PEELing: an integrated and user-centric platform for spatially-resolved proteomics data analysis

Summary Molecular compartmentalization is vital for cellular physiology. Spatially-resolved proteomics allows biologists to survey protein composition and dynamics with subcellular resolution. Here we present PEELing, an integrated package and user-friendly web service for analyzing spatially- resolved proteomics data. PEELing assesses data quality using curated or user-defined references, performs cutoff analysis to remove contaminants, connects to databases for functional annotation, and generates data visualizations—providing a streamlined and reproducible workflow to explore spatially-resolved proteomics data. Availability and Implementation PEELing and its tutorial are publicly available at https://peeling.janelia.org/. A Python package of PEELing is available at https://github.com/JaneliaSciComp/peeling/. Contact Technical support for PEELing: peeling@janelia.hhmi.org.


Introduction
Localization and function of proteins are always coupled.For instance, proteins for intercellular adhesion and communication are localized to the cell surface while many energy-producing enzymes stay in the mitochondrion.High-resolution, proteome-wide mapping of protein localization is of core importance for understanding cellular organization and processes.Emerging technologies for spatially-resolved proteomics, particularly by proximity labeling, make this possible and broadly applicable in cell biology (Roux et al., 2012;Rhee et al., 2013;Branon et al., 2018;Loh et al., 2016;Geri et al., 2020;Han et al., 2018;Qin et al., 2021).Like other enrichmentbased profiling methods, spatially-resolved proteomics can be interfered by ineffective enrichment, non-specific contamination, and other factors.Rigorous assessment of data quality and proper data processing is crucial for interpreting the results and for designing subsequent studies.However, this can be complex and overwhelming, particularly for biologists with limited proteomics experience or systems biology background.To address this, we built PEELing (proteome extraction from enzymatic labeling data), a platform integrating data quality checks, contaminant removal, functional annotation, and visualization into an automated workflow (Figure 1a).

