Computational approaches to identify sites of phosphorylation

Abstract

Due to their oftentimes ambiguous nature, phosphopeptide positional isomers can present challenges in bottom-up mass spectrometry-based workflows as search engine scores alone are often not enough to confidently distinguish them. Additional scoring algorithms can remedy this by providing confidence metrics in addition to these search results, reducing ambiguity. Here we describe challenges to interpreting phosphoproteomics data and review several different approaches to determine sites of phosphorylation for both data-dependent and data-independent acquisition-based workflows. Finally, we discuss open questions regarding neutral losses, gas-phase rearrangement, and false localization rate estimation experienced by both types of acquisition workflows and best practices for managing ambiguity in phosphosite determination.

Introduction:

Post-translational modifications (PTMs) are a significant regulatory mechanism of protein activity in cells and can operate on a much faster time scale than gene regulation. The addition of phosphates,¹ glycans,² methyls,³ acetyls,⁴ and even other proteins^5,6 can act as signals, activate or change the function of specific proteins, or mark proteins for degradation. Phosphorylation is a common form of cell signaling PTM, where cascades of kinases phosphorylating other kinases amplify signals from the cell surface to the nucleus. There are estimated to be over 500 unique kinases in the human genome, which phosphorylate over 200,000 unique phosphosites.⁷ While phosphorylation has been observed naturally on many residues including aspartic acid,⁸ arginine,⁹ cysteine,¹⁰ and histidine,^11,12 the most commonly observed phosphorylation sites in proteomics experiments are serine (S), threonine (T), and tyrosine (Y).

Given the physiological significance of PTMs, it is important to reliably determine which proteins within a sample are modified. While PTMs change a protein’s overall mass and charge, allowing for analysis with techniques like gel electrophoresis,¹³ liquid chromatography (LC) coupled tandem mass spectrometry (MS/MS) is the preferred technique for monitoring most PTMs.¹⁴ Unlike gel electrophoresis, MS/MS provides sequence-level information across a protein which can indicate the amino acid-specific site of the modification. Advancements in search engine and site localization algorithms, as well as data acquisition methods, have made it possible to identify tens of thousands of sites of phosphorylation in a single MS/MS acquisition. The review of these advancements is the focus of this work.

There are three general approaches to analyzing PTMs with LC-MS/MS: top-down, which measures PTMs on intact proteins, bottom-up (sometimes referred to as shotgun proteomics), which uses enzymatic digestion to process proteins into peptides before measurement, and middle-down, which is a hybrid of top-down and bottom-up. Top-down proteomics can identify the number of occupied phosphosites on a protein molecule, but it can sometimes be a challenge to determine the exact sites of phosphorylation without additional experiments.¹⁵ In contrast, bottom-up proteomics simplifies identifying sites of phosphorylation by analyzing shorter peptide sequences at the cost of context in multiply-phosphorylated proteins.¹⁶ In some experiments¹⁷ these techniques are used in conjunction to take advantage of both methods. Middle-down proteomics attempts to address the complications from both of these methods by generating large peptides using restricted enzymatic proteolysis.^18,19 While bottom-up methods can typically achieve a higher depth of measured phosphosites, top-down and middle-down make it easier to study interactions between multiple sites of phosphorylation or crosstalk between different types of PTMs.^20,21

Bottom-up proteomics methods for measuring phosphopeptides:

Figure 1 outlines a basic workflow of bottom-up proteomics for PTM analysis. As proteomics samples are processed, proteins are typically denatured and digested with an enzyme such as trypsin, which cuts each protein at regular sites. Due to the relatively low abundance of naturally occurring phosphopeptides, samples must first be enriched using techniques such as metal-affinity chromatography,^22,23 titanium dioxide,²⁴ strong cation exchange,²⁵ or antibodies^26,27 to produce quantifiable amounts of phosphopeptides from the sample. Phosphopeptides are then separated using an LC, ionized using electrospray, and taken into the MS where both precursor and sequence-specific fragment ions are measured. From those measurements, data analysis helps identify peptides and assign specific sites of modification.

Figure 1:

Outline of a standard bottom-up LC-MS/MS phosphoproteomics workflow. (a) Peptide samples are prepared by lysis, reduction, alkylation, and enzymatic digestion. (b) After digestion, phosphopeptides are enriched and prepared for LC-MS/MS measurement. (c) Peptides are chromatographically separated to reduce interference from competing signals and improve mass spectrum interpretation. (d) Eluting peptides are ionized and fragmented to produce sequence-level data. (e) Pattern recognition algorithms are used to identify phosphopeptides and fragmentation data is interpreted to identify the most likely site of phosphorylation in each peptide.

