Organelle landscape analysis using a multi-parametric particle-based method

Organelles have unique structures and molecular compositions for their functions and have been classified accordingly. However, many organelles are heterogeneous and in the process of maturation and differentiation. Because traditional methods have a limited number of parameters and spatial resolution, they struggle to capture the heterogeneous landscapes of organelles. Here, we present a method for multi-parametric particle-based analysis of organelles. After disrupting cells, fluorescence microscopy images of organelle particles labeled with six to eight different organelle markers were obtained, and their multi-dimensional data were represented in intuitive two-dimensional UMAP (uniform manifold approximation and projection) spaces. This method enabled visualization of landscapes of seven major organelles as well as the transitional states of endocytic organelles directed to the recycling and degradation pathways. Furthermore, endoplasmic reticulum–mitochondria contact sites were detected in these maps. Our proposed method successfully detects a wide array of organelles simultaneously, enabling the analysis of heterogeneous organelle landscapes.


Introduction
Organelles separated by biological membranes play vital roles in cellular processes.
Studies on the morphology, composition, and temporal changes of organelles have improved our understanding of the function of each organelle.Organelles are comprised of heterogeneous populations involved in various processes, including division, fusion, organelle-to-organelle contact, and maturation.Even minor components and intermediate structures can influence cellular functions 1 .To understand such heterogeneity of organelles, comprehensive yet intuitive methods are necessary.
Such methods require a sufficient number of parameters, high resolution, and unbiased sampling, but the methods commonly used in cell biology are inadequate.For example, biochemical methods such as immunoisolation and density gradient-based fractionation can analyze only a limited number of organelles at once, with unavoidable contamination between fractions.Fluorescence microscopy can be used to observe multiple labeled organelles, but the high density of cellular components makes it difficult to distinguish whether they are in contact or merely in close proximity 2,3 .Super-resolution microscopy could partially solve the resolution problem, but it is currently limited to four-color imaging 4 .Electron microscopy, particularly threedimensional volume electron microscopy, can provide more detailed information at high resolution but has a limited field of view 2 .Recently, more comprehensive methods, such as machine-learning-based multiple organelle segmentation at the whole cell level, have been developed; however, these methods require special equipment and skills, and the obtained data are difficult to interpret intuitively [5][6][7] .Furthermore, microscopy is generally limited in that only a small number of cells can be observed, leading to potentially biased data.Methods have been proposed to extract organelles as particles from cells and analyze them by flow cytometry [8][9][10][11][12][13][14][15] , which may improve the spatial resolution and allow for more unbiased detection of organelles.However, most of these methods remain limited to analyses of a single type or a few types of organelle because the sensitivity for detecting multi-color fluorescence of small particles is insufficient.
Although mass cytometry has been used for the detection of mitochondria, lysosomes, and autophagosomes, mass cytometers are not widely available 16 .Therefore, there is a trade-off between observing multiple organelles with sufficient resolution and unbiased sampling of organelles.
To overcome these limitations, in this study we developed a simple multiparametric particle-based method.In our proposed method, isolated organelle particles labeled with multiple markers are analyzed by fluorescence microscopy using spectral imaging.The obtained multi-dimensional data are then visualized as intuitive organelle landscape maps using dimension reduction techniques.Using these data, we were also able to successfully detect the endoplasmic reticulum (ER)-mitochondria contact sites and analyze the transitional states during endosome maturation.This technique is useful for elucidating the intracellular organelle landscape.

