Subcellular proteomics

Mass spectrometry-based methods

MS approaches offer accurate proteome-wide identification and quantification of proteins and proteoforms. MS-based workflows for subcellular proteomics use biochemical fractionation or proximity labelling methods to separate or discriminate specific subcellular compartments before MS analysis. We describe specific strategies for producing spatially informative samples for MS analysis and common quantitative MS techniques below. Note that we focus on centrifugation-based and detergent-based fractionation methods, although electrophoresis and affinity purification strategies have also been employed for organellar fractionation^22–28.

Fluorescent imaging methods

Imaging approaches give subcellular information of protein distribution in intact cells, and microscopy has been used for centuries to explore the cell interior. Recent advances in genome-wide RNA interference technologies and CRISPR–Cas9 (CRISPR-associated protein 9) have further contributed to imaging being a powerful approach for linking phenotype — including changes in subcellular location — to genotype^91,92. Two main fluorescent imaging approaches can be employed to capture subcellular information: live-cell imaging, where the protein of interest is modified to express a fluorescent tag to enable detection; and immunocytochemistry, where fluorescently tagged affinity reagents are applied to fixed material to bind specific proteins and allow their visualization. Antibodies are the most commonly used affinity reagents and can be labelled directly with fluorophores for detection of the bound protein (primary antibodies), or secondary antibodies conjugated to fluorophores can be used, which bind to unlabelled primary antibodies in turn. Alternatives to antibodies include nanobodies, affimers and aptamers^93–95. Several chemical dyes are available for live and fixed material to counterstain cellular structures such as the nucleus, cell membranes, mitochondria and actin filaments^95,96. We refer readers to a review on the range of available counterstains⁹⁷.

Shotgun MS proteomics data processing

The raw output of organellar fractionation and proximity labelling approaches consists of thousands of tandem mass spectra and it is therefore impractical to manually sequence peptides from spectra. Complex algorithmic database searching tools have been developed for DDA spectra, including Andromeda¹²⁵, Sequest¹²⁶, Mascot¹²⁷ and X!Tandem¹²⁸. These compare experimental peptide spectra with theoretical in silico spectra and generate peptide spectral match scores, which reflect the likelihood that the corresponding peptide was present in the sample. To ensure that peptide spectral match scores are reliable, spectra are searched against the target species peptide database and a decoy database that consists of a reversed or randomized version of the target database^129,130. This allows for the calculation of the percentage of peptide spectral match hits that are false positives, enabling estimation of the false discovery rate¹³¹. Database searching tools are often packaged into user-friendly software such as Proteome Discoverer and MaxQuant that allow for customizable and experiment-specific pipeline design¹³². For DIA spectra deconvolution, alternative software is required, such as OpenSWATH¹³³, Skyline¹³⁴ and DIA-NN¹³⁵, which are reviewed elsewhere¹³⁶. The output of these packages is a list of peptides and proteins alongside their quantitation across all samples, as well as other statistical and descriptive data. Statistical and visual analysis of this output can be performed using Perseus¹³⁷, Jupyter¹³⁸, Python¹³⁹ or R programming packages. We direct interested readers to a repository of open-source Python tools for proteomics analysis.

Analysing fractionation experiments

All organellar fractionation methods coupled with MS use ion intensities as a proxy for protein abundance. Proteins with similar abundance profiles or characteristics of defined organellar proteins are assigned to their respective organelles. As proteins are dynamic and can simultaneously reside in multiple cellular compartments¹⁴⁰, validation of organelle markers within different biological contexts is critical. Subtractive proteomics data must include an accurate reference set of known protein components of the target organelle and reference lists for protein contaminants from other subcellular compartments. These can be selected using Gene Ontology^141,142, although Gene Ontology is frequently not a robust way to define marker sets as terms can be broadly applied. Other databases can be used in conjunction with Gene Ontology that contain evidence-based information, such as COMPARTMENTS¹⁴³ and UniProt¹⁴⁴. Calculation of the abundance ratios of reference proteins against contaminant proteins allows proteins with enrichment comparable with the reference markers to be assigned to the target organelle and visualized in a ratio versus intensity plot (FIG. 5a). This analysis accounts for impure isolations and enables high-confidence assignment; however, proteins that also reside outside the target organelle can easily elude identification.

