Towards a Hypothesis-free Understanding of How Phosphorylation Dynamically Impacts Protein Turnover

Abstract

The turnover measurement of proteins and proteoforms has been largely facilitated by workflows coupling metabolic labeling with mass spectrometry (MS), including dynamic Stable isotope labeling by amino acids in cell culture (dynamic SILAC) or pulse SILAC (pSILAC). Very recent studies including ours have integrated the study of post-translational modifications (PTMs) at the proteome level (i.e., phosphoproteomics), with pSILAC experiments in steady state systems, exploring the link between PTMs and turnover at the proteome-scale. An open question in the field is how to exactly interpret these complex datasets in a biological perspective. Here, we present a novel pSILAC phosphoproteomic dataset which was obtained during a dynamic process of cell starvation using data-independent acquisition MS (DIA-MS). To provide an unbiased “hypothesis-free” analysis framework, we developed a strategy to interrogate how phosphorylation dynamically impacts protein turnover across the time series data. With this strategy, we discovered a complex relationship between phosphorylation and protein turnover that was previously underexplored. Our results further revealed a link between phosphorylation stoichiometry with the turnover of phosphorylated peptidoforms. Moreover, our results suggested that phosphoproteomic turnover diversity cannot directly explain the abundance regulation of phosphorylation during starvation, underscoring the importance of future studies addressing PTM site-resolved protein turnover.

1. Introduction

The turnover rate of proteins and proteoforms is a crucial concept in several disciplines including biochemistry, pharmacology, and systems biology. The measurement of diverse protein turnover and lifetime in cells has been greatly facilitated by the metabolic labeling coupled with mass spectrometry (MS), involving dynamic stable isotope labeling by amino acids in cell culture (SILAC) experiments ^[1], or pulsed SILAC (pSILAC) ^[2,3]. In a typical pSILAC workflow, cells are first cultured in the “normal” or light isotopic condition (L) and then switched to the heavy isotopically labelled medium (H; e.g., with Lysine and Arginine usually labeled with heavy stable isotopes of carbon and nitrogen) in a defined time course. Cells are then lysed, so that proteins are harvested, digested by trypsin, and identified by liquid-chromatography mass spectrometry (LC-MS) analysis. Importantly, during the time course the peptide H/L ratios can be quantified by MS to determine the incorporation rate of newly synthesized proteins (labeled by H). At the steady state (i.e., the total degraded and synthesized protein pools in the cells are balanced over time), the constant protein-specific turnover rate can be determined by calculating H/L exchanging rates using mathematical modeling ^[4–9], while in dynamic systems and in vivo the situation can be complicated by several aspects. Anyhow, protein turnover measurements have successfully shed light on many aspects of protein homeostasis and protein quality control processes in healthy and disease contexts ^[10–17]. For example, in aneuploid trisomy 21 cells, the accelerated protein turnover of protein complex subunits coded by chromosome 21 were confirmed to contribute to the gene dosage compensation and other proteotypes in human Down Syndrome ^[13]. Moreover, in the last five years there have been relevant advancements in the quantitative MS method for pSILAC, as we and others have demonstrated that the MS2-based data-independent acquisition mass spectrometry (DIA-MS) ^[18,19] can achieve higher reproducibility and accuracy than traditional MS1-based data-dependent acquisition (DDA) for quantifying SILAC ratios and protein turnover rates ^{[13,15,20,21]}. For more detailed discussion on pSILAC-based protein turnover measurement, refer to a number of wide-ranging recent reviews on this topic ^[8,22,23].

Protein post-translational modifications (PTMs) are considered essential for the functional diversity of the proteome. Hence, it is central to understand how these modifications may impact and possibly even regulate protein lifetime and turnover. For example, it is well-known that protein ubiquitination most commonly results in protein degradation via the ubiquitin-proteasome pathway ^[24]. In addition, protein glycosylation is known to be crucial for quality control of protein folding and endoplasmic reticulum (ER)-associated degradation ^[25]. Many other PTM types have been found to influence protein degradation and turnover dynamics ^[26]. Taking phosphorylation as an example, the phosphorylation of certain amino acids (or P-sites) in particular pathways, such EGFR/MAPK signaling ^[27,28] and cell-cycle control ^[29,30], frequently lead to active protein degradation, which were thus termed “phosphodegrons” ^[31,32]. On the other hand, a large number of rigorous studies has demonstrated that many site-specific phosphorylation can also increase protein stability by e.g., protecting them from proteolysis and suppressing their degradation (see examples ^[33–43]). Such P-sites, therefore, can function as “phosphostabilons”. The mechanisms underlying “phosphostabilons” often involve the phosphorylation-mediated change in protein-protein interaction (PPI) that leads to the stabilization of the target protein, although the exact molecular mechanisms remain to be fully explored. Altogether, several PTM types and sites, such as “phosphodegrons” and “phosphostabilons”, have been found with important functional consequences on protein degradation and turnover. At the same time only seldomly these studies have been systematic.

Despite the biological and translational significance of investigating how protein PTMs interact with protein turnover, previous studies mostly used low-throughput protein detection methods such as antibody-based western blotting (WB) combined with genetic approaches. Very recently, we and others have directly combined PTM proteomic approaches (such as phosphoproteomics) and pSILAC experiments to reveal the quantitative impact of PTMs on protein turnover on the proteome-scale ^[44,45] In both studies, the pSILAC data were processed using classic mathematical calculations to infer the protein half-live times, with the assumption that at the steady state the total protein amount (i.e., the light plus heavy labeled proteins) is equal across time points. The same assumption was applied to PTM-modified proteins. Notably, in these studies a site-resolved protein turnover profiling was used, so that the peptidoforms carrying particular PTM sites are compared to the unmodified peptidoform counterparts (i.e., DeltaSILAC in Wu et al. or SPOT in Zecha et al.) ^[44–46]. These studies derived useful biological insights. For example, in Wu et al. ^[44], we have discovered the majority of P-sites impaired protein clearance and verified several of them by traditional WB approach. At the same time, the application of the classic assumptions that the overall protein levels are not changed for inferring the turnover of modified protein forms, although useful and comprehensible, is a simplification introduced to facilitate the analysis of complex data at the steady state.

