000 cells per coverslip, and incubated at 37 °C with 5% CO2 for 48–72 h before fixing with 4% (vol/vol) paraformaldehyde (PFA; Electron Microscopy Sciences, 15714) in 1× PBS (ThermoFisher, AM9625) for 20 min. Cells were washed three times with 1× PBS, and permeabilized using 0.5% (vol/vol) Triton X-100 (Sigma, T8787) in 1× PBS for 10 min at room temperature. Cells were then washed three times with 1× PBS.
Encoding probe staining was performed as described previously (13, 14, 40). Briefly, cells were incubated for 5 min in a 30% formamide wash buffer, containing 2× saline-sodium citrate (SSC; ThermoFisher, AM9763) and 30% (vol/vol) formamide (ThermoFisher, AM9342) and then stained with encoding probes in encoding hybridization buffer, containing 2× SSC, 30% (vol/vol) formamide, 0.1% (wt/vol) yeast tRNA (Life Technologies, 15401-011), 1% (vol/vol) murine RNase inhibitor (New England Biolabs, M0314L), and 10% (wt/vol) dextran sulfate (Sigma, D8906), in a humidity-controlled 37 °C incubator for 36 h. Encoding probes were stained at a concentration of 100 µM unless otherwise specified for experiments in Fig. 1C. Where appropriate, the encoding probes were supplemented with 1 μM of anchor probe: a 15-nt sequence of alternating dT and thymidine-locked nucleic acid (dT+) with a 5′-acrydite modification (Integrated DNA Technologies). After staining, cells were washed two times for 30 min each with 30% formamide wash buffer at 47 °C.
Human lung fibroblast cells (American Type Culture Collection, IMR-90) were cultured, fixed, and stained following the same protocols described above for U-2 OS cells using 1 μM of a smFISH probe set targeting FLNA (Biosearch) described previously (14).
To anchor RNAs in place, the encoding-probe–stained samples were embedded in thin, 4% PA gels. Briefly, stained samples on coverslips were first washed for 2 min with a de-gassed PA solution, consisting of 4% (vol/vol) of 19:1 acrylamide/bis-acrylamide (BioRad, 1610144), 60 mM Tris⋅HCl pH 8 (ThermoFisher, AM9856), 0.3 M NaCl (ThermoFisher, AM9759), and either a 1:500 dilution of 0.1-µm-diameter light-yellow beads (Spherotech, FP-0245-2), when samples were used for four-color MERFISH measurements, or a 1:200,000 dilution of 0.1-µm-diameter carboxylate-modified orange fluorescent beads (Life Technologies, F-8800), when samples were used for two-color MERFISH measurements. The beads served as fiducial markers for the alignment of images taken across multiple rounds of smFISH imaging. Cells were then washed again for 2 min with the same PA gel solution supplemented with the polymerizing agents ammonium persulfate (Sigma, A3678) and TEMED (Sigma, T9281) at final concentrations of 0.03% (wt/vol) and 0.15% (vol/vol), respectively.
To cast a thin PA film, 50 µL of this gel solution was added to the surface of a glass plate (TED Pella, 26005) that had been pretreated for 5 min with 1 mL GelSlick (Lonza, 50640) so as not to stick to the PA. The sample on the coverslip, treated as described above, was aspirated to remove excess PA gel solution, then gently inverted onto this 50-µL droplet to form a thin layer of PA between the coverslip and the glass plate. The volume of this gel droplet could be used to control the thickness of this PA film. The gel was then allowed to cast for 1.5 h at room temperature. The coverslip and the glass plate were then gently separated, and the PA film washed twice with a digestion buffer consisting of 0.8 M guanidine-HCl (Sigma, G3272), 50 mM Tris⋅HCl pH 8, 1 mM EDTA, and 0.5% (vol/vol) Triton X-100 in nuclease-free water. After the final wash, the gel was covered with digestion buffer supplemented with 1% (vol/vol) proteinase K (New England Biolabs, P8107S). The sample was digested in this buffer for >12 h in a humidified, 37 °C incubator and then washed with 2× SSC three times. MERFISH measurements were either performed immediately or the sample was stored in 2× SSC supplemented with 0.1% (vol/vol) murine RNase inhibitor at 4 °C for no longer than 24 h.
