Peptide Retention Time Prediction in Hydrophilic ... - ACS Publications

Subscriber access provided by UB + Fachbibliothek Chemie | (FU-Bibliothekssystem)

Article

Peptide retention time prediction in hydrophilic interaction liquid chromatography: data collection methods and features of additive and sequence-specific models Oleg V. Krokhin, Peyman Ezzati, and Vic Spicer Anal. Chem., Just Accepted Manuscript • Publication Date (Web): 21 Apr 2017 Downloaded from http://pubs.acs.org on April 25, 2017

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

Analytical Chemistry is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Analytical Chemistry

Peptide retention time prediction in hydrophilic interaction liquid chromatography: data collection methods and features of additive and sequence-specific models Oleg V. Krokhin1,2*, Peyman Ezzati1, Vic Spicer1 1

Manitoba Centre for Proteomics and Systems Biology, University of Manitoba, 799 JBRC, 715

McDermot Avenue, Winnipeg, Manitoba R3E 3P4, Canada 2

Department of Internal Medicine, University of Manitoba, 799 JBRC, 715 McDermot avenue,

Winnipeg, Manitoba R3E 3P4, Canada

* - corresponding author Fax: (204) 480 1362, E-mail: [email protected]

ACS Paragon Plus Environment

1


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 2 of 22

ABSTRACT The development of a peptide retention prediction model for hydrophilic interaction liquid chromatography (XBridge Amide column) is described for a collection of ~40,000 tryptic peptides. Off-line 2D LC-MS/MS analysis (HILIC-RPLC) of S. cerevisiae whole cell lysate has been used to acquire retention information for a HILIC separation. The large size of the optimization dataset (more than two orders of magnitude compared to previous reports) permits the accurate assignment of hydrophilic retention coefficients of individual amino acids, establishing both, the effects of amino-acid position relative to peptide termini and the influence of peptide secondary structure in HILIC. The accuracy of a simple additive model with peptide length correction (R2-value of ~0.96) was found much higher compared to an algorithm of similar complexity applied to RPLC (~0.91). This indicates significantly smaller influence of peptide secondary structure and interactions with counter ions in HILIC. Nevertheless, sequence-specific features were found. Helical peptides that tend to retain stronger than predicted in RPLC exhibit negative prediction errors using an additive HILIC model. N-cap helix stabilizing motifs, which increase retention of amphipathic helical peptides in RP, reduce peptide retention in HILIC independently of peptide amphipathicity. Peptides carrying multiple Pro and Gly (residues with lowest helical propensity) retain stronger than predicted. We conclude that involvement of peptide backbone’s carbonyl and amide groups in hydrogen-bond stabilization of helical structures is a major factor, which determines sequence-dependent behavior of peptides in HILIC. The incorporation of observed sequence dependent features into our Sequence-Specific Retention Calculator HILIC model resulted in 0.98 R2-value correlation, significantly exceeding the predictive performance of all RP and HILIC models developed for complex mixtures of tryptic peptides.


2

Page 3 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


INTRODUCTION Rapid developments in the field of proteomics have fuelled recent progress in separation science of bio molecules, especially peptides. The vast majority of proteomic analyses nowadays are performed using the bottom-up approach: proteins of interest are digested with specific enzymes, resulting peptides are separated by high performance liquid chromatography (HPLC) or capillary zone electrophoresis (CZE) and analysed by mass spectrometry.1 This has brought more attention to accelerating the development of peptide separation techniques. These developments can be roughly divided into technological improvements of separation hardware, hyphenation techniques (with MS), and studies targeting better understanding of separation mechanisms. The former developments include ongoing improvements in peptide separation efficiency and selectivity, with the wide adoption of UPLC, core-shell sorbents, and introduction of stationary phases with unique selectivity being best examples. The latter targets the development of peptide retention prediction algorithms,2-4 which have found wide applications in improving the confidence of LCMS identification,5 and guiding method development for quantitative LC-MS analyses6 and multidimensional LC-MS protocols.7 Our laboratory has long standing interest in studying the mechanism of peptide reversedphase chromatography. Applying retention modeling to large retention datasets acquired using proteomic techniques has provided a significant impact. In our early work we used a dataset of 346 tryptic peptides to establish the effect of ion-pairing formation on apparent hydrophobicity of N-terminal residues.8 One of the latest discoveries – the description of N-cap stabilization of amphipathic helical peptides on C18 surface - required a collection of ~280,000 peptides with accurately measured retention properties.9 The incorporation of features describing peptide amphipathic helicity into our Sequence-Specific Retention Calculator (SSRCalc) is still ongoing: current SSRCalc database of ~1.5 million tryptic peptides separated using C18 – formic acid conditions provides solid support for it. Recently we applied the SSRCalc approach to CZE data using a large collection of peptides (~4,400) identified using CZE-MS/MS; all previously described models used less than 130 peptides. This thirty-fold increase in dataset size resulted in the discovery of novel sequence specific features that affect peptide electrophoretic mobility and yielded a significant improvement in model accuracy (~0.995 R2-value).10 We believe that this trend of using larger