Herein, we further aim to improve the data analysis strategy interrogating the association between PTM and protein turnover in dynamic systems. To be able to do so, we have introduced an analysis workflow accounting for the following four considerations. Firstly, as Hammarén et al. recently pointed out ^[47], since protein modifications generally happen during or immediately after translation, the turnover of unmodified and different modified proteins could be interlinked through a “proteoform modification network structure” ^[47]. Therefore, it might be nontrivial to directly apply the traditional modeling of pSILAC for PTM turnover measurement. Interestingly, however, the delaying effect of phosphorylation on protein clearance can be site-specific and of variable extent ^[44], and has been demonstrated with strong functional implications ^[33–43]. Furthermore, most current PTM proteomic workflows unfortunately do not directly measure proteoforms ( > 1 million in a given cell type ^[48]). In essence, only peptidoforms have been measured in bottom-up proteomic experiments, whereas one peptidoform carrying a particular PTM site can be derived from a series of variable proteoforms ^[46]. In simple words, the exact timing of phosphorylation and degradation events is not trivial to be inferred from these new bottom-up datasets. Secondly, it is well known that many biological processes, such as cell fate decision or cell death, can in some cases require few days or even weeks, whereas both short-term and long-term PTM regulations could carry distinctive biological functions. This consideration further complicates the temporal modeling of PTM site-specific turnover. Thirdly, certain pharmacological perturbations may strongly impact the protein synthesis or degradation processes and therefore nullify the basic assumption of protein turnover in steady state. Examples include the treatments of small-molecule inhibitors of protein synthesis (e.g., cycloheximide, puromycin) or degradation (e.g., MG-132, bortezomib). Finally, although a few approaches were previously developed to separate the contribution of synthesis to degradation to the measured protein turnover in the dynamic system (i.e., non-steady states) ^[9,10,49], these approaches currently do not allow to interrogate PTM site-resolved datasets. In the present study, we herein propose a “hypothesis-free” data analysis strategy following and building upon the DeltaSILAC concept ^[44,46]. Although preliminary, our analysis successfully elaborated the complexities of P-site specific turnover behaviors in a dynamic cellular system of starvation stress, providing a flexible framework for analyzing future experiments dealing with PTM site-specific turnover.

2. Materials and Methods

2.1. Materials and cell Culture

PC12 cells (ATCC, #CRL-1721) were cultured at 37 °C, humidified 5% CO₂ in complete RPMI-1640 medium containing L-glutamine (Gibco^™, #11875093), supplemented with 10% horse serum (Gibco, #16050122), 5% fetal bovine serum (Fisher Scientific, #10438026), 1 mM sodium pyruvate (Quality Biological, INC. #116-079-721EA),10 mM HEPES (Gibco, #15630080) and 1 x penicillin-streptomycin (Sigma-Aldrich, #P0781). Cells were cultured in flasks before seeding for the final experiment. For the pSILAC experiment, the SILAC RPMI 1640 media lacking L-Arginine and L-Lysine (Thermo Scientific, #88365) was used, supplemented with HEPES, sodium pyruvate, and penicillin-streptomycin solution. The heavy L-Arginine-HCl (13C6, 15N4, purity > 98%, #CCN250P1), and heavy Lysine-2HCl (13C6, 15N2, purity > 98%, #CCN1800P1) were purchased from Cortecnet and spiked into the culturing medium in the same manner as described previously for RPMI medium ^[50].

For dynamic SILAC or pSILAC experiments, cells were seeded on 100 mm Corning BioCoat collagen Ι coated dishes (Corning, #354450) in compete growth medium at a density of 2 × 10⁶ cells/dish. Forty-eight hours after seeding, the medium was exchanged to a serum-free medium, and the cells were starved for 16 hours. After synchronization, the light medium was switched to heavy serum-free medium, and the cells were harvested at 1, 4, 12, 24, and 48 hours in three biological replicates per time point. To harvest cells immediately at each time point, the medium was removed fast using vacuum aspiration and a snap-freezing step was performed by putting the dish on liquid nitrogen for 2 minutes. Then, the cell lysis buffer (10 M urea containing the cOmplete^™ protease inhibitor cocktail (Roche, #11697498001) and Halt^™ phosphatase inhibitor (Thermo Scientific, #78428)) was added into the dish directly. Cells were scraped from the surface using a scraper (RPI, #162424) and transferred into a 2 ml Eppendorf tube. The samples were stored at −80 °C.

2.2. Mass spectrometry sample preparation

As described in our previous proteome-level study ^[21], the cell samples were lysed by sonication at 4 °C for two cycles (1 min per cycle) using a VialTweeter device (Hielscher-Ultrasound Technology) ^[12,51] and then centrifuged at 20,000 x g for 1 h to remove the insoluble material. The protein concentration assay was performed using the Bio-Rad protein assay dye (Bio-Rad, #5000006), and a total of 700 μg protein mixture was transferred to new 2 ml Eppendorf tubes for the alkylation and reduction reaction. The protein mixture was reduced with a final concentration of 10 mM Tris-(2-carboxyethyl)-phosphine (TCEP) and incubated 1 hour at 37 °C. The alkylation was then performed with 20 mM iodoacetamide (IAA) in dark for 1 h at room temperature. After the reduction, five volumes of precooled precipitation solution containing 50% acetone, 50% ethanol, and 0.1% acetic acid were added to the protein mixture and kept at −20 °C overnight. The samples were centrifuged at 20,000 × g for 40 min to spin down the protein. The precipitated proteins were washed with 100% acetone and 70% ethanol with centrifugation at 20,000×g, 4°C for 40 min, respectively. The pellets were gently dried for 5 min with SpeedVac (Thermo Scientific). For trypsin digestion, 300 μL of 100 mM NH4HCO3 were added to all samples and digested with sequencing grade porcine trypsin (Promega) at a ratio of 1:20 overnight at 37 °C. The peptide mixture was acidified with 10% formic acid and then purified with C18 column (MacroSpin Columns, NEST Group INC). The amount of the final peptides was determined by Nanodrop (Thermo Scientific). 1 μg peptide/per sample was injected for MS analysis.

