Isoform‐resolved correlation analysis between mRNA abundance regulation and protein level degradation

Abstract Profiling of biological relationships between different molecular layers dissects regulatory mechanisms that ultimately determine cellular function. To thoroughly assess the role of protein post‐translational turnover, we devised a strategy combining pulse stable isotope‐labeled amino acids in cells (pSILAC), data‐independent acquisition mass spectrometry (DIA‐MS), and a novel data analysis framework that resolves protein degradation rate on the level of mRNA alternative splicing isoforms and isoform groups. We demonstrated our approach by the genome‐wide correlation analysis between mRNA amounts and protein degradation across different strains of HeLa cells that harbor a high grade of gene dosage variation. The dataset revealed that specific biological processes, cellular organelles, spatial compartments of organelles, and individual protein isoforms of the same genes could have distinctive degradation rate. The protein degradation diversity thus dissects the corresponding buffering or concerting protein turnover control across cancer cell lines. The data further indicate that specific mRNA splicing events such as intron retention significantly impact the protein abundance levels. Our findings support the tight association between transcriptome variability and proteostasis and provide a methodological foundation for studying functional protein degradation.


Supplementary Note 1: Step-by-step protocol for pSILAC data analysis in Spectronaut
A. Library Generation 1) Select files for library generation. Both DDA and DIA data are supported. For optimal results, consider adding SILAC labeled mass spectrometric measurements.
2) Select a FASTA database of protein sequences.
3) Select search settings. For label-free data, BGN Factory settings can be used. For pSILAC data, labels must be specified in the "Labeling" settings. 4) Optionally, previous searches can be added from Search Archives to further increase the library size, with the identification FDR globally controlled at the end. 5) Choose the settings for library generation. To ensure the complete labeling of the final assay library, the "In-Silico Generate Missing Channels" option must be enabled in the Workflow settings. The workflow needs to be specified as "label". 6) Check the analysis summary and run the library generation.

B. Targeted data search and extraction
1) Select DIA files to analyze and the previously created library.
2) Select the search and extraction settings for DIA analysis. The "Inverted Spike-In workflow" (ISW) is recommended for the pSILAC data analysis. To do so, in the Workflow settings, the "Spike-In" workflow is selected in Multi-Channel Workflow Definition. Both "Inverted" and "Reference based Identification" options are enabled. In the quantification settings, the "Interference correction" option is enabled with minimal 3 fragment ions to keep. The "Exclude All Multi-Channel Interferences" option has to be enabled to exclude ions that might interfere and cause ratio distortion in the labeled workflows. For example, this option enables the usage of b-ions falling in different isolation windows for pSILAC data quantification (as described in the Supplementary Note 2).
3) Select the same FASTA database of protein sequences, check, and run the analysis.
4) The "heavy" (target) and "light" (reference) fragment ion signals can be conveniently visualized in the Analysis perspective using the "XIC graphic" plot, which has been also optimized for labeled workflows. In this plot, the upper part corresponds to the "light" signal (reference) while the lower part shows the signal of the "heavy" (target) channel in retention time and within the peak boundaries identically defined for both channels. Interfering ions discarded by the quantification filtering are marked with dashed lines. The XIC graph can show either MS1 or MS2 quantification based on user preference (in the right-click menu). An example of a peptide quantified in three time points is given below. The intensity of the target channel (depicted below the zero line) is increasing over time while the "light" ion intensity is decreasing. 5) For the downstream analysis, the intensities of the "heavy" (target) and the "light" (reference) peptide precursors as well as the H/L ratio ("target/reference ratio") after the quantification data filtering can be conveniently reported by the pivot report in the Report perspective.
EG.PrecursorIdunique peptide precursor id EG.ReferenceQuantity(Settings) -"light" signal intensity EG.TargetQuantity(Settings) -"heavy" signal intensity EG.TargetReferenceRatio(Settings)the heavy/light ratio These values can be further used as the input values for turnover estimation and following bioinformatics analyses.