Development of a method for analyzing organelle particles
To detect multiple organelles with high resolution, we isolated organelle particles from cells and performed multi-color imaging.HeLa cells expressing markers for the endoplasmic reticulum (ER, mTagBFP2 (BFP)-SEC61B), mitochondria (GFP-OMP25 and SNAP-OMP25), and the Golgi (Venus-GS27) were used.Early endosomes were labeled by 5-min incubation with Alexa Fluor 647 (Alexa647)-conjugated epidermal growth factor (EGF) (hereafter, Alexa647-EGF) at 37°C.The plasma membrane was labeled by incubating cells for 15 min with Alexa405-conjugated N-hydroxysuccinimidyl (NHS) ester (hereafter, Alexa405-NHS) at 4°C just before cell homogenization.After homogenization by gentle sonication, peroxisomes and lysosomes were stained with Alexa594-conjugated anti-PMP70 antibody and Alexa680conjugated anti-LAMP1 antibody, respectively.Thus, the organelle particles were labeled with eight colors in total (Fig. 1a, Extended Data Fig. 1a).Correlative light and electron microscopy (CLEM) confirmed that membranous structures positive for the markers of these organelles, such as the mitochondria and ER, were contained in these samples (Fig. 1b).
While multi-colored particles can be detected by either flow cytometry or fluorescence microscopy, we chose fluorescence microscopy owing to its higher sensitivity for small particles.Using a confocal fluorescence microscope equipped with spectrometers (Extended Data Fig. 1b), lambda scanning was performed between wavelengths of 411 nm and 736 nm (Extended Data Fig. 1c).Single-color organelle particles labeled with different fluorescent dyes were used to acquire spectral data, and linear unmixing was performed to obtain eight-color fluorescence images (Fig. 1c).To identify organelle particles in the obtained fluorescence images, Gaussian mixture modeling was applied to the fluorescence intensity of each pixel, thus determining the threshold for separating organelle signals from the background.Using the determined threshold, each image was binarized, and the images of the eight colors were merged to obtain an image of all particles.
To analyze the fluorescence images, the sum of pixel intensities of each organelle marker was determined for each particle, and eight-dimensional data were thus acquired (Fig. 1a).Fluorescent signals of 36,195 particles from three independent experiments were combined and dimensionally reduced to be visualized in a twodimensional plane by principal component analysis (PCA) or uniform manifold approximation and projection (UMAP).While PCA resulted in one cluster overall (Extended Data Fig. 2a), UMAP separated the data into seven clusters (Fig. 2a).The numbers of particles classified in each cluster were as follows: Cluster 1, 21,522; Cluster 2, 5,028; Cluster 3, 2,517; Cluster 4, 2,481; Cluster 5, 1,806; Cluster 6, 1,702; and Cluster 7, 1,139.In UMAP, cluster size is dependent on various factors, including the number of data points, their variability, and hyperparameters (a parameter used to tune dimensionality reduction).The particles in each cluster showed different properties of organelle markers: Cluster 1 corresponded to the ER (SEC61B), Cluster 2 to peroxisomes (PMP70), Cluster 3 to mitochondria (GFP-OMP25 and SNAP-OMP25), Cluster 4 to early endosomes (EGF), Cluster 5 to the plasma membranes (Alexa405-NHS), Cluster 6 to the Golgi (GS27), and Cluster 7 to lysosomes (LAMP1) (Fig. 2b, c).
The clusters with high fluorescence intensity of the mitochondrial markers GFP-OMP25 and SNAP-OMP25 coincided with Cluster 3, verifying the specificity of this method.Additionally, as particles from three independent experiments were included in all clusters and no cluster unique to specific experiments was found, the classification of these particles was considered reproducible (Extended Data Fig. 2b).These results suggest that organelle particles retain their original membranous components and were accurately detected by this multi-parametric particle-based method.