There are several different ways that a precursor ion can be ionized and fragmented, which have additional considerations when analyzing PTMs. While most current experiments use positive ion mode for ionizing phosphopeptides, negative ion mode has historically been a powerful technique for identifying phosphopeptides²⁸ and the combination of both positive and negative ion modes can improve phosphopeptide detection confidence.²⁹ Phosphopeptides are commonly analyzed using resonance collision-induced dissociation (CID)³⁰ or beam-type CID³¹ (commonly referred to on Thermo instruments by the branded name, HCD³²). CID methods produce b- and y-type ions, where beam-type CID undergoes multiple fragmentation events and typically produces long runs of y-ions while resonance CID results in the most energetically favored fragmentation pathways producing both b- and y-type ions with similar efficiency. For PTMs, both approaches of ion generation have advantages and disadvantages for assigning phosphopeptides.³³ Additionally, CID methods impart energy into peptides directly through collision,^34,35 which can cause the loss of PTMs during fragmentation as neutral losses,³⁶ and complicate their interpretation. An alternative fragmentation approach uses electron transfer dissociation (ETD)^37,38 or electron capture dissociation (ECD).³⁹ ETD and ECD use free radicals to make peptides unstable, fragmenting them along the backbone while typically leaving PTMs intact.⁴⁰ However, these methods typically produces fewer intense fragment ions when compared with CID, which can result in lower numbers of peptide identifications.⁴¹ Tradeoffs between different fragmentation techniques suggest that using a combinatorial approach could help better assign phosphopeptides.^42,43

Once ions are fragmented, they can be analyzed in a variety of ways. Data-dependent acquisition (DDA)⁴⁴ uses precursor MS1 measurements to trigger selected (data-dependent) MS2s preferentially on the most abundant peptide signals. Instruments configured to collect DDA measurements typically use a top-N configuration, where the top-N most abundant precursor ions are selected for MS2 acquisition from the most recently collected MS1 spectrum. In general, this cycle is repeated every one to five seconds so that a new MS1 spectrum can indicate peptides that are newly eluting from the HPLC column into the mass spectrometer. In most DDA configurations, peptides elute over many cycles and a process called dynamic exclusion^45,46 is employed where recently measured precursor m/zs are put on an exclusion list for a specified duration such that precursors are not repeatedly re-measured in each cycle. As such, most peptides are identified by only a single MS2, allowing for more data acquisition time spent measuring low-abundance peptides. No matter what fragmentation technique is used, database search engines^47–52 such as Mascot⁵³ assign numerical scores to each potential peptide-spectrum match (PSM).⁵⁴ Search engines can be configured to search for variable PTMs,⁵⁵ which add the mass of modifications to user-specified amino acids.

Data-independent acquisition (DIA)^56,57 is an alternate acquisition method where MS2s are collected in a systematic manner regardless of precursor intensities. Systematic measurements from extracted fragment ion chromatograms can improve overall confidence in detecting specific phosphopeptides because each ion is measured multiple times over the elution profile of the peptide. However, precursor masses can be critical to assigning PTMs and peptide-centric DIA search engines⁵⁸ have to infer the specific precursor mass of a peptide from a wider precursor isolation window. As a result, care must be taken to either associate those fragment ions with precursors using tools like DIA-Umpire⁵⁹ or search for ions that are specific to the modified peptide. Targeted measurements, such as Parallel reaction monitoring (PRM),⁶⁰ can be thought of as a subtype of DIA where specific precursors are selected for fragmentation in targeted retention time windows. This approach marries the benefits of DIA with specific precursor information at the cost of measuring only a limited number of peptides per acquisition.

Challenges in interpreting phosphopeptide positional isomers:

The sheer number of kinases⁶¹ and prevalence of serines, threonines, and tyrosines means that many proteins can be phosphorylated in multiple locations, where those sites are frequently at neighboring amino acids.⁶² Modified peptides that differ only in which acceptor amino acid is phosphorylated are isomers because they have the same molecular formula, yet have different structures. As such, knowing the specific site of phosphorylation can be extremely informative, but these positional isomers can be very difficult to differentiate. For example, Figure 2 shows major phosphosites of Human CDK1 (cyclin-dependent kinase 1), an important regulator of the cell cycle by controlling cell division.⁶³ The three most commonly-cited phosphosites (T14, Y15, and Y19) all fall within the same tryptic peptide: IGEGTYGVVYK, which has exact conservation across human, mouse, rat, chicken, and fruit flies, with nearly identical homology in yeast (VGEGTYGVVYK). Each of these amino acids serves the same inhibitory function but differs in what upstream kinases phosphorylate them. T14 and Y15 can both be phosphorylated by MYT1,⁶⁴ while Y15 can also be phosphorylated by PKC⁶⁵ and WEE1.⁶⁶ Little is understood about kinases that target Y19. Palou et al62 observed that the SWE1 ortholog of WEE1 phosphorylates Y19 in S. cerevisiae, although this finding should be viewed skeptically as tyrosine phosphorylation signaling is rare or nonexistant in yeast⁶⁷ underlining the need to validate antibody-based phosphorylation assays with mass spectrometry.⁶⁸ Regardless, phosphorylation at each site produces the same CDK1 inhibition, yet observing each site indicates different upstream biology.⁶⁹

