Vascular proteome responses precede organ dysfunction in sepsis

Vascular dysfunction and organ failure are two distinct, albeit highly interconnected clinical outcomes linked to morbidity and mortality in human sepsis. The mechanisms driving vascular and parenchymal damage are dynamic and display significant molecular crosstalk between organs and tissues. Therefore, assessing their individual contribution to disease progression is technically challenging. Here, we hypothesize that dysregulated vascular responses predispose the organism to organ failure. To address this hypothesis, we have evaluated four major organs in a murine model of S. aureus sepsis by combining in vivo labeling of the endothelial proteome, data-independent acquisition (DIA) mass spectrometry, and an integrative computational pipeline. The data reveal, with unprecedented depth and throughput, that a septic insult evokes organ-specific proteome responses that are highly compartmentalized, synchronously coordinated, and significantly correlated with the progression of the disease. Vascular proteome changes were found to precede bacterial invasion and leukocyte infiltration into the organs, as well as to precede changes in various well-established cellular and biochemical correlates of systemic coagulopathy and tissue dysfunction. Importantly, our data suggests a potential role for the vascular proteome as a determinant of the susceptibility of the organs to undergo failure during sepsis.


Introduction
Sepsis is a dysregulated host response to infections that can rapidly evolve into a lifethreatening pattern of multiple organ dysfunction(1). The disease burden of sepsis has been estimated to ~50 million incident cases and ~11 million fatalities each year, which accounts for almost 25% of all annually reported global deaths (2). Despite these shockingly high levels of morbidity and mortality, no specific sepsis biomarkers are currently available. Treatments remain generic and limited to broad-spectrum antibiotics therapy, intravenous fluid resuscitation, and supportive care (3,4). The molecular basis of disease progression is also poorly understood, hindering the early recognition and management of critically ill patients in the intensive care units (ICUs).
Vascular dysfunction and organ failure are two highly intertwined clinical outcomes that determine most of the morbidity and mortality of human sepsis (5). Being a systemic syndrome, sepsis is characterized by extensive molecular crosstalk between dysregulated inflammatory, coagulopathic and metabolic processes, operating both at the local and systems levels, and compartmentalized across the blood-tissue interfaces of the organs.
Therefore, dissecting the individual contribution of vascular dysfunction to the development of organ damage is a difficult task. Nevertheless, understanding the spatiotemporal relationships between vascular and parenchymal tissues at the molecular level might unveil novel ways to impact sepsis outcomes. For example, pharmacological stabilization of the angiopoietin (Ang)/Tie2 axis reduces vasculopathy and inflammation in preclinical models of sepsis, directly implicating endothelial dysfunction in the etiology of organ failure and emphasizing the vasculature as a legitimate therapeutic target (6)(7)(8).
However, the temporal and causal links between vascular dysfunction and organ damage are still poorly understood, and new methodological frameworks are needed to capture disease progression in a compartmentalized and time-resolved fashion.
Sepsis induces dramatic perturbations in the plasma and organ proteomes of both humans and mice, highlighting the potential of proteomics as sensitive molecular readouts of disease progression (9)(10)(11)(12)(13)(14). However, most of the proteomics studies to date are limited to the analysis of single time points and/or single tissue compartments. In fact, despite the recurrent observation that vascular failure lies at the core of sepsis pathology, methodological limitations still preclude the detailed study of vascular proteome changes in sepsis, and their contribution to organ damage. Vascular proteome profiling in vivo is particularly challenging, and the suitability of in vitro cellular systems has been questioned in multiple studies (15,16). We have previously developed chemical-biology methods to tag vascular proteins in vivo, and reported that S. aureus sepsis triggers substantial changes in the murine vascular proteome (17,18). However, the temporal dynamics of the vascular proteome during infection, its molecular relationships with plasma and extravascular compartments, as well as its contribution to overall sepsis progression remain unaddressed.
