A tryptic digest of a cell lysate is overwhelmingly unphosphorylated peptide. Phosphorylation is often substoichiometric, which is why an enrichment procedure for phosphorylated peptides is treated as a necessary prerequisite for characterizing them by modern mass spectrometric methods [1]. Deep profiling of a single human cell line mapped more than 50,000 distinct phosphorylated peptides and detected more than three-quarters of cellular proteins as phosphoproteins, and determined very high stoichiometries in mitosis or growth factor signaling by label-free quantitation, while reporting that tyrosine phosphorylation is maintained at especially low stoichiometric levels in the absence of specific signaling events [2]. The modification is widespread across the proteome; the occupancy at any one site at any one moment usually is not.
Everything downstream inherits that ratio. What follows runs the chain from why the enrichment step exists, through the two dominant chemistries and the additives that make them selective, to the fragmentation and scoring decisions that determine whether you can name the residue carrying the phosphate.
The standard justification, and what a controlled test found
Enrichment from complex peptide mixtures is routinely justified on two grounds at once: it reduces sample complexity, and it reduces ionization suppression effects [3]. The first is straightforward bookkeeping. The second has been tested directly, and the result is more equivocal than the shorthand suggests. Steen and colleagues used a set of synthetic peptide/phosphopeptide pairs quantitated by amino acid analysis to examine three standard arguments under electrospray ionization: increased hydrophilicity of the phosphopeptide with a concomitant loss during loading onto reversed phase, selective suppression of phosphopeptide ionization in the presence of unmodified peptides, and lower ionization or detection efficiency for phosphopeptides compared with their unmodified cognates. Under the conditions applied, none of the experiments performed supported the first two notions, and for the third no evidence for decreased ionization or detection efficiency of phosphopeptides could be found [4]. That same report still names substoichiometric phosphorylation as an inherent problem, listed separately from the three arguments [4].
Read narrowly, as it should be — one ionization mode, synthetic pairs, defined conditions — the finding does not license the claim that suppression never occurs in a real lysate. What it does license is a change of emphasis. If phosphopeptides are missing, look first at how few molecules entered the instrument and at what the acquisition method chose to fragment, rather than assuming a per-molecule ionization penalty you cannot act on.
Two chemistries, four metals
Two families are in routine use: immobilized metal-ion affinity chromatography, in which a chelated metal ion is fixed to the support, and metal oxide affinity chromatography, in which the oxide itself provides the binding surface. One recent comparison ran Zr(IV)-IMAC magnetic microparticles against the more commonly used Ti(IV)-IMAC and TiO2 magnetic microparticles, and reported improved performance over HPLC-based Fe(III)-IMAC as well [5].
Format matters as much as metal. Moving Fe-IMAC out of batch and microtip formats into an HPLC column removed the bead-to-sample ratio problem and scaled linearly from 100 µg to 5 mg of digest; direct measurement identified more than 7,500 unique phosphopeptides at 90% selectivity with a median coefficient of variation of 15%, rising to more than 14,000 unique phosphopeptides when the IMAC eluate was subjected to a subsequent hydrophilic strong anion exchange separation [6].
Selectivity and recovery are different quantities, and only one of them is usually reported. Selectivity is the proportion of identified peptides that carry a phosphate, a property of the eluate. Recovery is the proportion of the phosphopeptides you loaded that you get back. In hydroxy acid-modified metal oxide chromatography, recovery varied greatly from peptide to peptide, ranging from a few percent to 100%, and the average was almost 50% [7]. A 90%-selective enrichment can be a 50%-recovery enrichment at the same time, which is why the absence of a peptide from the eluate is weak evidence that the protein was not phosphorylated there.
Competitive binders in the loading buffer
Both chemistries bind more than phosphate, and the fix in both cases is to put a competitor in the loading buffer that occupies the non-specific sites without displacing the analyte. Loading TiO2 in 2,5-dihydroxybenzoic acid produced a very efficient reduction in the binding of nonphosphorylated peptides while the resin retained its high binding affinity for phosphorylated peptides. Other substituted aromatic carboxylic acids were also capable of specifically reducing binding of nonphosphorylated peptides, whereas phosphoric acid reduced binding of both phosphorylated and nonphosphorylated peptides [1]. That last comparison is the design rule stated as an experiment: an additive that competes too well strips the analyte along with the background.
Aliphatic hydroxy acids do the same job through different chemistry. Lactic acid and β-hydroxypropanoic acid were used to reduce the interaction between acidic non-phosphopeptides and the metal oxides, with titania and zirconia respectively, and removed the vast majority of non-phosphopeptides from phosphoprotein standard digests [7]. In that work the overlap between the phosphopeptides enriched by the two hydroxy acid variants was 40%, and the combined result was 1,646 unique phosphopeptides from a HeLa cytoplasmic fraction [7].