Figure 2:

CDK phosphosites with at least 10 literature references. Highlighted phosphosites represent well-documented positional isomers. These positional isomers may serve similar functions and be located adjacent to one another, such as inhibition (T14, Y14, Y19), but can differ in what kinases can phosphorylate them. Phosphosite reference counts were derived from PhosphoSitePlus.⁷⁹

The easiest and most confident way to know the localization of peptides is to add stable isotope-labelled synthetic phosphopeptide standards to each sample.^70,71 However, the cost and scale of generating a massive library of synthetic phosphopeptide standards remains impractical,^72,73 requiring computational methods to assign sites of phosphorylation. Due to the prevalence of positional isomers within the proteome, the use of standard search engine scores is often not enough to confidently identify isoforms.⁷⁴ That said, simply considering the difference between the top two highest-scoring peptides can be an effective approach to localization. This concept was first considered in 1995 when Yates et al⁵⁵ used deltaCn, or the normalized difference between the best and second-best cross-correlation scores to identify phosphopeptides. The Mascot delta score (MD-score) approach⁷⁵ similarly considers the top two highest scores of a Mascot search, where the difference in Ions Scores relates to the separation between the quality of ions assigned to the first and second best peptides considered for that PSM. Since positional isomers have the same precursor mass and many of the same fragment ions, generally both the first and second-best peptides considered for a PSM are different positional isomers of the same peptide. As such, the MD-score frequently becomes a measure of how much better the top positional isomer scores above the other isomeric forms. Similarly, the SLIP score⁷⁶ for Protein Prospector^77,78 computes the difference between expectation values, analogous to probability values adjusted for the number of precursors considered within a given mass tolerance.

While search engine scores consider the presence or absence of every potential sequence-specific fragment ion, only some ions change between positional isomers. Figure 3 shows a spectrum from the Phosphopedia database⁸⁰ associated with each singly-phosphorylated form of the IGEGTYGVVYK peptide from CDK1 discussed previously in Figure 2. Notably, each positional isomer has the same precursor mass, but different fragmentation patterns. For example, y2 to y5 (as well as b6 to b10) fragment ions are mass-shifted between the pY15 and pY19 isomers. These peaks are called site-determining ions and can be used to differentiate the isoforms of a specific peptide. In some cases, only a few site-determining ions are produced, such as only y6 and b5 differentiating between pT14 and pY15. Notably, the overall relative intensities of the fragmentation patterns,³⁴ as well as the LC retention times,⁸⁰ can differ between serine/threonine and tyrosine phosphorylated peptides, which can be used as an additional source of localization information. However, serine and threonine phosphorylated peptides generally have more similar relative fragmentation patterns and retention times due to the closeness of the chemistries of those amino acids. Spectra are considered “ambiguous” if no site-determining ions are present or if those ions cannot be distinguished from noise.⁸¹

Figure 3:

Resonance CID MS2 spectra from the CDK1 peptide IGEGTYGVVYK from the Phosphopedia library.⁸⁰ Each spectrum corresponds to one of three positional isomers. Peaks are colored based on what isomer they can be used to identify. (a) The pT14 isomer; pink b/y-type ions uniquely identify this positional isomer. (b) The pY15 isomer; this particular species does not possess any unique site-determining ions and can only be identified in comparison to pT14 or pY19, with C-terminal ions shared with pT14 (blue) and N-terminal ions shared with pY19 (green). (c) The pY19 isomer; orange b/y-type ions uniquely identify this positional isomer.