Applications for the detection of minor organellar populations: ER-mitochondria contact sites
Most organelle markers were enriched in one of the clusters, but the ER marker SEC61B was mixed into several clusters, including mitochondrial clusters (Cluster 3) (Fig. 2b).This small population was posited to correspond to the ER-mitochondria contact sites.We therefore sought to visualize the ER-mitochondria contact sites using a split-GFP-based reporter 17 (Fig. 3a).This reporter consists of the 1st to 10th β-sheets of GFP (GFP1-10) fused to the N terminus of the ER membrane protein ERj1 and the 11th β-sheet of GFP (GFP11) fused to the N terminus of the mitochondrial outer membrane protein TOMM70.When GFP1-10 and GFP11 were associated at the ER-mitochondria contact sites, GFP fluorescence was emitted (Extended Data Fig. 3a).We introduced the fluorescent markers for the ER (BFP-SEC61B) and mitochondria (SNAP-OMP25) into cells expressing this contact site reporter and labeled early endosomes and the plasma membrane with Alexa647-EGF and Alexa405-NHS, respectively.After the preparation of organelle particles, peroxisomes and lysosomes were labeled with Alexa594conjugated anti-PMP70 and Alexa680-conjugated anti-LAMP1 antibodies, respectively, and seven-color fluorescence images were thus obtained (Fig. 3a, Extended Data Fig. 3b,c).
To obtain clusters of the six markers (except the split-GFP-based contact site reporter), we used the six-color data set obtained in the previous eight-color analysis (BFP-SEC61B, SNAP-OMP25, Alexa647-EGF, Alexa405-NHS, Alexa594-anti-PMP70 antibody, and Alexa680-anti-LAMP1 antibody) as reference data.We embedded these reference data in a new two-dimensional plane using UMAP, resulting in five clusters (Fig. 3b).Examination of the fluorescence intensity of each marker in the UMAP space revealed that the clusters were formed to contain different organelle markers (Extended Data Fig. 4a).These clusters were detected in all three independent experiments, validating the reproducibility of this experiment (Extended Data Fig. 4b).
When the query data from cells expressing the ER-mitochondria contact site reporter were annotated using metric learning with the reference data, all five of the clusters were mapped with the query data (Fig. 3c).Monitoring the fluorescence intensity of each marker in the plotted query data revealed that each cluster primarily contained a distinct marker, as observed among the reference data (Extended Data Fig. 5a).As these clusters were also detected in all three independent experiments, the particle classification was considered reproducible (Extended Data Fig. 5b).
We then plotted the GFP signals derived from the ER-mitochondria contact sites onto the UMAP space.Particles with high GFP intensities were detected within both the mitochondrial and ER clusters (Fig. 3d).These particles were found in the areas with high BFP-SEC61B signal in the mitochondrial cluster and high SNAP-OMP25 signal in the ER cluster (red arrows in Extended Data Fig. 5a).Thus, those GFP-positive particles were considered to be ER-mitochondria contact sites.Moreover, in our CLEM data on organelle particles, ER fragments were often detected on the mitochondrial surface (Fig. 1b).In most cases, ER fragments were much smaller than associated mitochondria particles.Therefore, it is reasonable that the signals of the ERmitochondria contact sites were detected mainly in the mitochondrial cluster.In summary, the present method can be applied to detect small organelle populations like organelle contact sites in the organelle landscape.