3


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 4 of 22

MS-acquired datasets is applicable to any peptide separation technique. Similar to RPLC and CZE, the increase in the size of the data for any peptide separation mode will shed a new light on details of separation mechanisms and lead to significantly more accurate retention modeling. Hydrophilic interaction liquid chromatography (HILIC) is one of the most popular peptide separation techniques in proteomics.11,12 HILIC separation carries characteristics of other major HPLC techniques: normal-phase, reversed-phase and ion-exchange.13 Its history started back in 1970s,14 and the HILIC acronym was introduced by Alpert,15 who also studied this separation technique in great detail. HILIC provides separation efficiency comparable to RPHPLC along with unique separation selectivity, allowing the separation of hydrophilic analytes that aren’t retained in RP systems.16 Not surprisingly, it drew a lot of attention of peptide separation specialists upon the arrival of proteomic bottom-up techniques.11 Significant efforts have been applied to establish an optimal combination of stationary/mobile phases to improve both efficiency of separation and sensitivity of ESI MS detection.17,18 Additionally, HILIC exhibits separation selectivity sufficiently orthogonal to RP to prompt its use in multidimensional separation schemes.12,19 Wider application of HILIC separation mode in proteomics called for the development of HILIC peptide retention prediction models – similar to RP-HPLC in the early 2000s. Yoshida developed the first additive model to predict peptide retention on TSK gel Amide-80 column using a 121 synthetic peptide dataset (2-54 residues long, 9.3 residue length on average) with ~0.94 R2-value prediction accuracy.20 Gilar et al. used retention datasets of ~150 tryptic peptides to study the contribution of individual residues in peptide retention on 3 different HILIC columns at different pH values by optimizing additive prediction models with a logarithmic length correction.21 The accuracies of resulting models varied between 0.92 and 0.97 R2-value. Harscoat-Schiavo et al.22 separated 58 synthetic peptides (2-11 residues, 4.6 on average) on a TSK Gel Amide 80 column and used retention times to build their model based on amino acid composition (0.97 correlation). Le Maux et al.23 used a dataset of 153 peptides (2-4 amino acids long) with a resulting 0.992 R2-value correlation; to achieve this accuracy the authors introduced separate sets for retention coefficients for N-, C-terminal and internal residues (i.e. the model had 60 parameters). Badgett et al.24 used a HILIC-ESI MS of tryptic digest of 8 purified proteins to generate a dataset of 118 peptides for model optimization; they selected peptides shorter than 15 residues and applied an additive model with correction for N-terminal


4

Page 5 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


positions of six residues. They found that the correlation of 0.96 “is on the higher end of previous RP and HILIC peptide retention prediction models”. This review of the literature on peptide retention prediction in HILIC shows that datasets of ~150 peptides or less were used in all of them. According to ours25 (and others’4) observations, this count is very close to the lower limit that provides sufficient information to accurately derive retention coefficients of individual amino acids in additive (20 parameter) models without being subject to model over-fitting. The accuracy of resulting HILIC algorithms varied from 0.92 to 0.97. This also could be a consequence of small size of the datasets and significant differences in the length of peptides. The superior correlation demonstrated by Le Maux et al. (0.992 R2value)23 serves as the best example of that. Very small size of peptides (2-4 residues) and subsequent absence of any secondary structure helped achieve this high prediction accuracy. Taken together, these studies led us to the conclusion that a significant increase in the size and complexity of optimization dataset for HILIC separation may help in solidifying assignment of retention coefficients of individual amino acids and finding novel sequence-specific features of HILIC separation. All previously reported HILIC models used retention times accurately determined using ESI MS21-24 or UV20 detection. In our laboratory we often apply a simplified approach based on fraction collection and subsequent analysis of these fractions by either MALDI MS8 or LC-ESI MS (as a second dimension separation).26 Our first version of reversed-phase SSRCalc model was based on RP-HPLC – MALDI MS analysis of a 17 protein digest.8 We used only 40 fractions, with most of the 346 peptides found within a 25 fraction wide window. Nevertheless, it was sufficient to establish sequence-specific effect of ion-pairing at peptide N-terminus. Since this study, the throughput of mass analyzers improved significantly. Application of 2D LC-MS with 30-40 fractions in the first dimension and 60-90 min runs in the second dimension can generate tens of thousands of unique peptide identifications – hundreds times more than used in any of the previous reports on modeling HILIC separation of peptides. The goal of this study was to develop approach for high-content retention data collection in HILIC (Waters XBridge Amide) mode using 2D LC-MS (HILIC-RP, ~40 fractions in the first dimension) of a complex tryptic digest. The resulting retention data would provide a solid background for confident assignment of retention (hydrophilicity) coefficients of individual amino acids through the development of additive retention prediction model and help in


5


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 6 of 22

establishing sequence-specific features of HILIC separation. So far these features have not been sufficiently described due to small number of peptides used in previous studies. We anticipate that the fine details of separation mechanisms discovered in this work will be applicable to other HILIC separation conditions (columns) we plan to study.