When a phosphopeptide has only two acceptor sites, each positional isomer has an equal number of unique ions that could distinguish that isomer. In peptides with more than two acceptor sites, the internal acceptor sites will not have any unique site-determining ions that can be used to definitively identify it. For instance, in the peptide IGEGTYGVVYK, the pT14 isomer can be distinguished from both pY15 and pY19 via the unique b5 and y6 fragment ions. Similarly, pY19 can be distinguished from both pT14 and pY15 by the isomer-specific b6–b9 and y2–y5 fragment ions. In contrast, the phosphopeptide pY15 shares the shifted/unshifted C-terminus fragments with pT14 and the shifted/unshifted N-terminus fragments with pY19. This can create an interpretation challenge if the pT14 and pY19 isomers were to co-elute, where pY15 can only be confidently identified if the other isomers can be ruled out. As such, most approaches to identifying the site of phosphorylation consider the delta between the best and second-best matches. In this case, the pY15 isomer can be localized compared to pT14 or pY19, but not both simultaneously. If the delta between the top and second-best match is sufficiently large, that result would imply that any further result (third-best species and so forth) would not score better than the second-best match, allowing them to be discounted from consideration.

The variety and scope of positional isomers in phosphopeptide-enriched datasets are most easily visualized by considering precursor-extracted ion chromatograms across retention time. Figure 4a presents a three-dimensional landscape of precursor intensities in a human phosphopeptide-enriched sample downloaded from the MassIVE repository (ProteomeXchange dataset PXD042974). By focusing on a narrow 800.5 to 803 m/z range from 15 to 26 minutes, this plot shows the isotopic structure of peptides in this range, indicating charge, as well as peptides with the same precursor mass and isotopic pattern, suggesting the same molecular formula. Three peptides have precursor mass patterns suggesting positional isomers, where the yellow peptide was identified as singly phosphorylated SRTESITATSPASMVGGKPGSFR from IRS1. Based on the precursor data, this peptide has two positional isomers. In Figure 4b and the corresponding blowup Figure 4c, the ambiguous isomer (pS307|pT309) is identified by the blue ions while the ambiguous isomer (pT311|pS312) is identified by the red ions, but pS307 and pT309 as well as pT311 and pS312 cannot be uniquely distinguished from each other due to missing ions. In situations with localization ambiguity, existing literature can act as a powerful tool⁸² where pS307 and pS312 have been measured in 64 and 87 studies, respectively, from data in PhosphositePlus. Meanwhile, the other acceptor sites in SRTESITATSPASMVGGKPGSFR have only been observed in a handful of high-throughput studies. This result suggests the presence of IRS1 phosphorylation at S307 and S312, which are both phosphorylated by p70-S6 kinase.⁸³ The purple peptide (m/z=801.733, z=3) and red peptide (m/z=800.867, z=2) are unassigned but show similar precursor patterns.

Figure 4:

Positional isomers observed in a human phosphopeptide-enriched sample. (a) A three-dimensional precursor intensity landscape, where highlighted precursors are suggestive of positional isomers producing multiple peaks with the same m/z value and isotopic distribution, but at different retention times. The yellow peak corresponds to the singly phosphorylated IRS1 peptide SRTESITATSPASMVGGKPGSFR while the red and purple peaks are unassigned. If the retention time between these peaks is not sufficiently different enough, dynamic exclusion in DDA may cause only one isoform to be analyzed. (b) Extracted chromatograms of the peptide SRTESITAPASMVGGKPGSFR, where site-determining ions specific to the ambiguous SRTE(pSIT)ATSPASMVGGKPGSFR (e.g., either pS307 or pT309) are highlighted in red and site-determining ions belonging to SRTESITA(TpS)PASMVGGKPGSFR (e.g., either pT311 or pS312) are highlighted in blue. Ions belonging to both peptides are dashed gray lines. (c) A zoomed-in view of (b) showing lower-intensity peak shapes in greater detail.

While dynamic exclusion is a key tool for measuring low-abundance peptides with DDA, the approach can create complications when analyzing phosphopeptide positional isomers. Because positional isomers have the same molecular formula, they also have the same precursor mass. As such, if those positional isomers elute in close proximity to each other, the precursor mass from analyzing the first eluting positional isomer may still be on the instrument exclusion list during the time the second isomer elutes. In this case, the second positional isomer may never trigger an MS2 spectrum, even if its overall signal is higher than the first isomer.⁸⁴ Similarly, inherent stochasticity caused by the data-dependent nature of DDA may cause the first isomer to be missed if it is low in abundance, leaving room to collect an MS2 spectrum of the second isomer. Left unchecked, these characteristics of DDA cause a dramatic drop in the reproducibility of DDA-based phosphoproteomics experiments relative to DDA experiments of unmodified peptides.⁸⁵ Furthermore, positional isomers with the same precursors and nearby elution times can complicate precursor-based quantification and quantification approaches that use “match-between-runs”⁸⁶ where MS2-based detections in one injection can be transferred to other injections based on retention time and precursor mass matching alone. These factors underline the need for using isomer-specific fragment ions for detecting positional isomers.