Supplementary Note 2: Optimized pulsed SILAC DIA-MS data analysis A. Library selection for pulsed SILAC data analysis
This Note further supports the Supplementary Note 1. Spectronaut currently can combine label-free data and isotopically-labeled data, either acquired by using DDA-MS or DIA-MS, to generate a hybrid sample-specific and heavy-labeled library in one step with proper global FDR control. Thus, the coverage of the library can be increased by using comprehensive sample-specific assays. First, data are searched using Pulsar and "directDIA" with appropriate settings for the label-free and labeled data.
Then, a complete labeling of the hybrid library is achieved through the "In-Silico Generate Missing Channels" option, which detects the labeling setup of an experiment and applies this labeling to the whole assay library to ensure each one of the peptides is represented by both "light" and "heavy" transitions. Another option is to generate a library using label-free DIA and DDA data only and label the library in-silico using the "Generate Labeled Library" function in the library perspective of Spectronaut.
To test whether the hybrid library performs better than the label-free-only library in our pSILAC data, we performed a systematic evaluation of four different libraries (Table 1). Libraries 1 and 2 (L1 and L2) were created using label-free DIA and DDA data only. On the other hand, for the libraries 3 and 4 (L3 and L4) we also included pSILAC data raw files to create a "hybrid" library. Furthermore, in L2 and L4 only y-ions were used while all the b-ions were removed. Thus, the libraries differed in two main parameterswhether the labeled data were included and whether the b-ions were kept or removed. For the purpose of the library and workflow testing, fifteen raw files were included.
HeLa cell lines 1, 2, 3, 4, and 5 were used in all three time points of a pSILAC experiment (1, 4.5, and 11 hours, or T1, T4, and T11).   Secondly, two libraries were used for targeted pSILAC data analysis, L2 and L4, to evaluate which libraryin-silico labeled or hybrid -provides the best identification and quantification for the downstream analysis. For data extraction, the Inverted Spike-In workflow (ISW) was used. The number of valid target/reference ratios (or H/L ratios) reported was higher in all 15 raw files subjected to the analysis independently on the cell line or time point (Figure 2A). Consequently, the number of calculated peptide and protein kloss values was larger: the number of peptide kloss values increased by 6.2% whereas the number of protein kloss increased by 3.9% (both reported values are average from the five cell lines) (Figure 2B and 2C). Thus, the sample-specific hybrid library provided better quantification results and was used for the data analysis presented in the main text.

B. Selection of the workflow for pulsed SILAC data analysis
We also tested the usage of the Inverted Spike-In workflow (ISW) and the "Labeled" workflow (LBL) for pSILAC data analysis. Notably, the peak picking and identification in the LBL is performed by using both "light" and "heavy" precursors while ISW (with "Reference based Identification" option selected) does both the peak-picking and identification based on the "light" precursors. Moreover, all labeled workflows assemble the heavy-versus-light elution by using the same peak boundaries before DIA identification, so that no post feature alignment (Rost et al, 2016) is required for labeling DIA analysis.
Due to the peak picking algorithm, ISW can improve the reliability of detection of newly synthesized "heavy" protein in the early time points during labeling, when the "heavy" signals are dramatically lower than the "light" ones from those pre-existing proteins. To test whether this feature can improve pSILAC data analysis, we further evaluated the outcome of LBL and ISW workflows in the five HeLa cell lines (Figure 3). The hybrid Library 4 was used for this comparison. As expected, the ISW significantly outperformed the LBL workflow in the two earlier time points ( Figure 3A).
Overall, ISW also provided kloss values for 6.2% more peptides and 6.5% more proteins (Figure 3B and 3D). We thus used ISW for pSILAC data analysis.