Applications to organelles in transition: An intuitive view of endosomes
The successful detection of the ER-mitochondria contact sites suggested that not only typical organelles but also organelles during transition or maturation could be investigated with this method.Accordingly, we analyzed the endocytic pathway, which contains various organelles in transition, after incorporation of EGF (a cargo representative of lysosomal degradation) and transferrin (a cargo representative of recycling to the plasma membrane).To assign organelles more precisely, we used fluorescent cargos in addition to static organelle markers.GFP-RAB5 (early endosome marker), Venus-RAB11 (recycling endosome marker), and SNAP-RAB7 (late endosome marker) were expressed in HeLa cells.These HeLa cells were incubated with Alexa594-conjugated transferrin (hereafter, Alexa594-transferrin) and Alexa647-EGF at 4°C for 30 min.Then, excess cargos were washed away, while the remainder was internalized by endocytosis at 37°C (Fig. 4a).At each time point of endocytosis, we collected organelle particles, stained them with antibodies against LAMP1, and obtained six-color fluorescence images (Extended Data Fig. 6a, b).
To analyze the temporal changes of the features of only EGF-or transferrinpositive particles, we strictly identified these particles as follows (because many endosomes contained neither EGF nor transferrin).We measured the background intensities of particles derived from cells not treated with Alexa594-transferrin and Alexa647-EGF and set the 99th percentile point for signal strength as the threshold for the Alexa594-transferrin and Alexa647-EGF signals (Extended Data Fig. 6c, bottom panels).We then extracted particles that were positive for EGF or transferrin and tracked the EGF and transferrin signals in this population at each time point.A bright EGF population appeared 15 min after culture conditions were shifted to 37°C, possibly owing to the fusion of early endosomes, which increased the EGF fluorescence intensity per particle 18 (Fig. 4b, left panel).The bright population decreased after 30 min, suggesting that Alexa647-EGF was degraded in lysosomes.This observation was consistent with previous reports that EGF reaches lysosomes approximately 30 min after incorporation 19 .In contrast, transferrin showed only a slight decrease after 10 min.This probably indicates a low homotopic fusion rate in RAB11-positive recycling endosomes 20 and recycling into the plasma membrane, which is consistent with the reported 10-min half-life of transferrin recycling 21 (Fig. 4b, right panel).
We then integrated the time-course data from three independent experiments, reduced the dimensions of the fluorescence intensity data of four endosomal markers (excluding EGF and transferrin) into a two-dimensional UMAP plane, and performed clustering.This resulted in seven clusters (Fig. 4c), all of which were detected in three independent experiments, supporting the reproducibility of this particle classification (Extended Data Fig. 7).These clusters roughly reflected the fluorescence intensity of each endosomal marker (Fig. 4d, e).Cluster 1 was considered to comprise early endosomes based on high RAB5 fluorescence intensity, and Clusters 2 and 3 were considered to comprise recycling endosomes based on high RAB11 fluorescence intensity.Cluster 5, with high RAB5 and RAB7 signal intensity, was considered to be a population of endosomes undergoing RAB conversion, during which RAB5 is replaced by RAB7.Clusters 4, 6, and 7, with high signal intensities of RAB7 and/or LAMP1, were considered late endosomes and lysosomes.
Next, we extracted the fluorescence intensities of EGF and transferrin from the original data and plotted them with the embedded data at each time point (Fig. 4f).EGF was predominantly present in Clusters 1 and 5 at 0-10 min and moved to Clusters 4, 6, and 7 after 15 min.Transferrin was predominantly present in Clusters 1 and 2 at 0 min and migrated to Cluster 3 at 5-10 min.After 20 min, there were fewer particles with high transferrin signals.These results suggest that our method successfully visualized the continuous diagram of the endocytic process of EGF and transferrin in an intuitive manner.
To further characterize each cluster, we analyzed the temporal transition of the two cargos.We counted the number of EGF-or transferrin-positive particles in each cluster and plotted the proportion over time (Fig. 5a, b).Regarding EGF-positive particles, Clusters 6 and 7 showed the highest proportions at 15 min (Fig. 5a).These clusters were considered to be late endosomes positive for RAB7, as they did not show high RAB5 signal intensity (Fig. 4d, e).The proportion of Cluster 4 increased over time, indicating that it represents lysosomes (Fig. 5a).In Cluster 4, LAMP1 was more enriched than RAB7 (Fig. 4d, e).The proportion of transferrin-positive particles in Clusters 6 and 7 increased at 10 min, and that of Cluster 4 increased at 15 min (Fig. 5b).
These data are consistent with Clusters 6 and 7 representing late endosomes and Cluster 4 representing lysosomes because some population of transferrin receptors are directed to lysosomal degradation [22][23][24] .By tracking the intensities of EGF and transferrin, we were able to classify endocytic organelles in detail, which could not be determined simply by detecting organellar markers (Fig. 5c).

Discussion
We developed a multi-parametric particle-based method that simultaneously detects multiple organelles with sufficient resolution.This method also minimizes observer biases because it analyzes organelle particles from a large number of cells.With this method, we can monitor the landscape of organelles in UMAP space in an intuitive manner.This multi-parametric particle-based analysis is effective when a large number of (typically more than four or five) markers are observed simultaneously.In addition, normalization of signal intensity across different datasets is straightforward, enabling the combined analysis of multiple datasets, as in our analysis of ER-mitochondria contact sites and endocytosis.We were able to exploit this advantage to extract and define intermediates of transitional organelles such as endosomes (Fig. 5c).While population analysis for EEA1, APPL1, RAB5, and RAB7 combined with CLEM has been reported, only two endosome markers were used for each electron microscopy image 25 .Therefore, experiments must be performed separately for each of these marker pairs.For late endosomes and lysosomes, counting structures positive for RAB7 and LAMP1 after dextran incorporation has been reported 26 .In contrast, our method using four endosomal markers and two representative cargos (for recycling and degradation, respectively) can monitor endocytosis processes from early endosomes to recycling endosomes, late endosomes, and lysosomes all at once.This technique is applicable not only to fluorescence microscopy but also to flow cytometry.By combining flow organelle sorting and highly sensitive proteomic or lipidomic analysis, it would be possible to identify new organelle entities.However, the detection limit of particle size by flow cytometry is approximately 100 nm in diameter, hindering the detection of small organelles 27 .In addition, multiple organelles can be included in a single droplet during sorting, making purification difficult.If highthroughput, multi-parametric microparticle sorting becomes possible with microfluidics 28 , it may be applicable to organelle isolation.
One clear limitation of this technique is that the original morphology and spatial information are lost.To compensate for this shortcoming, combined applications with fluorescence microscopy or CLEM of intact cells are necessary.Furthermore, the fact that only a small number of antibodies are effective for staining intact organelle particles limits the number of parameters.If organelle particles can be detected using only antibodies without exogenous expression of organelle markers, this method would become particularly applicable to physiological and pathological studies.Just as flow cytometry revolutionized immunology and hematology, multi-parametric organelle particle analysis would provide more detailed information about organelles, thereby deepening our understanding of cell biology.

