Multiplexed labeling of cellular proteins with split fluorescent protein tags

Self-complementing split fluorescent proteins (split FP1-10/11) have become an important labeling tool in live-cell protein imaging. However, current split FP systems to label multiple proteins in single cells have a fundamental limitation in the number of proteins that can be simultaneously labeled. Here, we describe an approach to expand the number of orthogonal split FP systems with spectrally distinct colors. By combining rational design and cycles of directed evolution, we expand the spectral color palette of FP1-10/11. We also circularly permutate GFP and synthesize the β-strand 7, 8, or 10 system. These split GFP pairs are not only capable of labeling proteins but are also orthogonal to the current FP1-10/11 pairs, offering multiplexed labeling of cellular proteins. Our multiplexing approach, using the new orthogonal split FP systems, demonstrates simultaneous imaging of four distinct proteins in single cells; the resulting images reveal nuclear localization of focal adhesion protein Zyxin. Tamura et al. expand the split fluorescent protein system to multi-colour imaging (six-colours). Their multiplexing approach, using the orthogonal split FP systems, allows simultaneous imaging of four distinct proteins in single cells.

I n the self-complementing split GFP system, super-folder GFP is split between β-strands 10 and 11, rendering 214-amino acid and 16-amino acid fragments 1 . The short fragment, GFP 11M3 OPT , acts as an epitope tag when inserted into a gene of interest 2 . When expressed in the same cell, the GFP 1-10 D7 and GFP 11M3 OPT fragments (hereafter referred to as GFP 1-10/11 ) spontaneously interact with each other to form a functional GFP (Supplementary Fig. 1). The GFP 11 fragment has been used in numerous biological studies 3,4 : targeting nanomaterials in cells 5,6 , forming protein oligomeric structures 2,7 , verifying aggregation processes 8 , and imaging protein localization in living cells 9 . By introducing additional substitutions into the GFP 1-10 fragment, cyan and yellow spectral variants were previously created and used to visualize localization patterns of cellular proteins 2,10 . The majority of substitutions which lead to the spectral shifts in these variants are located within the large fragments (i.e., CFP 1-10 and YFP [1][2][3][4][5][6][7][8][9][10]. These fragments retain the ability to bind to the identical GFP 11 fragment, so that reconstitution with GFP 11 produces a functional cyan or yellow FP. These self-complementing split GFP variants have already become a powerful and versatile tool for various imaging applications. In particular, endogenously tagged cell lines can be produced by the efficient introduction of the short fragment (GFP 11 ) into a genomic locus without perturbing local genomic structure 2,11 . Additionally, we have been able to generate a library of human cells with GFP 11 -tagged endogenous proteins via CRISPR/Cas9-mediated homology-directed repair (HDR), and demonstrate that GFP 11 -tag is compatible with a wide range of cellular proteins such as enzymes, receptors, transport proteins, and structural proteins 12 . However, labeling multiple proteins simultaneously in single cells has been challenging. Multiplexed visualization is tremendously beneficial for simultaneous comparisons of protein dynamics. Recently, great advances have been made in split super-folder Cherry (sfCherry 1-10/11 ) as a second, orthogonal split FP system 2,13,14 . The GFP 11 and sfCherry 11 fragments allow simultaneous labeling of two different proteins. Although this multicolor approach has expanded the potential of split-FP labeling, it has a bottleneck in multiplexing because of the limited number of available orthogonal split FP systems with different colors.
In this report, we expand the color palette of self-associating split FPs. We have introduced rational mutations into the amino acid sequence of EBFP2 through site-directed mutagenesis and generated two blue-colored split FPs, EBFP2 1-10/11 , and Capri 1-10/ 11 . We have also engineered self-associating fragments of mRuby3 (mRuby3 is a red-colored FP with a shorter-wavelength than sfCherry) 15 . We have evolved mRuby3 1-10 by a directed evolution strategy to increase its complementation with mRuby3 11 . Our final optimized construct, split mRuby4, becomes a fusion pair when expressed in human cells. In addition, we propose a new approach to generate more orthogonal split FPs using circularly permutated FP fragments. This approach can potentially overcome multiplexing limitations of split-FP labeling. Finally, as a proof-of-concept experiment, we applied our technique to visualize differential distribution of four proteins in single human cells and found that focal adhesion protein Zyxin sometimes accumulated in the nucleus.