Correlation profiling data do not require a contaminant reference and consist of multiple organelle profile plots, where the abundance of the protein is plotted across the biochemical fractionation gradient (FIG. 5b). To reduce complexity and improve visualization, dimension reduction methods such as principle component analysis (PCA) can assess subcellular resolution as proteins associated with the same subcellular niche should form defined clusters in PCA plots¹⁴⁵ (discussed further in BOX 2). Supervised, semi-supervised or unsupervised machine learning algorithms can also be used for correlation profiling; supervised and semi-supervised algorithms require the spatial profile of known organelle marker proteins to assign proteins with unknown localizations to subcellular niches^44,146–149 and include support vector machines, neural networks, random forest models and naive Bayes algorithms. The choice of machine learning algorithm is dependent on the available computation power, processing time frames, experience and what assumptions about the data are appropriate to make. We refer readers to a review of common machine learning algorithms in proteomics¹⁵⁰.

Unsupervised clustering algorithms, such as k-means clustering or DBSCAN, are useful when training data are scarce, for example when investigating non-model organisms with limited marker proteins^45,151,152. These methods are most suited to static and single locale assignment of proteins. Dynamic classification of proteins and classification of MLPs using these methods is challenging, but has been performed in previous studies^{44,47,56,89,153–155}. Bespoke pipelines that use training data have recently facilitated dynamic classifications and MLP classifications, including T-augmented Gaussian mixture model approaches able to quantify the posterior probabilities of proteins residing in multiple organelles and TRANSPIRE (Translocation Analysis of Spatial pRotEomics), which models synthetic translocations from experimental organelle marker profiles to train a classifier to identify true protein dynamic events^{52,54,149,156,157}.

Proximity labelling data

The selection of benchmarking baits as organellar markers and appropriate controls informs downstream analytical pipelines. Negative controls must identify likely contaminants⁷¹, which include endogenously biotinylated proteins and proteins that become biotinylated by BirA* or APEX2 in a dose-dependent manner irrespective of the bait used⁷¹. These controls may include samples omitting the substrate or activator (for example, biotin or biotin-phenol/peroxide) or the fused enzyme, or samples where the labelling enzyme is fused to an irrelevant protein and this fusion protein is expressed at the same level as the bait fusions in the experimental samples^60,71. Input of bait and control data into computational tools such as SAINT (Significance Analysis of INTeractome) allows the probability of bona fide protein–protein interactions to be derived¹⁵⁸. Depending on the questions being asked, inclusion of additional controls assists in organellar proteome definition on a case-by-case basis and can help to discriminate between adjacent compartments. For example, to define the proteome of the mitochondrial intermembrane space, a ratiometric strategy based on SILAC was used to distinguish intermembrane space resident proteins from those in the cytoplasm⁶⁷.

When analysing compartments that are not separated by membranes, using components of adjacent organelles or structures as controls can also help compartment annotation. This was exemplified by use of the SNARE VAMP8 as an adjacent-compartment bait to identify proteins enriched with the endolysosome-specific VAMP7 bait¹⁵⁹. The use of multiple baits — labelling both the desired organelle and its neighbours — has been exploited to define the composition of membrane-less organelles such as the centrosome⁶⁸, RNA granules and processing bodies⁶⁹. Sub-organellar protein organization and contact sites between organelles and other structures can be resolved by interrogating multiple organellar baits, as was recently demonstrated within the mitochondria⁷⁰. For these data sets, which typically contain data from at least 20 baits, prey-wise analysis using factorization or correlation approaches aids in organellar inference^69,70,160. Factorization approaches are most useful for the definition of distinct groups of proteins associated with sets of structures, whereas correlation approaches produce continuums of scores displaying the relationships between pairs of tested proteins. In either case, the integration of these scoring tools into visualization and analysis schemes helps to explore localization assignments and relationships between proteins predicted to co-localize^161–163.

