A high-throughput colocalization pipeline for quantification of mitochondrial targeting across different protein types

Abstract

Efficient metabolic engineering and the development of mitochondrial therapeutics often rely upon the specific and strong import of foreign proteins into mitochondria. Fusing a protein to a mitochondria-bound signal peptide is a common method to localize proteins to mitochondria, but this strategy is not universally effective with particular proteins empirically failing to localize. To help overcome this barrier, this work develops a generalizable and open-source framework to design proteins for mitochondrial import and quantify their specific localization. This Python-based pipeline quantitatively assesses the colocalization of different proteins previously used for precise genome editing in a high-throughput manner to reveal signal peptide-protein combinations that localize well in mitochondria.

Graphical Abstract

INTRODUCTION

Synthetic biologists increasingly leverage natural mitochondrial protein import pathways for compartmentalized metabolic engineering^1-4 and the development of molecular therapeutics^5-8. These approaches require efficient targeting of proteins of interest (POIs) to the mitochondria. The most common strategy to achieve such localization is by fusing a mitochondrial targeting sequence (MTS), typically a short and positively charged signal peptide, to the N-terminus of the POI. MTSs are recognized by translocases on the outer and inner mitochondrial membrane (TOM/TIM23 complex) that import it through both mitochondrial membranes before the protein is released into the mitochondrial matrix following cleavage of the N-terminus MTS from the POI^9,10. Hundreds of MTSs have been identified from natural proteins using proteomic approaches^11,12.

However, attachment of an individual MTS to a given POI does not always guarantee efficient import into mitochondria^5,13-17. Instead, researchers often empirically test multiple MTSs before finding one that results in their specific POI localizing in mitochondria^18,19. To help establish a more rapid and quantitative assessment of mitochondrial localization in mammalian cells, we developed a high-throughput imaging-based pipeline to measure colocalization of POI with mitochondria that we validated by screening combinations of three different N-terminus MTSs and POIs from five protein families relevant to mitochondrial gene editing.

RESULTS

High-throughput Microscopy Workflow Enables Visualization of Single-Cell Localization

To rapidly assess the localization of different combinations of MTS and POI, we developed a high-throughput microscopy workflow and analytical pipeline to quantify how efficiently different engineered proteins localize within mitochondria or nuclei.

We first generated 66 protein cassettes containing a combination of localization signal and POI, followed by a HA tag on the C-terminus. Localization signals included three MTSs derived from proteins that naturally localize to the mitochondria—COX8 (human cytochrome c oxidase subunit VIII)²⁰, Su9 (Neurospora crassa ATPase subunit 9)²¹, and ATG4D (Atg4 cysteine protease)²². We chose these MTSs specifically to test the constraints of our pipeline; while both COX8 and Su9 are commonly used MTSs^19,23-26, ATG4D is a cryptic MTS²² that we hypothesized would have a lower targeting potential than the other MTSs. The use of no localization signal or a nuclear localization signal (NLS) served as negative controls. Nineteen POIs were chosen across five different protein classes that have been used as components of precise gene editing technologies: Class I CRISPR systems (Cas3/CASCADE)²⁷, Class II CRISPR/Cas nucleases²⁸, RecTs, single-stranded binding proteins (SSBs)^29-32, and retron reverse transcriptases (RTs)^33-36. As a positive control, we used the construct mito-APEX2, a protein that contains an MTS derived from the mitochondrially imported COX4 fused to APEX2, which has been shown to localize to the mammalian mitochondrial matrix using immunocytochemistry and proteomic mapping that found mito-APEX2 in close proximity to mitochondrial matrix proteins³⁷.

To engineer mammalian cell lines expressing a given protein cassette, each construct was cloned into a PiggyBac vector under the control of a doxycycline-inducible promoter adjacent to a constitutive puromycin resistance gene. These cassettes were randomly integrated into the genome of HEK293T cells using the PiggyBac transposase system and selected with puromycin (Fig 1a). Biological replicates of a given construct were defined as either individual clones derived from a single bulk transposase integration or multiple parallel transposase integrations. We screened the localization of each cassette by seeding cells into 96-well plates, expressed each protein for 24 hours under an inducible promoter, performed immunocytochemistry, and imaged each well using a high-throughput confocal microscope (ImageXpress Micro Confocal High-Content Imaging System). Specifically, each cell line was imaged for nuclei using Hoescht, POI using an antibody against HA, and mitochondria using an antibody against the mitochondrial marker TIM23 (Fig 1b).