Results and discussion
Rationally designed variants of split BFP and CFP. To expand our color palette of split FPs, we split EBFP2 at the same site as GFP 1-10/11 (note that EBFP2 is 4-fold brighter and >500-fold more photostable than EBFP 16 ). While the short fragment is identical to the amino acid sequence of GFP 11 , six substitutions have been introduced into the large fragment through site-directed mutagenesis (N40I/T106K/E112V/K166T/I167V/S206T; the numbering of amino acids follows that of EBFP2). These substitutions have been previously shown to enhance complementation efficiency of GFP 1-10 variants 1 . To verify in vivo complementation between the two fragments, we used GFP 11tagged β-actin and histone 2B. Co-expressing each one with EBFP2 1-10 in HeLa cells, we observed blue fluorescence in images of the actin stress fibers and the nucleoplasm (Fig. 1a, b). In some cases (e.g., the actin image), autofluorescence limits the usefulness of this split construct because its overall fluorescent signal is extremely weak. In fact, high autofluorescence background with UV light is often observed in the perinuclear region (Supplementary Fig. 2). To improve its overall brightness, we decided to add six more substitutions to EBFP2 1-10 (S65T/Q80R/F99S/ V128T/M153T/V163A; some of these have previously been characterized to promote the stability and folding rate of GFP 1,17 ). This new split FP, termed split Capri for its cyan-blue color, has the same absorption spectrum as split EBFP2 (Supplementary Fig. 3). The emission spectrum, however, is redshifted from split EBFP2 by 20 nm (λ abs /λ em = 379/469 nm). Furthermore, its peak extinction coefficient of 37,300 M −1 cm −1 and quantum yield of 0.13, are greater than those of split EBFP2 (Supplementary Table 1). When associated with GFP 11 -tagged βactin or H2B, Capri 1-10 exhibits very bright fluorescence in HeLa cells (Fig. 1c, d). To assess the improvement in the resulting brightness, we co-expressed GFP 11 -H2B in HEK 293T cells with either EBFP2 1-10 or Capri 1-10 . Quantifying the fluorescence intensity of cells by flow cytometry, we found that Capri 1-10/11 had a four-fold brighter fluorescence than EBFP2 1-10/11 (Supplementary Fig. 4).
In addition to BFP variants, cyan-colored FPs have been widely studied. When we introduced substitutions into GFP 1-10 (Y66W to make CFP 1-10 ), complementation fluorescence was observed for GFP 11 -β-actin or GFP 11 -H2B fusions co-expressing with CFP 1-10 in HeLa cells ( Supplementary Fig. 5). However, it is noticeable in the figure that the overall brightness of CFP 1-10 is relatively weak, making it difficult to visualize thin actin filaments (A recent in vitro assessment also reported that split CFP has a low brightness 10 ). Therefore, we sought to produce a cyancolored FP that has enhanced brightness. A recent improvement of full-length CFP, named Cerulean, increases the brightness bỹ 1.6 times 18 . Because the known substitutions are located only on GFP 1-10 (Y66W/S72A/Y145A/H148D for Cerulean), Cerulean 1-10 can associate with GFP 11 . To evaluate the enhancement in its overall brightness for cellular microscopy, we prepared plasmids encoding Cerulean 1-10 or CFP 1-10 . We co-transfected each one of these plasmids with a GFP 11 -H2B plasmid in HEK 293T cells. Imaging by confocal microscopy, we quantified the signal level of these split FPs. We found that Cerulean 1-10 signal was~1.7 times brighter than that of CFP 1-10 (p < 0.0001, Student's t-test; Supplementary Fig. 6). We next assessed the performance of Cerulean 1-10 when used as a fusion tag. We used GFP 11 -fused βactin or H2B and co-expressed each one with Cerulean 1-10 in HeLa cells (Fig. 1e, f). Although we observed complementation of Cerulean 1-10 in the appropriate locations, some cells exhibited thicker actin bundles, which we have never seen in cells expressing a full-length Cerulean fusion ( Fig. 1e and Supplementary Fig. 7). Because this artifact is common for dimeric or tetrameric FPs when they are targeted to two-dimensional structures 19,20 , we suspect that Cerulean 1-10/11 is an oligomeric split FP. Nonetheless, an in-depth investigation is required to validate such a property in split Cerulean and under more various experimental conditions.