EXPERIMENTAL Materials and digest preparation. Deionized water and HPLC-grade acetonitrile were used for preparation of the eluents. All chemicals were sourced from Sigma Chemicals (St. Louis, MO) unless otherwise noted. Sequencing grade modified trypsin (Promega, Madison, WI) and 15mL Amicon centrifugal filter units (Merck Millipore, Ireland) were used for the digestion. Siliconized 1.5 mL tubes (BioPlas, San Rafael, CA) were used to handle the fraction collection. The custom designed standard peptides P1-P627 were synthesized by BioSynthesis Inc. (Lewisville, TX). A tryptic digest of S. cerevisiae was prepared using the FASP protocol scaled up for 15mL centrifugal filter units.28 The digest (~1 mg of peptides) was acidified with trifluoroacetic acid, purified by reversed-phase C18 chromatography, then aliquoted into vials with approximately 100 µg (according to NanoDrop 2000 (ThermoFisher Scientific)) of digest in each, and finally lyophilized. A tryptic digest of bovine serum albumin was used for the initial selection separation conditions (determining the optimal gradient slope). This digest was prepared using a standard in-solution digestion of 2 mg/ml solution of BSA: reduction / alkylation with iodoaceatmide (with quenching), followed by trypsin digestion. The resulting digest was purified by RPLC, aliquoted (50 µg each vial) and lyophilized.

First dimension separation conditions. An Agilent 1100 series HPLC system with UV detector (214 nm) and 50 uL injection loop was used for HILIC and RPLC separations. A 3 mm x 50 mm XBridge Amide 3.5µm column (Waters, Milford, MA) was used with 300 µL/min flow rate for HILIC separations. Both eluents: A (water) and B (9:1 acetonitrile:water) contained 10 mM ammonium formate pH 4.5). These were prepared by 1:10 dilution of 100 mM ammonium formate pH 4.5 with water and acetonitrile, respectively. Optimized separation conditions to fit a ~40 min separation window used a 0.7% per minute increase of water content (10 to 60 %). The gradient program consisted of following steps. Starting conditions: 100% eluent B (10% water). Linear decrease of B from 100 to 44.4 % (60% water) in 71.43 min. Linear portion of the


6

Page 7 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


gradient was followed by 5 min wash with 90% of eluent A and an equilibration step with 100% B. One-minute fractions were collected within the expected interval of peptide elution (8-46 minutes), lyophilized and dissolved in 30 µl of buffer A for the second dimension.

Second dimension LC-MS/MS. The 2D LC Ultra system (Eksigent, Dublin, CA) delivered buffers A and B through a 100µm x 200mm analytical column packed with 3 µm Luna C18(2) (Phenomenex, Torrance, CA) at 500nL/min flow rate. Approximately 1/3 of each collected fraction (10 µL) was spiked with standard P1-P6 peptides (~200 fmole per injection) and was loaded on a 300 µm x 5 mm PepMap 100 trap-column (ThermoFisher). The gradient program consisted of following steps: a linear increase from 0.4 to 31% buffer B (acetonitrile) in 77 minutes, 5 minutes at 80% B and then 8 minutes at 0.4% B for column equilibration (90 min total analysis time). Both eluents A (water) and B (acetonitrile) contained 0.1% formic acid. Data-dependent acquisition using a TripleTOF5600 mass spectrometer (Sciex, Concord, ON) in standard MS/MS mode was used. The following settings were applied: 250 ms survey MS spectra (m/z 370-1500) followed by up to 20 MS/MS measurements on the most intense parent ions (300 counts/sec threshold, +2 through +5 charge state, m/z 100-1500 mass range for MS/MS, 100 ms each). Previously targeted parents ions were excluded for 12 seconds from repetitive MS/MS acquisition.

Data Analysis and retention time assignment. Raw spectra files were converted to Mascot Generic Format files for protein/peptide identification by the X!Tandem algorithm. The following search parameters were applied: 20 ppm and 50 ppm mass tolerance for parent and daughter ions, respectively; constant modification of Cys with iodoacetamide; the list of potential modifications included: oxidation of Met, Trp; N-terminal cyclization at Gln and Cys, N-terminal acetylation, deamidation (Asn, Gln). All non-modified tryptic peptides with Log (e) < -1 confidence score were considered for modeling. All identifications with low confidence values -3 < Log (e) < -1 were additionally filtered using retention time prediction. The existing SSRCalc retention prediction model for formic acid conditions, along with stored retention values (Hydrophobicity Index, HI) for yeast peptides from SSRCalc Database, and a first rough approximation of HILIC model were used for the peptide retention filtering in both dimensions.


7


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 8 of 22

Retention times of peptides in the second dimension were assigned as the time of acquisition of the most intense MS/MS spectra, and converted into HI (% acetonitrile) units using the established retention values of the standard peptides.27 Retention times in the first (HILIC) dimension were assigned as being equal to the fraction number in which this peptide was found. When the peptide signal was distributed between two or more fractions, an intensity weighted average fraction number was used.

RESULTS AND DISCUSSION Selection of chromatographic conditions (gradient slope) in HILIC mode. A review of the literature showed that linear acetonitrile:water (10-50% water) gradients are typically used for HILIC separation of peptides,20-24 while pH of the eluent varies. Our preliminary experiments with separation of synthetic peptides (not shown here) with formic acidic based eluents showed significant peak tailing; these observations were supported by a number of previous publication.17,23 Subsequently, we selected 10 mM ammonium formate with pH 4.5 as an eluent additive. It showed good peak shape and reproducible separation over time. Another important parameter to consider is a gradient slope: it has to provide sufficient peptide separation to fit the expected ~40 minute (fractions) elution window. Figure 1 shows the optimization of experimental gradient slope for HILIC separation using a BSA tryptic digest, chosen to represent typical collection of tryptic peptides. HILIC separation with 1% per minute increase of water starting with 10% showed that majority of tryptic peptides from BSA elute between 10 and 30 minutes (Figure 1A). Thus the gradient slope was adjusted to 0.7% per minute, which gave the desired length of the separation window (Figure 1B). Separation of the ~100 µg of yeast tryptic digest is shown in Figure 1C. As expected, we did not observe wellseparated high intensity peaks, as this mixture contains thousands of peptide species of greatly varying abundance.