In this study, we hypothesize that dysregulated vascular proteome responses predispose the organism to organ failure. Using chemical-proteomics and an integrative computational pipeline, we generated a large-scale molecular description of the proteome trajectory of sepsis, in a compartmentalized and time-resolved fashion. We provide evidence for the presence of organotypic proteome changes triggered by S. aureus sepsis across the blood-tissue interfaces of four major murine organs (SI appendix, Fig. S1).
These proteome responses are largely compartmentalized, temporally coordinated, and significantly correlated with the progression of the disease. Vascular proteome alterations were also found to precede changes in well-established cellular and biochemical correlates of sepsis, opening a window for the discovery of new diagnostic markers and novel therapeutic targets. More importantly, our data suggests a potential role for the vascular proteome as a determinant of the susceptibility of the organs to develop coagulopathy and tissue dysfunction.

Results
A two-staged model of the plasma proteome response to sepsis. To start tracking the progression of the host response to S. aureus bacteremia, mice were intravenously challenged with a LD100 bacterial dose, and their plasma proteomes were analyzed at 0h, 6h, 12h and 24h post infection. To facilitate accurate time-resolved quantitative proteomics, a data-independent acquisition (DIA) workflow based on Sequential Window Acquisition of all Theoretical Mass Spectra (SWATH) was applied (19). DIA-SWATH-MS uses broad isolation windows to measure almost all MS detectable peptides in a biological sample. This results in an increased number of protein identifications and more robust quantification compared to approaches based on data-dependent acquisition (DDA).
Initially, 1μL of non-depleted unfractionated plasma from S. aureus infected or healthy control mice was subjected to proteomics analysis. Protein intensities were extracted from the DIA data using predicted spectral libraries generated through deep neural networks in the software suite DIA-NN (20), and further constrained by experimental information from public databases (Material and Methods). As shown in Fig 1A, the plasma proteome is significantly altered during infection, and clustering analysis segregates the samples into early and late responses (Fig. 1B). To uncover the main protein changes driving the partitioning of the samples, the data were deconvolved through non-negative matrix factorization (NMF) (21). Three NMF clusters or "protein signatures" were successfully identified and condensed into discrete proteome trajectories (i.e., temporal intensity patterns) (Fig. 1C). The NMF Signature 1 identified the early (6h) accumulation of inflammatory proteins, complement, coagulation factors and acute-phase reactants (APRs), including the C-reactive protein (CRP), the serum amyloid protein 2 (SAA2), and soluble CD14, as well as antimicrobial factors from the cathelicidins family (Fig. 1D). The NMF Signature 2 was characterized by the plasma accumulation of liver metabolic enzymes (Fig. 1E), whereas the NMF Signature 3 identified the depletion of hepatic factors involved in lipid metabolism (Fig. 1F). These two last signatures were linked to the late time points (12h and 24h), and were indicative of the occurrence of acute liver failure.
Based on these data-driven patterns, the plasma proteome analysis reflected a twostaged model of the progression of sepsis, each stage associated with specific inflammatory and metabolic alterations. (Fig. 1G).
Vascular proteome responses are temporally coordinated in an organ-specific fashion. Vascular surface proteins are located at the blood-tissue interface of the organs, which makes them potential targets of early molecular perturbations induced by blood infections. To examine whether sepsis triggers a similar staging pattern in the vascular proteome as observed in the plasma proteome, infected mice underwent systemic perfusion using Sulfo-NHS-biotin to tag proteins located on the vascular surfaces (17). This technique greatly enriches for proteins accessible to the blood, with very little penetration into the parenchyma based on streptavidin staining of tissues (17). Biotinylated proteins were enriched on streptavidin columns and analyzed by DIA-SWATH-MS. Thousands of proteins were accurately quantified across the vasculature of the organs ( Fig. 2A), resulting in a substantial increase in the number of identifications compared with our previous studies using regular shotgun proteomics (SI appendix, Fig. S2, Supplementary   table 1). Additionally, statistical analysis indicated that a large fraction of the vascular proteome was significantly perturbed during infection (Fig. 2B). Clustering of the data shows that the septic vascular proteome has a more complex kinetics than the plasma proteome, displaying clusters with similar temporal behaviors across the organs, and protein clusters exhibiting organ-specific changes (Fig. 2C).
