Multiplexed triage of candidate biomarkers in plasma using internal standard triggered-parallel reaction monitoring mass spectrometry

Despite advances in proteomic technologies, clinical translation of plasma biomarkers remains low, partly due to a major bottleneck between the discovery of candidate biomarkers and downstream costly clinical validation studies. Due to a dearth of multiplexable assays, generally only a few candidate biomarkers are tested, and the validation success rate is accordingly low. Here, we demonstrate the capability of internal standard triggered-parallel reaction monitoring (IS-PRM) to prioritize candidate biomarkers for validation studies. A 5,176-plex assay coupling immunodepletion and fractionation with IS-PRM was developed and implemented in human plasma to quantify peptides representing 1,314 breast cancer biomarker candidates. Compared to prior approaches using data-dependent analysis, IS-PRM showed improved sensitivity (912 vs 295 proteins quantified) and precision (CV 0.1 vs 0.27) enabling rank-ordering of candidate biomarkers for validation studies. The assay greatly expands capabilities for quantification of large numbers of proteins and is well suited for prioritization of viable candidate biomarkers.


Introduction
Blood plasma is an easily accessed biofluid that reflects the physiological state of a patient; thus, it remains an attractive source of clinical biomarkers 1,2 . Despite considerable investment and advances in liquid chromatography mass spectrometry-based (LC-MS/MS) proteomic technologies that allow for deep coverage and quantification of proteins 3,4 , translation of biomarker discoveries to clinical use remains slow, tedious, and generally disappointing 5,6 . A large factor contributing to this state of the field is the mismatch between the large number of potential biomarkers identified and the resources required for their validation. A method to prioritize amongst candidate biomarkers to identify those with the greatest probability of clinical utility would allow clinical validation efforts to focus on the subset of candidates most likely to succeed 7 .
The emergence of targeted mass spectrometry-based proteomics approaches (e.g., multiple reaction monitoring (MRM) and parallel reaction monitoring (PRM) [8][9][10] ) enables highly sensitive, specific, and multiplexable assays that can be implemented with relatively low cost (compared to traditional immunoassays) 11 . These approaches have been incorporated into pipelines for biomarker evaluation 7,[12][13][14] and have been applied to biomarker development efforts 15,16 . However, even with optimized parameters and careful attention to method details (e.g., tight retention time windows, elimination of overlapping interfering transitions, enrichment and/or fractionation for low abundance targets), it is a challenge to measure more than a few hundred proteins and maintain high analytical performance using these approaches [17][18][19][20] . Thus, there is a need for a method capable of measuring a thousand or more biomarker candidates in plasma with high sensitivity and specificity. Such an approach could be used in the biomarker development pipeline to triage a large list of candidates down to a smaller number that can be applied with more quantitative rigor at increased throughput.
The recent development of internal standard triggered-parallel reaction monitoring mass spectrometry 21 (IS-PRM-MS, or SureQuant) has allowed for high multiplexing with the benefits of the performance of PRM 22,23 . The IS-PRM method expands the capacity of the PRM method without relying on retention time windows or co-isolation of target peptides by instead relying on added internal standards to trigger the real-time measurement of endogenous peptides. Upon detection of the internal standard, quantification is performed by PRM, allowing for highly sensitive and specific measurements.
In this study, we evaluated the effectiveness of IS-PRM in the context of prioritizing candidate biomarkers for more costly validation studies. As a test case, we developed an IS-PRM method to quantify 5,176 peptides representing 1,314 candidate breast cancer plasma biomarker proteins identified using a novel strategy leveraging preclinical patient derived xenograft (PDX) mouse models. We hypothesized that the IS-PRM method could quantify these candidates in human plasma with high specificity and precision to enable the rank ordering of candidate biomarkers for further investment of resources to perform validation studies in large patient cohorts. The methodology developed herein presents a significant advance in reliable quantification and verification of large numbers of plasma-based biomarker candidates, and the approach is generally applicable to other diseases or translational studies requiring highly precise relative quantification of large sets of proteins.