Stable expression in HeLa cells by retrovirus-mediated transfection
To prepare the retrovirus solution, HEK293T cells were transfected with the pMRX-IP-based retroviral plasmid, along with pCG-gag-pol and pCG-VSV-G (a gift from Prof. T. Yasui), using Lipofectamine 2000 (Thermo Fisher Scientific, 11668019) for a period of 4-6 hours.Following transfection, the cells were cultured in DMEM for 2-3 days.The medium containing the retrovirus was collected and filtered through a 0.45-μm filter unit (Ultrafree-MC; Millipore).The filtered medium was then added to HeLa cells for infection, along with 8 μg/ml polybrene (Sigma-Aldrich, H9268).Viral infections were conducted to obtain cells expressing single or multiple organelle markers.Multipleorganelle marker co-expressing cells were isolated using a cell sorter (MoFlo Astrios EQ, Beckman Coulter).

Preparation of organelle particles
Eight-color typical organelle particles were prepared as follows.HeLa cells coexpressing mTag2BFP2-SEC61B, EGFP-OMP25, Venus-GS27, and SNAP-tag-OMP25 were subconfluently cultured in a 6-cm dish.The cells were washed three times with serum-free DMEM and then incubated at 37°C for 5 min with 400 ng/ml Alexa647-EGF.Subsequently, the cells were immediately washed three times with serum-containing DMEM, followed by three washes with ice-cold PBS.Afterward, the cell membrane was labeled by incubating the cells at 4℃ for 15 min with 8 μg/ml Alexa405-NHS.The cells were then washed twice with ice-cold PBS and with HEPES buffer (20 mM HEPES-KOH, pH 7.4, 250 mM sucrose, 1 mM EDTA) and were then collected by adding 1 ml of HEPES buffer with protease inhibitor (Nacalai Tesque, 03969-34).The cells were sonicated three times for 1 s each using a sonicator (TAITEC, VP-050N).The supernatant was obtained by centrifugation at 1500 × g for 10 min, repeated twice.To this supernatant, 0.3% BSA, 125 nM SNAP-Cell TMR-star, Alexa594-conjugated anti-PMP70, and Alexa680-conjugated anti-LAMP1 antibodies were added and incubated at 4°C for 1 h.The multi-color organelle suspension was mounted on poly-L-lysine-coated coverslips (Matsunami, C1110) with ProLong Gold Antifade Mountant (Thermo Fisher Scientific, P36934).For the experiment investigating mitochondria-ER contact sites, HeLa cells co-expressing mTagBFP1-SEC61B, TOM70(1-70)-3×FLAG-GFP11, ERj1(1-200)-V5-GFP1-10, and SNAPtag-OMP25 were prepared for a six-color organelle suspension according to the methods described above.For the experiment examining the process of endocytosis, HeLa cells co-expressing EGFP-RAB5, Venus-RAB11, and SNAP-tag-RAB7 were cultured overnight in seven dishes with serum-free DMEM.In six dishes, 1 μg/ml Alexa647-EGF and 5 μg/ml Alexa594-transferrin were added and incubated at 4°C for 30 min.After two washes with ice-cold PBS and with serum-free DMEM, the cells were incubated at 37°C for 0, 5, 10, 20, 30, and 40 min.The cells in the seven dishes were washed twice with ice-cold PBS and twice with HEPES buffer.Cell lysis was obtained as described above, followed by incubation of the lysates at 4°C for 1 h with 125 nM SNAP-Cell TMR-star and Alexa680-conjugated anti-LAMP1 antibody.The six-color labeled organelle suspension was mounted on coverslips using the same procedure.Sample preparations under these respective conditions were repeated three times.As reference samples for spectral imaging and linear unmixing, each single-color organelle sample was prepared using the same methods, either with cells expressing single-organelle markers or with cells not expressing any exogenous organelle markers.