8

Page 9 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 1. Selection of the gradient slope for HILIC separation and fraction collection for 2D-LC MS/MS analysis of yeast digest. A and B – HILIC separation of the BSA digest using 1 and 0.7 % water per minute gradients (10 mM ammonium formate, pH 4.5), respectively; C – HILIC separation of S.cerevisiae whole cell digest using 0.7 % gradient.

LC-MS/MS analysis in the second dimension: identification output. One third of each fraction (< 1 µg of peptides) was injected in the second dimension LC-MS/MS analysis. The 38 fractions were analyzed (8-46 min separation window in Figure 1C), corresponding to 57 hours of instrument time. This resulted in the acquisition of 389,917 MS/MS spectra, identification of 207,357 spectra corresponding to 44,489 unique peptides (Log (e) < -1) and 4218 proteins (Log (e) < -3). This identification output is similar to our standard 2D LC-MS/MS protocol with high pH – low pH (RP-RP)26 separation scheme with fractions concatenation, both showing a ~55% MS/MS identification rate (Table 1, Figure S-1). Table 1. Identification output of 2D (HILIC-RP) LC-MS/MS and 2D (RP-RP)-LC MS/MS for the analysis of whole cell yeast tryptic digest. Separation mode

Number of fractions

Total LCMS time (hr)

Amount injected (µg)

# of MS/MS

# of identified peptides

HILIC-RP

38

57

~30

389917

207357

# of nonredundant peptide IDs 44489

# of protein IDs 4218

RP-RP* 20 30 ~30 226386 126705 34621 4093 * - a standard 2D LC-MS/MS (high pH – low pH) with fraction concatenation applied in our lab.26


9


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 10 of 22

Retention time prediction filtering. Using high quality retention data is a key to success in developing retention prediction models. Application of synthetic peptides or digests of purified proteins with known sequence is the preferable option, but is time/cost prohibitive when larger datasets are required. Our experience shows that 2D LC-MS/MS analysis of complex digests with retention time prediction filtering in both dimensions represents a compromise between quality and size of retention dataset.26 The steps we apply for dealing with separations with novel selectivity in one of the separation dimensions are shown in Figure S-2. 0.7% (297 peptides) of all identifications were excluded. The remaining population of 40,290 species was used for model optimization. Tryptic peptides in this dataset were 6-51 residues long (14.1 on average).

Optimization and major features of additive HILIC model. Due to the specific chemical heteropolymeric nature of peptides, all peptide retention prediction models have a major components based on the summation of retention coefficients (RC) of individual amino acids.29, 30 These values are usually determined through linear multiple regression analysis with the goal of maximizing the correlation between predicted and observed retention values across all peptides. A 20-parameter model should be supported by a sufficiently large retention dataset to avoid overfitting. We have found empirically that ~100 peptides are needed to ensure a reliable assignment of retention coefficients in additive models: i.e a ~5:1 points to parameters count ratio.25 As seen from our overview of the literature, most of the prior models barely meet this requirement. Some of the models introduce separate retention coefficients for terminal positions of the residues-- this further increases the number of required data points. Additionally, all residues should be well represented in the training set sequences for confident assignment of RC; this can be achieved easily using synthetic peptides. Using real protein digests predetermines the representation of each residue according to the natural abundance of amino acids. Therefore specific care should be taken to ensure that a sufficient number of "rare" amino acids (Trp, Cys) are present. The predictive accuracy for our additive HILIC model with peptide length correction was 2

0.96 R -value (Figure 2). This is significantly better compared to additive RP models (~0.91) for the datasets of the same complexity, indicating a significantly simpler separation mechanism in HILIC.


10

Page 11 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 2. Workflow for optimization of the SSRCalc HILIC model.

Sequence-specific features of HILIC separation. There are several features defining the separation mechanisms specific to peptides. We observed these during the development of SSRCalc RP,2 and believe that similar corrections may be relevant to any peptide retention prediction model: i) distance of the residue from N- or C-termini; ii) propensity to form helical structures; iii) clustering of hydrophobic or hydrophilic residues within peptide sequence; iv) nearest neighbour effects of basic/acidic residues.

Position dependent retention coefficients. The influence of amino acid position relative to the peptide ends is a unique feature of peptide separation modeling. Terminal positions are more accessible for the interaction with stationary phase, and should exhibit higher retention coefficients for hydrophobic and hydrophilic residues in RP and HILIC modes, respectively. The first sequence-specific feature of RP separation we established during SSRCalc development was related, however, to ion-pairing at positively charged peptide N-termini.8 It showed an effect opposite to what one might expect – terminal residues exhibited lower hydrophobic contributions compared to internal ones. The hydrophilic counter-ions such as formate shield and reduce apparent hydrophobicity of the most hydrophobic residues when they are located closer to Nterminus; this effect decreases with the distance from a peptide end. Therefore all SSRCalc