Computational approaches to localize phosphates to specific sites in phosphopeptides:

Ascore⁷⁴ is the earliest method that utilizes site-determining ions to localize phosphate groups. The Ascore algorithm consists of three major steps. First, a set of fragment ions is generated for all known positional isomers from a given sequence. Second, these peaks are allocated into a series of 100 m/z wide bins and a binomial probability scorer is used to generate a Peptide Score for each isoform based on the number of matching peaks. The peak depth, or the ‘n’ most intense peaks within each 100 m/z window where n is an integer from 1 to 10 is determined by the largest delta Peptide Score for a given value of n. The last step is to recalculate the Peptide Score for the two highest scoring isomers, but considering only site-determining ions. In this scorer, the probability P(x) for PSM x is calculated as:

$$P\left(x\right)={\sum }_{k=n}^N\frac{N}{k}{p}^k{\left(1-p\right)}^{N-k}$$ (Eq. 1)

where N is the number of possible ions that could differentiate between the two isomers and n is the number of those ions present in the spectrum. Assuming unit resolution, p is set to the peak depth divided by 100 (as the windows are 100 m/z wide). This formulation can be interpreted as the integration of the binomial distribution for the probability that the same or higher number of ions observed could be observed by random chance. Finally, the Peptide Score is set to −10*log₁₀(P(x)) to ensure increasing numbers for more significant matches. The Ascore is the difference between these two site-specific Peptide Scores. This approach for scoring only site-determining ions for localization is demonstrably better than calculating the delta score between first and best-scoring peptides when considering all possible fragment ions, as demonstrated by Ascore’s precision and sensitivity improvements over both the MD-score and deltaCn.⁷⁴

The fundamental Ascore algorithm using the binomial distribution for calculating localization probabilities has been implemented in several other contexts and tools.⁸⁷ For example, the Phosphate Localization score (PLscore)⁸⁸ simplifies the algorithm by utilizing a single peak density and extending the principles of Ascore to InsPecT⁸⁹ results. PhosphoRS⁹⁰ expands the idea of peak depth to allow for variable amounts of peaks within each window, thus accounting for an unequal distribution of peaks across m/z windows in a single spectrum. SLoMo (site localization of modifications)⁹¹ is designed to allow for ETD/ECD data to be analyzed, while the Cscore (complementary score)⁴² allows for both ETD/ECD and CID data to be utilized concurrently, increasing the confidence of peptide site localization. PhosCalc⁹² can be used to analyze MS3 spectra⁹³ in addition to MS2 spectra, which can help assist in CID-based experiments where backbone fragmentation may be poor. Another method for utilizing MS3 in conjunction with MS2 spectra is the PTM score used in MSQuant,⁹⁴ where localization is performed using the same approach as Ascore but limiting peak depth to four.

Other algorithms have been developed to score site-determining ions without assuming peaks are measured at random in a binomial distribution. For example, LuciPHOr⁹⁵ considers the likelihood odds ratio that a fragment peak would be assigned correctly or randomly and computes the cumulative sum of the log odds to assess each considered positional isomer, where the confidence of the reported isomer is the delta of the top two permutations. PTMProphet⁹⁶ uses a Bayesian approach to assess the confidence at each possible site using mixture models. This approach allows for ambiguous results to be returned when the exact site cannot be determined, but it can be limited to specific amino acids. PTMiner⁹⁷ is an approach for open-searching, where a database search engine is allowed to detect unanticipated modifications considering large mass tolerances. This strategy uses an empirical Bayesian-based approach to iteratively learn the prior probabilities for observing each type of modification on any given amino acid. Similar to the open-searching approach, pSite⁹⁸ uses a support vector machine to consider the confidence of each individual amino acid in de novo-derived sequences that have been assigned to proteins using sequence alignment.^99,100

While Ascore and related algorithms primarily utilize database searching, the use of spectral libraries allows for incorporating sequence-specific features such as intensity values, neutral losses, and unusual fragments into a localization analysis process.¹⁰¹ Recent methods allowed for the development of computationally predicted spectral libraries, where non-phosphorylated spectra are shifted by +80 Da to simulate phosphorylation.^102,103 First developed for Orbitraps, this method was expanded upon by Suni et al,^101,104 using a two-step process of enzymatic dephosphorylation followed by in silico rephosphorylation and has been extended to ToF platforms.¹⁰⁵ Dephosphorylated peptides share many of the same peaks as phosphorylated peptides but are dramatically easier to detect. As such, library searching for “re-phosphorylated” forms of dephosphorylated peptides greatly improves phosphopeptide detection using SpectraST.¹⁰⁶ In this approach, localization is performed by recalculating the original deltaDot (delta score of the dot product) for the first and second-best matches that correspond to the same peptide sequence. DeepFLR¹⁰⁷ also utilizes library searching to assess site localization where phosphopeptide libraries are built from deep learning-based predictions. It is important to note that library searching localization is based on changes in relative peak intensity, and these algorithms may not require the presence or absence of any site determining ions to localize peptides. A summary of these algorithms can be found in Table 1.