Fluorescence spectral imaging and linear unmixing
Fluorescence spectral images were acquired using an Olympus FV3000 confocal laser microscope equipped with a four-channel cooled GaAsP photomultiplier and a diffraction grating.The microscope was equipped with a 60× oil-immersion objective lens (NA 1.42, UPLAXAPO; Olympus).Fluorescent images were captured using FLUOVIEW software (FV31-SW, ver.2.4.1.198, Olympus).All fluorophores were excited sequentially using 640 nm, 488 nm, and 561 nm lasers as well as a 405 nm laser with a 405/488/561/640 nm multi-band beam splitter.Fluorescent signals were separated with dichroic mirrors (SDM400-540, SDM400-470, SDM400-620) and were collected onto four GaAsP detectors with grating in lambda mode every 5 nm with a 10nm bandwidth [13 steps from 411 to 481 nm (spectrometer 1), 11 steps from 494 to 554 nm (spectrometer 2), 12 steps from 569 to 634 nm (spectrometer 3), and 16 steps from 646 to 731 nm (spectrometer 4), as shown in Extended Data Fig. 1b].We obtained images of organelle particles labeled with a single fluorophore (each color used for labeling multi-colored organelles) using the same image acquisition settings as for multi-colored organelles (including laser power and detector settings).The images captured across a total of 52 steps were integrated using the "Append Images" function in FLUOVIEW to create a single lambda series.The multi-color images were subjected to linear unmixing using the "Normal Unmixing" function in FLUOVIEW.Reference spectra generated from the images of single-color organelle using the "Spectral Image Unmixing" function in FLUOVIEW were employed.
The organelle particles were fixed in 2% paraformaldehyde (Nacalai Tesque, 26126-54) and 0.5% glutaraldehyde (TAAB, G018/1) in 0.1 M phosphate buffer at pH 7.4 for 1 hour at 4℃. Fluorescence spectral imaging and linear unmixing were performed as described above.Following fluorescence image acquisition, the organelles were embedded in resin and subjected to trimming, slicing of 25-nm thick serial sections, electron microscopy imaging, and image processing, following previously established protocols 32 .

Particle extraction
Data processing and analysis were performed mostly using ImageJ (v1.54f) in Fiji 33 and R (v4.2.1) in RStudio (v2023.03.1).After linear unmixing, the images were exported in 16-bit grayscale TIFF format using FLUOVIEW (FV31S-SW, Evident).The TIFF images were imported using the EBImage (v4.38.0) package and compiled for each replicate, and the intensity of all pixels was recorded for each channel.For each channel, pixel intensities were classified using a Gaussian Mixture Model as implemented in the R package mclust (v6.0.0).The number of clusters was selected by calculating the Bayesian information criterion (BIC) and selecting the smallest value (in practice, the largest value was selected if the BIC calculation using the GMM function in mclust yielded negative values).Pixel intensities were assigned to clusters by taking the cluster with the highest responsibility for each data point.By assigning the signal intensity of each pixel to clusters, intensities at which the cluster switches were set as crossing points.In all cases, the threshold between particles and the background was determined as the highest crossing point.After determining the thresholds for all channels, particles were extracted using Jython in Fiji (ImageJ).A binary image was created for each channel using the above thresholds, and all eight channels were merged to create a binary image of the particles.The fluorescence intensity of the particles was determined using the pixel intensity from the pre-binary image, using the total sum of pixel intensities contained within the particles.
When extracting particles from endocytosis experiments, RAB11 signals were excluded from the creation of a binary image of the particles, because the RAB11positive signal foci were so numerous that the clusters were divided into only two clusters based on the presence or absence of RAB11, making it difficult to analyze endocytic particles.