Processing imaging data

Imaging data give spatial information regarding target proteins in intact cells. A trained human eye can distinguish between more than 20 subcellular structures using no more than three standard markers¹¹; however, reference proteins or dyes are needed to stain and distinguish between organelles with similar distributions and size, for example cytosolic bodies and vesicles. Adequate controls and replicates are crucial to distinguish true localization from false staining patterns caused by non-specific binding or artefacts from sample preparation. Negative controls can include a sample lacking the affinity reagent towards the protein target, or a sample lacking the expression of the protein target, such as a knockout cell line to allow for background subtraction^164,165. Affinity reagents targeting specific cellular structures, such as cytoskeletal components or the intranuclear domain, should be included to evaluate reproducibility, fixation and sufficient permeabilization. Affinity reagents can be validated by evaluating signal loss following partial or complete knockdown of the target protein, comparing the staining pattern with the same, endogenously tagged protein or examining correlations between signal intensity and target protein expression levels according to RNA sequencing data across samples. To date, manual pattern recognition has been the primary approach to assign protein subcellular localization simply by inspecting images using the microscope software or any software supporting the image file format. Popular open-source software for image analysis include ImageJ, CellProfiler, QuPath, ilastik and Orbit^166–170.

A significant proportion of proteins localize to multiple cellular structures¹¹ and populations of genetically identical cells show variability in their protein expression levels and localization^11,171–173. Quantitative analysis techniques are needed to reveal subtle cell to cell variations and partial translocations between organelles across cellular states. This requires cells, and sometimes also subcellular structures, to be segmented by measuring the signal intensity for a structure-specific marker and determining the boundaries of the structure with respect to a threshold for defining a positive signal from the background. Once the objects are defined, fluorescence intensities from any stained marker can be measured and compared across segmented objects.

Machine learning and especially deep learning methods are becoming increasingly popular for image analysis owing to their superior performance and scalability, which make them attractive for analysing large data sets^174,175. These involve training a prediction model on an existing data set of images representing known protein localization patterns defined by the user — referred to as the ‘ground truth’. The model extracts features from this training data set and through iterative cycles of model fitting and improvement can predict the subcellular localization of a protein based on the staining pattern in sample data. Machine learning techniques such as k-nearest neighbour classifiers, artificial neural networks and support vector machines use feature sets to capture protein location in relation to cellular morphology, as displayed by reference markers for certain organelle structures^176–179. As many proteins localize to more than one organelle, models must accurately assign complex staining patterns to several fluorescent labels representing different organelles. This challenge was addressed in the Kaggle challenge for multi-label classification of cell organelles using the proteome scale collection of immunofluorescence images from the Human Protein Atlas (HPA) to generate a model that could assign subcellular locations for proteins in various cell lines displaying different morphologies¹⁸⁰. Machine learning techniques have also been applied as high-content screening approaches for identifying changes in phenotypes in response to small molecules^176,180,181. Model generation and training analysis software include WEKA¹⁸² and Scikit-learn¹⁸³.

Characterizing organelle proteomes

Early organelle-centric purification methods generally isolated target organelles at a high purity before systematically identifying organellar proteins using LC-MS^184–189. These approaches established the protein inventories of organelles, although they could not discriminate between targeted organellar constituents and co-purified proteins originating from non-target subcellular compartments. Typically, fewer than 1,000 proteins were measured over as few as four fractions in these early studies^190,191 — a decade later, more than 8,000 proteins were identified in four distinct fractions from mouse lungs following the rapid development of liquid chromatography and MS technology¹⁹².

Subtractive proteomics methodologies using differential separation technologies were first applied to define the nuclear envelope proteome from mouse liver³⁹. Subsequently, the quantitative comparison of a crude preparation of mitochondria against purified mitochondria from ten different mouse tissues greatly contributed to the mitochondrial compendium, MitoCarta³⁶. Subtractive proteomics has helped establish the protein inventories of peroxisomes^193–196, autophagosomes¹⁹⁷ and lysosomes¹⁹⁸. A landmark single-organelle spatial proteomics paper described the systematic study of the human centrosome using intensity-based, label-free protein correlation profiling of consecutive density gradient fractions³⁰. Several adaptations of subtractive proteomics have used SILAC ratios produced by spiking a differentially labelled internal standard into each fraction^{33,34,199,200}; this approach, with or without stable isotope labelling, has been used to characterize the proteome of peroxisomes^38,201, nuclei²⁰², autophagosomes^33,199, lipid droplets³⁴, vesicles²⁰⁰ and mitochondria^35,203.