Cultured-cell samples were imaged on a home-built imaging platform with minor modifications from that described previously (13, 40). Briefly, this microscope was built using an Olympus IX-71 body and a 1.45 NA, 100× oil-immersion objective. Illumination in 750, 641, 561, and 488 nm were provided using solid-state lasers (MPB communications, VFL-P500-751; MPB communications, VFL-P500-642; Coherent, 561-200CWCDRH; and Coherent, 1069413/AT) for excitation of readout probes labeled with Alexa750, Cy5, ATTO565 and Alexa488, respectively. For two-color MERFISH measurements using Alexa750-labeled and Cy5-labeled readout probes, the 561-nm laser was used to excite the orange fiducial beads. A 405-nm solid-state laser (Coherent, Cube) was used to excite the nuclear stain DAPI, where appropriate. For four-color MERFISH measurements using Alexa750-, Cy5-, ATTO565-, and Alexa488-labeled readout probes, the 405-nm light was also used to excite the light-yellow fiducial beads. All laser lines were combined with a custom dichroic (Chroma, zy405/488/561/647/752RP-UF1), and the emission was filtered with a custom penta-notch filter (Chroma, ZET405/488/561/647–656/752m). Fluorescence was imaged with an EMCCD camera (Andor, iXon-897). The pixel size for the EMCCD camera was determined to correspond to 167 nm in the sample plane. Each FOV was imaged at a single z-plane with a 100-ms exposure for each color channel.
Tissue slices were imaged on a second home-built high-throughput imaging platform as described previously (14). Briefly, this microscope was constructed around an Olympus IX-71 microscope body and a PlanApo, 1.3 NA, 60× silicone–oil-immersion objective (Olympus, UPLSAPO 60×S2). Illumination in 754, 647, 561, and 405 nm was provided using solid-state lasers (Toptica, DL100/BoosTA; MBP Communications, F-04306-113; Crystalaser GCL-150-561; Coherent, Cube 405). These laser lines were used to excite readout probes labeled with Alexa750 and Cy5, orange fiducial beads, and DAPI, respectively. The illumination profile was flattened with a square multimode fiber (Andor, Borealis). The fluorescence emission from the sample was separated from the laser illumination using a penta-band dichroic (Chroma, zy405/488/561/647/752RP-UF1) and imaged using a scientific CMOS camera (sCMOS; Andor, Zyla 4.2) after passing through two duplicate custom penta-notch filters (Chroma, ZET405/488/561/647–656/752m) to remove stray excitation light. The pixel size for the sCMOS camera was determined to correspond to 109.2 nm in the sample plane. During the imaging of tissue, z-stacks consisting of seven, 1.5-µm-thick optical sections were collected in each color channel at each FOV so as to image the entire volume of the tissue. Each exposure was 500 ms. The z-steps were controlled via an objective nanopositioner (Mad City Labs, NanoF200).
On both set-ups, sample position was controlled via a motorized microscope stage (Marzhauser, SCAN IM 112 × 74) and focus was maintained via a custom focus-lock system, realized through a feedback system between an objective nanopositioner (Mad City Labs, NanoF200) and the reflection of an IR laser (Thorlabs, LP980-SF15) onto an inexpensive CMOS camera (Thorlabs, uc480). The sample coverslip was held inside a flow chamber (Bioptechs, FCS2), and buffer exchange within this chamber was directed using a custom-built automated fluidics system described previously (13, 14, 40), controlling three eight-way valves (Hamilton, MVP and HVXM 8–5) and a peristaltic pump (Gilison, Minipuls 3).
Samples were hybridized with readout probes and imaged following protocols similar to those previously described (14), with slight adjustments to readout hybridization buffer composition and flow times. Readout hybridization buffer was composed of 2× SSC, 10% (vol/vol) ethylene carbonate (Sigma-Aldrich, E26258), 0.1% (vol/vol) murine RNase inhibitor in nuclease-free water, and 3 nM of the appropriate readout probes. Previously, we used dextran sulfate in this buffer to increase the rate of readout probe hybridization; however, we have found that the same hybridization kinetics can be achieved without dextran sulfate by increasing the readout probe concentrations from 1 nM to 3 nM. Removing dextran sulfate from the readout buffer dramatically reduced the viscosity of this buffer, and this reduction in buffer viscosity, in turn, effectively eliminated the occasional flow failures that arose from the high pressures required to pull high-viscosity buffers through the fluidics system.
Two different configurations of readout probes were used: for two-color MERFISH measurements, two readout probes, one conjugated to Cy5 and the other to Alexa750 via a disulfide bond were used in each round of hybridization; and for four-color MERFISH, four different readout probes each conjugated to one of Alexa750, Cy5, ATTO565, or Alexa488 via a disulfide bond were used in each round of hybridization. Table S1 contains the readout probe sequences and dye combinations used for both two- and four-color measurements. All readout probes were purchased from Biosynthesis.