Implementation, Benchmarks, and Results
In Figure 1, we used a published cell-surface proteome of mouse developing Purkinje cells (Shuster et al., 2022) (Figure 1b and Supplementary File 1) to demonstrate the functionalities of PEELing.In this study (Shuster et al., 2022), cell-surface proteins were chemically labelled and isolated ("labelled" groups, hereafter) using horseradish peroxidase (HRP) mediated cell-surface biotinylation (Loh et al., 2016;Li et al., 2020;Shuster et al., 2022).Non-labelled control groups ("control") were included to capture contaminants such as endogenously biotinylated proteins and non-specific binders to isolation reagents.PEELing uses a ratiometric strategy as previously described (Hung et al., 2014) in which the labelled-to-control ratio of each protein reflects whether this protein is cell-surface enriched or not.A bona fide cell-surface protein should exhibit a high ratio because it should be enriched in the labelled group relative to the control group.A contaminant should have a low ratio since it should be captured similarly in both labelled and control groups.
To assess data quality, PEELing started with pairwise correlation analysis to check whether biological replicates exhibited consistency (Figure 1c).To determine whether cell-surface proteins were enriched, PEELing ranked all detected proteins in descending order based on the labelled-tocontrol ratio and scanned through them to mark curated cell-surface proteins (true positives, TPs; Supplementary File 10) and intracellular contaminants (false positives, FPs; Supplementary File 12).As shown in Figure 1d, true positive rate (TPR, blue), false positive rate (FPR, orange), and their difference (TPR-FPR, green) were calculated and plotted against ratio-based ranking.TPR increased quickly while FPR rose slowly, which led to a single peak of TPR-FPR, revealing a high ranking and an effective enrichment of cell-surface proteins.This is further illustrated by a receiver operating characteristic (ROC) curve bending towards the left-upper corner and exhibiting an area under the curve (AUC) of 0.79 (Figure 1e).
To test whether PEELing detects data of poor quality, we generated a pseudo dataset with random values mimicking a failed enrichment (Supplementary File 2 and Supplementary Figure 1).TPR, FPR, and ROC curves all followed the diagonal line (Supplementary Figure 1b,c,e,f), showing a complete mix of cell-surface and intracellular proteins without any enrichment (Supplementary Figure 1d,g).We then tested reference specificity by analyzing the cell-surface proteome (Shuster et al., 2022) (Supplementary File 1) with nuclear and mitochondrial references (Supplementary File 14-17).No enrichment of proteins of unintended cellular compartments (Supplementary Figure 2) showed the quality of the example data as well as the necessity of data-reference matching in PEELing analysis.
All analyses above used the labelled-to-control ratio as the enrichment indicator.Protein abundance is widely used in proteomic data quantification; however, it is not a reliable enrichment indicator in spatially-resolved profiling compared with the labelled-to-control ratio (Supplementary Figure 3, compared with Figure 1).From the same cell-surface proteome dataset (Shuster et al., 2022), we obtained protein abundance values of the labelled groups (Supplementary File 3).Despite comparable correlations based on abundance (Supplementary Figure 3a) and labelled-to-control ratio (Figure 1c), cell-surface protein enrichment was not observed when ranking the proteome by abundance (Supplementary Figure 3b,c,e,f).Intracellular proteins were highly enriched instead (Supplementary Figure 3d,g).Notably, a contaminant (e.g., a non-specific binder to the isolation reagent) can be abundant while a cellsurface protein may have a low expression level.As long as the control group captures the contaminant but not the cell-surface protein, these proteins will be ranked correctly by the labelledto-control ratio instead of being ranked reversely by abundance.Therefore, labelled-to-control ratio is the preferred input data for PEELing.
Following data quality checks, PEELing performed cutoff analysis to remove contaminants.For each labelled-to-control ratio, PEELing found the ranking position where TPR-FPR was maximal, as indicated by the peak of the green line in Figure 1d and the red dot in Figure 1e, and retained proteins ranked above this position.The "TPR-FPR maximum" cutoff provides two key benefits: 1) The cutoff position is determined by data quality rather than an arbitrary value, allowing for unbiased assessment of the proteomic results.If the data is of high quality with sparse false positives, TPR-FPR will peak later in the ranking, resulting in the retention of more proteins.
If the data is heavily contaminated with false positives, TPR-FPR will peak earlier, leading to the retention of fewer proteins.2) Any protein ranked above the cutoff position is retained, regardless of how it is annotated by a database.Therefore, missing annotations or occasional incorrect annotations in databases will not impact the analysis, as long as they are largely accurate and comprehensive.As illustrated in Supplementary Figure 4, 10-fold reduction of reference coverage, in either TP/FP or both, did not impair the analysis of the cell-surface proteome of Purkinje cells (compared with Figure 1).Additionally, the Drosophila proteome has less complete annotations than those of mouse and human, leading to jagged TPR, FPR, and ROC curves (Supplementary Figure 5c,d).Nevertheless, it did not interfere with the analysis of a Drosophila neuronal surface proteome (Li et al., 2020) (Supplementary Figure 5 and Supplementary File 4).Importantly, in certain physiological or pathological contexts, intracellular proteins can be transported to the cell surface-for instance, RNA helicase U5 snRNP200 in acute myeloid leukemia (Knorr et al., 2023).PEELing retains these proteins if they are ranked highly, potentially enabling researchers to discover novel biomarkers and cellular processes.Despite its robustness to varied reference coverages, we note that PEELing relies on high-quality references (see Supplementary Materials -Method Details on how to create references).When references are not possible to obtain, commonly used statistical analysis is an alternative approach.
PEELing conducted cutoff analysis on all submitted labelled-to-control ratios individually and, for the final proteome, retained only those proteins that passed the cutoff of all ratios, which further removed contaminants.As shown in Figure 1f, PEELing displayed the post-cutoff proteome and provided information of the top 100 most enriched proteins for each labelled-tocontrol ratio (Figure 1g).On the website, each UniProt accession number is a clickable link to the corresponding UniProt protein page (Consortium et al., 2023).PEELing then transmitted the postcutoff proteome to the Panther server (Mi et al., 2019;Thomas et al., 2022) for over-representation analyses on protein localization (Figure 1h), function (Figure 1i), and pathway (Figure 1j) and revealed an enrichment of cell-surface proteins related to cell adhesion and neuronal development, perfectly matching the dataset-a cell-surface proteome of developing Purkinje cells.
For custom TP and FP references, PEELing requests two tab-separated value (.tsv) files from the user, each containing one column of UniProt accession numbers (e.g., Supplementary File 10-17).Despite the robustness of this algorithm and its tolerance to imperfect references (Supplementary Figure 4), constructing high-quality references is essential for proper data filtering and interpretation.The TP reference should contain proteins known to be in the designated cellular compartment while the FP reference should contain proteins that are, to one's best knowledge, not in this compartment.Curated Swiss-Prot/UniProt and Gene Ontology Cellular Component (GOCC) databases, as well as relevant literature, provide resources for creating the TP and FP references.More TP and FP examples and their designing rules can be found in (Hung et al., 2016) and (Cho et al., 2020).