For the present phosphoproteome analysis, 250 μg purified peptides were used for phosphopeptide enrichment with the High-Select^™ Fe-NTA kit (Thermo Scientific, #A32992). The enrichment protocol was the same as described previously ^[52]. Briefly, the resins of one tube in the kit were used to five samples equally. The peptide-resin mixture was incubated for 30 min at room temperature and gently shaken for 10 min and then transferred into the filter tip (TF-20-L-R-S, Axygen) to remove the supernatant by centrifugation. Then the resins binding phosphopeptides were washed sequentially with 200 μL × 3 washing buffer (80% ACN, 0.1% TFA) and 200 μL × 3 H₂O to remove nonspecifically peptides. The phosphopeptides were eluted from the resins by 100 μL × 2 elution buffer (50% ACN, 5% NH₃•H₂O) and dried with SpeedVac (Thermo Scientific) immediately and resuspended with buffer A for MS analysis.

2.3. DIA Mass Spectrometry

All the proteomic and phosphoproteomic samples were measured by a DIA-MS method as described previously ^[44,50,53]. LC separation was performed on an EASY-nLC 1200 system (Thermo Scientific, San Jose, CA) using a self-packed analytical PicoFrit column (New Objective, Woburn, MA, USA) (75 μm × 50 cm length) using C18 material of ReproSil-Pur 120A C18-Q 1.9 μm (Dr. Maisch GmbH, Ammerbuch, Germany). A 150-min measurement with buffer B (80% acetonitrile containing 0.1% formic acid) from 5% to 37% and corresponding buffer A (0.1% formic acid in H₂O) during the gradient was used to elute peptides from the LC. The flow rate was kept at 300 nL/min with the temperature controlled at 60 °C using a column oven (PRSO-V1, Sonation GmbH, Biberach, Germany). The Orbitrap Fusion Lumos Tribrid mass spectrometer (Thermo Scientific) instrument coupled to a nanoelectrospray ion source (NanoFlex, Thermo Scientific) was calibrated using Tune (version 3.2) instrument control software. Spray voltage was set to 2,000 V and heating capillary temperature at 275 °C. All the DIA-MS methods consisted of one MS1 scan and 33 MS2 scans of variable isolated windows with 1 m/z overlapping. The MS1 scan range was 350 – 1650 m/z and the MS1 resolution was 120,000 at m/z 200. The MS1 full scan AGC target value was set to be 2.0E6 and the maximum injection time was 50 ms. The MS2 resolution was set to 30,000 at m/z 200 with the MS2 scan range 200 – 1800 m/z and the normalized HCD collision energy was 28%. The MS2 AGC was set to be 1.5E6 and the maximum injection time was 52 ms. The default peptide charge state was set to 2. Both MS1 and MS2 spectra were recorded in profile mode.

2.4. DIA mass spectrometry data procession

All the raw DIA-MS data analyses were performed using Spectronaut version 13 ^[54,55] by a library-based searching. For the pulsed SILAC DIA library generation, the labels were specified in the “Labeling” setting, the “Labeling Applied” option was enabled, the “Arg10” and “Lys8” were specified as SILAC labeling in the second channel, and the “In-Silico Generate Missing Channels” and the “label” workflow setting was selected ^[15]. The Rattus norvegicus protein sequence database was downloaded from UniProtKB with 29,969 entries (04/10/2018) ^[56]. Methionine oxidation was set as variable modification, and cysteine carbamidomethylation was selected as fixed modification. For the phosphoproteome data search, the phosphorylation at S/T/Y was enabled as variable modification for the generation of the DIA library. For the data search of pulsed SILAC datasets, the Inverted Spike-In workflow (ISW) pipeline was used as described previously ^[15]. The “Qvalue” was selected for all data filtering. Both peptide and protein FDR cutoff were controlled at 1%. The probability of post-translational modifications cutoff was set at 0.75 and 0 to report two datasets, respectively, and then select the overlapping hits, as described ^[57], to restrict the dataset to confidently localized PTMs while improving the data completeness ^[57]. All the other Spectronaut settings were kept as default.

2.5. maSigPro and other bioinformatics analysis

The “hypothesis-free” differential analysis was performed in R studio with the package “maSigPro” (https://bioconductor.org/packages/release/bioc/html/maSigPro.html) ^[58,59]. The Heavy/Light (H/L) ratios were log2-transfromed for time-series analysis using maSigPro. There were five time points in our study, and the polynomial regression model (degree = 3) was used to find the differential results. The FDR of differential profiles was set as 0.05, and the minimal observation value was 10. The “backward” of step.method was used to find the phosphorylation and non-phosphorylation differences. The cut-off value for the R-squared of regression model was set 0.8. The number of hierarchical clusters K was set to be 6. The design table for the maSigPro input can be found in Table S1. Only those heavy peptidoforms showing a Pearson correlation higher than R = 0.5 for both phosphorylated (P) and non-phosphorylated (NP) peptides with real temporal hours (i.e., 1, 4, 12, 24, and 48 hours) were included for the time series analysis (i.e., 2,597 of 3,755 pairs of P and NP peptides, Table S2).