Engineering of a red-colored split FP variant based on mRuby3. Although developmental efforts are ongoing to improve the brightness of split sfCherry 2,13,14 , having spectrally distinct split red FPs would foster the gross usefulness of FP 11 -tags. Since split sfCherry2 has a far-red shifted emission peak at 610 nm, we sought to explore the evolution of orange-red FPs such as mKO2, mRuby3, mApple, and mScarlet-I 15,21-23 in E. coli. Following the previously established approach 13 , we inserted a 30 amino-acid spacer between the 10th and 11th β-strand of the four FPs. The long spacer insertion greatly diminished colony fluorescence of mKO2, mApple, and mScarlet-I, while colonies expressing spacer-inserted mRuby3 remained fluorescent ( Supplementary  Fig. 8). To improve the brightness of the spacer-inserted mRuby3, we mutagenized it using error-prone PCR and then transformed into E. coli; the three brightest candidates were pooled and subjected to another round. After three rounds, brightness of the best candidate revamped six-fold relative to that of spacer-inserted mRuby3 ( Supplementary Fig. 9). We found seven substitutions in mRuby3 1-10 (M15T/Q27H/T31I/V106I/S113C/R126S/A154V) and termed this variant split mRuby4 (λ abs /λ em = 557/592 nm; see also Supplementary Fig. 3 for its absorbance spectrum). Compared to split mRuby3, we created a particularly bright variant that has a higher extinction coefficient and increased quantum yield (Supplementary Table 1).
To assess whether split mRuby4 could fluoresce in human cells, we over-expressed mRuby4 11 -β-actin with mRuby4 1-10 in HeLa cells. We observed that complemented split mRuby4 has a bright signal in fluorescent images of actin and various fusion proteins ( Fig. 2a-f). To determine the signal level of split mRuby4, we performed a cellular fluorescence measurement by flow cytometry and compared the signal to full-length mRuby3. With expression of spacer-inserted mRuby4 in HEK 239T cells, we found that its signal level became around 69% of full-length mRuby3 (Supplementary Fig. 10). We have also demonstrated that mRuby4 1-10/11 has sufficient efficiency to detect proteins expressed at endogenous levels. We employed CRISPR/Cas9-mediated HDR and introduced a 200-nucleotide ssDNA donor into the HIST2H2BE locus of HEK 293FT cells expressing mRuby4 [1][2][3][4][5][6][7][8][9][10] . Subsequently, we found that split mRuby4 complementation had a prominent signal in images of the mRuby4 11 knock-in ( Supplementary  Fig. 11).
As shown in Fig. 2g, the emission peaks for split mRuby4 and split sfCherry2 are only 20 nm apart yet still visually distinguishable ( Supplementary Fig. 12). To further evaluate how many split FPs could be simultaneously visualized in different cells, we performed spectral imaging of HEK 293 cells expressing H2B fused proteins (Fig. 2g). We co-cultured six types of HEK 293 cells, each of which expressed H2B labeled with either EBFP2 1-10/ 11 , Capri 1-10/11 , Cerulean 1-10/11 , GFP 1-10/11 , mRuby4 1-10/11 , or sfCherry2 1-10/11 . After we synchronized at the G2/M phase by release from a cyclin-dependent kinase inhibitor, we imaged the cells (Supplementary Fig. 13). Within a couple of hours, 20% of the cell population was in cytokinesis, which is consistent with previous literature 24 . We then captured cells with each split FP fusion at different stages of mitosis (Fig. 2h). Overall, these experiments illustrate six-color spectral imaging of cellular proteins.