Figure 1.

High-throughput microscopy workflow enables visualization of single-cell localization. (a) 66 genetic cassettes containing a combination of localization signal and POI, followed by a C-terminus HA tag, were synthesized then randomly integrated into the genome of HEK293T cells using a piggyBac transposase system. Following co-transfection of the cassette in a piggyBac vector and a plasmid constitutively expressing piggyBac transposase, cells were selected for at least one week using the antibiotic puromycin. To image each cell line, expression of a given cassette was induced for 24 hours using doxycycline before being fixed and stained prior to imaging using a high-throughput confocal microscope (ImageXpress Micro Confocal High-Content Imaging System). (b) Engineered cell lines were stained with Hoescht (blue), an antibody against HA (green), and an antibody against the mitochondrial marker TIM23 (red). Shown are representative images from the positive control (red background) and LbCas12a fused to a MTSs(purple), NLS (blue), or no localization tag (white). (c) A machine learning algorithm was implemented to automate labeling of cells containing cassette protein. Scale bar = 25 μm. Unprocessed images from each fluorescent channel are fed into a Python-based analysis pipeline to segment individual cells. Shown under “Experiment” are representative images from the cell line Su9-LbCas12a (top, HA; middle; Hoescht, bottom; TIM23). The Hoescht and TIM23 channels are merged (“Segment,” top) prior to being fed into a custom, retrained neural network (arrow), resulting in automated labeling of each cell found within the image. Images below arrow in “Segment” column; top image shows Cellpose-generated mask of segmented cells, where each yellow line indicates a cell boundary. Bottom image in “Segment” column shows mask overlaid atop merged Hoescht/TIM23 channels. Cellpose-segmented cells are filtered to keep only cells which express the cassette protein. Otsu thresholding is applied to the HA channel (“Filter,” top) to determine a threshold separating true fluorescent signal from background noise. Segmented cells containing at least 50 pixels with signal are kept to perform further colocalization measurements. Images below arrow in “Filter” column; top image shows segmented cells following filtering. Bottom image shows filtered mask overlaid atop HA channel.

Crucially, we found that expression and localization were variable between individual cells of a given condition so we built an analysis pipeline that quantifies colocalization at the level of single cells (Fig 1c). Images corresponding to the nuclei and mitochondria for each biological replicate cell line from each condition were fed into a Cellpose-based machine learning model^38,39 to label individual cells. Next, cells were filtered using Otsu thresholding to remove any cells with no detectable protein expression.

To ensure that images with such low fluorescence that signal was effectively indistinguishable from noise did not bias the final colocalization scores, any images in which Otsu thresholding did not separate signal from noise in each cell, as defined by the majority of filtered cells in an image failing to show a non-Gaussian intensity distribution typical of true fluorescent signal, were discarded (Fig S1). In some cases, only a few images for a cell line were eliminated, although—in cases where a clonal or transfected line suffered from minimal cassette expression—the entire biological replicate was removed from analysis.

Automated, quantitative, and open-source analysis pipeline quantifies colocalization of cassette proteins with mitochondria or nuclei

To better compare localization differences exhibited by MTS-POI combinations, we developed an unbiased Python-based analysis pipeline to quantify the mitochondrial and nuclear import between our dozens of cassettes.

After selecting a population of filtered cells to further analyze for a given protein cassette, colocalization between the cassette protein and either mitochondria (Fig 2a,b) or nuclei (Fig 2d,e) was measured on a per cell basis using Pearson’s correlation coefficient (PCC)⁴⁰. High colocalization scores indicate that a protein is collocated with a given organelle while low colocalization scores suggest little to no specific colocalization between a POI and a given organelle occurred.

Figure 2.