For proximity labelling approaches, the initial BioID study in which BioID was fused to lamin A identified new components of the nuclear lamina⁶³. The first APEX study targeted the peroxidase to the mitochondria to define the proteome of the mitochondrial matrix and inner mitochondrial space⁶⁷. Proximity labelling has now contributed to the characterization of various organelles and some of the more recent studies have employed multiple baits to provide extensive coverage of specific structures, in some cases revealing their organization^{68–70,204–212}. Proximity labelling is typically performed on whole cell lysates; upstream fractionation steps can also be coupled with the capture of biotinylated proteins to assist in defining specific proteomes, although these approaches can have issues with post-lysis artefacts and imperfect fractionation^209,211.

Imaging approaches have helped characterize the proteome of many organelles, such as the centrosome and nucleolus²¹³, and revealed novel structures including cytoplasmic rods and rings^11,214 and the nucleoli rim²¹³. Complete characterization of the organelle proteome using affinity reagents or fluorescent protein fusions is both expensive and laborious; these approaches are complementary to MS-based approaches and valuable for validation of subsets of proteins identified in MS data.

Cell-wide correlation profiling

Cell-wide mapping of proteins using correlation profile approaches has been achieved in many cell types^10,44,47, tissues and even whole organisms. Models include Arabidopsis root-derived cells⁴², chicken DT-40 cells¹³, mouse neuronal cells¹⁴, rat liver^40,45, Saccharomyces cerevisiae¹⁵, Drosophila embryos¹⁶ and the photo-synthetic bacterium Synechocystis¹⁸. A recent study achieved the first subcellular-resolution proteome of an apicomplexan parasite, Toxoplasma gondii¹⁷; T. gondii is a widespread parasite that infects 30% of the human population, leading to various pathologies in infected individuals²¹⁵. Using a variant of the LOPIT workflow known as hyperplexed LOPIT (or hyperLOPIT) along with image-based validation^10,216, this study mapped two-thirds of the predicted proteome to 26 subcellular and sub-organellar niches in the tachyzoite form of the parasite. High-resolution isoelectric focusing LC-MS (HiRIEF-LC-MS)²¹⁷ coupled with differential detergent solubility and TMT 10-plex labelling was recently used to generate comprehensive subcellular maps with an 8,140-protein overlap between replicates across five human cell lines⁴⁷. Using support vector machines, 9,594 proteins were classified into 15 specific compartments and 12,125 proteins were classified into 4 subcellullar ‘neighbourhoods’ — groups of functionally related subcellular compartments such as those involved in the secretory pathway. Cell-wide correlation profiling can also be used to characterize protein complexes (BOX 3).

The imaging equivalent of these cell-wide organellar fractionation methods to achieve cell-wide distribution of proteins is the work by the HPA project. This is the largest collection of subcellular imaging data to date, with 12,813 human proteins mapped to 30 different subcellular structures on a panel of 22 different human cell lines using immunofluorescence and confocal microscopy as of January 2021. This project has predicted that up to 50% of human proteins localize to multiple compartments, with 17% showing single-cell variations¹¹. Several studies have investigated the sources of this single-cell heterogeneity^11,218,219. Characterizing single-cell heterogeneity has implications in health and disease; for example, many proteins are expressed heterogeneously across tumours, which has presented a challenge for developing treatments^220,221. Imaging further enables analysis of spatio-temporal human proteome reorganization over the course of the cell cycle or as a response to metabolic change^222,223.