These are not marginal adjustments. Optimizing sample-loading conditions alone increased phosphopeptide recovery by 22%, 24% and 35% for Zr-IMAC, Ti-IMAC and TiO2 workflows respectively, and the optimized Zr-IMAC protocol then identified up to 23% more phosphopeptides than Ti-IMAC, TiO2 and HPLC-based Fe(III)-IMAC [5]. Those two numbers are not the same quantity — the first is a gain in recovery, the second a gain in identified phosphopeptides — but their similar size is worth holding in mind before a lab changes resin to chase what the loading buffer might have delivered.
The multiply phosphorylated fraction
IMAC and TiO2 do not sample the phosphoproteome identically, and the difference concentrates in peptides carrying more than one phosphate. In a multi-step comparison from whole cell lysate, with peptide-to-bead ratios optimized for each material and three rounds of enrichment each, both enriched similar amounts of phosphopeptide with comparable enrichment efficiency. But the phosphopeptides unique to IMAC enrichment showed a higher percentage of multi-phosphopeptides, as well as a higher percentage of longer, basic and hydrophilic phosphopeptides, and the two procedures clearly enriched phosphopeptides with different motifs [3]. The same study reports that further enriching with two rounds of TiO2 from the supernatant after IMAC, or two rounds of IMAC from the supernatant after TiO2, does not fully recover the phosphopeptides that the corresponding multi-step enrichment missed [3]. Cross-chemistry mop-up is not a substitute for getting the first enrichment right.
The alternative is to separate the two populations deliberately. Sequential elution from IMAC splits monophosphorylated from multiply phosphorylated peptides so that each pool can be analyzed with mass spectrometric parameters differentially optimized for its properties. From 120 µg of human mesenchymal stem cells, that scheme yielded more than double the total number of identified phosphorylation sites obtained with an optimized titanium dioxide chromatographic method, primarily from a 3-fold increase in recovery of multiply phosphorylated peptides [8].
Not all of the bias sits in the resin. In the same work that introduced DHB loading, LC-ESI-MS/MS was biased toward monophosphorylated peptides whereas MALDI MS was not [1]. A multiply phosphorylated peptide can survive enrichment and still be under-reported by the front end you chose.
Complementarity, or capacity
The standard reading of these differences is that the materials are orthogonal. Assessed on aliquots of a tryptic digest of the cytosolic fraction of Drosophila melanogaster Kc167 cells, phosphoramidate chemistry, IMAC and titanium dioxide each isolated phosphopeptides reproducibly, but the methods differed in their specificity of isolation and, notably, in the set of phosphopeptides isolated; the results were taken to suggest that the three detect different, partially overlapping segments of the phosphoproteome and that, at that time, no single method was sufficient for a comprehensive phosphoproteome analysis [9].
A later study that revisited the column format, and tested binding directly by re-enriching flow-throughs, reached a different one. Permutation enrichments of flow-throughs showed that all materials largely bound the same phosphopeptide species, independent of physicochemical characteristics, while binding capacity and elution efficiency did profoundly differ among the enrichment materials and formats. On that basis the often-quoted orthogonality of the materials has to be called into question, and the authors state that their results strongly suggest that insufficient capacity, inefficient elution and the stochastic nature of data-dependent acquisition are the causes of the experimentally observed complementarity [6].
Both papers point at the same bench test. Re-enrich the flow-through on the same material before introducing a second chemistry. A large second crop means you were capacity-limited; a small one means the difference you are chasing is in the chemistry.
What CID does to a phosphoserine peptide
Once the peptide reaches the instrument the problem changes shape. The gas-phase reactivity of phosphoester bonds leaves CID spectra dominated by phosphoric acid neutral loss events, which suppress informative peptide backbone cleavages [10]. Loss of H3PO4 was observed as a dominant reaction pathway on MS/MS of singly and doubly charged phosphoserine- and phosphothreonine-containing peptide ions, and the magnitude of that loss was highly dependent on the proton mobility of the precursor ion [11].
The mechanism is not the one usually quoted. Work with regioselectively and uniformly deuterated derivatives demonstrated that the loss of H3PO4 does not predominantly occur via a charge-remote beta-elimination reaction. Product ions corresponding to loss of formaldehyde (30 Da) from phosphoserine peptides and acetaldehyde (44 Da) from phosphothreonine peptides, seen on MS3 dissociation of the [M+nH-H3PO4] product ions, provided evidence for a charge-directed SN2 neighboring group participation reaction forming a cyclic product ion, and those diagnostic MS3 product ions may potentially provide additional information for characterizing phosphopeptides containing multiple potential sites [11].