C. Improving the identification and quantification by including the b-ions for the pSILAC data analysis
In our previous work, we strictly excluded the b-ions during pSILAC or SILAC data analysis to avoid interferences and corresponding ratio distortion caused by b-ions not carrying the "heavy" labels in case the "light" and "heavy" peptide precursors are fragmented in the same isolation window Liu et al, 2019). However, in case the b-type fragment ion either contains a heavy-label (e.g., due to a missed cleavage) or the "light" and "heavy" precursors fall in a different isolation window, these ions theoretically could be still used to improve identification and quantification.
To enable the usage of all fragment ion types in the isotopically-labeled data analysis, a new "Exclude All Multi-Channel Interferences" option is now available in Spectronaut v13. This function automatically recognizes ions that show channel interference in the labeled workflows, and these are completely removed from quantification. The function was evaluated in a smaller test (only HeLa 1 data were included). By enabling the b-ions, the number of peptide and protein ids increased by 14% and 11.2%, while the quality of the data (i.e., H/L ratio distribution in different time points) seems to be not affected when compared to the results of the analysis completely excluding b-ions (Figure 4).

Figure 4: The usage of b-ions together with the "Exclude All Multi-Channel Interferences" option improves pSILAC data identification and quantification. (A) The Venn-diagram shows the number of peptide ids shared between and exclusive for two data sets generated either with b-ions completely removed from the library and analysis ("b-ions excluded") or with all ions included in the library and analysis ("all ions"). Log2 H/L ratio distribution of peptides across three time points in the "b-ions excluded" (B)
and "all ions" (C) data sets.

D. Summary
• The sample-specific hybrid library provided better identification and quantification than the in-silico labeled library generated from label-free data exclusively.
• The "Inverted Spike-In" workflow outperformed the "Labeled" workflow, especially in the early time points of the pulse-chase SILAC labeling.
• A new option is now available in Spectronautv13, called "Exclude All Multi-Channel Interferences", so that the users can benefit from using the b-ions in their analysis.

Supplementary Note 3: Comparison of pSILAC-DIA to pSILAC-MS1 and pSILAC-TMT approaches for protein turnover analysis.
Below we compared our pSILAC-DIA data to two alternative methods for the proteome-wide turnover measurement by pSILAC experimentnamely pSILAC-MS1 and pSILAC-TMT. Both methods are often used. For example, Savitski and colleagues are the major developers of the methods and software for the multiplexed TMT-based workflow analyzing protein thermal stability (Franken et al, 2015) and dynamics. In 2018, they published a systematic analysis on protein turnover in primary cells using an optimized pSILAC-MS1 workflow (Mathieson et al, 2018) and a study uncovering proteolysis targeting chimeras (PROTACs) effects using a pSILAC-TMT workflow .

pSILAC-DIA vs. pSILAC-MS1
In pSILAC-MS1 approach, the same pSILAC experiment is performed, whereas the SILAC heavy/light ratios are inferred from MS1 isotopic pairs following a shotgun proteomic workflow. In our pSILAC-DIA, correspondingly to above, a) With the new Appendix Figure S2, we have shown that MS2 based DIA quantification has much less and more uniformly distributed charge states (83.08% are charge 1, and the other 15.92% are charge 2, Figure S2A) than the MS1 results ( Figure S2B), which greatly reduced the isotope-fitting difficulty for calculating theoretical isotopic envelope. b) Furthermore, because all high-resolution DIA-MS2 peak groups (with 30k in Orbitrap) are aligned and identified at MS2 level, the "heavy" and "light" signals are well matched. Importantly the DIA isolation window schema and the MS2 level acquisition greatly reduced the noise background from a full mass-range MS1 level acquisition. These factors together ensure that the pSILAC-DIA data has a significantly better quantitative accuracy and reproducibility than the direct MS1-based quantification (even at 120k) (as compellingly shown in Appendix Figures S3 and S4), especially for early pulse-chase time points which are known to heavily impact the protein turnover calculation (Claydon & Beynon, 2012).
In summary, the pSILAC-DIA quantification successfully improved the quantitative accuracy and reduced the MS data procession difficulty in pSILAC experiment, as compared to pSILAC-MS1 approach.