To determine if vascular proteome trajectories shared across the organs reflect the regulation of similar proteins, all vascular clusters were extracted, and both their compositions, sharedness degree and functional enrichment were evaluated. Four clusters (heart cluster-5, spleen cluster-2, liver cluster-3 and kidney cluster-3; Blue) shared a significant number of proteins (~15%), and common biological pathways, including neutrophil and platelet degranulation, activation of coagulation and complement, and regulation of leukocyte-endothelial interactions (Fig. 2D-E). The abundance of all proteins in these clusters was linearly increased over the time of infection (Fig. 2F). Additionally, three other clusters (spleen cluster-6, liver cluster-1 and kidney cluster-1; Green) also shared many proteins (~14%), as well as pathways related to phagocytosis and the adaptive immune response, particularly receptor mediated interactions ( Fig. 2E and 2G).
Interestingly, the abundance of all proteins in these clusters dramatically dropped at 6h post infection, suggesting the rapid downregulation and/or shedding of a large fraction of the vascular proteome. In sharp contrast to these general responses across the organs, evidence of tissue-specific changes was also detected. For example, some specific clusters (heart cluster-3, spleen cluster-5, liver cluster-4 and kidney cluster-4; Orange) shared only a few proteins (~2%), suggesting the induction of organotypic responses (Fig.   2D). Functional enrichment linked these clusters to ion transport and homeostasis (heart cluster-3), metal sequestration by antimicrobial proteins (spleen cluster-5), cellular responses to labile heme (liver cluster-4) and mitochondrial dysfunction (kidney cluster-4). Notably, despite the marked differences in their protein contents, these clusters  3A), whereas the heart exhibited a more delayed response (no notable difference between 0h and 6h, but clear differences between samples at 6h and 12h) (Fig. 3B). The kidney and the spleen were also quick responders, although the spleen showed the largest parenchymal proteome changes of all organs, with >2000 proteins significantly altered during sepsis (Fig. 3C-D). Interestingly, the analysis of the vascular proteomes shows a rather uniform pattern ( Fig. 3E-H), in contrast to the heterogeneity of the temporal patterns of the parenchymal responses. All vascular proteome datasets displayed a distinct separation of the 0h and 6h time points, and clustering of the 12h and 24h time points, independently of the organ. Despite this similar behavior, the number of dysregulated proteins that were shared across the organ vascular fractions at 6h was very low (41, ~1.4%). These findings suggest that sepsis induces early changes in the vascular proteome of all organs, but each proteome response is most likely shaped by organ specific factors.
Building the proteome trajectory of sepsis through a higher-order networking approach. To reconstruct the disease trajectory of sepsis using the host proteome response as a readout, we developed a novel computational pipeline to integrate the data across time points, organs, and tissue compartments. Correlative Stratification of Proteome Trajectories (COR-SPOTS) was designed as a higher-order networking approach to store, organize and integrate multiple data layers based on temporal and functional relationships between the data points (see material and methods). Significantly dysregulated proteins from each organ and compartment (plasma, vascular, and parenchymal) were parsed through STRING-DB to group proteins based on functional associations (22). Highly interconnected clusters of nodes were then identified through the Louvain algorithm for community detection, and collapsed into new single nodes. We then computed ~24 million pairwise spearman correlations over all combinations of detected community nodes, which resulted in ~3.7 million significant correlations (false discovery rate; FDR<0.10) that were used to define the linking edges of the network. Approximately an equal number of significant positive and negative correlations were observed. Other similarity relationships between the nodes such as the content overlap and the functional enrichment were also stored within the network, resulting in a compiled and fully searchable small database containing all proteomics results and metadata from this study, which can be found on the online repository the Network Data Exchange (NDEx) with UUID: 45474980-9d56-11eb-9e72-0ac135e8bacf (Fig. 4A).