Automated, quantitative, and open-source analysis pipeline quantifies colocalization of cassette proteins with mitochondria or nuclei. (a) Representative image of a cell line (Su9-LbCas12a) with clear mitochondrial expression of its cassette protein, with mask in yellow overlaid on top. Numbers refer to three representative cells for which data is shown in (b). Scale bar = 25 μm. (b) Heatmaps depicting the relationship between Su9-LbCas12a pixel intensity and organellar pixel intensity (mitochondria on top in red; nuclei on bottom in blue) for each pixel within a representative cell from (a) (left; cell #1, middle; cell #2, right; cell #3). Color depicts the number of pixels. The strength of the linear relationship between pixel intensities, or colocalization, within each cell is calculated using PCC, and the result depicted on top its respective heatmap. (c) Histogram depicting the all the colocalization scores for all the cells for one clonal line expressing Su9-LbCas12a. Mitochondrial PCC scores are shown in red, while nuclear PCC scores are shown in blue. Dotted lines depict the median colocalization score for mitochondria (red) and nuclei (blue). (d) Representative image of a cell line (NLS-LbCas12a) with clear nuclear expression of its cassette protein, with mask in yellow overlaid on top. Numbers refer to three representative cells. Scale bar = 25 μm. (e) Heatmaps depicting the relationship between NKS-LbCas12a pixel intensity and organellar pixel intensity (mitochondria on top in red; nuclei on bottom in blue) for each pixel within a representative cell from (d) (left; cell #1, middle; cell #2, right; cell #3). Color depicts the number of pixels. The strength of the linear relationship between pixel intensities, or colocalization, within each cell is calculated using PCC, whose result is on top its respective heatmap. (f) Histogram depicting the all the colocalization scores for all the cells for one clonal line expressing NLS-LbCas12a. Mitochondrial PCC scores are shown in red, while nuclear PCC scores are shown in blue. Dotted lines depict the median colocalization score for mitochondria (red) and nuclei (blue). (g) Mitochondrial colocalization of positive control mito-APEX2 (red) and LbCas12a fused to different localization tags, as measured using the described experimental and analytic workflow. There is a significant effect of localization signal (one-way ANOVA, P<0.0001), where ATG4D (P=0.0021), no signal (P<0.0001), and NLS (P<0.0001) are all significantly different from mito-APEX2, but Su9 (P=0.9831) and COX8 (P=0.376) are not (Dunnett’s corrected). Open circles are biological replicates; closed circles are average of all biological replicates. (h) Nuclear colocalization of mito-APEX2 (red) and LbCas12a fused to different localization tags measured using the described experimental and analytical workflow. There is a significant effect of localization signal (one-way ANOVA, P<0.0001), where NLS (P<0.0001) is significantly different from mito-APEX2, but Su9 (P=0.8587), COX8 (P=0.9930), ATG4D (P=0.9998), and no signal (P=0.0.8825) are not (Dunnett’s corrected). Open circles are biological replicates; closed circles are average of all biological replicates. Additional statistical details in Supplementary Table 1. (i) Mitochondrial (red) and nuclear (blue) colocalization scores of 66 different protein cassettes. Proteins were ranked based on average mitochondrial colocalization score, from highest mitochondrial PCC to lowest. Open circles are biological replicates; closed circles are average of all biological replicates.

For our analysis, we considered individual cells that survived quality filters from a single transfection or clone as technical replicates and summarized the overall colocalization score for a single biological replicate of each protein cassette by taking the median of all the individual cell colocalization scores for mitochondria (Fig 2c) or nuclei (Fig 2f). We replicated our experiments using at least three different transfections or five clonal lines as biological replicates.

We generally found low variability within our biological replicates, suggesting that protein import is a fairly reliable phenomenon. The positive control mito-APEX2 obtained an average score of 0.63 +/− 0.03 (mean +/− std. dev), a high PCC value that strongly implies mito-APEX2 is imported into the mitochondrial matrix. Similar to our previous qualitative assessments (Fig 1b), we found that colocalization scores, even across a single POI, vary depending on localization signals. The colocalization scores of LbCas12a fused to the Su9 or COX8 MTSs were not statistically different from mito-APEX2, suggesting mitochondrial import had occurred. In comparison, when LbCas12a was instead fused to ATG4D, no localization signal, or NLS, colocalization scores dropped significantly, indicating less mitochondrial import occurred (Fig 2g). Moreover, when comparing nuclear colocalization scores (Fig 2h), all cell lines except NLS-LbCas12a showed a consistent, low nuclear colocalization score, suggesting little to no nuclear import. As expected, only the cell line fused to a NLS had a high nuclear colocalization score indicating high nuclear import.