Which sequences lose the phosphate hardest has been mapped across a collection of more than 35,000 unique phosphopeptide MS/MS spectra. Neutral loss of H3PO4 is affected by both proximal and distal sites, most notably basic residues and the peptide N-terminal primary amine. Contrary to the proposal that protonated basic residues catalyze the loss through direct interaction with the phosphate, nearby basic groups were found to decrease neutral loss regardless of mobility class, an effect only seen once spectra were stratified by charge mobility, with the most inhibitory bases those immediately N-terminal to the phosphate. Proline positioned between the phosphate and a possible charge donor can dramatically reduce the presence of neutral loss, and in mobile proton spectra the N-terminus is the strongest predictor of high neutral loss [10].
The practical consequence is that identification and localization fail independently. A spectrum whose base peak is the neutral-loss channel can still identify the peptide from the precursor mass and a handful of surviving fragments while lacking the ions that discriminate between two candidate serines three residues apart.
Radical-driven fragmentation, and where it wins
Electron transfer dissociation was introduced as a technique providing more comprehensive coverage of peptide sequences and posttranslational modifications. In an early global comparison on human embryonic kidney 293T cells, ETD identified 60% more phosphopeptides than CID, with an average of 40% more fragment ions that facilitated localization of phosphorylation sites, and the two modes were run alternately for more comprehensive analysis; that study reported 1,435 phosphorylation sites, of which approximately 80% had not previously been described [12].
The comparison is not settled by activation type alone. Using 180 individually synthesized phosphopeptides with precisely known phosphorylation sites, the highest number of correctly assigned phosphopeptides at 1% false localization rate was achieved by higher energy collision induced dissociation in combination with an Orbitrap mass analyzer, followed very closely by low resolution ion trap spectra obtained after electron transfer dissociation, and both of those were significantly better than low resolution spectra acquired after collision induced dissociation and multi-stage activation [13]. Analyzer and activation method set the localization ceiling together, not activation alone.
Combining the two activation types in a single event is a further step. Assessed against methods employing either ETD or HCD alone, using a defined synthetic phosphopeptide mixture and a larger data set of Ti(IV)-IMAC enriched phosphopeptides from a tryptic human cell line digest, and scored with a modified version of phosphoRS, EThcD generated richer and more confidently identified spectra in the majority of cases, resulting in superior phosphosite localization scores [14].
Localization scores, and what they are calibrated against
A localization score answers a narrower question than a database search does: given that this peptide sequence is correct, which residue carries the modification? The Ascore measures the probability of correct phosphorylation site localization based on the presence and intensity of site-determining ions in MS/MS spectra, the fragments that exist for one positional isoform and not the other [15]. Applied in a fully automated fashion to nocodazole-arrested HeLa cell lysate, the workflow reported in that paper returned 1,761 nonredundant phosphorylation sites from 491 proteins with a peptide false-positive rate of 1.3% [15]. That 1.3% is a peptide-level error rate determined by the target-decoy approach. It is not a statement about how many of those sites sit on the right residue.
The probability-based tools are explicit about needing calibration. phosphoRS calculates site probabilities for all potential phosphorylation sites in a peptide, with scoring parameters optimized on tandem mass spectra from synthetic phosphopeptides across all commonly used fragmentation techniques, and probability calculation adapted both to the fragmentation method and to the maximum mass deviation of the analysis. Validation on mixtures of synthetic phosphopeptides produced a false localization rate versus site probability cutoff characteristic in which the empirically obtained precision was higher than the applied probability cutoff indicated [16]. The Mascot Delta Score was likewise calibrated per fragmentation method, with thresholds taken from simple calibration functions that were stable over replicate analyses of the same phosphopeptide set; it outperformed the Ascore for tyrosine phosphorylated peptides, and the ability to call sites correctly increased with increasing distance of two candidate sites within a peptide sequence [13]. Two serines one residue apart are a harder problem than two serines eight residues apart carrying the same score.
The calibration is the weak joint. These scores have generally been calibrated using synthetic datasets, and their statistical reliability on real datasets is largely unknown, which risks studies reporting incorrectly localized phosphosites through inadequate statistical control. One route to an independent estimate is to score modifications on a decoy amino acid, one that cannot be modified, and read the global false localization rate off the decoy assignments. Across synthetic and real data sets the selection of that amino acid can make a substantial difference to the estimated global FLR, and the most reliable FLR results were achieved using alanine and leucine [17].
Three failure modes look identical in a results table. A site can be absent because it never bound the resin or never eluted; absent because the precursor was never selected for fragmentation; or present but placed on the wrong residue because the spectrum carried no site-determining ion. Loading buffer and capacity address the first, fractionation and acquisition scheme the second, activation method and an explicit FLR the third. State which one you controlled.