A strategy to create new orthogonal split FPs using circularly permutated FP fragments. In order to provide more variants of split FPs orthogonal to existing FP 1-10/11 , we took advantage of circularly permutated GFP variants 25 . By linking the N-termini and C-termini and cutting out a single β-strand, any one of the eleven β-strands could be a split GFP-tag. We chose to measure complemented signal of the β-strands 7, 8, 9, and 10. (The βstrands 1-6 were excluded because the complementary fragments of these strands are unlikely to be water-soluble 26 ). To this end, we prepared DNA constructs encoding each of the β-strands fused to β-actin and measured the overall complemented signal of each construct in HEK 293T cells by flow cytometry. We observed fluorescence signal reconstituted from the β-strands 7, 8, and 10 (hereafter named GFP 8-6/7 , GFP 9-7/8 , and GFP 11-9/10 ) with their corresponding partners (Fig. 4a). These split GFPs retained 7-57% brightness of GFP 1-10/11 , albeit leaving room for improvement. To validate protein labeling using the β-strands, we generated constructs encoding various cellular proteins fused with GFP 8 and co-expressed each one of them with GFP 9-7 in HeLa cells. For three proteins tested, we observed their expected localizations (Fig. 4b-d).
Next, we assessed the binding specificities of GFP 8-6/7 , GFP 9-7/8 , GFP 11-9/10 , and GFP 1-10/11 . We performed the flow cytometry assay conducted in a grid format as described earlier. Either GFP  , GFP 9-7 , GFP 11-9 , or GFP 1-10 was co-expressed with β-actin fused with the β-strands 7, 8, 9, or 11 in HEK 293T cells. In this experiment, each of the β-strands only binds to its corresponding partner (Fig. 4e). For instance, GFP 8 interacts with GFP 9-7 , but not GFP 1-10 . This orthogonal interaction was validated by dual-color imaging of U2OS cells, in which GFP 11 -H2B and GFP 8 -Lamin A/ C were co-expressed with Capri 1-10 and GFP  . We observed the exclusion of GFP 8 -Lamin A/C from the nucleoplasm where GFP 11 -H2B predominately localized (Fig. 4f). Taken together, circularly permutated FP fragments can be used to generate additional orthogonal pairs for multiplexed split FP-labeling. (h) HEK 293 cells expressing H2B labeled with EBFP2 1-10/11 , Capri 1-10/11 , Cerulean 1-10/11 , GFP 1-10/11 , mRuby4 1-10/11 , or sfCherry2 1-10/11 were co-cultured in the same plate. Spectrally unmixed images at the different stages of mitosis are represented (see also Supplementary Fig. 14). Scale bars,10 μm. Fig. 3 Characterizing the binding specificities of available FP 1-10/11 pairs. a Characterizing the binding specificities of GFP 1-10/11 , sfCherry2 1-10/11 , mNeonGreen2 1-10/11 , and mRuby4 1-10/11 by flow cytometry (see also Supplementary Fig. 15). Each of the FP 11 fragments was tested for complementation to all of the FP 1-10 fragments. Complementation is indicated as the percentage of fluorescent cells by a color scale and the number in each block. b, c Dualcolor fluorescence images of HeLa cells expressing GFP 1-10/11 -β-actin and Zyxin-mRuby4 1-10/11 (b), and mNeonGreen2 1-10/11 -β-actin and mRuby4 1-10/11 -Clathrin (c). d This dendrogram is based on the similarities of the following fluorescence protein sequences: EBFP2, Capri, Cerulean, CFP, GFP, mNeonGreen2, mRuby4, and sfCherry2. Proteins that share sequences are separated by smaller branch lengths. Scale bar, 20% dissimilarity. The dendrogram was constructed using MEGA 7.0 software. Scale bars, 10 μm. Multicolor images reveal nuclear localization of Zyxin. Finally, we assessed the potential of split FP systems for multiplexed labeling of proteins in single cells. As a proof-of-principle, we used four orthogonal split FPs that we thoroughly investigated in this report (Capri 1-10/11 , GFP 9-7/8 , mNeonGreen2 1-10/11 , and mRuby4 1-10/11 ), which are distinguishable from each other by using spectral imaging methodology (Fig. 4g). U2OS cells were transfected to express these split FPs targeted to four distinct proteins (H2B, Lamin A/C, β-actin, and Zyxin), and we observed their correct localization. Strikingly, a few cells displayed some portion of Zyxin proteins localized to the nucleus, although the proteins predominantly localized at focal adhesions in these cells. By visual inspection of a total of 145 cells, we found that 37 of these cells exhibited nuclear localization of Zyxin during interphase ( Supplementary Fig. 18). Because Zyxin is a relatively large molecule (>80 kDa) but does not have a nuclear localization signal, Zyxin must enter the nucleus in contact with other components. We observed a similar nuclear localization pattern of Zyxin tagged with a full-length FP tag in U2OS cells (50 out of 174 cells), and found that this observation has also been confirmed in other cell lines [27][28][29] .