Functional annotation analysis was carried out in David Functional Analysis Tool v6.8 (https://david.ncifcrf.gov/tools.jsp) ^[60,61]. For the sequence motif analysis and kinase linear motifs enrichment analysis, the ± 14 amino acids sequence windows surrounding the identified phosphorylation sites were extracted using motifeR (https://www.omicsolution.org/wukong/motifeR/) ^[62]. The linear motifs were matched using the “Add linear motifs” tool in Perseus v.1.6.14.0 ^[63]. For OmniPath kinase-substrate interaction enrichment analysis, the identified rat P-sites were first mapped to human P-sites using motifeR 2.0 (https://github.com/wangshisheng/motifeR2.0/) ^[62] with the following parameters: expectation value threshold = 0.0001, exact matching, sequence window similarity = 7. Next, the known human kinase targets were downloaded from the OmniPath database ^[64–66] and matched to the rat dataset using the human P-sites. The enrichment analysis was performed in Perseus v.1.6.14.0 ^[63] using Fisher’s exact test. The IceLogo was used for peptide sequence motif enrichment ^[67]. The circular plot in Figure 3 was generated using unique protein IDs per cluster in Metascape ^[68]. The other plots were generated by R packages (“ggplot” and ”pheatmap”) or GraphPad Prism version 9.0.0 (GraphPad Software, San Diego, CA, USA).

Figure 3: Functional characterization of phosphorylation sites and phosphoproteins in six turnover clusters identified by maSigPro.

(A) The Circos plot depicts the overlap of protein identities between the six turnover clusters (i.e., maSigPro Clusters 1–6). (B) Selected linear kinase motifs (HPRD) and kinase-substrate interactions (OmniPath) significantly enriched in the clusters are shown. The P values were estimated using Fisher’s exact test. (C) Sequence motif enrichment analysis. Amino acids significantly (P < 0.05) overrepresented or underrepresent in the sequences surrounding the phosphorylated amino acid are shown above or below the zero line, respectively, with size indicating the percentage difference of the frequency of an amino acid compared to the analysis reference set. (D) Selected gene ontology biological processes and Swiss-Prot keywords significantly enriched in the clusters. P values were visualized by a red-to-blue bar. The analysis was performed using unique protein identities (non-overlapping between clusters) using DAVID.

For phosphorylation (H+L) abundance hierarchical clustering analysis and enrichment analysis (Figure 5), the total intensities of the matched phosphorylated and non-phosphorylated peptidoforms were calculated by summing the corresponding light and heavy signals (H+L), and the resulting matrices were log2 transformed and normalized by LOESS normalization ^[69]. Next, the intensities of the non-phosphorylated peptidoforms were subtracted from the intensities of the phosphorylated peptidoforms to reflect the potential protein level relative abundance changes during the time course. Hierarchical clustering analysis was performed using the resulting matrix with the “pheatmap” R package after row scaling (distance = “euclidean”, method = “complete”) resulting in an identification of five row clusters. The data were annotated using protein-level annotation terms (GOBP_DIRECT, GOCC_DIRECT) from DAVID database ^[60,61], and the enrichment analysis was performed using Fisher’s exact test in Perseus v1.6.14.0 ^[63].

Figure 5: Phosphoproteomic abundance regulation during starvation tends to be not correlated with turnover diversity.

(A) Log2 ratio distributions of relative phosphorylation change between the late (48 h) and early (1 h) timepoints of the pulsed labeling experiment in each turnover cluster identified by maSigPro. Statistical analysis was performed using pairwise Wilcoxon test. (B) Hierarchical clustering analysis of normalized intensities of the phospho-peptidoforms (intensities of the corresponding non-phosphorylated peptidoforms were subtracted to normalize for the protein abundance changes) across the five timepoints. Five clusters were identified (“H+L”_Clusters 1–5”) and subjected to an enrichment analysis. (C) Selected examples of gene ontology terms (GO DAVID Direct) significantly overrepresented in the five clusters shown in (B). Only those phosphorylated peptidoforms present in the H+L cluster in which a term was significantly overrepresented are shown. The percent stacked barplots on the right of the heatmap indicate the relative proportion of these peptidoforms in the 6 turnover clusters identified by maSigPro. The percentage indicates the proportion of these phospho-peptidoforms from all phospho-peptidoforms belonging to the maSigPro clusters to account for the different size of the maSigPro clusters.

3. Results and Discussion

3.1. A hypothesis-free, time series comparison of turnover behaviors between modified and unmodified peptidoforms.

As illustrated in Figure 1, previous data analysis in DeltaSILAC-type experiments ^[44–47,70] assumed that the total proteoform amounts (i.e., the sum of L and H) remain constant over the time course. In the present study, we first generated a novel non-steady state phosphoproteomic pSILAC dataset using the synchronized rat pheochromocytoma PC12 cells. This cell system was effectively “dynamic”, since after an initial 16-hour synchronization, the serum-free light medium was replaced by SILAC (K8R10) heavy medium that was still serum-free (see Methods), representing a gradual cell starvation process ^[21] (Step 1, Figure 1). In addition to samples of the total proteome ^[21], the corresponding phosphoproteome samples at the respective labeling time points were harvested at 1, 4, 12, 24, and 48 hours and analyzed by a 2.5-hour single-shot DIA-MS analysis ^[44,50,53]. From the proteomic and phosphoproteomic datasets, the phosphorylated (P) and non-phosphorylated (NP) peptidoforms were identified by Spectronaut (both peptide- and protein- FDRs < 1%, PTM localization score > 0.75). A total of 3,755 pairs of P and NP peptides (harboring identical amino acid sequences) were detected, with both H and L versions and quantified in each time point (Step 2, Figure 1). Herein, the signal sum of “H+L” describes the peptidoform abundance change, whereas the ratio of “H/L” depicts the peptidoform turnover dynamics.