Sub-organellar resolution

The application of tailored and combinational biochemical separation approaches, such as centrifugation and affinity purification, has enabled the definition of the sub-organellar proteomes of lysosomes³², nuclear envelopes^37,39, chloroplast envelopes³¹ and sub-compartments of mitochondria^{29,35,224,225}. For example, the incorporation of organelle purification, proximity labelling and subtractive methods effectively achieved sub-organellar resolution of the nuclear envelope, lipid droplets and centrosomes in Arabidopsis^37,68,211. To map the nuclear envelope, subtractive proteomics was used to distinguish proteins associated with the nuclear envelope from microsomal membrane proteins, followed by BioID2 with several bait proteins to provide complementary data for the characterization of the plant nuclear membrane proteome and nuclear pore complex. Subsequent studies used LOPIT in a targeted manner alongside electrophoresis-based protein separation to resolve and identify previously unknown endoplasmic reticulum proteins and cis, medial and trans-Golgi proteins in Arabidopsis^41,51,226; these findings have implications for understanding the role of Golgi proteins in the cell wall synthesis pathway and the development of biofuels²²⁷. A recent large-scale subtractive proteomics study used a combination of label-free and SILAC-based MS profiling to profile the sub-organellar compartments of mitochondria in S. cerevisiae cultures grown in different carbon sources. Integration of these multiple MS data sets resulted in the high-confidence identification of 901 mitochondrial proteins, including numerous proteins of previously unknown mitochondrial residency and some with multiple cellular locations²⁹.

As mentioned above, the breadth and precision of organellar and sub-organellar proteomics by proximity labelling can be improved by selecting multiple baits for organelle profiling. In a ground-breaking 2015 study, 58 baits mapping to the centrosome and primary cilia were systematically profiled to define the composition of these structures using bait–prey hierarchical clustering, revealing different groups of proteins that performed distinct functions⁶⁸. This study set the stage for mapping organelles with more extensive sets of baits. Recently, the steady-state composition and organization of RNA stress granules, processing bodies⁶⁹ and mitochondria⁷⁰ were defined using more than 100 baits. In each of these studies, clustering of preys rather than baits provided a high-resolution map of these structures, in some cases suggesting direct protein–protein interactions within the structures that could be experimentally validated using immunoblotting and microscopy⁶⁹. A comprehensive map of the human cell made using ~190 baits spanning most organelles¹⁶⁰ provided a global reference map. As these resources become increasingly available, selection of ideal baits to profile an organelle across different cell types or following different treatments will permit more in-depth investigation of spatio-temporal changes in organellar composition and organization. As imaging allows the study of proteins in their native cellular environment without the need for cell lysis, fine sub-organellar structures can be preserved in situ to enable detection of dynamic and MLP behaviour across the sub-organellar proteome¹¹. This trait of immunofluorescence has enabled the spatio-temporal characterization of the nucleolar proteome and its relocalization to the mitotic chromosomes during mitosis²²⁸. Further, MS analysis has exposed diverse extracellular vesicle subpopulations with roles in important biological functions and disease, such as angiogenesis and cancers^229–232. Multiple strategies to isolate extracellular vesicles can be used; some are similar to subtractive proteomics or correlation profile strategies, such as density or differential centrifugation, and others use alternative biochemical isolation strategies, such as size exclusion, microfluidics and immune purification. These methods are detailed and compared in a recent review²³³.

Multi-localized proteins

Owing to the dynamic nature of proteins during signalling and heterogeneity between different cell types, there is no consensus on exactly what proportion of the proteome is multi-localized. Predictions of MLPs vary and have been reported as 50% of the human proteome by hyperLOPIT studies and the HPA¹¹ but as little as 10% or less by other correlation profiling studies^47,53. This discrepancy could be owing to low confidence measurements, the type of classification tool used or biological variation^47,53 and demonstrates the difficulty of accurately measuring MLPs.

There is a growing repertoire of tools for describing MLPs. An unorthodox citizen science-based deep learning approach was able to create a robust tool for overcoming the challenges of automated subcellular allocation of MLPs for the analysis of the HPA¹². This approach robustly classified proteins into 29 subcellular compartments across 17 human cell lines. Despite inaccurate annotation of structurally similar subcellular features such as centrosomes and microtubule organization centres, unusual features such as blebs and condensed chromosomes were successfully profiled using this approach¹². The identification of MLPs in T. gondii using T-augmented Gaussian mixture model–Markov-chain Monte Carlo models demonstrated that these proteins are primarily associated with organelles associated with protein transportation, such as the plasma membrane and Golgi. Organelles known to be relatively static such as rhoptries, micronemes and dense granules showed more proteins with single assignments¹⁷.