For the initial demonstration of multiplexed labeling, split FPs were over-expressed as fusions to target proteins in single cells (Fig. 4g). However, these fusion proteins might be subject to limitations because of the potential for overexpression artifacts (e.g., aberrant organelle and/or cellular morphology). To further verify our observation in the future, this approach will be extended to label endogenous proteins by methods such as CRISPR/Cas9-mediated gene knock-in 12 . Because a split FP tag is 42-63 nucleotides long (which is~10 times smaller than the size of an intact FP), a short donor oligo can be directly synthesized, making this a cloning free approach (see also Supplementary  Fig. 11) 2,12 . In addition, a small tag such as GFP 11 -tag can be introduced into a host cell genome at high homologous recombination efficiencies 11 . Such a simple and efficient approach would facilitate the generation of multiple insertions in single cells. Thus, the sequences for multiple split FP tags could then be inserted either sequentially or simultaneously into targeted loci in individual cells stably expressing the complementary fragments, enabling multiplexed visualization of endogenous proteins.
GFP permits circular permutation of the amino acid sequence 25 . By linking the N-and C-termini and cutting out a single β-strand, any one of the eleven β-strands would become a new split GFP-tag. Huang et al. 26 previously investigated all possible β-strands, and measured the solubility and relative reconstituted fluorescence intensity of each split GFP construct in E. coli. We tested Huang's design of the β-strands 7, 8, 9, or 10 system in human cells. We constructed plasmids encoding each of the β-strands fused to β-actin. We used mEmerald-Actin-C-18 (Addgene #53978) as the template for PCR amplification of ACTB. The ACTB gene was amplified using primers, in which DNA sequences encoding the βstrands were included in part (For the sequence information of the primers, see Supplementary Data 3). The resultant PCR products were cloned into the KpnI/ EcoRI sites of pcDNA3.1.
Mutagenesis and screening of libraries. When engineering split orange-red FP variants, we adopted a complementation assay previously described to optimize split mCherry2 in E. coli 13 . We inserted a 30-aa spacer (GGGGSEGGGSGGPGSG GEGSAGGGSAGGGS) between the tenth and eleventh β strands of each of the following fluorescent proteins; mKO2, mRuby3, mApple, and mScarlet-I 15,21-23 . The corresponding DNA sequences were directly synthesized and then cloned into the BamHI/XhoI sites of the E. coli expression vector pET28a (Novagen).
The longer spacer insertion eliminated colony fluorescence of mKO2, mApple, and mScarlet-I, whereas colonies expressing spacer-inserted mRuby3 gave low signal ( Supplementary Fig. 8). To improve the brightness of spacer-inserted mRuby3, we mutagenized it by using a GeneMorph II Random Mutagenesis Kit (Agilent). Mutants were expressed and screened in pET28a. Plasmids were transformed into E. cloni EXPRESS Electrocompetent Cells (Lucigen). Transformation was performed by the Gene Pulser Electroporation Systems (BioRad). Colonies were grown on LB agar media (30 μg/mL Kanamycin) at 37°C for 24 h and for additional 12-48 h at 37°C after induction with 1 mM IPTG. For each round of mutagenesis, the number of colonies screened was at least 1 × 10 4 . Colonies expressing spacer-inserted mRuby3 variants were screened for fluorescence with the ChemiDoc Imaging System (BioRad). The imaging system was equipped with an Epi-green 520-545 nm excitation source, a Green Epi 605/50 filter, and a cooled CCD camera.