11


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 12 of 22

models have separate sets of retention coefficients for all 20 amino acids across the first five terminal positions, plus an internal position. This automatically increases the number of retention coefficients to 220, and requires significantly larger datasets. Using 40,000+ peptides is sufficient for accurate modeling: e.g. the least abundant amino acids (His, Cys and Trp) occupy the Nterminal position in our dataset 721, 411 and 408 times, respectively. Optimization of separate retention coefficients for each of the 11 positions relative to peptide termini (Table S-1) further improved correlation to 0.976 (Figure 2). It was interesting to investigate the correlation between RC (internal retention coefficients) and RN1, RN2 (first and second N-terminal) observed for both the HILIC and RP separations. Figure 3 B-E shows that the effect of formate counter ions is much more subtle in HILIC. The correlation coefficients RC vs. RN1 and RN2 were found 0.95 and 0.98, respectively (Figure 3 B,C). The same correlations for RP showed 0.865 and 0.95 – indicating the much greater effect of counter ions (Figure 3 D,E).

Figure 3. Retention coefficients in HILIC and RP separation modes. A – comparison of internal (RC) retention coefficients for HILIC (pH 4.5) and RP (C18, formic acid) conditions. B, C – correlations between internal RC and first two terminal retention coefficients (RN1 and RN2) in HILIC; D, E – the same correlations applied to RPLC.


12

Page 13 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Table S-1 shows both the slope and correlation values in a similar fashion for all positiondependent coefficients. This provides additional insight into the influence of terminal positioning in HILIC. All positions, except for N1, show slightly higher slopes (i.e. higher contribution to retention) compared to internal RC. The slope gets closer to 1.0 and the correlations improve as we move deeper into the peptide sequence, with negligible differences from RCs by position #5. Also the values of N-terminal and C-terminal coefficients behave similarly, indicating that the orientation of the peptide relative to the sorbent surface does not play important role in HILIC. Alpert et al.31 showed that peptide orientation is important in ion-exchange chromatography-- the positively charged peptide N-terminus comes in contact with negatively charged cation-exchange surface first, reducing the contribution of C-terminal groups to overall peptide retention. In our case, the neutral character of the surface of the amide bonded phase leads to no effect from peptide orientation. It is interesting to note that the effect of formate counter ions in HILIC is similar to RP, but has smaller amplitude and of opposite sign for the major contributors. Compact, hydrophobic amino acids (Leu, Ile, Val, Met, Ala) are affected the most in RP (labeled in Figure 3 D, E) – a decrease in hydrophobicity compared to internal position. Smaller size of these residues promotes the shielding effect of hydrophilic counter ions. Conversely, bulky hydrohobes (Trp, Phe, Tyr) are less affected —positioned above the trend line (Figure 3 D, E). The situation is opposite in HILIC, the same compact and bulky hydrophobic residues are located slightly above and below the correlation plot (Figure 3 B, C), respectively. Some relatively hydrophilic residues (Asn is shown) increase their hydrophobic contribution at the N-terminal position in RP. Conversely, the same residues exhibit decreased HILIC retention when placed in the N-terminal position. All together this suggests that the effect of counter ions on N-terminal residues is of similar nature for RPLC and HILIC. However the direction and amplitude of this effect is dependent on hydrophobicity of counter ion relative to stationary phase and size of N-terminal residues. Similar to other studies,20-22 retention coefficients in HILIC show anti-correlation compared to RCs in RPLC (Figure 3 A). The largest contributors to peptide retention at pH 4.5 on neutral HILIC matrix are the charged residues: Lys, Arg, His (+) and Asp, Glu (-).

The effect of peptide helical propensity is the major sequence-dependent feature, which defines anomalous behavior of some peptides in HILIC. Table 2 shows several examples of extreme


13


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 14 of 22

deviations from the model, and provides possible explanations of such behavior. Our descriptions of possible mechanisms of interaction of helical peptides in HILIC are supported by a comparison of the behavior of the same species in the RP system (Table 2), a computation of peptide helical propensity using AGADIR,32 and interpretation of their axial helical projections (Table S-2). Overall, the effect of peptide helicity in HILIC is smaller in magnitude and opposite compared to RP: higher helical propensity leads to reduced retention. The largest negative deviations were observed for peptides carrying multiple N-cap sequences and helical peptides with multiple Ala residues. The former contain -(N/G/S/T/D)P- or -(N/G/S/T/D)XX(E/Q/D) motifs, but do not exhibit amphipathic character (Tables 2 and S-2). It is known that specific hydrogen bonding patterns make these motifs a signal sequence for helix nucleation both in proteins33 and peptides.32 (N/G/S/T/D) represent preferential N-cap residues, while Pro and (E/Q/D) tend to stabilize helix when placed in the first and third position from Ncap. Inside of the helix stabilization is achieved by hydrogen bond interactions between amide and carbonyl oxygen, separated by one helical turn. This excludes these groups from possible hydrophilic interaction with the HILIC phase, leading to reduced retention. The latter group has a well-defined hydrophobic surface (but not amphipathic: YATASAIAATAVASLVLAR) and contains multiple Ala, which is known for its highest propensity to be in helical conformations. This stabilizes the helical structure and excludes backbone’s polar groups from interaction. Amphipathic peptides with highest helical propensity tend to exhibit extremely high prediction errors in RP.9 For example, TIAETLAEELINAAK is amphipathic (Table S-2) and starts with a characteristic N-cap box motif TIAE-. All these features explain higher than predicted retention in RP mode (~6.7 % acetonitrile). The corresponding negative prediction error in HILIC is only moderate: -2.9% of water – similar to other examples of this type in Table 2. The reason for this discrepancy could be interplay between interaction of hydrophilic face of the amphipathic molecules with HILIC surface and helix stabilization via hydrogen bonds, which drive retention in different directions. The former was suggested as a possible mechanism of interaction of amphipathic peptides with cation-exchange surface in HILIC/CEX mode. Mant et al.34 demonstrated increased interaction of synthetic amphipathic peptides with five positively charged residues in hydrophilic face of the helix with cation-exchange sorbent. Such kind of interactions will be much weaker for neutral HILIC phase. Nevertheless, this should offset the influence of peptide helicity, which reduces peptide retention.