Table 1:

DDA-Based Localization Algorithms

Name	Category	Summary Description	Reference
Ascore	Binomial	Uses sequence information to predict fragmentation patterns; looks for the presence or absence of site-determining ions.	74
Cscore	Binomial	Implementation of Ascore that utilizes both ETD and CID data concurrently to increase localization confidence.	42
DeepFLR	Library Matching	Constructs a spectral library using deep learning; localization scores are calculated using delta cosine similarity.	107
LuciPHOr	Probabilistic	Computes the likelihood odds that a given peak is matched at random before calculating a cumulative sum of the log odds to represent the entire peptide. The score itself is the delta of the two most likely isoforms.	95
MD-Score	Delta Score	Difference between the first and second best results from the MASCOT search engine for positional isomers.	75
PhosCalc	Binomial	Implementation of Ascore that can be applied to both MS2 and MS3 spectra.	92
PhosphoRS	Binomial	Implementation of Ascore that uses variable peak densities per 100 m/z window to account for unequal peak distribution in a spectrum.	90
Phosphorylation Localization Score (PLS)	Binomial	Extends Ascore method to InsPecT; uses a single peak density to simplify the scoring process.	88
PTMiner	Binomial	Designed for high mass-tolerance open searching; uses an empirical Bayesian-based iterative learning strategy for site localization.	97
PTMProphet	Bayesian	Uses a Bayesian mixture model to evaluate confidence at each possible site.	96
PTM score (MSQuant)	Binomial	Implementation of Ascore that allows for the analysis of MS3 spectra along with MS2 spectra; only considers the top 4 most intense peaks per 100 m/z window.	94
pSite	Bayesian	Site-based approach that first evaluates individual site confidence using support vector machines for de novo derived sequences assigned through sequence alignment.	98
SLIP	Delta Score	Designed to work with Protein Prospector using expectation values. The SLIP score is the difference between the first and second-best expectation values.	76
SLoMo	Binomial	Implementation of Ascore that allows ETD and ECD data to be used.	91
Suni et al.	Library Matching	Constructs a spectral library by applying mass shifts to enzymatically dephosphorylated peptide spectra; localization scores are recalculated deltaDot scores from SpectraST using these libraries.	101

Phosphoproteomics analysis methods for data-independent acquisition:

While DDA typically attempts to fragment only a single peptide in each MS2 spectrum, DIA actively generates highly multiplexed MS2 spectra where multiple peptides are co-fragmented at the same time and extracted computationally.¹⁰⁸ Specter¹⁰⁹ is one approach to deconvolve this data using a non-negative least squares approach to find the simplest linear combination of library spectra that interpret a given window. Given complete data, this approach is able to distinguish highly similar library spectra including phosphopeptide positional isomers. However, this approach uses intensity alone to identify positional isomers, with no consideration for site-determining ions. Time alignment is another method of grouping fragment ion signals. For example, PIQED¹¹⁰ leverages the DIA-Umpire⁵⁹ tool to match time-extracted fragment ions with specific precursor ions in MS1 spectra to create pseudo-DDA spectra that can be searched using standard search engines. PIQED creates a pipeline that processes pseudo-DDA spectra through search engines, such as Comet,¹¹¹ using statistical analysis with the TPP,¹¹² including PTMProphet.⁹⁶

Other approaches to analyzing phosphopeptides measured with DIA are built into DIA-specific processing tools. Inference of peptidoforms (IPF)¹¹³ is a Bayesian hierarchical model integrated into the OpenSWATH¹¹⁴ workflow that determines the most likely positional isomer (peptidoform) at a given retention time based on the elution pattern of site-determining ions. Thesaurus⁸⁴ is built into EncyclopeDIA¹¹⁵ and uses an Ascore-like approach to model the probability of observing site-determining ions by chance. DIA MS2s are typically far more complex than DDA MS2s, and thus it is incorrect to assume the probability of observing an ion by chance is related to the number of peaks considered per m/z window. As a result, Thesaurus constructs a background distribution for each m/z window based on the frequency that any given m/z value is observed. Finally, site localization has been built into Spectronaut through directDIA^116,117 using a deconvolution approach similar to DIA-Umpire. A summary of these algorithms can be found in Table 2.