Figure 1: Hypothesis-free time course analysis combined with DeltaSILAC enables an unbiased classification of the impact of PTMs on turnover in a non-steady state system.

Step 1: In the present experiment, PC12 cells cultured in a serum-free SILAC light (L) media for 16 hours were transferred into a serum-free SILAC heavy (H) media and harvested after indicated time points of the pulse SILAC labeling. Step 2: The total proteome and phosphoproteome analyses were performed using DIA-MS enabling accurate quantification and matching of the phosphorylated (P) and nonphosphorylated (NP) peptidoforms, based on the DeltaSILAC workflow and bottom up proteomics. Step 3: In an analysis of a “dynamic” system such as cells undergoing starvation stress, the assumptions used in a “classic” steady-state analysis may not hold true. Also, the DeltaSILAC approach would facilitate the PTM-site specific analysis. Thus, instead of curve-fitting effort to infer half lifetime, the hypothesis-free investigation applying a time-course analysis algorithm would enable an unbiased analysis in a “dynamic” system. For those PTM sites exerting a significant effect on peptidoform turnover, further clustering analysis and functional annotation can be performed.

The inference of half-life time for each peptidoform in the non-steady state can be rather problematic due to the considerations mentioned in the Introduction, especially for those peptidoforms carrying PTMs. Since the time-series data analysis algorithms are well-established for comparing expression of the same genes between conditions ^[59,71–73], we reasoned that we can turn the H/L curve-fitting problem into a time-series differential analysis with no prior assumption on PTM site-resolved turnover dynamics (Step 3, Figure 1). Thus, for the present PC12 dataset, after filtering out noisy signals, we performed a comparison against all H/L ratios for the paired P and NP peptidoforms. Using an available time-series tool maSigPro tool and its statistical framework ^[58,59], we identified 1,531 P-sites showing significantly differential profiles of H/L between P and NP peptidoforms (P < 0.05, Benjamini-Hochberg corrected, n = 3 biological replicates), revealing the causal or non-causal impacts of site-specific phosphorylation on protein turnover. In summary, we proposed and applied a time-series analysis to interrogate the de facto abundance and turnover dynamics of site-specific phosphorylation during cell starvation.

3.2. Complex relationship between phosphorylation and protein turnover.

Previously, only two categories, namely “faster” or “slower”, were used to describe how the average H/L exchanging rates of P differ from NP peptidoforms ^[44,46]. Here, to further scrutinize the heterogeneity of temporal H/L patterns between the 1,531 P-sites, we utilized the hierarchical clustering function of maSigPro and identified six distinctive clusters (i.e., Clusters 1–6 respectively representing 349, 194, 337, 94, 395, and 162 P-sites, Figure 2), which uncovered complex and dynamic relationships between phosphorylation and protein turnover measured in starved PC12 cells. Firstly, we found that in most of the six clusters, the H/L deviation between P and NP peptidoforms is not consistent over 48 hours of labeling (Figure 2A–B). For example, in Cluster 2, the H/L ratios of P were slower than NP versions in the first 12 hours, but the H/L exchange rate of P kept increasing and eventually surpassed NP peptidoforms turnover. Thus, the terms of “faster” or “slower” were indeed insufficient for describing H/L profiles for all P-sites. Secondly, in many clusters, the early labeling time point of 1 hour showed a significant extent of H/L deviation between P and NP pairs, as compared to later time points. For example, on average, the 1-hour H/L turnover of P was ~8 times slower than that of NP for Cluster 6 P-sites, but ~2 times higher in Cluster 3. Notably, the larger deviation of 1-hour H/L ratio cannot be simply explained by the technical factors, because the 4-hour H/L data showed a consistent changing direction (despite the dampened extent) in most clusters and an even greater deviation in Cluster 1. Previously, it was demonstrated that the probability of protein degradation may vary depending on its molecular age ^[74]. Thus, our results might suggest that the newly synthesized and phosphorylated proteins can undergo extremely dynamic and divergent clearance at the nascent stage of the protein lifespan. Thirdly, the averaged deviating patterns of P from NP (Figure 2B) indicated that Clusters 3 and 5 harbored mostly “faster” cases (i.e., the P peptidoform displaying a higher H/L exchanging ratio than NP over the entire or most time course, or “phosphodegrons”), whereas most P-sites in Clusters 1, 2, and 6 could have been inferred as “phosphostabilons” in traditional analysis. And the number of P-sites of “phosphostabilons” clusters are more than that of “phosphodegrons” – this overall trend is consistent to the theoretical deduction ^[47] and previous experiment datasets ^[44,46]. Fourthly, we found that, interestingly, the NP peptidoforms showed inconsistent H/L profiles between clusters (e.g., Cluster 1 versus 2) or even between NP peptidoforms of the same protein (Figure S1). This result agrees with our previous observation ^[44] and emphasized that the NP peptidoforms cannot be simply used as a “unmodified” reference at the protein level, because a particular NP peptidoform can be possibly derived from other phosphorylated proteoforms or other PTM modified proteoforms, in which the modifications just happen at the other sites. Only the P and NP peptidoforms matching would provide the accurate P-site turnover investigation ^[46]. Fifthly, it is not surprising that only 94 P-sites showed quite close H/L profiles between P and NP after maSigPro analysis. Indeed, even for Cluster 4 P-sites, there seems to be a small delaying effect on H/L turnover around 12 hours.

Figure 2: Hierarchical clustering analysis of differential profiles between phosphorylated and non-phosphorylated peptidoforms identified by maSigPro.