pSILAC-DIA vs. pSILAC-MS1 and pSILAC-TMT
One interesting advantage of pSILAC-DIA over both pSILAC-MS1 and pSILAC-TMT is the availability of multiple quantitative data points. This is because, for a given peptide (

pSILAC-DIA vs pSILAC-TMT
The hyper-multiplexing approach combining "pSILAC" and "MS2-tag based quantification" is not actually new, and has been already proposed in 2010 in which pSILAC labeling was combined with iTRAQ quantification (i.e., pSILAC-iTRAQ) (Jayapal et al, 2010;Hinkson & Elias, 2011 However, in our hands pSILAC-DIA achieved comparable proteome coverage to pSILAC-TMT with a similar total MS measurement time. This could be due to the facts that e.g., there is no fractionation needed in pSILAC-DIA and that extra time for MS2 scans (that are not used for quantitation) was spent during SPS-MS3 TMT analysis. The comparable coverage between DIA and TMT approaches is consistent to a previous report comparing the performance of the two (Muntel et al, 2019). Besides the reduced ion complexity and more quantitative data points mentioned above (Appendix Figures S2, S5, and S6), in particular, we also summarize other advantages of pSILAC-DIA over pSILAC-TMT below. Therefore, we suggest all these considerations together make pSILAC-DIA a competitive, if not a better, method than pSILAC-TMT.
1) Even the current most optimized version of SPS-MS3 based pSILAC-TMT seems to (only) provide a similar quantitation performance of turnover rates to pSILAC-MS1, but not better. This may be not surprising as SPS-MS3 only largely reduces the co-isolation problem but does not totally eliminate it (Muntel et al, 2019), which stems from the TMT design. Thus, in the previous publications, the pSILAC-TMT (MS2) generated less-precise turnover rates than pSILAC-MS1 (Welle et al, 2016), whereas the pSILAC-TMT (SPS-MS3) with a further bioinformatic correction dealing with the normalization bias (Zecha et al, 2018), makes the approach become close or comparable to pSILAC-MS1 results in precisely determining the H/L ratios and thus, kloss.
In contrast, again, with Appendix Figures S3 and S4, we show compelling results that pSILAC-DIA can generate significantly better quantification accuracy and precision than pSILAC-MS1 (see above).
2) The problem with large-scale, multi-batch TMT data was recently reported by experienced labs (Brenes et al, 2019), suggesting the pSILAC-TMT approach might not be ready to be used by every lab, and importantly, to be used to compare protein turnover across multiple cell lines (like our application shown in this study) and conditions. Indeed, false positives, batch effects, missing values, and other issues were reported in a multi-batch TMT experiment despite the use of SPS-MS3 (Brenes et al, 2019).
3) Experimental design flexibility. In a pSILAC experiment, a 10-time point sampling during pulse-chase process does not have to be always necessary (10 TMT channels were used in (Zecha et al, 2018)). More importantly, the pSILAC-DIA offers a greater flexibility in trouble-shooting a proteomic experiment (e.g., if one of the samples has some technical or experimental problem, in pSILAC-DIA workflow the problematic sample can be easily re-processed and measured with new data combined, whereas in pSILAC-TMT the whole sample set has to be re-measured).
4) Machine flexibility. Currently, the SPS-MS3 seems to be essential for pSILAC-TMT to ensure the quality of turnover estimation to be comparable to pSILAC-MS1 (Zecha et al, 2018). This function, however, is only provided in Fusion-type Orbitrap analyzers from Thermo Scientific. pSILAC-DIA has a much broader application potential in this regard. 5) pSILAC-DIA might request lower cost for proteomics labs, considering the cost of TMT reagents for lots of samples.
6) Potential of pSILAC-DIA in analyzing peptide PTM isoforms (peptidoforms). DIA-MS includes extra information of the perfect co-elution behavior between all fragment ions of the same peptide isoform along the liquid chromatography. This feature will offer much higher ability in discriminating SILAC H/L ratios for discriminating PTM isoforms (Rosenberger et al, 2017). Instead, in pSILAC-TMT, those different PTM isoforms of the same peptide backbone eluting together (or, arising from other pSILAC channels of a different time point) can be easily co-isolated for MS2 and SPS-MS3 analysis, and thus, interfere with each other. 7) Potential of analyzing "heavy" and "light" signals separately for non-steady state measurement.
Although for the protein turnover measurement a general assumption is applied that cells are growing in the steady state (i.e., the sum of "heavy" and "light" signals is stable during labeling), DIA data provides the possibility to directly and robustly analyze "heavy" and "light" version of proteins separately. This ability will be useful for studying dynamic processes e.g., systems under a perturbation.
To achieve such a task, pSILAC-TMT analysis would have to rely on a complex statistical approach .