Reference-based data quality check and cutoff analysis
For each enrichment index (e.g., each labelled-to-control ratio), PEELing first ranks all proteins in descending order according to this index.For each protein on the ranked list, accumulated true positive count and false positive count above this ranking position are calculated to obtain true positive rate (TPR), false positive rate (FPR), and their difference (TPR-FPR) at this ranking position (visualized in Figure 1d and other plots of TPR, FPR, and TPR-FPR).A receiver operating characteristic (ROC) curve is produced accordingly (Figure 1e and other ROC curves).
For each enrichment index, the cutoff is set where TPR-FPR maximizes, representing the largest segregation of intended signal (TP) and unintended noise (FP).PEELing conducts cutoff analysis on all submitted enrichment indexes individually and, for the final proteome, retains only those proteins passing the cutoff of all or multiple indexes.Both the PEELing website and command line program offer the optional "Tolerance" setting, enabling users to control the stringency of the cutoff.By default, it is set to 0, meaning that a protein must pass the cutoff of all indexes to be included in the final proteome.If it is set to n, a protein can fail the cutoff in up to n indexes and still be included in the final proteome.Despite the flexibility, we recommend setting the tolerance value to a small number to better filter out contaminants.

Protein annotation
From the post-cutoff proteome, PEELing retrieves from UniProt and displays gene names, protein names, organisms, and protein lengths of the top 100 enriched proteins based on each index.For protein ontology, PEELing sends the post-cutoff proteome to the Panther server (Thomas et al., 2022;Mi et al., 2019) for over-representation analysis.The annotation datasets used include: GO slim cellular component, for protein localization; GO slim biological process, for protein function; and reactome pathway, for signaling pathway.Results are ranked in ascending order by false discovery rate (FDR).Top 10 terms are listed along with their FDRs.green) plotted against Random_1 (b) and Random_2 (e) based ranking (x-axis).(c,f) Receiver operating characteristic (ROC) curves, based on Random_1 (c) and Random_2 (f), respectively.AUC, area under the curve.(d,g) Subcellular localization annotation did not enrich any cell surface related terms.d, Random_1; g, Random_2.FDR, false discovery rate.
(a-f) Analysis of a mitochondrial nucleoid proteome of human embryonic kidney 293 cells, profiled by (Han et al., 2017) using the peroxidase APEX2 (Supplementary File 7).(g-l) Analysis of a mitochondrial matrix proteome of human embryonic kidney 293 cells, profiled by (Branon et al., 2018)

Figure 1 .
Figure 1.PEELing workflow and the analysis of a mouse cell-surface proteome.