Results
Overview. We evaluated the utility of a highly multiplexed IS-PRM assay for prioritizing a large list of candidate plasma biomarkers of breast cancer (identified in breast cancer PDX models; see assay quantifies endogenous ("light") peptide after first observing and identifying its cognate spiked-in isotope-labeled ("heavy") internal standard peptide. After a positive identification is confirmed, quantification is achieved by performing targeted PRM on the endogenous peptide. This method accomplishes high sensitivity and specificity with improved multiplexing (required to triage large numbers of biomarker candidates) by using fast MS scans for identification and maximizing the time devoted to quantitative scans. In addition, the inclusion lists employed by the method can survey for tens of thousands of target precursors making the method easier to implement because it does not require characterization and monitoring of retention time windows.
As a test case for employing IS-PRM in triaging biomarker candidates, we used the method to rank order a list of 1,314 candidate biomarker proteins (Supplementary Fig. 1) by quantifying the differences in expression in pooled plasma samples from women diagnosed with breast cancer vs plasma samples from women diagnosed with benign breast lesions. The 1,314 breast cancer biomarker candidates were identified as human proteins in the plasma of patient-derived xenograft (PDX)-bearing mouse breast cancer models, where proteins leaked, secreted, or shed from the transplanted human breast tumors were the exclusive source of human proteins in the plasma. Of note, 1,179 (90%) of the candidate biomarkers were previously observed in proteomic profiles of human breast cancers 24 .
Heavy stable isotope-labeled standard peptides (SIS) were synthesized for each target peptide (Supplementary Table 1) and used to determine the optimum precursor m/z for the IS-PRM inclusion list, the fragment ions to be used for peptide identification, and the intensity thresholds for triggering the identification scan. To optimize these parameters, a mixture of the 5,176 heavy peptides (~500 fmol) was spiked into 200 g of trypsin/LysC-proteolyzed yeast lysate. The mixture was fractionated by basic reverse phase (bRP) liquid chromatography into 12 fractions. Each fraction was then subjected to data dependent acquisition (DDA) LC-MS/MS using an inclusion list (i.e., directed DDA) containing m/z values for +2 and +3 charge states for each SIS peptide. Spectra matching the heavy peptide sequences were used to identify the six most intense fragment ions for each precursor, and to identify the precursor charge state that had the most intense sum of these six transitions. Using these empirical data, we set the intensity threshold for triggering the identification scan at 2% of the maximum height of the MS1 intensity of the precursor to maintain the highest sensitivity without triggering on noise. The precursor m/z and intensity thresholds are listed in Supplementary Table 2, and fragment ions used for identification and quantification are listed in Supplementary Table 3.
The analytical performance of the IS-PRM method was characterized using a response curve consisting of a ten-fold serial dilution of human cell (MCF10A) lysate into yeast lysate (100% MCF10A to 0.1%, blanks were prepared using 0% MCF10A). The MCF10A concentration levels corresponded to an approximate MCF10A cell count of 200,000 to 200 cells, respectively. Each concentration point underwent proteolytic digestion, addition of SIS peptides, and separation into 6 bRP fractions. Each fraction was analyzed by the IS-PRM method using triplicate injections on the LC-MS (Fig. 2a). For a peptide to be classified as quantified, we imposed the following requirements which had to be satisfied in at least two of the three replicates: (i) at least 4 transitions (light endogenous peptides) or 5 transitions (heavy SIS peptides) were present in the MS2 spectra, (ii) the ratio dot product of MS2 spectra from heavy and light peptides was > 0.98, (iii) at least 5 points across the peak were profiled in the chromatogram, and (iv) the peak area was > 5,000. All integrations were manually checked, and 93 (~2%) peptides had a fragment ion with interference in either the heavy or light peptide removed from the analysis.