Through library screening, we obtained an extremely bright variant of spacerinserted mRuby3, which we named spacer-inserted mRuby4 ( Supplementary  Fig. 9). The mRuby4 1-10 sequence of spacer-inserted mRuby4 was amplified by PCR and cloned into the KpnI/EcoRI sites of pcDNA3.1. For the nucleotide sequence of mRuby4 1-10 , see Supplementary Data 1.
Protein production and characterization of FPs. For spectral characterization of FPs, we produced and purified recombinant proteins: full-length EBFP2, spacerinserted EBFP2, spacer-inserted Capri, full-length mRuby3, spacer-inserted mRuby3, and spacer-inserted mRuby4 (the amino acid sequences of those were listed in Supplementary Data 2). We designed pET plasmids such that recombinant proteins were labeled at the C termini with poly-histidine tags. The plasmids were introduced into BL21(DE3) Competent E. coli cells (NEB) via transformation. Cells were grown in 250 mL LB medium at 37°C for 6 h (OD 600 = 0.5), induced with IPTG (1 mM) for 4 h, and harvested by centrifugation. Cell pellets were lysed by French press. His-tagged proteins were purified with HisPur Cobalt Resin (Pierce). Proteins were further desalted into PBS pH7.4 using a GE Healthcare illustra NAP column (GE Healthcare). Extinction coefficients were calculated using Beer-Lambert law 10 . Quantum yields were determined using EBFP2 16 , and Rhodamine B (Wako) as reference fluorophores. The absorbance signals of samples and reference were measured using a microreader (Biotek Synergy 2). Diluted samples and reference were added into a quartz fluorescence cuvette (Thorlabs), and their integrated fluorescence intensities were measured by a fluorescence spectrophotometer (Hitachi F-7100). With the quantum yield of reference to be known, the final quantum yields of samples were attained using: Qs ¼ Qr ðAr=AsÞ ðEs=ErÞ ðns=nrÞ 2 ½r and s refer to the reference and samples; where Q is the quantum yield, n is the refractive index, A is the absorbance of solution, and E is the integrated fluorescence intensity of emitted light.
For spectral imaging of multicolor H2B fusions ( Fig. 2h and Supplementary  Figs. 13 and 14), we used a Cdk1 inhibitor (10 μM of RO-3306, Sigma-Aldrich) to synchronize HEK 293 cells. HEK 293 cells were treated with the inhibitor for 18 h, blocked in the G2/M phase. For release from the inhibitor, we washed the culture five times with prewarmed culture media. Released cells returned to normal cell cycle progression, and were eventually fixed with 100% ice-cold methanol and mounted with PBS for microscopy.
Knock-in cell creation. For knock-in of mRuby4 11 into the HIST2H2BE locus, we ordered 200-nt HDR templates in single-stranded DNA (5′-gcccggcgagctggccaagca cgccgtgtccgagggcaccaaggcggtcaccaagtacaccagctccaagGGTGGCGGCGAAACCTAC GTAGTGCAAAGAGAAGTGGCAGTTGCCAAATACAGCAACtgagtccctgccggga cctggcgctcgctcgctcgagtcgccggctgcttgactccaaaggctcttttcagag-3′, Integrated DNA Technologies). Cas9 protein was expressed in E. coli and purified by the Kipreos laboratory at UGA as described previously 11 . sgRNA and Cas9/sgRNA ribonucleoprotein complexes were prepared as described before 12 . After the treatment of HEK293 FT cells with nocodazole (200 ng/mL, Sigma-Aldrich) for 16 h, we performed electroporation on an Amaxa Nucleofector 2b device with Nucleofector Solution V reagents (Lonza).
Nocodazole-treated cells were resuspended at a concentration of 1 × 10 4 cells/μL in 100 μL of Nucleofector Solution V. We added cells to the RNP/donor template mixture (50 μL), electroporated using the Q-001 program, and quickly transferred to 12-well plates with pre-warmed media. Electroporated cells were cultured for 2-5 days and transfected with mRuby4 1-10 plasmid.
Statistics and reproducibility. All experiments for the measurement of signal levels were replicated multiple times independently. Statistical analyses were performed using GraphPad Prism 7. Error bars in all figures refer to the standard error of the mean.