Vascular proteome responses in the liver precede coagulopathy and hepatic damage. We have previously shown that the murine model of S. aureus sepsis suffers from a dominant coagulopathic phenotype in the liver (17). To demonstrate the utility of COR-SPOTS to dissect molecular changes across the blood-tissue interfaces, we ranked the nodes and selected the hepatic vascular, parenchymal and plasma nodes with the highest inter-correlation. Proteins stored within these highly correlated nodes displayed a bimodal network topology characterized by the presence of two distinct anti-correlated clusters of proteins (Fig. 4B). Cluster 1 contained most of the plasma and parenchymal proteins linked to triglyceride and phospholipid metabolism, whereas cluster 2 was enriched in vascular proteins linked to coagulation (Fig. 4C). Many proteins involved in lipid metabolism were rapidly diminished from the hepatic vasculature, including multiple apolipoproteins, angiopoietin-like proteins, and the glycosylphosphatidylinositol-anchored high-density lipoprotein-binding protein 1 (HDBP1). Some of these factors were also reduced in the liver parenchyma and plasma during infection, but these changes were  (Fig. 3B). Inspection of the cardiac nodes in the network showed a distinct separation of the parenchymal nodes from the vascular nodes, distinguishing the heart from the other organ proteomes (Fig. 4A). We hypothesize that this disparity in the cardiac protein response was linked to a difference in the kinetics of the disease trajectory of sepsis in the heart, and perhaps linked to the presence or absence of major cellular events, such as bacterial invasion and/or leukocyte infiltration of the organs. The bacterial burden of the organs was in line with this hypothesis, showing that despite the presence of bacteria in the liver, kidney, and spleen already by 6h post infection, no bacteria were found in the heart until 12h after inoculation  5B). Histological analysis confirmed that neither bacterial microabscesses nor immune infiltrates were observable in heart tissue until 12h post infection (Fig. 5C). Taken together, these data suggests that in this model of S. aureus sepsis, bacterial dissemination and leukocyte infiltration are temporally delayed in the heart compared to the other organs, opening an interesting window to capture very early molecular changes in the vasculature as a response to systemic inflammation.
Next, we focused on the changes in the heart vasculature at 6h to identify molecular alterations that precede bacterial and leukocyte invasion of the tissue (Fig. 5D). Multiple vascular adhesion proteins were rapidly induced during infection, such as CADM1, ICAM1 and VCAM1. Upregulation of junctional endothelial proteins such as JAM3 was also observed, together with an increase in the levels of ADAM17, a protease that has been linked to most of the early shedding of the vascular surfaces upon inflammation. Changes in the glycocalyx were also noticed, including upregulation of heparan sulfate proteoglycans such as GPC4, and very robust induction of the hyaluronic acid synthase HAS1. In fact, early upregulation of HAS1 was consistently observed across all vascular surfaces, directly implicating HAS1 in glycocalyx remodeling during sepsis (Fig. 5E).