Data analysis
Each experimental dataset was independently tested three times, and all results were used together for data analysis.All data were converted to cell_data_set objects using the monocle3 34 (v1.2.9) package, followed by dimension reduction using principal component analysis (PCA), batch processing using the align_cds function, and embedding on a two-dimensional plane using UMAP.Clustering was tested using k = 60.
Data analysis of particle data obtained from cells labeled with split-GFP for analysis of ER-mitochondria contact sites and embedded by metric learning was performed as follows.First, six-dimensional data obtained from the previous eight-color analysis, excluding GS27 and GFP-OMP25 signals, were embedded in a twodimensional plane using UMAP, from which reference data were obtained.These data were then used to map six-dimensional query data obtained from cells labeled with split-GFP for ER-mitochondria contact sites, excluding split-GFP contact site markers, onto UMAP using metric learning.The brightness intensity of GFP was plotted on twodimensional data embedded in UMAP after converting the seven-dimensional data back to the cell_data_set format, dimension reduction with PCA, batch processing with the align_cds function, and normalization with size factors.Clustering was tested using k = 20, thus obtaining nine clusters.BFP-SEC61B-positive clusters were grouped as one cluster manually, and five clusters were thus obtained.For visualization, the square root values of GFP signals were plotted.
For particle analysis of endocytosis, the top one percentile of the background brightness intensity of Alexa647 and Alexa594 channels extracted from cells without addition of Alexa647-EGF and Alexa594-transferrin was used as a threshold to extract particles from each time point post-stimulation.In addition, all points with a value of 0 for either RAB5, RAB7, or RAB11 were removed.All of the above data were processed in the cell_data_set format, reduced in dimension using PCA, batch processed using the align_cds function, and further reduced using UMAP.Clustering was tested using k =

Use of large language model (LLM) and other AI tools
We utilized ChatGPT, DeepL, and DeepL Write to translate the text from Japanese into English and improve the English text.

20 .
Particles with mean fluorescence intensity values above the aforementioned thresholds, which were used for determining particle extraction, were used for counting EGF-or transferrin-positive particles.For visualization of EGF or transferrin transition in endocytic pathways, the square root values of Alexa647-EGF or Alexa594-transferrin signals were plotted.

Fig. 1 |Fig. 2 |
Fig. 1 | Multi-parametric single-particle analysis.a, Workflow of multi-parametric single-particle analysis of typical organelles.Markers used for labeling organelles are indicated in red.A405, A594, A647, and A680 indicate the fluorescent dyes Alexa Fluor 405, Alexa Fluor 594, Alexa Fluor 647, and Alexa Fluor 680, respectively.Fluorescence intensities of organelle particles for each marker were obtained as eight-dimensional data, as shown in the matrix, which were subjected to dimension reduction for visualization in a two-dimensional map.NHS, N-hydroxysuccinimidyl esters; EGF, epidermal growth factor.b, Correlative light and electron microscopy (CLEM) of organelle particles.Magnified images are shown in the right panels.Blue, GFP-OMP25 (mitochondria); green, BFP-SEC61B (ER); and red, Alexa405-NHS (plasma membrane).Red arrows indicate endoplasmic reticulum fragments associated with mitochondria.Scale bar, 10 µm and 2 µm (magnified images).c, Unmixing of eight-color fluorescent spectral images and their merged image.Scale bar, 50 µm

Fig. 5 |
Fig. 5 | Changes in the distribution of clusters during the endocytic process.a, b, Stacked bar graphs showing the proportion of clusters in EGF-positive particles (a) and transferrin-positive particles (b) at each time point (after the shift to 37°C).c, Characters of the clusters and putative pathways depicted in uniform manifold approximation and projection (UMAP) space.The magenta arrow indicates the putative recycling pathway, and the green arrow indicates the degradation pathway.

Fig. 1 |Fig. 2 |
Fig. 1 | Multi-parametric single-particle analysis.a, Workflow of multi-parametric single-particle analysis of typical organelles.Markers used for labeling organelles are indicated in red.A405, A594, A647, and A680 indicate the fluorescent dyes Alexa Fluor 405, Alexa Fluor 594, Alexa Fluor 647, and Alexa Fluor 680, respectively.Fluorescence intensities of organelle particles for each marker were obtained as eight-dimensional data, as shown in the matrix, which were subjected to dimension reduction for visualization in a two-dimensional map.NHS, N-hydroxy-succinimidyl esters; EGF, epidermal growth factor.b, Correlative light and electron microscopy (CLEM) of organelle particles.Magnified images are shown in the right panels.Blue, GFP-OMP25 (mitochondria); green, BFP-SEC61B (ER); and red, Alexa405-NHS (plasma membrane).Red arrows indicate endoplasmic reticulum fragments associated with mitochondria.Scale bar, 10 µm and 2 µm (magnified images).c, Unmixing of eight-color fluorescent spectral images and their merged image.Scale bar, 50 µm