Table 2:

DIA-Based Methods

Name	Category	Summary Description	Reference
Inference of Peptidoforms (IPF)	Bayesian	Applies a Bayesian hierarchical model to spectral library data to determine the most likely positional isomer or ‘peptidoform’ at a given retention time based on elution pattern.	113
PIQED	Workflow	Sequences DIA-Umpire to deconvolve DIA data through the creation of pseudo-DDA spectra that can be localized through TPP-compatible methods like PTMProphet.	110
Spectronaut	Closed Source	Allows for analysis of DIA without libraries through directDIA using deconvolution methods similar to DIA-Umpire.	117
Thesaurus	Probabilistic	Creates a background distribution using differential frequencies of peptides in a dataset and then calculates probabilities based on whether site-determining ions can be matched in that background distribution by random chance.	84

Estimating false localization rates across datasets:

Several experimental factors, such as data acquisition parameters, chromatography, instrument resolution, peak selection, and even the localization approach used, can cause results between algorithms to vary greatly.¹¹⁸ For normal proteomics experiments, false discovery rate (FDR) estimation can control for false positives within a set of PSMs.^119,120 The target-decoy approach uses the score distribution of decoy peptides to model a null distribution from which to evaluate target peptide matches.¹²¹ This principle works because the decoy-based null distribution closely approximates incorrect target matches, where: i) all decoy matches are incorrect, ii) all correct matches are from the target database, and iii) incorrect matches are equally likely to be from the target or decoy database.

A similar concept for analyzing the accuracy of site localization algorithms is the false localization rate (FLR). However, the FLR approach cannot make use of standard decoys because they model incorrect peptides, not incorrect localizations.⁸¹ One key difference is that incorrect localizations still score similarly to correct peptides with only a few site-determining ions mismatched. As a result, FLR estimations lack a standardized metric, resulting in arbitrary cutoffs being used to evaluate the performance of localization algorithms.¹¹⁸

Chalkley et al.⁸¹ assert that the true FLR rate of a dataset can only be truly measured with prior knowledge of correct localizations, which may be modeled with synthetic peptides or singly-phosphorylated peptides. Most current methods of FLR estimation utilize biologically improbable amino acids to represent decoys, such as with PhosphoFLR,¹²² which considers decoy localizations to alanine and leucine. DeepFLR¹⁰⁷ additionally tested an alternate decoy generation approach that randomly swapped a phosphorylated amino acid with an unphosphorylated amino acid, which maintains biologically plausibly modified amino acids at the cost of introducing sequence order changes where the resulting peptide is not necessarily present in the sample. The authors determined that this approach was more effective at estimating the real FLR than moving the phosphate to a randomly selected, biologically improbable amino acid. One possible explanation of this result is that exchanging amino acids is better suited for deep learning prediction tasks, which may account for the improvement in accuracy. However, localization engines that do not use fragment intensities likely will not benefit from this technique and may produce worse results due to introducing sequence alterations.

Other methods for FLR estimation are integrated into localization algorithms themselves. SLIP scoring⁷⁶ takes advantage of amino acid distribution and constructs decoys from glutamine and proline localizations, which are residues commonly observed near actual phosphosites but may vary from dataset to dataset. LuciPHOr⁹⁵ uses a Bayesian approach similar to PeptideProphet¹²³ to consider score distributions of S/T/Y localizations to other amino acids. PTMProphet⁹⁶ computes probabilities for each potential site of phosphorylation and calculates the localization mean best probability as a reasonable estimate for FLR.

Benchmarking studies can also be helpful in validating FLR estimates between algorithms. Locard-Paulet et al.¹²⁴ exhaustively evaluated the performance of 22 different proteomic pipelines using synthetic phosphopeptide data with known phosphosites. These pipelines were composed of search engines, localization algorithms, and validation strategies embedded within different phosphoproteomics software tools. The direct comparison of FLR used in this study serves as a useful resource for adapting the score thresholds between pipelines, helping to mitigate the variation in FLR between algorithms at the same threshold.