Compartmentalization of acute-phase reactants and immunoglobulin shedding are early vascular responses to S. aureus sepsis. Acute phase reactants (APRs) are inflammation markers that exhibit rapid and substantial changes in serum concentration during inflammation. Most APRs are synthesized in the liver as a response to both local and systemic inflammation, and some of them get associated with the endothelium, where they exert multiple functions. We used COR-SPOTS to map the dynamics of APRs during infection across organs and compartments. Although the dynamics of the APRs in plasma followed a linear increase, as has been previously described, when mapping the levels of APRs associated with parenchymal and vascular compartments, we noticed at least two distinct patterns. Type I APRs experienced rapid induction in all organs and compartments, and steadily accumulated over time. This pattern was characteristic of APRs such as SAA1, SAA2, HPT and LBP (Fig. 5F-I). However, we also identified a second group or Type II APRs, whose levels were compartmentalized and sensitive to the disease status of the organs. For example, A1AG2 and SAMP levels in liver parenchyma and vascular fractions followed the same kinetics as in plasma (Fig. 5J-K). Interestingly, their levels in cardiac parenchyma and vasculature remained unchanged at 6h, followed by a linear increase at 12h and 24h. Other APRs such as SAA3 followed a similar trend, except for a temporal difference between the cardiac vascular proteome response (up at 12h) and the cardiac parenchymal response (up at 24h). CD14 was also increased in the liver and in the heart vasculature by 6h, but increased cardiac parenchymal levels were only detected by 24h. We have previously shown that PRG4, a protein that behaves as an APR in many inflammatory models gets deposited in the liver vasculature in this sepsis model of S. aureus bacteremia by 24h. Looking at the dynamics of PRG4 across organs and compartments, it became apparent that vascular accumulation starts early in the liver and increases linearly, but association with the heart vasculature could not be detected until the late time points. Finally, we also observed that in contrast to the increasing levels of many vascular adhesion molecules and APRs, the most dramatic downregulation taken place in the cardiac vasculature at 6h was the shedding of many immunoglobulins (IgGs), probably loosely associated with the vascular wall, without changing IgG total plasma levels. All these changes are consistent with a massive remodeling of the vascular proteome during sepsis, even before bacteria or infiltrating immune cells are observed.
These changes are expected to significantly alter the physicochemical properties of the vascular surfaces, and their organotypic responses to systemic infection.

Discussion
Understanding the relationships between vascular failure and organ damage has important implications for sepsis therapeutics and diagnostics. Even though many patients quickly respond to standard care, an unacceptably high fraction of cases still progress into intractable shock responses with a high risk for long term morbidity and deadly outcomes.
Vascular stabilization as a therapeutic goal in sepsis has been discussed for years, although few preclinical studies have been conducted to assess the impact of vasculoprotective approaches on relevant sepsis outcomes, and so far, none of them have succeeded in making a translational jump from the bench to the bedside. Part of the problem lies on the enormous heterogeneity of the organ vasculatures, and a lack of detailed knowledge on the molecular composition of the vascular surfaces and how they are affected by sepsis.
Here, we demonstrate that sepsis triggers a series of well-coordinated proteome changes across the blood-tissue interfaces of the organs, but the nature of those changes and their specific temporal patterns are largely organotypic. These organ-specific patterns seem to be partially linked to the occurrence of major sepsis "checkpoints", such as systemic vascular activation or the actual invasion of the organs by pathogens and infiltrating immune cells. Interestingly, the liver undergoes a rapid reprogramming of its vascular proteome by downregulating lipid metabolic functions and becoming primed towards a prothrombotic and anti-fibrinolytic state, long before signs of organ dysfunction and actual thrombosis are observable. This finding is remarkable, firstly because it suggests that vascular proteome alterations can have a yet unrealized potential as a source of new generation diagnostic and prognostic markers. Secondly, it also opens the possibility that timely therapeutic interventions targeting the vasculature might prevent irreparable organ damage. For example, we observed that hepatic thrombosis is preceded by early depletion of vascular TFPI, upregulation of VWF and intravascular platelet deposition.
Alterations in the extrinsic pathway of coagulation were also reported in baboon models of gram-negative sepsis, and exogenous TFPI administration rescues the organs from coagulation-induced damage (23,24). However, there can be difficulties in translating some of the findings from preclinical models to therapy (25,26). For example, results from clinical trials to evaluate the efficacy and safety of Tifacogin, (recombinant TFPI) to treat sepsis have so far been disappointing, emphasizing the complexity of the host response to systemic infection.