14

Page 15 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Our findings were partially supported by calculation of in-solution peptide helical propensity using the online AGADIR algorithm.32 Table 2 shows high helicity values for amphipathic and Ala-rich molecules and low values for peptides carrying N-cap motifs but with no amphipathic features (all featuring negative prediction errors in HILIC). Peptides showing high positive prediction errors have low helical propensity (Table 2). But plotting a correlation between AGADIR helicity and the prediction error of the model across the whole dataset did not show a discernable correlation that could be directly incorporated into the model (not shown here). This shows once again that calculation of peptide helicity upon interaction with chromatographic stationary phases (both hydrophobic and hydrophilic) is more complicated than calculations done for peptides in solution.32

Peptides with large positive prediction errors are characterized by their low helical propensity and show no correlation with prediction errors in RPLC (Table 2). The presence of multiple Pro and Gly residues leads to increased peptide retention. This observation is consistent with the low helical propensity of these residues, which keeps these peptides in a random-coil conformation and makes backbone polar groups accessible for interaction with the hydrophilic stationary phase. Clusters of hydrophobic residues lead to higher retention, too; this effect is opposite to what is observed in RPLC. We also observed that peptides with positive prediction errors often contained a large number basic amino acids. This can be explained by their additional interaction electrostatic with the residual silanol groups of the modified silica surface.


15


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 16 of 22

Table 2. Typical examples of peptides with largest deviations from HILIC prediction model and possible sequence-specific features, which explain such behavior. Peptide

HILIC prediction error (% RP prediction error (% Agadir helicity water) acetonitrile)* Negative prediction errors in HILIC Peptides with N-cap motifs (not amphipathic) SSLDGDPYR** -6.3 -0.51 0.01 SILQCNPLDPTNTTR -5.5 3.1 0.01 GFSGGPLDPR -5.1 1.6 0 SILNPYCVIDPR -5.1 1.8 0.01 VHYDPNGILNPYK -3.1 0.8 0.01 LVSPSDPTSYMK -3.0 1.2 0.03 Ala-rich helical peptides YATASAIAATAVASLVLAR -5.5 3.8 0.81 ANVADILVATAVAAR -3.6 3.4 0.82 Amphipathic helical peptides (extremely high retention in RPLC) NIKTIAETLAEELINAAK -3.8 8.9 3.41 TIAETLAEELINAAK -2.9 6.7 2.23 SSILETLVGR -2.9 3.4 1.04 MTVAHLIECLLSK -1.3 8.0 0.56 Positive prediction errors in HILIC Peptides with multiple Pro, Gly, hydrophobic clusters, multiple positively charged residues KFVFNPPKPR 3.8 1.2 0 KQIAFPQRK 3.8 0.6 0 GSNFQGSSRPPRR 3.8 0.7 0.01 GINKIPPKPR 3.4 0.5 0 DYVVEDGDIIYFR 3.3 -2.6 0.1 YGGVYVGTLSK 3.5 -0.5 0.01 KMSWAAIATPKPK 3.0 -1.2 0.15 SGGGNGGSGVAIR 2.5 -0.5 0.02 * RP-LC model of similar complexity (additive model with separate N- and C-terminal retention coefficients); ** N-cap motifs are underlined

Introducing helical features into SSRCalc modeling was approached in a simple format. The precise incorporation of helicity into prediction models is extremely challenging, and has yet to be completely addressed in RPLC mode-- which has been studied in much greater detail compared to HILIC. We applied simple corrections using an additive approach, optimizing weighting variables with the goal of increasing the predicted versus observed correlation across the training dataset. For example, the corrective algorithm counted the number NP, GP, SP, TP, DP motifs in the sequence, and added their respective weighting values to the overall peptide


16

Page 17 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


hydrophilicity (a negative contribution in this case). Similarly, corrections were introduced for clusters of hydrophobic residues, peptide charge, and multiple Pro and Gly instances. Resulting R2-correlation of the model improved slightly from 0.976 to ~0.980 (Figures 2 and 4A).

Expression of peptide hydrophilicity in HILIC separations. All steps of model optimization were performed using unitless hydrophilicity retention values (Figure 2). Practical implementation of retention prediction in HILIC would benefit from expression in more “user-friendly” format. Similar to RP-HPLC, which often uses acetonitrile percentage as indication of peptide hydrophobicity,27,30 we propose to use water percentage at which peptide is eluting from the HILIC column as a measure of peptide hydrophilicity – Hydrophilicity Index (HII). Retention times (fraction #) were converted into water % using the known values of delay time of this LC system (3.5 minutes at 300 uL/min) and experimental gradient slope. The unitless output of the predictive model was converted into HII units by introducing mapping slope and intercept values into the final reporting calculation (Figure 2).