Supplementary Note 4: Comparison between the relative isotope abundance (RIA)-based and normalized light intensities (NLI)-based kloss calculations
The RIA algorithm used in this study determines the protein turnover rate in the steady state considering both "light" (L) and "heavy" (H) intensities (see Methods). However, the high accuracy of our DIA-MS measurement actually enables measuring L and H peptide intensities separately in the pSILAC experiment. A simpler approach to determine a de facto protein degradation rate would be to directly calculate the rate of loss from the L (i.e., unlabeled peptide) intensities. Therefore, in this Supplementary Note, we performed such a calculation termed normalized light intensity (NLI)-based method (see Methods) and compared the results to results determined by the RIA algorithm.
Regarding the NLI calculation process, firstly, as in all label-free quantification experiments, we normalized the total identified DIA signals based on the sum of the heavy and light channels across time. Then, we extracted the light channel quantities (which showed a perfect pattern of gradual loss of the light intensity over time) by fitting the desired curve on each peptide. Finally, we aggregated peptide precursor kloss values to the protein AS isoform groups level. For the purpose of comparison, the calculation was performed for the same data set (twelve HeLa cell lines), using the same Spectronaut results as an input for the algorithm, and the same parameters for peptide kloss averaging as in the case of the RIA-based data presented in our study.

Figure 5: Correlation between RIA-and NLI-based log2 kloss and corresponding mRNA-kloss values. Correlation between log2 kloss absolute values (average of all HeLa cell lines; A) and relative values (six HeLa Kyoto cell lines/ six HeLa CCL2 cell lines log2 fold change; B). Correlation between the across-cell lines protein AS isoform-specific mRNA-kloss correlation (C) and corresponding mRNA-kloss distributions (D). Rho indicates Spearman's correlation.
We first assessed whether the RIA and NLI methods provided comparable results. The log2 scale kloss correlation analysis based on averaged value of all HeLa cells demonstrated a high correlation (rho = 0.95; Figure 5A). The relative correlation analysis between HeLa Kyoto and CCL2 samples also revealed that consistent results were largely achieved between RIA and NLI (rho = 0.77; Figure 5B).
This highlights the possibility of directly using light signals in a two-channel pSILAC-DIA experiment for kloss calculation. Next, we calculated the proteome-wide, protein AS isoform-specific mRNA-kloss correlation between all twelve HeLa cells using NLI-derived kloss values. We found that despite this process involved independent calculation in each of twelve HeLa cells, the correlation coefficients still significantly correlated (Figure 5C), which is as strong as the relative ratio-based correlation. Finally, we assessed whether majority of the NLI-derived mRNA-kloss rho remained positive (i.e., above 0; Figure 5D). Indeed, although NLI generated a lower median (0.08, compared to 0.13 in RIA), 57% of the values were still above 0. This result thus supports the notion that the cells are likely to use protein degradation to fine-tune mRNA variation.