After validating the analytical pipeline using LbCas12a, we used this workflow to quantify the mitochondrial and nuclear import of all 66 different protein cassettes (Fig 2i; S2). Interestingly, mitochondrial colocalization scores did not cluster bimodally into high and low scores. Instead, scores were distributed continuously, suggesting that different cassettes have varying capabilities to drive POIs to the mitochondria. These findings suggest that our workflow is able to compare import efficiencies across different combinations of MTS and POI for multiple organelles.

DISCUSSION

Here, we design a high-throughput imaging-based workflow to quickly screen the subcellular localization of a tagged protein. This method enabled us to determine the best MTS-POI combination across three commonly used MTSs and five different protein classes that are components of gene editing technologies. Our pipeline accurately quantifies the efficiency of localization of cassette proteins that almost uniformly are imported into mitochondria, cassette proteins that are only partially imported into mitochondria, and cassette proteins where there is no mitochondrial import. One advantage of our pipeline is that it avoids additional burdensome control experiments, such as expressing a cytosolic variant of each cassette protein to measure overall expression level, by relying on intensity-insensitive metrics. For example, our pipeline uses PCC, a metric that controls for a given signal level, to measure colocalization⁴⁰.

As a tool for the field, our Python-based workflow implemented in an annotated Jupyter notebook can be reused for future experiments, including more generally to other screens using different proteins, organelles of interest, or cell lines. Although we used a high-throughput confocal microscope in our own workflow, other confocal microscopes could easily be used, depending on the necessary throughput or number of samples. Computationally, the analytic pipeline uses Python, an open-source programming language, and relies on pixel-based colocalization analyses and quality check steps to eliminate cells not expressing a protein of interest that can be universally applied regardless of differing expression levels or phenotypes⁴⁰. Therefore, others should be able to easily apply this analytical framework to quantify colocalization within their own fluorescent images.

Methods

Constructs and strains

Protein cassettes were constructed by amplifying localization signals and POI nucleotide sequences using PCR from synthesized gBlocks (IDT) or existing plasmids. Complete protein cassettes were cloned into a PiggyBac integrating plasmid for doxycycline-inducible human protein expression (TetOn-3G promoter) using Gibson assembly. Alternatively, some cassettes were synthesized into the same custom PiggyBac integrating plasmid by Twist Bioscience (see Supplementary Table 2).

Stable mammalian cells for imaging were generated using the standard Lipofectamine 3000 transfection protocol (Invitrogen) and a PiggyBac transposase system. T12.5 flasks with 50-70% confluent HEK293T cells were transfected using 1.6 μg POI cassette expression plasmid and 0.8 μg PiggyBac transposase plasmid (pCMV-hyPBase). Stable cell lines were selected using puromycin for at least one week.

Clonal lines were generated by growing individual cells into separate cell populations. Specifically, stable cell lines were serially diluted to a final concentration of 2.5 cells per mL media then seeded into a 96-well plate using 100 μL/well. Wells that received a single cell had media refreshed weekly until a clonal line proliferated to ~40% confluency, at which point a clonal line was passaged to a larger flask for further experiments.

Immunocytochemistry

96-well glass bottom plates with #1 cover glass (Cellvis, catalog # P96-1-N) were coated with a mixture of 50% poly-D-lysine (ThermoFisher Scientific, catalog #A3890401) and DPBS (ThermoFisher Scientific, catalog #14040133) for 30 minutes at room temperature. Wells were washed three times with distilled water and left out to dry for at least 2 hours prior to seeding.

Cells were seeded at a density of 10,000 cells per well. The following day, doxycycline was added at a final concentration of 1 μg/mL to induce expression of the protein cassette. At 24 hours post-induction, cell nuclei were stained using a final concentration of 10 μM Hoescht for at least 5 minutes prior to fixation.