Response curve results for the IS-PRM assay are summarized in Fig. 2 and data are provided in Supplementary Table 4. The IS-PRM method triggered the quantification of most of the light peptides ( Fig. 2b). Endogenous signals were quantified for nearly half of the peptides (n=2,443) at the highest concentration of MCF10A (Fig. 2b), resulting in quantification of 953 proteins (73%) of the targeted proteins (Fig. 2c). Decreasing the percentage of MCF10A cells resulted in the expected decrease in proteins quantified. The method exhibited excellent analytical precision, with a median coefficient of variation (CV) of 7.7% across all concentration points (Fig. 2d). To estimate the sensitivity of the method for detection of low abundance proteins, we used the number of proteins expressed per cell reported in Ly et. al. 29 . Fig. 2e shows a histogram of proteins detected by the IS-PRM method in each dilution point versus the number of proteins per cell. As expected, as the MCF10A cells were diluted, the histogram curve shifts to those proteins that were most abundant.
We next compared the sensitivity and precision of the IS-PRM method to directed DDA, which is also capable of highly multiplexed identification and quantification and which has been previously used to prioritize candidate biomarkers 7,30 . The target peptides were measured using both IS-PRM and directed DDA in a common trypsin/LysC-proteolyzed pool of human plasma depleted of abundant proteins and fractionated by bRP chromatography in the same manner as the yeast sample above (Fig. 3a). While both directed DDA and IS-PRM use an inclusion list to target precursors for MS2 fragmentation, the method of quantification differs considerably. Directed DDA uses the MS1 scans to create an extracted ion chromatogram for quantification, whereas IS-PRM quantifies the target peptide based on the intensities of selected fragment ions in the MS2 scan. The IS-PRM assay triggered quantification of more light peptides (4674 vs 3991) and exhibited better sensitivity (Fig. 3b) compared to directed DDA, as shown by quantifying 1683 endogenous peptides in IS-PRM versus 436 in directed DDA. The difference in peptide quantification rate was even more pronounced at the protein level (Fig. 3c), where IS-PRM quantified 912 proteins versus 295 for directed DDA. The precision of the approaches was estimated by calculating the CV from peak area ratios measured in the neighboring bRP fraction as technical replicates of the LC-MS/MS measurements. As expected, the median CV was lower in results from the IS-PRM analysis compared to the directed DDA analysis (0.10 versus 0.27; Fig. 3d). This is a result of improved signal-to-noise and reduction in interferences in using the MS2 signals for quantification in PRM compared to MS1 signals in directed DDA.
Evaluation of the IS-PRM method for highly multiplexed quantification of candidate biomarkers in human plasma. We next applied the 5,176-plex IS-PRM assay to quantify the biomarker candidates in 3 pools of human plasmas from women diagnosed with breast cancer and 3 pools of human plasmas from women diagnosed with benign breast lesions (Fig. 4a), with a goal of rank-ordering the 1,314 candidate biomarkers to identify those meriting further evaluation in larger, case-control validation studies. Each pool represented 19-20 women. Two hundred micrograms of each plasma pool underwent reduction, alkylation, and proteolytic digestion. The digested plasma pools were desalted, spiked with all 5,176 SIS peptides (~500 fmol) and fractionated into 24 fractions using bRP chromatography.