The sample was stained with readout probes by first flushing the sample chamber with 2 mL of readout hybridization buffer over the span of 5 min to fully exchange buffers. Then, an additional 2 mL of buffer was flown across the sample for 6 min. The sample was then washed by flowing 2 mL of readout wash buffer, containing 2× SSC and 10% (vol/vol) ethylene carbonate, over a span of 9 min. Finally, 2 mL of imaging buffer, containing 2× SSC, 50 mM Tris⋅HCl pH 8, 10% (wt/vol) glucose, 2 mM Trolox (Sigma-Aldrich, 238813), 0.5 mg/mL glucose oxidase (Sigma-Aldrich, G2133), 40 μg/mL catalase (Sigma-Aldrich, C30), and 0.1% (vol/vol) murine RNase inhibitor, was flown across the sample for 6 min, after which the flow was halted and ∼400 FOVs were imaged. This imaging buffer was used to decrease the effect of photobleaching (44). Imaging buffer was stored under a layer of mineral oil (Sigma-Aldrich, 330779) throughout the measurement as a barrier against oxygen. Because glucose oxidase was determined to contain trace amounts of RNase, the imaging buffer also contained 0.1% (vol/vol) murine RNase inhibitor. We replaced ribonucleoside vanadyl complex (New England Biolabs, S1402S), which was used previously (13, 14, 40), with Murine RNase inhibitor because the citrate in 2× SSC was found to significantly reduce the effective lifetime of ribonucleoside vanadyl complex.
After each round of imaging, the fluorescent dyes were removed from readout probes by reductive cleavage of the disulfide bond conjugating these dyes to the probes. Three milliliter of cleavage buffer comprising 2× SSC and 50 mM of the reducing agent Tris(2-carboxyethyl)phosphine (TCEP; Sigma, 646547) was flown across the sample over the course of 15 min. After cleavage, the chamber was flushed with 2 mL of 2× SSC for 4 min to flush any residual cleavage buffer from the sample before the introduction of the subsequent hybridization buffer. All buffers were freshly prepared before each experiment using nuclease-free water.
After the final round of hybridization and imaging, the sample was stained with DAPI at a concentration of 1 μg/mL in 2× SSC for 10 min to mark nuclei, and then imaged at 405 nm.
Registration of images of the same FOV across imaging rounds as well as decoding of the RNA barcodes was conducted using a previously described analysis pipeline (14). Briefly, the locations of the fiducial beads in each round of imaging were found via a Gaussian fitting routine (45), and these locations were used to create affine transformations that correct offsets between images in each imaging round. Additional corrections to account for minor chromatic aberrations were not applied because the offsets in the centroid of RNAs labeled simultaneously with Alexa750, Cy5, ATTO565, and Alexa488 were not substantial. Images were then high-pass filtered to remove background, deconvolved to tighten RNA spots, and then low-pass filtered so as to connect RNA centroids that differ slightly in location between images, a property that we have previously observed (13). Individual pixels were then assigned to barcodes by comparing the intensity of each pixel across the 16 images collected across all color channels and all hybridization rounds (two color channels in eight hybridization rounds or four color channels in four hybridization rounds) to each of the different barcodes. Specifically, the set of 16 intensities for each pixel derived from each of the 16 images were used to define a vector that was normalized to unitary magnitude (i.e., by dividing by the L2 norm). A unit vector was similarly defined for each of the 140 barcodes. The Euclidean distance was then calculated between each pixel vector and each of the barcode vectors. A pixel was assigned to a barcode if the Euclidean distance separating it from a barcode was smaller than a given threshold. This distance threshold was determined from the largest Euclidean distance between each normalized barcode and the set of normalized barcodes formed from all single-bit errors to that barcode. Conceptually, this distance defines a 16-dimension sphere that contains all possible modifications to a barcode that correspond to a single-bit error to that barcode, and this decoding approach can be thought of as assigning pixels to a given barcode based on whether they fall within this sphere for a given barcode. Pixels with vectors that do not fall within one of these 140 spheres are left unassigned. Contiguous pixels assigned to the same barcode were combined to form a single RNA. Each RNA was then identified as requiring error correction (or not) by comparing the average pixel vector across all pixels assigned to that RNA to the set of unitary vectors defined by all single-bit errors to the assigned barcode. If the average pixel vector was closer to a vector corresponding to a single-bit error than it was to the correct barcode, the RNA was marked as requiring error correction.