(A-B) A total of 1,531 P-sites with differential temporal H/L profiles identified by maSigPro were further classified into six distinctive clusters. (A) Average profiles are shown for matched phosphorylated (red) and non-phosphorylated (black) peptidoforms in each cluster. (B) Average difference between the matched phosphorylated and non-phosphorylated peptidoforms in each cluster. The data represent the mean and standard deviation of all log2 H/L ratios (n = 3 biological replicates). (C) The Heavy/Light (H/L) ratios of the selected P-sites (of particular proteins) with their phosphorylated (red) and non-phosphorylated (black) peptidoforms quantified per time point, as examples representing the H/L profiles for each cluster in (A) and (B). The mean and standard deviation of H/L are shown (n = 3 biological replicates).

It should be stressed that, while Figure 2A–2B used equal y-axis tick intervals to emphasize the H/L profiles at early time points, Figure 2C displayed the actual temporal intervals. Taken together, if the term “stabilon” and “degron” can be used to describe the stabilizing or destabilizing effect for a particular P site on peptidoform turnover, the diversity profiles between clusters could be approximately summarized as “constant stabilons” (Cluster 1, see Figure 2B), “late degrons” (Cluster 2), “early degrons” (Cluster 3), “local effectors” (Cluster 4), “constant degrons” (Cluster 5), and “early stabilons” (Cluster 6). We also provided examples of H/L profiles per each cluster in Figure 2C, and all the H/L ratios quantified in Table S2. In summary, a time-series data analysis followed by clustering illustrated previously uncharted temporal variabilities of how phosphorylation can be linked to protein synthesis and degradation during the protein lifespan.

3.3. Biological features associated with the peptidoform turnover of phosphorylation

Subsequently, we explored whether and how Clusters 1–6 have potential biological implications. According to Metascape mapping ^[68], we noted that about half of the phosphoprotein identities (IDs) are shared by two or more clusters (Figure 3A). This result strongly indicates a site-specific, rather than a protein-specific effect of phosphorylation on H/L turnover. We therefore performed P-site-level annotation and enrichment analysis using kinase~ substrate relationships in OmniPath resource ^[65] and the linear Human Protein Reference Database (HPRD) motifs ^[63] (See Methods). Very few kinases were found to be uniquely enriched in each of the clusters (Figure 3B). Interestingly, CDK1 substrates were significantly enriched in Cluster 5 (P = 5.15E-06), the “constant degrons” (Figure 2B). Previously, we also reported that P-sites activated by CDK1 have a drastically shorter lifetime in HeLa cells ^[44]. This observation thus aligned well with previous reports that the phosphorylation coupling degradation is essential in cell division ^[32]. Moreover, the AMP-activated protein kinase (AMPK) motifs tended to be enriched in Cluster 3 (“early degrons”). The phosphorylase kinase (PhK), a kinase that plays the important role of stimulating glycogen breakdown into free glucose-1-phosphate, seemed to be enriched in Cluster 2 (“late degrons”). Whether the enrichment of AMPK or PhK substrates in the respective clusters is linked to cellular energy homeostasis during starvation remains to be studied in the future. Next, we used IceLogo and assessed the frequency of amino acids (A.A.) surrounding the P-sites for each cluster. Using the total 2,597 P-site analyzed by maSigPro as the reference, we did not detect pronounced motif patterns known for specific kinases (Figure 3C). However, the Cluster 3 (“early degrons”) and 6 (“early stabilons”) exhibited a reversed enrichment of a proline at +1 position, which is consistent to their H/L ratio profiles that are also reversed (Figure 2B). Moreover, we noticed a depletion of acidic A.A. (aspartic acid D and glutamic acid E) in Cluster 5 “constant degrons”. This pattern partially aligns with our previous observation that the enrichment of E as a “stabilon” feature for P-sites in growing cells of steady state ^[44].

Because the P-site annotation is rather sparse (e.g., OmniPath only annotated ~16% of all human P-sites with a known kinase-substrate relationship), we complemented the above site-specific analysis with protein-level annotation. To this end, we removed all the overlapping protein IDs shared between clusters. Based on unique IDs per cluster, enrichment test against GO Biological processes (BP) and Swiss-Prot keywords was performed using David Bioinformatics ^[60,61]. As expected, the “Phosphoprotein” item was universally enriched among all clusters (P = 1.727E-05 in Cluster 6 to P = 8.650E-14 in Cluster 1). Nevertheless, specific items were preferably enriched in each one of clusters. For example, “SWI/SNF complex” (P = 4.38E-04) and “cellular response to epidermal growth factor stimulus” (P = 8.25E-03) were enriched in Cluster 1 (“constant stabilons”); “negative regulation of translation” (P = 4.87E-03) and “cell division” (P = 9.96E-03) were enriched in Cluster 2 (“late degrons”); “K48-linked polyubiquitin binding” (P = 2.23E-05), “focal adhesion” (P = 1.09E-04), “proteasome complex” (P = 3.43E-03), and “protein stabilization” (P = 5.15E-03) were enriched in Cluster 3 (“early degrons”); “guanyl-nucleotide exchange factor activity” (P = 0.00103) was enriched in Cluster 4 (“local effectors”); and “metal ion binding” was enriched in Cluster 6 (“early stabilons”). Notably, Cluster 5 (“constant degrons”) enriched more distinctive crucial functions, such as “GTPase activity” (P = 6.25E-04), “Bromodomain” (P = 2.97E-03), “circadian rhythm” (P = 3.43E-03), “SH2 domain binding” (P = 4.68E-03), “Wnt signaling pathway” (P = 6.87E-03), “MAPK cascade” (P = 0.0185), and “protein autophosphorylation” (P = 0.0187), suggesting the functional importance of P-sites in Cluster 5. (See selected items in Figure 3D). Altogether, the P-site and protein level enrichment analysis suggest that the diversity of phosphopeptidoform turnover can be potentially linked to biological functions.