The host response to sepsis is overwhelmingly complex since it encompasses crosstalk between multiple systems, and a time-dependent staging, where both pro-and antiinflammatory components might be operating at different time points. Predicting the disease trajectory of sepsis is therefore a most sought-after clinical goal. Here we show that the host proteome response to sepsis is a powerful molecular readout of the septic response. Using a novel networking approach, COR-SPOTS, we were able to identify very early vascular changes that even precede bacterial and leukocyte infiltration of the tissue.
Some of these responses have been previously reported, such as the upregulation of vascular adhesion molecules and glycocalyx components, whereas other were completely unexpected, such as the compartmentalization of the acute-phase reaction and the massive shedding of vasculature-associated IgGs. Future work should explore how these early vascular events are coordinated, for example by focusing on transcriptional and/or posttranslational mechanisms of proteome regulation. Similarly, our approach might also facilitate the quick evaluation of theragnostic markers that are sensitive to treatment options, in order to find novel indicators that reflect the status of critically ill patients and to accelerate the process of clinical decision-making in intensive care units.

Materials and Methods
Bacterial preparation and infection. S. aureus strain USA300/TCH1516 was grown at 37 °C in liquid cultures of Todd-Hewitt broth (THB, Difco) with agitation (200 rpm), and later incubated in 5 mL of fresh THB overnight. Roughly, 400μL of the overnight culture was inoculated into 6 mL of fresh THB and incubated until an OD600 = 0.4 was reached.
Bacteria were centrifuged, washed twice, and resuspended in PBS. 8-10-week-old C57Bl/6 mice were intravenously infected through the retroorbital sinus with 5 × 10 7 CFU of the bacteria culture in 100μL PBS, or just with 100μL PBS in the control group. Animals were euthanized at 6h, 12h or 24h using isoflurane. One group of animals was immediately subjected to chemical biotinylation perfusions as described below, and another group was subjected to cardiac puncture to collect blood samples. After these procedures, multiple organs were collected from both groups (liver, kidney, heart and spleen). All animals were housed in individual ventilated cages in vivaria approved by the Association for Assessment and Accreditation of Laboratory Animal Care at the School of Medicine, UC San Diego. All experiments followed relevant guidelines and regulations consistent with standards and procedures approved by the UC San Diego Institutional Animal Care and Use Committee (protocol #S99127 and #S00227M).

Systemic chemical perfusions. Chemical perfusions were performed as previously
reported (17). Briefly, mice were subjected to a median sternotomy and the left ventricle of the heart was punctured with a 25-gauge butterfly needle (BD Vacutainer). A small cut was made in the right atrium to allow draining of perfusion solutions. All ice-cold perfusion reagents were infused using a perfusion pump (Fischer scientific). Blood was washed out with PBS for 5 min at a rate of 5 mL/min. A solution containing 100mM EZ-link Sulfo-NHSbiotin (Thermo Fischer) in PBS, pH 7.4 was perfused at a rate of 3 ml/min for 10 min.
Control animals were perfused with PBS only.
Organ preparations. Collected organs were homogenized in a buffer containing 5M urea, 0.25M NaCl and 0.1% SDS. Samples were briefly centrifuged for 5 min and the supernatant was transferred to new tubes. Protein was quantified by standard BCA assay (Thermo Scientific) as per manufacturer instructions. Samples were stored at −80 °C until further analysis. In parallel, a piece of each organ was homogenized in 1 mL ice-cold PBS, and the samples were plated on Todd-Hewitt Agar for CFU analysis.
Blood Chemistry and Coagulation Assays. Whole blood was collected via cardiac puncture and placed in a pro-coagulant serum tube (BD Microtainer #365967) for 4 hours at room temperature. Serum was isolated by spinning the tubes at 2000xg and collecting the supernatant. All samples were frozen and thawed once before analysis. Blood chemistry parameters were measured on a Cobas 8000 automated chemistry analyzer (Roche) with a general coefficient of variance of <5%. All samples were frozen and thawed no more than two times before analysis. Blood coagulation factor assays were performed as previously described (27).  (28), and from two published large-scale DIA proteomics studies (11,29). The compiled library with a final number of 667,455 precursors was used for DIA data extraction with a protein q-value of 0.01 and RT-profiling enabled.