Open questions when analyzing phosphoproteomics datasets:

Regardless of the data acquisition method used, several key challenges still persist within the site localization process. The use of resonance or beam-type CID-derived fragmentation techniques produces b- and y-ions for site identification,¹¹⁸ however, this technique creates additional dataset-complicating ions. These new ions include neutral losses and gas-phase re-arrangements, which may result in failure to detect or correctly localize the phosphosite.¹²⁵ Efforts to incorporate and adjust for these complications within scoring techniques likewise face similar challenges, particularly with regard to creating a standardized method of reporting false localization rates (FLRs).⁸¹

In CID-based mass spectrometry, the phosphate group of a phosphopeptide can often be lost in the form of phosphoric acid (H₃PO₄) in positive ion mode or metaphosphoric acid (HPO₃) in negative ion mode.^126,127 Serine and threonine residues undergo these neutral losses through beta elimination at a rapid rate due to the labile nature of the phosphate group,¹²⁸ while tyrosine is typically unable to undergo neutral losses. However, isobaric tag labeling with tandem mass tags (TMT) has been observed to have phosphotyrosine neutral losses.¹²⁹ For example, in Figure 3a, the CDK1-pT14 isomer has several neutral loss peaks (y7–98 = 811.435 m/z, y8–98 = 868.456 m/z, y9–98 = 997.499 m/z, and y10–98 = 1054.520 m/z) that are absent in the pY15 and pY19 spectra. These losses can undercut localization by shifting the mass of modified residues since the beta elimination neutral loss of phosphoric acid from phosphorylated serine and phosphorylated threonine produces the same fragment mass as beta elimination of water from an unmodified serine or threonine.¹¹⁸ Some site-localization algorithms still attempt to use phosphate neutral loss ions as site-determining ions, but this has been demonstrated to increase localization error rates.¹⁰¹ Library searching may provide an avenue to consider phosphoric acid neutral losses by leveraging differences in the observed intensity for each ion type.

During the fragmentation process of CID-based experiments, the same gas phase beta-elimination reactions that induce neutral losses can result in the translocation of the phosphate group to a nearby acceptor site.^125,130 Gas-phase rearrangements result in chimeric spectra¹³¹ that appear to be the product of both isomers but at the elution time of the original isomer. While these gas-phase products are typically low in abundance, it has been estimated that as many as 37% of positional isomers that appear to be identically co-eluting are actually side-products formed in the gas-phase,⁸⁴ matching estimates from Plaumbo and Reid¹²⁵ showing that 36% of tested phosphopeptides underwent some degree of site rearrangement. Plaumbo and Reid suggest that CID-based proteomic data should be re-evaluated before making conclusions regarding biological significance, especially if localization software considers the possibility of multiple positional isomers from the same MS2 spectrum. In these cases, positional isomers can be validated by considering known kinase motifs, comparisons to similar data produced using ETD or ECD fragmentation, or by comparison to CID MS2 spectra of synthetic phosphopeptide standards.¹²⁵ While gas-phase rearrangement is reproducibly common across resonance and beam-type CID instruments, Aguiar et al¹³² argue that from a practical perspective, it may have a limited impact on phosphopeptide localization. This study considered two combinatorial libraries of singly phosphorylated peptides with two acceptor sites and found that rearrangement products were rarely observed to decrease localization accuracy in +2H MS2 spectra. The authors further argue that using ETD, which does not produce rearrangement products, did not dramatically affect localization confidence in Ascore. While combinatorial libraries contain significantly more phosphopeptides than the pool of phosphopeptides studied by Palumbo and Reid, the limited combinations of chemistry restrain the practical conclusions that can be drawn from this result, leaving the real effects of gas-phase rearrangement on proteomics studies still somewhat unresolved.

Regardless, chimeric spectra containing fragments produced by multiple isomers resulting from gas-phase rearrangement or simply shortened LC gradients are a major challenge for detecting middle positional isomers. Since middle isomers such as IGEGTpYGVVYK from CDK1 (Figure 3) also produce site localizing ions that could be the result of combinations of other isomers (i.e., IGEGpTYGVVYK and IGEGTYGVVpYK), it can be additionally challenging to confidently confirm these sites. In these cases, changes in retention time, fragment intensity, or quantification are critical to confidently validate these peptides. The effective and automated use of these factors to improve site localization, in combination with literature mining or evolutionary conservation to assess the prior probability of observing a phosphosite, remain fruitful computational problems for future researchers.

Conclusion:

The field of phosphoproteomics has enjoyed rapid growth over the last twenty years as proteomics instrumentation and methodologies have improved. While mass spectrometry is a powerful method for detecting phosphopeptides, challenges remain regarding determining the exact sites of phosphorylation, which are needed to improve our understanding of the biological context surrounding these modifications. New approaches to making mass spectrometry measurements with data-independent acquisition are improving our ability to detect and quantify phosphosites. Despite these advances, neutral losses, gas-phase rearrangements, and lack of standard FLR metrics remain open challenges in both DDA-based and DIA-based proteomic workflows. Representing ambiguity in incomplete phosphosite localization, both in the literature as well as in databases and data standard formats such as mzIdentML,^133,134 will be critical to our ability to interpret and reuse results from phosphoproteomics studies.