Fig. 3 |Fig. 4 |
Fig. 3 | Multi-parametric single-particle analysis of the ER-mitochondria contact site.a, Schematic illustration of cells labeled with the ER-mitochondrial contact site reporter and six organelle markers (left) and the split-GFP-based ER-mitochondria contact site reporter (right).ERj1(1-200)-V5-GFP1-10 and TOMM70(1-70)-3xFLAG-GFP11 are only assembled on the ER-mitochondria contact site to form GFP. Thus, those reporters that are not assembled on the ER-mitochondria contact site do not produce GFP signals.b, Uniform manifold approximation and projection (UMAP) embedding and clustering of the data obtained from fluorescent images of particles labeled with six organelle markers as references.The numbers of particles classified in each cluster were as follows: ER, 22,233; peroxisome, 5,088; endosome and lysosome, 3,583; mitochondria, 2,243; the plasma membrane, 1711.c, Plot of the data of the experiments with the ER-mitochondrial contact site marker as query using metric learning with the UMAP results in b as reference.The numbers of particles plotted on the query were 17,479, and those of references were 34,858.d, Plot of the GFP signal intensity derived from the ER-mitochondria contact site reporter.The maximum fluorescence intensity of GFP was set to 100%.For visualization, the square root of % max has been plotted.

Fig. 5 |. 1 |
Fig. 5 | Changes in the distribution of clusters during the endocytic process.a, b, Stacked bar graphs showing the proportion of clusters in EGF-positive particles (a) and transferrin-positive particles (b) at each time point (after the shift to 37°C).c, Characters of the clusters and putative pathways depicted in uniform manifold approximation and projection (UMAP) space.The magenta arrow indicates the putative recycling pathway, and the green arrow indicates the degradation pathway.

b. 4 |. 5 |. 6 |
Extended Data Fig. 2 | Dimension reduction of the eight-color data.a, Principal component analysis (PCA) plots of the data obtained from eight-color fluorescent images of organelle particles derived from HeLa cells.Particles were colored according to the fluorescence intensity of each marker.The maximum fluorescence intensity of each marker was set to 100%.b, The data from three independent experiments were plotted on uniform manifold approximation and projection (UMAP) spaces.The numbers of particles plotted in each experiment were as follows: Experiment 1, 10,749; Experiment 2, 14,286; and Experiment 3, 11,160.Distribution of organelle markers of the reference data in uniform manifold approximation and projection (UMAP) space.a, Intensities of fluorescent markers from the reference data shown in Fig. 3. Particles were colored according to the fluorescence intensity of each marker.The maximum fluorescence intensity in each marker was set to 100%.b, UMAP embedding of the data of the reference obtained from three independent experiments.The numbers of particles plotted on each experiment were as follows: Experiment 1, 10,387; Experiment 2, 13,830; and Experiment 3, 10,641.Distribution of organelle markers of the query data of Fig. 3 in uniform manifold approximation and projection (UMAP) space.a, Intensities of the fluorescent markers from the query data.Red arrows indicate split-GFP-positive particles.b, UMAP embedding of the data of the query obtained from three independent experiments.The numbers of particles plotted in each experiment were as follows: Experiment 1, 72,45; Experiment 2, 4,160; and Experiment 3, 6,065.Spectral imaging and linear unmixing of the images of organelle particles labeled with endocytosis-related markers.a, Montage of fluorescence images obtained by spectral imaging of fluorescently labeled endocytic particles.Images were acquired and are shown as in Extended Data Fig. 1c.b, Unmixing results of the fluorescent spectral images in a. c, Histogram of the signal intensity of particles from cells treated or untreated with A647-EGF and A594-transferrin.The 99th percentile point for untreated samples is indicated by the red dashed line.

. 7 |
Reproducibility of the data and marker intensities in each cluster in the endocytosis analysis.a, Particles from each replicate separately plotted on the uniform manifold approximation and projection (UMAP) data shown in Fig. 4c.The numbers of particles plotted on each experiment were as follows: Experiment 1, 2,893; Experiment 2, 5,791; and Experiment 3, 8,429.b, Distribution of each replicate in Clusters 1-7.