Bacterial Colony Forming Units
Statistical Analysis. All statistical methods were implemented using Python 3.6.10.
Proteomics results from the DIA-SWATH-MS analysis were filtered using a one-way ANOVA followed by a Benjamini-Hochberg procedure to control for a false discovery rate (FDR) < 0.10. Statistically significant identifications were further subjected to principal component analysis (PCA). Proteins were given a standardized score using a z-score normalization. Proteomics results were separately analyzed using Welch's t-test to generate volcano plots and heatmaps.

Clustering.
Non-negative matrix factorization: Plasma DIA proteomics was analyzed using nonnegative matrix factorization (NMF). Following pseudocode of Lee and Seung's multiplicative update rule, the distance between the dataset matrix (D) and the dot product of hypothetical partitions (W and H) was minimized using a coordinate descent algorithm (21). The two resulting matrices were constrained during the optimization to allow for better interpretability. First, the predictions in the matrix H must fall within the range 0.0 to the maximum observed log2(relative abundance). Second, the distributions in the W matrix must fall in the range 0.0 to 1.0 and must sum across signatures to 1.0. In order to determine the number of predicted proteome signatures, k, for each value k, 200 random initializations were made. The average Pearson correlation between the ground truth hierarchy learned from D and the weights hierarchy learned from W and average distance scores (||D -WH||F) are plotted for visual inspections at each k. Cluster number k was finally selected using the elbow selection method as well as meaningful investigation of correlation and proteome signature content.
Abundance Clustering: Significant proteins were grouped based on their relative abundances using k-means clustering. K selection was done through the visualization of distances to cluster centers across k values up to 20 and selecting based on the elbow method (32).
Functional Clustering: Protein identifications were parsed through the medium confidence (PPA>0.4) STRING-DB to build preliminary networks based on physical or functional associations. These networks were subjected to the Louvain method for community detection, a modularity based clustering algorithm. The Louvain clustering was implemented using the Python packages NetworkX and python-Louvain with default settings (33). To circumvent the heuristic property of the Louvain algorithm, ~1000 parallel random starts were initially performed. Using the combined results from these 1000 iterations, a matrix of cluster clarity was built to store the average number of times two proteins appear in the same community, divided by the number of iterations. Using this community clarity matrix, a k-means clustering was finally performed with k being the average number of communities found across the 1000 runs.
Cluster Annotations: The Jaccard similarity index was also used to determine the overlap of proteins contained between two clusters by quantifying the intersection over the union.
A ranked list of functional associations for each cluster were defined by using a hypergeometric test (SciPy v1.4.1). Using the Tau-b statistic from the Kendall rank correlation test (SciPy v1.4.1), functional terms between two clusters were compared by ranked enriched terms across GO terms, Reactome, and KEGG pathways.
Higher-order Molecular Networking. All significant protein identifications from all datasets (i.e., different time points, organs, and compartments) were parsed through the functional clustering pipeline described in the previous section. Each of these higher-order community nodes contained a variable number of proteins (ranging from just a few to hundreds of proteins), operating within similar biological processes. Finally, we computed pairwise spearman correlations over all combinations of detected community nodes, which resulted in detection of significant correlations that were used to define linking edges between the nodes. Benjamini-Hochberg procedure was used to control for a false discovery rate (FDR) < 0.10. The multi-edge was then constructed using three pieces of relational information. Contained in the multi-edge are the scores: the proportion of significant pairwise spearman correlations, the Jaccard similarity index, and the Tau-b statistics previously defined. Visualization and analysis of the network layers were conducted through Cytoscape (34).

Code Availability
We provide the code library in Python described in this work through Github: https://github.com/LewisLabUCSD/CORSPOTS. We provide jupyter notebooks in python to generate our figures and analysis.

Declaration of interests:
The authors declare that they have no conflict of interest.