For fixation, media was aspirated from each well and replaced with a solution of 4% paraformaldehyde (PFA) created fresh by fixing a 1 mL 16% (w/v) PFA ampule (ThermoFisher Scientific, catalog #28906) with 3 mL PBS. Cells were fixed for 30 minutes at room temperature prior to three 5-minute washes with PBS. Following fixation, cells were permeabilized and blocked for an hour at room temperature using blocking buffer made fresh with the following ingredients: PBS containing 10% donkey serum (Sigma-Aldrich, catalog #D9663), 10% Triton X-100 (Sigma-Aldrich, catalog #X100), and 100 mg BSA (Sigma-Aldrich, catalog #A9418) per 10 mL solution. Next, cells were incubated overnight at 4°C in blocking buffer with the antibodies anti-HA tag conjugated to DyLight 550 (ThermoFisher Scientific, catalog #26183-D550) and anti-TIM23 (Abcam, catalog #ab230253) each added at a 1:100 dilution. After performing three more 5-minute washes, cells were incubated with a secondary antibody goat anti-rabbit conjugated to DyLight 650 (ThermoFisher Scientific, catalog #84546) at a 1:500 dilution in blocking buffer for 3 hours. Following secondary antibody incubation, three more 5-minute washes were performed prior to the addition of 30 uL antifade mountant (ThermoFisher Scientific, catalog #S36967) per well.

Plates were wrapped in aluminum foil to avoid light and either stored temporarily at 4°C or at −20°C for longer-term storage prior to imaging.

Imaging

Stained cells were imaged using an ImageXpress Micro Confocal High-Content Imaging System (Molecular Devices) using a 40X water immersion objective by taking a 7-layer Z-stack, with each layer spaced 0.3 μm apart, at four different sites per well.

Colocalization Image Analysis Pipeline

A colocalization image analysis pipeline was made using jupyter-notebook in Python 3^41,42, and uses the following packages: numpy, pandas, scipy, skimage, tdqm, and tifffile. Additionally, the pipeline requires the Cellpose code library^38,39 along with these additional packages: numba, opencv, and pytorch. Using the Cellpose GUI also requires PyQt and pyqtgraph⁴³. Code is available on the Shipman Lab GitHub in the “MtProtein_Coloc” repository.

TIFF files consisting of merged nuclear and mitochondrial channels were created using a custom function and fed into a neural network retrained according to the instructions for the Cellpose GUI³⁹. Briefly, the “CP” model from the Cellpose model zoo was initially used to segment all images. Afterwards, about five images with poor initial segmentation were chosen for manual annotation. The CP model was then retrained using the corrected labels and the new model was re-run on all images.

To remove segmented cells that did not contain expression of the cassette protein, Otsu thresholding was performed on the HA channel to determine a pixel intensity threshold separating signal from background for each image. Only segmented cells containing at least 50 pixels of cassette protein signal, referred to as filtered cells, were kept for further analysis.

Two additional functions to ensure quality-check steps were also implemented. First, any image containing fewer than six filtered cells was automatically removed from further analysis. Second, since the overall expression of a cassette protein can vary between different cell lines, a function was written to ensure that the filtering step effectively distinguished between cells that did or did not express a cassette protein. Individual cells containing noise, rather than signal, exhibit a Gaussian distribution of protein cassette pixel intensities. In contrast, cells with signal tend to exhibit non-normal or skewed pixel intensity distributions. Thus, for every image, a “non-Gaussian” test was performed on each filtered cell by testing for normality. If over 60% of filtered cells failed the “non-Gaussian” test, then this result suggests that the majority of filtered cells within the image do not contain true expression of the cassette protein, thus that specific image would be removed from further analysis.

Afterwards, a custom function was built to calculate PCC between the HA channel pixel intensities and either the mitochondrial or nuclear channel pixel intensities for every filtered cell related to a given biological replicate. Due to the skew present in most PCC distribution, these results were summarized by taking the median of all the filtered cells for a given biological replicate.

Statistics

ANOVA and post-hoc analyses were performed using GraphPad Prism v9.4.1.

ABBREVIATIONS

MTS

mitochondrial targeting sequence

POI

protein of interest

CRISPR

clustered regularly interspaced short palindromic repeats