The IS-PRM method was applied to each of the 24 fractions (per pool), using a rigorous QC program to avoid any system degradation during the analysis (Supplementary Fig. 3). All peak integrations were manually reviewed, and interferences removed from 193 (~4%) peptides. Summed transition areas and number of transitions and points per peak are reported in Supplementary Table 5. In addition to the quantification criteria used for the response curve (above), we required endogenous signal to be >2x the signal from blank runs (Supplementary Table 6). On average, endogenous signals were measured for 1,708 (33%) of the target peptides ( Fig. 4b) across the pools, corresponding to 760 (58%) proteins ( Fig.   4c). The sum of all proteins quantified across the plasma pools was 893 (68%). Analytical reproducibility was determined using the endogenous measurements in the neighboring bRP fraction as a technical replicate (n=2). The overall distributions of the 893 candidate biomarker protein abundances across the six plasma pools, shown in Fig. 5a, varied widely. To determine if the candidate biomarker protein signals were higher in the cancer plasma pools, we tested the proteins for significant differences (p < 0.001) in each cancer pool compared to at least 2 of the confounding (i.e., benign) control plasma pools. To allow for variability in the control samples, we used a regression trend approach, which accounted for measurements that were lowest in the non-proliferative control, increasing in the proliferative and the atypia controls, and reaching a maximum in the cancer subtype sample (i.e., candidates whose plasma levels progressively increased as the biology of the breast lesions became more aggressive). An example of an individual protein featuring this trend is shown in Fig. 5b, the results for all proteins are reported in Supplementary   Table 7. Two peptides for PZP show consistent relative quantification (Fig. 5b), where the lowest measurement is seen in the non-proliferative control, followed by the proliferative and atypia controls, and finally the TNBC cancer subtype sample shows the highest levels. Overall, there were 162 candidate proteins showing significant differences in at least one of the three cancer subtypes (triple negative, HER2 positive and ER positive/HER2 negative), and 22 were significant in all three (Fig. 5c). The distribution of measured abundances for the 22 overlapping proteins (Fig. 5d) reflects an improvement in differentiating the cancer pools from control (compared to total proteins measured). Compared to a random sampling of 22 proteins, the candidates overlapping from all three subtypes show a better differentiation from controls (Fig. 5e).

Discussion
With the application of modern approaches, many hundreds of candidate protein biomarkers of disease are readily identified, but few of these candidates are ever carried forward to clinical validation studies 32,33 , and almost none are clinically translated. This chasm between biomarker discovery and validation is largely caused by a paucity of validated multiplexable assays for quantification of protein biomarker candidates in clinical validation studies, leading to a largely arbitrary selection of a few candidate biomarkers for which commercial assays are available for advancement to clinical studies. An empirical method to rank order large lists of candidate biomarkers to identify those meritorious enough to warrant an investment in assay development and clinical validation could improve success rates for translating novel biomarkers into clinical use.
We demonstrate that IS-PRM can be deployed on plasma samples to credential candidate plasma biomarkers for follow up validation studies. The method was capable of targeting >1,300 proteins in a highly reproducible manner, measuring >800 proteins in human plasma over several orders of magnitude with high specificity and sensitivity. Endogenous measurements across six human plasma pools included 200 proteins with reported plasma concentrations <1 ng/mL. The data demonstrate, perhaps not surprisingly [34][35][36] , that differences in protein expression levels between case and control pools were relatively small, highlighting the need for highly precise measurements (and perhaps multi-protein panels and longitudinal sampling of individual patients over time) 37 to provide clinical diagnoses. IS-PRM quantification showed excellent analytical precision in both the triplicate analysis of a response curve (median CV of 7.7%) and the analysis of neighboring fractions of the human plasma pools (median CV across fractions of 11.0%), proving the method is capable of high precision.
One challenge in measuring low abundance plasma proteins is the extensive sample preparation required. In this study we incorporated abundant plasma protein depletion and bRP fractionation, which limited the analytical throughput for analyzing large numbers of samples and necessitated the use of plasma pools instead of individual plasma samples. One limitation of this study design is that a single outlier patient in one plasma pool can skew the biomarker results from that pool, a situation that can be corrected by analyzing multiple independent pools, each of which includes multiple patients, and aggregating results. This workflow was able to triage the list of candidates to a number more amenable to workflows like multiplex immuno-MRM 11,38-40 , which can be used to support clinical validation studies in a high throughput manner for the most promising candidates.
Follow-up studies will be required to determine if the PDX model serves as a viable conduit for discovering clinically translatable biomarkers. Patient-derived xenografts of human cancer have emerged as powerful tools for clinical/translational science due to their recapitulation of many aspects of the biology of tumors derived from patients, including treatment responses, genomic mutation and copy number alterations, as well as RNA and protein expression [41][42][43] . This high degree of biological consistency with clinical samples may make PDX-bearing mice a potential discovery platform for identification of tumor-derived proteins in plasma since human sequences can be distinguished from mouse peptides using mass spectrometry 44,45 .