Peptide retention prediction filtering in 2D LC (HILIC-RP) systems. Figure 4 shows application of peptide retention filtering in both dimensions to the whole dataset of non-modified S.cerevisiae peptides identified in our 2D LC-MS/MS analysis. The 297 peptides excluded during original data filtering are shown in red on both the HILIC and RPLC plots (Figure 4 A and B). These plots show another advantage of expressing peptide hydrophobicity/hydrophilicity in % or acetonitrile/water units. Slopes of resulting correlation plots are reciprocal to experimental gradient slopes settings 0.7% per minute in HILIC and 0.4% per minute in RP dimensions. Developing peptide retention/migration prediction models makes possible the in-silico evaluation of separation space coverage and uniformity of peptide distribution through it. Thus, using in-silico digest of S.cerevisiae (> 4 residues, 1 allowed missed cleavage) showed that all 441,664 peptides are amenable to separation in CZE,10 while 14.2% of peptides with HI < 0% acetonitrile will not be retained in RPLC system with formic acid as ion pairing modifier (Figure 4 C). A similar calculation for HILIC showed that only 0.27% of tryptic peptides have HII < 10% water and will not retain under the separation conditions used.


17


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 18 of 22

Figure 4. Peptide retention time prediction in 2D LC (HILIC-RPLC). A – SSRCalc HILIC peptide retention time prediction in first dimension; B – similar plot for RPLC (C18, formic acid). Correct identifications are shown in green for HILIC and blue for RP (40,290 peptides). Excluded false positives are shown in red; C – in-silico distribution of tryptic peptides from S.cerevisiae (more than 4 amino acid long, 1 missed cleavage allowed) across separation space in RPLC and HILIC. Peptide populations that will not retain under the separation conditions used are indicated in red.

CONCLUSIONS High-content proteomic analysis has allowed us to access the retention characteristics of ~40,000 tryptic peptides in HILIC using 58 hrs of instrument LC-MS time, to build a peptide retention prediction model (SSRCalc HILIC), and to characterize the major additive and sequence specific features driving the separation mechanism. Limiting the size of collected fractions (38 one minute fractions) led to some uncertainty in assigning retention times (≤30 sec). However, this still was sufficient for correct delineating major retention trends: average absolute retention prediction error of the final model was 0.57 min. Optimization dataset contained longer peptides (14.1 residues on average) and was more than 2 orders of magnitude larger than that used for all previously reported HILIC models. We applied our proven semi-empirical approach to model optimization, using previously acquired knowledge about peptide secondary structure effects in RPLC. This led to the development of the most accurate prediction model for separation of tryptic peptides reported to date (among both RPLC and HILIC). Higher predictive accuracy in HILIC was demonstrated previously only on a collection of peptides 2-4 amino acids long.23 The accuracy of current SSRCalc models for complex mixtures of tryptic peptides decreases in the order: CZE (0.995 R2-value) > HILIC (0.98) > RPLC (~0.965). We attribute this


18

Page 19 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


to the significantly more complicated mechanisms in sorbent-based separation techniques compared to CZE. We found that the contributions of peptide helicity and interactions with counter ions are smaller in HILIC when compared to RPLC. This resulted in higher prediction accuracy for SSRCalc HILIC, seen even in a first attempt to optimize the HILIC model. RPLC systems are the most studied, however the profound effects of amphipathic helicity on peptide retention still hamper efforts of separation scientists to build highly accurate predictive models. Overall, HILIC demonstrates opposite retention trends compared to RPLC. This is true for the retention coefficients of individual amino acids, peptide helicity, presence of hydrophobic clusters within peptide sequence, etc. At the same time, our observations indicate the mechanisms of interaction of helical peptides with hydrophobic (RP) and hydrophilic (HILIC) chromatographic supports are fundamentally different. Peptide amphipathic helicity and hydrophobic interactions of amino acid side chains with C18 sorbent dominate in RPLC. In HILIC, accessibility of hydrophilic amide and carbonyl groups on the peptide backbone plays an important role. Helical structures are stabilized by =C=O…HN= hydrogen bonds, thus excluding them from possible interactions with HILIC sorbent in helical peptides, reducing peptide retention. The drastic difference in the polarity of the mobile phase is another possible factor driving effects of peptide helicity in RPLC and HILIC—still has to be explored in greater detail.

Acknowledgements This work was supported by grant from the Natural Sciences and Engineering Research Council of Canada (RGPIN-2016- 05963; O.V.K.). The authors thank Dr. M. Gilar for providing HILIC columns. The authors also thank to Dr. D. Court and S. Shuvo for providing S.cerevisiae samples. The authors declare no competing financial interest.

Supporting Information ac-2017-005375-SI.pdf – Supporting Information containing: Figure S-1. Distribution of number of detectable features across fractions for 2D LC-MS acquisitions; Figure S-2. Workflow for retention data filtering and optimization of the model; Table S-1. Position dependent retention coefficients; Table S-2. Axial helical projections for peptides in Table 2.