Dynamic studies

A study to chart the mitochondrial importome in Trypanosoma brucei employed an approach known as importomics to identify substrates of glycosomal and mitochondrial import pathways^155,234. Importomics approaches can assess changes in protein abundance in a target organelle following RNAi-mediated knockdown of a central component of the protein import machinery, to determine potential substrate proteins that are trafficked to the organelle and imported via this machinery^155,234. Importomics is not limited to use with organelle-enriched fractions and can be used to globally study organellar protein import using cell lysates^155,235. A more simplistic importomics study design for investigating the nucleus used nucleocytoplasmic fractionation coupled with quantitative MS to study mechanisms underlying nucleocytoplasmic partitioning in Xenopus laevis eggs²⁰². Knockdown of the nuclear export receptor exportin 1 largely did not affect nucleocytoplasmic partitioning, suggesting that protein distribution between the nucleus and cytoplasm is largely maintained through passive processes rather than active trafficking²⁰².

Cell-wide dynamic studies studying trafficking proteins on the proteome scale using correlation profiling have given unprecedented insight into EGFR-related protein trafficking events⁴⁷. The combination of CRISPR–Cas9-based gene knockouts with dynamic organellar maps has been used to assess AP-4 vesicles¹⁵³ and AP-5 cargo¹⁵⁴ and perform the first cell-wide, dynamic correlation profiling experiment intended to define drug mechanisms²³⁶. Recently, LOPIT was coupled with a technique where selective capturing of vesicles destined for the Golgi were relocated to mitochondria by replacing the Golgi targeting domains of golgins with a mitochondrial transmembrane domain⁵⁵. This enabled characterization of vesicle cargo and regulatory proteins by redirecting them to the mitochondria, bypassing technical issues that arise owing to the genetic redundancy of golgin proteins and their transient interaction with vesicles.

Several studies have been published assessing organelle remodelling during viral infection and the distribution of virus and host proteins. Organellar fractionation followed by MS has been used to quantify temporal changes in plasma membrane and organelle proteomes during human cytomegalovirus infection^53,237,238. Analysis of organelle remodelling during the virus replication cycle demonstrated that alterations in organelle composition can reflect changes in the structure and function of an organelle, as peroxisomes shift between antiviral and proviral functions^239,240. Coupling of density and differential centrifugation has also been utilized to explore the mitochondrial-associated endoplasmic reticulum membrane, the endoplasmic reticulum and cytosol proteomes during infection with hepatitis C and Sendai virus²⁴¹. The above studies have described virus manipulation of host protein subcellular localization, allowing for the increased understanding of virus–host interactions and identification of potential therapeutic targets.

Generally, MS techniques are only able to capture averaged information for heterogeneous populations in samples, losing dynamic events occurring in subgroups or single cells. Imaging studies have reported that heterogeneity in cell samples could be the result of sudden stresses, cell crowding, cell–cell contacts, cell differentiation and different metabolic states^{218,242–245}. Several imaging studies have reported a spatio-temporal resolved map of the human proteome over the course of a cell cycle using imaging^227,245,246 and similar techniques have been applied to disease settings. Indeed, mislocalization of proteins and variations in protein expression levels have been shown to be associated with diseases such as cancer, neurodegenerative diseases and cardiac diseases^247–249. Combining imaging with computational analysis enables the identification of proteins with similar relocalization and changes in expression levels, which implies functional connectivity for such groups of proteins^101,250,251. Imaging-based approaches allow quantification of intensity levels in different cellular compartments that can better indicate whether a protein translocates from one compartment to the other, or merely changes in abundance.