3.4. Phospho- and nonPhospho- peptidoforms of varied abundances tend to display different turnover dynamics

We next asked whether the phosphorylation stoichiometry of P-sites differ among Clusters 1–6. We herein summed and normalized the H and L DIA-MS intensities for both P and NP peptidoforms at each time point across 1–48 hours (Figure S2). This H+L sum essentially quantifies the total peptidoform abundance. We found that the global DIA-MS intensity of P seems to be slightly lower than NP peptidoforms; however, surprisingly, there is a drastic heterogeneity of H+L abundance between all clusters. Firstly, unlike other clusters, in Cluster 2, the P peptidoforms even showed higher H+L signals than the NP versions (Figure 4A), indicating these “late degrons” had the relatively highest phosphorylation stoichiometry among the six clusters. In contrast, the NP peptidoforms in Cluster 3 (“early degrons”) demonstrated the highest NP expression among all clusters and the least phosphorylation stoichiometry. Indeed, Cluster 2 and 3 differed significantly based on the P/NP ratio, a proxy for quantifying relative phosphorylation stoichiometry (P < 1E-15, Wilcoxon test, Figure 4B). The “constant degrons” in Cluster 5, however, did not manifest any extreme phosphorylation stoichiometry compared to other clusters. Secondly, Cluster 5 and 6 showed similar H+L distributions for both P and NP versions and comparable phosphorylation stoichiometry; but their H/L ratios exhibited very different temporal profiles (i.e., constant degrons” vs. “early stabilons”, Figure 2B–C). Thirdly, the average P and NP abundances can differ up to 3.073 (Cluster 2 vs. Cluster 3) and 3.863 folds (Cluster 4 vs. 6) and did not rank similarly among clusters (Figure 4A, right panel). Also, the phosphorylation stoichiometry proxy frequently yielded statistically significant differences between clusters (Figure 4B). Altogether, these results exemplified a striking, inter-cluster difference of phosphorylation extent and peptidoform abundance that is tightly associated with (but cannot be ascribed to) the H/L temporal clusters. It is interesting to note that, for example, Cluster 3 has the most abundant NP forms, so the early H/L turnover rates of P might have to be fast (a unique feature of Cluster 3 in Figure 2B) to exert their impact. However, more datasets and insights are needed in the future to explain these results.

Figure 4: Phosphorylated and non-phosphorylated peptidoforms of varied abundances tend to be classified into different turnover clusters (i.e., maSigPro Clusters 1–6).

(A) The abundance of the total signal (i.e., log2 L+H intensities) of the phosphorylated and nonphosphorylated peptidoforms in each turnover cluster is shown. The right panel shows the abundance of the peptidoforms in each cluster ranked by the abundance of the phosphorylated (top) or non-phosphorylated (bottom) peptidoforms. (B) Log2 ratio of the matched phosphorylated and non-phosphorylated versions was used as a proxy for phosphorylation stoichiometry distribution per cluster. Statistical analysis was perform using the Wilcoxon test.

3.5. Phosphoproteomic abundance regulation during starvation tends to be not correlated with turnover diversity

As a final step, we analyzed how phosphoproteomic abundance regulation (i.e., based on H+L sum) is coupled with peptidoform turnover regulation (i.e., based H/L ratio) during cell starvation. To study the abundance regulation per P-site, we determined the H+L abundance change of P peptidoforms between 48 hours and 1 hour and normalized them by subtracting the corresponding NP peptidoforms change between 48 hours and 1 hour. This normalized “48h/1h” ratio essentially would deliver the identical result to a routine label-free quantification studying phosphoproteomic changes after starvation, since the heavy SILAC labeling should not affect the cell status. We analyzed this “48h/1h” ratio to the Clusters 1–6 from maSigPro above. Herein, Cluster 4 (“local effectors”) showed a slight but significant upregulation of P-site abundance compared to other clusters (Figure 5A). Moreover, Cluster 5 (“constant degrons”) exhibited a down-regulation of “48h/1h” ratio compared to Cluster 1 (“constant stabilons”; P = 0.0135), as one would expect. However, “48h/1h” ratio values of all the clusters do not have a substantial deviation from zero.

To further inspect the functional regulation during cell starvation, we performed a hierarchical clustering analysis of normalized P/NP abundances (using H+L) across the five timepoints (Figure 5B). The general protein-level annotations were performed using David database ^[60,61]. Five clusters emerged from this analysis (hereafter, “H+L”_Clusters 1–5), showing biological regulations during cell starvation that are largely expected (Table S3). For example, “Signal transduction” (P = 8.43E-03) and “Cell division” (P = 0.0478) were enriched in “H+L”_Cluster 2, with most P-sites down-regulated during the cell starvation. “Response to nutrient” (P = 4.05E-03) and “cell-cell adhesion” (P = 4.39E-03) were enriched in “H+L”_Cluster 5 containing P-sites down-regulated at 48 hours post a long-term serum deprivation. In comparison, “Lysosomal membrane” (P = 2.59E-03) was enriched in “H+L”_Cluster 3, in which most P-sites were upregulated along with starvation. Still, no obvious enrichment of P-sites in any maSigPro cluster (i.e., Cluster 1–6 in Figure 2B) was notable in H+L_clusters (Figure 5B). Examples of GO BP or cellular component (CC) items of selected significant “H+L”_Clusters also did not display obvious association between the “H+L”_clusters and maSigPro clusters in these functional items (Figure 5C and Figure S3). Previously, in the same PC12 system, the degradation profiling at the protein level was also demonstrated to provide valuable biological insights that were not evident by analyzing protein abundance alone ^[21]. Taken together, these results thus indicate that the phosphorylation “degrons” or “stabilons” did not strongly lead to a directional, end-point phosphorylation abundance alteration post a dynamic process of starvation stress.