Regardless of the source of biomarker candidates, IS-PRM is well suited to the challenges posed by biomarker prioritization studies, and application of the method is generally applicable to any large-scale protein quantification study. Considerations for the sample complexity, dynamic range of targeted proteins, extent of sample preparation necessary, throughput required, and the availability of internal standards may dictate which proteomic method is most appropriate for protein quantification in each application. cleavages, oxidized methionine set as a variable modification, and carbamidomethylated cysteine set as a static modification. Peptide MH+ mass tolerances were set at 20 ppm. The overall FDR was set at ≤1%. Results from the search against the combined human/mouse databased allowed categorization of peptides into 3 classes: i. human-specific, ii. mouse-specific, and iii. ambiguous (mouse or human). All spectra that were categorized as human specific in this search that also returned an identification in the search against the mouse-only database were filtered out to ensure candidate biomarker proteins were human in origin.

Preparation of lysates for method characterization. Yeast cells (Saccharomyces cerevisiae) were
harvested and lysed using a previously described method 51  Identification was considered successful if the ratio dot product of the transition intensities between the heavy and light peptides was > 0.98. Quantification was considered successful if the PRM results contained peak areas from at least 4 transitions (light endogenous peptides) or 5 transitions (heavy SIS peptides), had at least 5 points across the peak and had a peak area greater than 5,000. All quantifications were manually checked and any one transition with interference in either the heavy or light peptide was removed from the analysis. Peptide concentrations are reported as the peak area ratio of the light and heavy peptides.
Verification of candidate biomarkers. Peptide peak area ratio (PAR) from the individual plasma pools were filtered to include only those that were greater than two-fold greater than the maximum PAR reported in the three yeast blank samples. For the three breast cancer subtypes, PAR were compared to that in the proliferative and non-proliferative control pools. A weighted z-score for each protein was derived based on joint evidence from multiple peptides of the protein. For each peptide, a z-score was calculated by standardizing the mean intensity difference between the cancer subtype pool and the 2 control pools with the empirical standard deviation of the peptide abundances across normal pools. If there were missing results in the control pools, the standard deviation across all pools was used and the peptide z-score was weighted by 0.5. The weighted z-score for a protein was calculated as weighted sum of z-scores of peptides mapping to the protein, approximated weighted z-scores by a normal distribution and p-values were obtained from a right tailed test. The proteins were screened for markers by fitting a regression model. Relative PAR for each protein was  into the mixture. In step 1, the peptide mixture is analyzed by liquid chromatography-mass spectrometry, using an inclusion list containing one precursor (peptide) and six transition (fragment) m/z associated with each heavy standard. An MS1 scan is performed to look for precursor m/z on the inclusion list. Once a heavy precursor m/z has been observed above a given intensity threshold, step 2 is initiated, where a low-resolution MS2 scan is triggered to identify transition fragment ions associated with the targeted heavy peptide. If 5 out of 6 transitions are observed, then a high-resolution MS2 scan is initiated for the light version of the peptide in step 3. Finally, in step 4, parallel reaction monitoring (PRM) using the high resolution MS2 scans of light and heavy peptides allows for conclusive identification of the peptide sequence and relative quantification of the endogenous "light" peptide, which is reported as the peak area ratio (PAR) of the light peptide intensity over the heavy peptide intensity. b Candidate protein biomarkers were identified by LC-MS/MS analyses of plasma from mice harboring human patient-derived xenografts (PDX) of breast cancer to identify human proteins secreted, leaked, or shed from tumors (5,498 unique human-specific peptides, mapping to 1,314 human proteins). As a test case for prioritizing the candidates for further up validation studies, the IS-PRM assay was used to interrogate differential expression in plasma from breast cancer patients versus benign controls. 138 human plasma samples were individually depleted of high-and mid-abundance proteins and combined accordingly to create cancer and control pools. The heavy isotope-labeled standard peptide mix was spiked into the digested pools prior to fractionation by basic reverse phase (bRP) liquid chromatography. Each of the 24 bRP fractions was subjected to the IS-PRM method using a Thermo Orbitrap Eclipse Tribrid mass spectrometer. The peak area ratio of endogenous light peptides relative to heavy spiked peptide was used to quantify the relative expression between pools.