Mass spectrometry-based methods

As mentioned above, MS-based proteomics data are prone to missing values owing to stochastic extraction and solubilization of peptides in sample preparation and during MS acquisition. Despite these issues, BioID and affinity purification MS can typically achieve high reproducibility (R² > 0.9) as the complexity of MS samples in these methods is low compared with the complexity of whole cell lysates²⁵². Reproducibility and coverage is usually more achievable when performing LC-MS/MS with less complex peptide samples, as in these samples there are fewer precursor ions per isolation window in the MS1 step and fewer ions are therefore omitted for subsequent sequencing in MS2. Labelling methods can reduce missing values, although cell-wide correlation profiling methods that allow for multiplexed characterization of many organellar fractions in each replicate still suffer from losses in protein identification of ~40–50% across three biological replicates. Unlabelled methods, by contrast, have attrition rates of ~60%^14,43. There are no tools to assess correlation profile reproducibility, only those to assess and compare the quality of organellar separation between cell-wide spatial proteomics experiments⁸⁸. DIA provides improved protein identification and quantification as it involves less stochastic ion sampling during MS acquisition than DDA.

For analyses of data collected after using biochemical separation techniques such as density centrifugation, losses in protein identification are proportional to the number of organellar fractions and replicates analysed. Performing reproducible subcellular fractionation is partly dependent on the simplicity of the fractionation procedure. Differential centrifugation and sequential detergent preparations are more reproducible than density equilibrium centrifugation owing to the intricacy of forming and collecting density equilibrium gradients, although spatio-temporal analysis using density gradients has proven to be effective^{17,18,45,53,89}.

For any MS-based proteomics experiment, the Minimum Information About a Proteomics Experiment (MIAPE) guidelines should be adhered to when publishing. The principles of MIAPE and the basic criteria are shown in BOX 4. For MIAPE compliance, the location of raw mass spectra files should be published; files can be stored in a public repository, such as the PRoteomics IDEntifications Database (PRIDE). There is no overarching deposition resource for data from all subcellular MS methodologies, although bespoke platforms and pipelines are available^47,156,253 (TABLE 2). The creation of guidelines to ensure efficient capture and reporting of crucial metadata associated with spatial proteomics data will be essential to enable data reanalysis and align the field with Findability, Accessibility, Interoperability, and Reusability (FAIR) principles²⁵⁴.

Imaging

Immuno-imaging data can be difficult to reproduce because of biological variation — for example, variation between cell types — and technical variation introduced by different sample preparation procedures or affinity reagents. Both the HPA and the independent antibody validation initiative antibodies-online found that less than 50% of commercially available antibodies are appropriate for studying protein distribution, owing to specificity issues in the context of intact tissues and cells^255,256. Major research and awareness efforts are improving this situation by using alternative antibody production and product validation procedures (see the Limitations and optimizations section).

As each microscopy vendor has their own internal imaging format and software for analysing and visualizing data, exporting files into standardized formats is important to disseminate scientific discoveries. The Open Microscopy Environment is an open-source informatics framework for biological microscopy experiments, designed to support various imaging applications. Most imaging platforms and open-source software are compatible with the Open Microscopy Environment^{166,170,257,258}. Publishing imaging data in a suitable data repository for data collection and mining is encouraged and is a prerequisite for publishing in many journals. Providing the original raw images is crucial, along with metadata accompanying each individual image containing all hardware and software settings applied upon image acquisition. When raw images are processed — for example, to increase intensity or contrast — these processes must be described. Upon submission of images to most journals there are clear guidelines regarding what image processing is accepted and part of the review process is sometimes dedicated to tracking image manipulations. Public repositories for imaging experiments are noted in TABLE 2 (REF.²⁵⁹).

Mass spectrometry-based methods

Imaging

Although use of a fixative allows cells to be preserved at precise time points, it can cause high background and signal loss, or disruption of subcellular structures and protein macromolecules if the fixation protocol used is inappropriate for the protein of interest²⁷⁶. An optimized single fixation protocol can be used for proteome-wide compatibility for studies such as the HPA¹⁰⁷; although this protocol has proven to work for a large variety of proteins representing >30 cellular structures, the best staining conditions for each protein depend on the sample context and the combination of the target protein and antibody used for detection.

Antibodies used in imaging studies are subject to biological batch-to-batch variability, with inconsistent binding properties between batches. Every antibody batch needs careful validation for their prospective application and sample type prior to use. More commercial suppliers are using recombinant means to manufacture their antibodies over traditional immunization, providing more sustainable and reproducible products. Awareness of careful antibody selection, validation and optimization has improved significantly over recent years. In subcellular analysis, the potential for high local concentrations of any cross-reacting proteins increase the need for thorough validation.