4. Conclusion

In conclusion, we herein provided a novel and high-resolution phosphoproteomic pSILAC dataset under the gradual starvation in synchronized mammalian cells. Importantly, we propose a hypothesis-free data analysis framework which embraces the well-established statistical methods developed for time-series experiments. The site-resolved PTM turnover profiling such as DeltaSILAC has been newly enabled by the latest high-resolution mass spectrometry and bioinformatic advances. Our analysis workflow now encompasses dynamic measurements, complementing previous approaches that were previously only performed in the steady states ^[44–47,70]. The traditional turnover modeling approaches that were used could have been insufficient to understand this new type of peptide-protein turnover data and the biology behind it ^[47], especially in dynamic processes. Our study thus presents a flexible, descriptive analysis framework for future DeltaSILAC-type experiments.

Among the limitations of our work, we would like to underline some aspects that will require additional future developments both from a technological and a biological point of view. Although it would be interesting to speculate about the biological relevance of the modules that we identified, we would like to stress that our approach has discovery purposes and raises several questions without yet providing causal relationships. Some of these questions are at the center of the field that explores the link between PTMs and turnover at the proteome-scale. For example, we still do not understand why and how the relative and absolute abundances of P and NP affect the turnover behaviors of phosphorylation. If the major purpose of regulating the turnover of peptidoforms is not for the desirable end-point abundance regulation, what are the mechanisms that determine the diverse phosphorylation turnover we observed? One could hypothesize that in some cases the difference in protein stability observed in the presence of a defined phosphorylation state has a causative effect. And as such whenever possible it will be necessary to further test these possibilities with orthogonal methods such as site-specific mutagenesis, WB, and imaging. Moreover, eventually, the context of the modifications needs to be considered for each proteoform. In other words, the surrounding “PTM code” might influence the biological “meaning” of each single PTM site and their possible role in e.g., protein stabilization, turnover, and protein-protein interaction. In practice, to address this question one would need to integrate the information from the surrounding peptides and modification sites, which is a rather difficult task for bottom-up proteomics as the context is usually lost due to the proteolysis step. As such one important future development for this field is to combine the peptidoform level measurements with approaches that provide information at the level of the whole protein polypeptide by using e.g., top-down proteomics ^[75]. Another limitation concerns the issue of linking peptide modifications and turnover with protein structure and function ^[76]. In addition to a larger phosphoproteomic coverage required, such a major gap that may need to be addressed by large collaborative efforts combining functional, phenotypical data with protein structural information and biological knowledge, as well as mathematical modeling, cross-validation experiments, and even other types of data such as such as metabolomics. Importantly, all these questions were not easily addressed for pSILAC experiments previously, as the site-resolved PTM analysis across the labeling samples with sufficient reproducibility has been available only very recently due to the resolution and selectivity offered by the latest mass spectrometry methods ^{[8,15,17,21,44–46]}.

While technical challenges remain, in our analysis, both H+L and H/L profiles acquired could be determined by DIA-MS. This might be indispensable for certain cell systems. For example, in drug-perturbation experiments, it is widely acknowledged that the early abundance regulation of phosphorylation is regulated by kinases and phosphatases, without changing the synthesis and degradation rates of the proteins due to the short term. Therefore, in the early time points (e.g., 1 or 4 hours), the quantified L version of phosphorylation peptidoform could be affected by both phosphorylation reaction and phosphoprotein turnover. In this regard, our hypothesis-free approach does not require prior knowledge of a system being “steady” or “dynamic”, which enables a turnover investigation integrating the information of temporal expression regulations for both proteins and P-sites.

Significance statement.

The turnover measurement of proteins and proteoforms has been crucial for biochemistry, pharmacology, and systems biology. Previously, detailed studies on individual proteins and phosphosites have established “phosphodegrons” and “phosphostabilons” with important functional consequences on protein degradation and turnover. However, so far, it remains underexplored whether and how pSILAC labeling coupled with the state-of-the-art mass spectrometry such as DIA-MS could contribute to the large-scale discovery of the interdependency between protein PTMs and protein turnover. This is an essential question in the field, as these proteome-scale measurements could clarify the general principles that regulate protein degradation dynamics, and how PTMs might dynamically play a role in this context. More in general, such PTM peptide-protein turnover experiments could be applied to study dynamic biological processes without the limitation of the steady state assumption that was used in many previous protein turnover studies. Essentially, the accurate profiling of the heavy and light peptidoforms for both phosphorylated and non-phosphorylated proteins allow to gather a new type of datasets that were virtually unexplorable with older proteomic techniques. At the same time, the biological understanding of these more recent datasets poses a real challenge. We herein propose a flexible hypothesis-free analysis framework combining “peptidoform-matching” and “time-course analysis” strategies that is applicable for both steady and non-steady state systems. Our results emphasized a dynamically variable effect of phosphorylation on peptidoform turnover and urges future studies on PTM turnover.

Abbreviations

A.A

amino acid

BP

biological process

CC

cellular component

DDA

data-dependent acquisition

DIA-MS

data-independent acquisition mass spectrometry

FDR

false discovery rate

H

Heavy channel (SILAC

IDs

identities

LC

liquid chromatography

L

Light channel (SILAC)

MS

mass spectrometry

NP

non-phosphorylated

P

phosphorylated

P-site

phosphorylation site

PPI

protein-protein interaction

pSILAC

dynamic Stable isotope labeling by amino acids in cell culture

PTM

post-translational modification

SILAC

Stable isotope labeling by amino acids in cell culture

WB

western blotting

Data availability statement

The mass spectrometry-based phospho-DIA datasets have been deposited to PRIDE ^[77] with the dataset identifier PXD037360.