19


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 20 of 22

REFERENCES 1. Bensimon, A.; Heck, A. J.; Aebersold, R. Annu. Rev. Biochem. 2012, 81, 379-405. 2. Krokhin, O. V. Anal. Chem. 2006, 78, 7785-7795. 3. Petritis, K.; Kangas, L. J.; Yan, B.; Monroe, M. E.; Strittmatter, E. F.; Qian, W. J.; Adkins, J. N.; Moore, R. J.; Xu, Y.; Lipton, M. S.; Camp, D. G., 2nd; Smith, R. D. Anal. Chem. 2006, 78, 5026-5039. 4. Moruz, L.; Tomazela, D.; Kall, L. J. Proteome Res. 2010, 9, 5209-5216. 5. Strittmatter, E. F.; Kangas, L. J.; Petritis, K.; Mottaz, H. M.; Anderson, G. A.; Shen, Y.; Jacobs, J. M.; Camp, D. G., 2nd; Smith, R. D. J. Proteome Res. 2004, 3, 760-769. 6. Lange, V.; Picotti, P.; Domon, B.; Aebersold, R. Mol. Syst. Biol. 2008, 4, 222. 7. Spicer, V.; Ezzati, P.; Neustaeter, H.; Beavis, R; Wilkins, J. A; Krokhin, O. V. Anal. Chem. 2016, 88, 2847-2855. 8. Krokhin, O. V.; Craig, R.; Spicer, V.; Ens, W.; Standing, K. G.; Beavis, R. C.; Wilkins, J. A. Mol. Cell. Proteomics 2004, 3, 908-919. 9. Spicer, V.; Lao, Y. W.; Shamshurin, D.; Ezzati, P.; Wilkins, J. A.; Krokhin, O. V. Anal. Chem. 2014, 86, 11498-11502. 10. Krokhin, O. V.; Anderson, G.; Spicer, V.; Sun, L.; Dovichi, N. J. Anal. Chem. 2017, 89, 2000-2008. 11. Boersema, P. J.; Mohammed, S.; Heck, A. J. Anal. Bioanal. Chem. 2008, 391, 151-159. 12. Di Palma, S.; Hennrich, M. L.; Heck, A. J.; Mohammed, S. J. Proteomics 2012, 75, 37913813. 13. Buszewski, B.; Noga, S. Anal. Bioanal. Chem. 2012, 402, 231-247. 14. Linden, J. C.; Lawhead, C.L. J. Chromatogr. 1975, 105, 125-133. 15. Alpert, A. J. J. Chromatogr. 1990, 499, 177-196. 16. Jandera, P. Anal. Chim. Acta 2011, 692, 1–25. 17. Yang, Y.; Boysen, R. I.; Hearn M. T. W. J. Chromatogr. A 2009, 1216, 5518–5524. 18. Simon, R.; Enjalbert, Q.; Biarc, J.; Lemoine, J.; Salvador, A. J. Chromatogr. A 2012, 1264, 31-39. 19. Gilar, M.; Olivova, P.; Daly, A. E.; Gebler, J. C. Anal. Chem. 2005, 77, 6426-6434. 20. Yoshida, T. J. Chromatogr. A 1998, 808, 105–112. 21. Gilar, M.; Jaworski, A. J. Chromatogr. A 2011, 1218, 8890-8896.


20

Page 21 of 22

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


22. Harscoat-Schiavo, C.; Nioi, C.; Ronat-Heit, E.; Paris, C.; Vanderesse, R.; Fournier, F.; Marc, I. Anal. Bioanal. Chem. 2012, 403, 1939–1949. 23. Le Maux, S.; Nongonierma, A. B.; Richard J. FitzGerald, R. J. Food Chem. 2015, 173, 847– 854. 24. Badgett, M. J.; Boyes, B.; Orlando, R. Chromatogr. Today 2015, Nov-Dec, 39-42. 25. Shamshurin, D.; Spicer, V.; Krokhin, O. V. J. Chromatogr. A, 2011, 1218, 6348– 6355. 26. Dwivedi, R. C.; Spicer, V.; Harder, M.; Antonovici, M.; Ens, W.; Standing, K. G.; Wilkins, J. A.; Krokhin, O. V. Anal. Chem. 2008, 80, 7036-7042. 27. Krokhin, O. V.; Spicer, V., Anal. Chem. 2009, 81, 9522-9530. 28. Wisniewski, J. R.; Zougman, A.; Nagaraj, N.; Mann, M. Nat. Methods 2009, 6, 359-362. 29. Meek, J. L. Proc. Natl. Acad. Sci. U.S.A. 1980, 77, 1632-1636. 30. Guo, D.; Mant, C. T.; Taneja, A. K.; Parker, J. M. R.; Hodges, R. S. J. Chromatogr. 1986, 359, 499-517. 31. Alpert, A. J.; Petritis. K.; Kangas, L.; Smith, R. D.; Mechtler, K.; Mitulovic, G; Mohammed, S.; Heck, A. J. Anal. Chem. 2010, 82, 5253–5259. 32. Lacroix, E.; Viguera, A. R.; Serrano, L. J. Mol. Biol. 1998, 284, 173-191. 33. Richardson, J. S.; Richardson, D. C. Science 1988, 240, 1648-1652. 34. Mant, C. T.; Litowski, J. R.; Hodges, R. S. J. Chromatogr. A, 1998, 816, 65–78.


21


1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Page 22 of 22

TOC graphics


22

Peptide Retention Time Prediction in Hydrophilic ... - ACS Publications

Recommend Documents