Figure 6: Comparison of RIA-and NLI-based kloss precision across the HeLa cell lines. Coefficient of variation (CV) was calculated for each protein AS isoform group within six HeLa Kyoto or six HeLa CCL2 cell lines. A threshold of CV < 0.2 (red line) was highlighted.
We next asked whether NLI derives precise measurement of kloss similar to RIA results. We performed a comparison of CVs of protein-level kloss values generated by both methods, across replicates of CCL2 and Kyoto cells. Interestingly, we found RIA still generated more precise quantification with a statistical significance (Wilcoxon test, p < 1 x 10 -16 for both CCL2 and Kyoto HeLa cells; Figure 6). This seems to suggest the calculation of kloss using ratio of H and L signals (i.e. RIA) is still more stable than single L channel-based determination (i.e., NLI), possibly because the decrease of light signals is small in early time points (e.g., at 1 hour or 4.5 hours), which essentially presents a challenging case of relative quantification even for DIA-MS.
Following, we repeated the analysis presented in Figure 3D-G (main text in the article), but limited to kloss values with a CV < 0.2 for both RIA and NLI, to achieve a fair comparison. The results were consistent to Figure 3D-G in the main text. As shown in Figure 7, the absolute mRNA-kloss correlation was determined to be ρ = -0.22 by RIA and -0.26 by NLI (Figure 7A&B; n = 2,513; p < 0.0001) in all protein AS isoform groups (including both UQ and SM), and the absolute protein-kloss correlation was even more negative (ρ = -0.40 by RIA and -0.42 by NLI; n = 2,513; p < 0.0001). Thus, the degradation seems to be slower for genes with higher expression on the absolute scale. However, when the Kyoto/CCL2 fold changes are analyzed, slight but significant positive across-gene mRNAkloss correlations were obtained, which is identical to Figure 3 in the main text ( Figure 7C&D; ρ = 0.13 for RIA, ρ = 0.09 for NLI; n = 2,513; p < 0.0001). Altogether, NLI and RIA were in a good agreement about absolute and relative correlations between mRNA and kloss.

Figure 7: Absolute and relative correlation analysis using RIA-and NLI-based kloss values. (A-B) Across-gene Spearman's absolute correlation between indicated values (average from all cell lines). (C-D) Spearman's correlation between indicated values using relative quantification data (Kyoto/CCL2 fold change).
Finally, we assessed the overall protein buffering significance in the NLI-and RIA-based data (Kyoto and CCL2 CV < 0.2 in both) to confirm whether the conclusions remain the same when kloss is directly derived from the loss of the light, unlabeled, peptide. According to Figure 3H in the main text of the manuscript, we assessed the mRNA-protein and the mRNA-kloss correlations for protein AS isoform groups (including both UQ and SM) that were either differentially expressed (t-test, Benjamini-Hochberg FDR < 0.01) between Kyoto and CCL2, or not ( Figure 8A). Consistently with Figure 3H in the main text, we observed for both kloss calculation methods proteins significantly regulated between the cell lines show a significantly lower mRNA-kloss correlation than proteins which were not differentially expressed.
In Figure 8B, we provide mRNA-kloss correlation for both methods distributed to protein complex participation (i.e., according to CORUM annotation, proteins known to be the subunits of a stable protein complex are marked as "Complex IN", and proteins without a protein complex annotation are marked as "Complex OUT"). Because protein complex stoichiometry is a general mechanism that is well known for protein buffering, this analysis confirms mRNA-kloss correlation is indeed a good measure. The better mRNA-kloss correlation tend to result in more constraints at the protein level and thus worse mRNA-protein correlations (Figure 8B). Compellingly and conceivably, proteome wide mRNA-kloss is a good indicator of protein abundance buffering mechanism independently on the method of kloss calculation.