This decoding approach assumes that the brightness of each RNA spot is identical between imaging rounds. To correct for differences in the brightness between color channels, images were initially normalized by equalizing their intensity histograms, as described previously (14). This normalization was then refined via an iterative process, again as described previously (14). A background of spurious RNAs were removed with thresholds on the brightness of the RNA (i.e., the L2 norm of the pixel vector) and the number of pixels combined to form that RNA (i.e., its area), as described previously (14). For tissue imaging, this pipeline was modified to accommodate z-stacks. Because each z-stack was separated by a distance slightly larger than the axial extent of the point-spread function, each stack was decoded independently of the others. Nuclei were identified and counted via intensity thresholding of the DAPI images, as described previously (14).
Computations were split between the Odyssey cluster supported by the FAS Division of Science, Research Computing Group at Harvard University and a desktop server that contained two 10-core Intel Xeon E5-2680 2.8-GHz CPUs and 256 GB of RAM.
U-2 OS or IMR-90 samples were stained with smFISH probes or MERFISH-encoding probes as described above in Cell Culture and Fixation and Encoding Probe Staining. Samples were then stained with readout probe 1 (Table S1) as in MERFISH Imaging, above. Samples were imaged, and then treated for 30 min with 1% (vol/vol) RNase A (Qiagen, 19101) in 2× SSC, and then reimaged.
U-2 OS cells were cultured, fixed, and labeled with smFISH probes or MERFISH-encoding probes as described in Cell Culture and Fixation and Encoding Probe Staining, above. Samples were then either matrix-imprinted and -cleared as described in Sample Embedding and Clearing, above, or stored at 4 °C in 2× SSC. Cells were stained with a 1:10 dilution of Krypton Fluorescent Protein Stain (ThermoFisher, 46629) in 2× SSC at room temperature for 15 min and washed once in 2× SSC at room temperature for 15 min. One-hundred FOVs were imaged with the 561-nm laser. Samples for lipid staining were prepared in the same fashion but stained with a 1:200 dilution of Vybrant DiD Cell-Labeling Solution (ThermoFisher, V22887) in 2× SSC at room temperature for 15 min and washed once briefly with 2× SSC. One-hundred FOVs were imaged with the 641-nm laser. Imaged samples were quantified by averaging the observed fluorescence across all FOVs, and this value was then averaged across three biological replicates.
MERFISH imaging in tissue was performed on 10-μm-thick cryosectioned mouse hypothalamus slices. Whole-brain tissue was removed from mice (C57BI6/J) killed using CO2, and immediately frozen in optimum cutting temperature compound (Tissue-Tek O.C.T.; VWR, 25608–930). Frozen blocks were coarsely sectioned to the hypothalamus region, trimmed to an area of roughly 3 mm × 3 mm, and sectioned at a thickness of 10 μm at −18 °C on a cryostat (MICROM, HM550). Sections were collected on silanized coverslips coated with PDL prepared following protocols described in Silanization of the Coverslips and Cell Culture and Fixation, above. These sections were then immediately fixed in 4% (vol/vol) PFA in 1× PBS for 12 min at room temperature and washed with 1× PBS for 5 min three times. The samples were then partially cleared by treating them with 4% (wt/vol) SDS in 1× PBS for 2 min with gentle agitation at room temperature. After this treatment, samples were washed three times with 1× PBS for 5 min, and then immersed in 70% (vol/vol) ethanol and stored at 4 °C for at least 18 h.
Tissue samples were then stained and cleared following the protocols described in Encoding Probe Staining and Sample Embedding and Clearing, above. Tissue samples were measured using 2-color MERFISH as described in MERFISH Imaging, above. Four cryosections were imaged in a single MERFISH experiment on the high-throughput imaging platform described in Imaging Platforms, above.
We thank Hazen Babcock, Guiping Wang, and Chenglong Xia for helpful discussions and assistance in experiments. This work was supported in part by NIH Grants R01HD082131 (to C.D.) and R01MH113094 and R01MH111502 (to C.D. and X.Z.). C.D. and X.Z. are Howard Hughes Medical Institute investigators.
↵1J.R.M. and J.H. contributed equally to this work.
Author contributions: J.R.M., J.H., D.B.-M., T.L., C.D., and X.Z. designed research; J.R.M., J.H., D.B.-M., and T.L. performed research; J.R.M., J.H., and T.L. analyzed data; and J.R.M., J.H., D.B.-M., T.L., C.D., and X.Z. wrote the paper.
Reviewers: G.D., University of Texas Southwestern Medical Center; and T.H., The Johns Hopkins University.
Conflict of interest statement: X.Z., J.R.M., J.H., and T.L. are inventors on patents applied for by Harvard University that cover the multiplexed error-robust FISH and matrix-imprinting–based clearing methods.
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1617699113/-/DCSupplemental.
Freely available online through the PNAS open access option.