Comparison between MS1 and MS2-based SILAC quantification
For the analysis presented in Figures S2-S6, the same data analysis workflow was used as for the data described in the Methods section in the main text. Briefly, the same library (i.e., a hybrid library containing 159,963 peptide precursors corresponding to 11,623 proteins, including the Top 6 fragment ions per peptide and excluding the b-type ions), the same FASTA protein sequence database, and workflow (i.e., the Inverted Spike-In workflow) were used for targeted data extraction of three raw files from HeLa 7 (time points 1, 4.5, and 11 hours). The interference correction was applied to both MS1and MS2-based quantification with minimum of two values for MS1 quantification (the mono-isotopic precursor and isotopic envelope) and a minimum of three fragment ions for MS2 quantification.
For fragment ion-based ratio quantification, the fragment ion-level report was generated in Spectronaut, and the SILAC H/L ratios were further calculated, processed, and visualized in R. The ratios were calculated using the matched "light" and "heavy" fragment ion intensities (F.PeakArea). The F.ExcludedFromQuantification column was used to exclude potentially interfering ions that were filtered using the interference filtering. F.Charge column was used to report the fragment ion charge state.
For peptide-precursor MS-1 and MS2-based quantification, the fragment group-level report was generated in Spectronaut, and the ratios were further calculated, processed, and visualized in R. The ratios were calculated using the matched "light" and "heavy" (FG.IsotopeLabelType) fragment ion group intensities for MS1-(FG.MS1Quantity) and MS2-based quantification (FG.MS2Quantity). Thus, we used all peptide precursors quantified by Spectronaut with no additional quantification filtering applied. FG.Charge column was used to report the fragment ion charge state.
Thus, in the analyses described below, we compared four data sets below, each one of them containing HeLa 7 H/L ratio quantification in three labeling time points.
1) Top 6this data set contained all H/L ratios computed using individual fragment ions, with no interference filtering applied.
2) Top 6 + interference filteringthis data set contained all H/L ratios computed using individual fragment ions; both "light" and "heavy" fragment ions were required to pass the interference filtering.
3) MS2this data set comprised peptide precursor H/L ratios quantified using MS2-based quantification in Spectronaut (i.e., the sum fragment ions after interference filtering).
4) MS1this data set comprised peptide precursor H/L ratios quantified using MS1-based quantification in Spectronaut (i.e., the sum of min two MS-1 signals after interference filtering).
These data sets were used for the following analyses: 1) All H/L ratios of fragment ions/peptide precursors distribution visualization across different labeling time points.
2) Protein H/L ratio summary: for each protein AS group a median protein ratio and a standard deviation were calculated using all values (fragment ions/peptide precursors H/L ratios) assigned to a protein AS group. Furthermore, a correlation (Spearman's rank correlation) between all SILAC "light" and "heavy" intensities for every time point was calculated for each protein group.
3) For every protein AS group, we created a visualization showing correlation between "light" and "heavy" channels in three different time points, and a boxplot summarizing corresponding heavy/light ratios. Selected examples from these analyses are presented in the figures.

Comparison of within-and between-protein alternative splicing (AS) groups peptide profile kloss variability
In Figure 4, certain protein AS isoforms of the same gene may have differential degradation rates or get differentially degraded between Kyoto and CCL2 cell lines. We plotted the across-cell line variability of peptide profiles within genes, within protein AS isoform groups and between protein AS groups of the same gene in additional analyses (shown in Figure S7). 1) Using the peptide level k loss data, we estimated a standard deviation of all peptide log2 k loss values assigned to each gene (n = 2,390, npeptides > 1) and to each protein AS group (n = 3,848, npeptides > 1). We further visualized distributions of the standard deviations for genes and protein AS groups, and using Wilcoxon test, we tested whether the observed difference was statistically significant. This analysis was performed separately for the HeLa 1 log2 kloss value, for HeLa Kyoto average log2 kloss value, and for the HeLa CCL2 average log2 kloss value.
2) Similar to a gene/protein-specific correlation between different layers, we calculated a correlation coefficient (Spearman's rho) between peptide log2 kloss profiles across all HeLa cell lines. For every protein AS group (with at least two peptides quantified, n = 3,848), the correlations for all binary peptide comparisons were calculated (i.e., within-protein AS groups). For every gene (with two to four protein AS groups quantified with at least two peptides, n = 1,477), the correlations were calculated for all binary peptide comparison between peptides corresponding to different AS isoforms of the same gene (i.e., betweenprotein AS groups). We then visualized the distribution of all within-and between-protein AS groups correlations, and estimated statistical significance of the difference using Wilcox test. The same analysis was performed for a subset of 30 genes and the corresponding 60 protein AS groups for which we reported differential degradation between HeLa CCL2 and Kyoto (p < 0.05, only genes with 2 protein AS groups were used for this analysis).

Supplementary Figures Captions
Appendix Figure S1: Principles and assumptions underlying protein degradation quantification using pSILAC-DIA. (A) Routine design of a simple, dynamic SILAC labeling experiment. Steadystate cells are grown in a culturing medium containing the "light" versions of lysine and arginine. At time_0, the medium is exchanged for a cell culture medium containing the ""heavy"" versions of the amino acids ("Arg10" and "Lys8"). The cells are then harvested at several different time points, when different fraction of the proteome is increasingly labeled by the ""heavy"" amino acids. Under the hypothesis of "steady-state", the seeding density and the havest time points must be carefully selected to ensure exponential proliferation of the cells during the whole time course of the experiment to maintain the proliferation rate. (B) DIA-MS data of a pSILAC experiment. The "light" and "heavy" peptide precursors (MS1) and fragment ions (MS2) co-elute, but can be distinguished by a mass shift (m/z) due to the SILAC labeling. The "light" and "heavy" channels are aligned by Spectronaut and can be accurately quantified by setting the same peak boundaries for "light" and "heavy" ions, further facilitating the detection of the low intensity "heavy" signal in Correlation between "heavy" and "light" SILAC intensities. Spearman's rank correlation between all "light" and "heavy" peptide-precursors quantified within each time point was calculated for each protein AS group.
Appendix Figure S4: Comparison between MS1-based and MS2-based quantification of peptide precursor ratios using pSILAC-DIA. Similar to Figure 2C-F, for the selected example proteins in (A-C), the scatter plots show the correlation between SILAC "light" and "heavy" peptide-precursor intensities quantified in either MS1 or MS2 from the example data in Figure S3; a linear regression line  A-B), the scatter plots show the correlation between Top 6 SILAC "light" and "heavy" fragment ions, between Top 6 "light" and "heavy" fragment ions that passed the interference filtering step, and between "light" and "heavy" peptide precursors quantified using the MS2 quantification (i.e., the sum of the Top 6 fragment ions that passed the interference filtering). The standard deviations were estimated using all peptide log2 kloss values assigned to either a gene (n = 2,390) or a protein AS group (n = 3,848). (B) Correlation (Spearman's rho) between peptide log2 kloss profiles across HeLa cell lines. For every protein AS group (with at least two peptides quantified, n = 3,848), the correlations for all binary peptide comparisons were calculated (i.e., within-protein AS groups). For every gene (with two to four protein AS groups quantified with at least two peptides, n = 1,477), the correlations were calculated for all binary peptide comparison between peptides corresponding to different AS isoforms of the same gene (i.e., between-protein AS groups). (C) The same analysis was performed for a subset of 30 genes and the corresponding 60 protein AS groups for which we previously reported differential degradation between HeLa CCL2 and Kyoto. p values were estimated using the Wilcox test.
Appendix Figure S8: Intensity distributions of protein and mRNA measurements corresponding to mRNA-kloss quintiles. Log2 protein intensities (A) and mRNA intensities (B) of protein AS isoform groups included in mRNA-kloss quintiles shown in Figure 5D.