Mass spectrometry‐based high‐throughput proteomics and its role in biomedical studies and systems biology

There are multiple reasons why the next generation of biological and medical studies require increasing numbers of samples. Biological systems are dynamic, and the effect of a perturbation depends on the genetic background and environment. As a consequence, many conditions need to be considered to reach generalizable conclusions. Moreover, human population and clinical studies only reach sufficient statistical power if conducted at scale and with precise measurement methods. Finally, many proteins remain without sufficient functional annotations, because they have not been systematically studied under a broad range of conditions. In this review, we discuss the latest technical developments in mass spectrometry (MS)‐based proteomics that facilitate large‐scale studies by fast and efficient chromatography, fast scanning mass spectrometers, data‐independent acquisition (DIA), and new software. We further highlight recent studies which demonstrate how high‐throughput (HT) proteomics can be applied to capture biological diversity, to annotate gene functions or to generate predictive and prognostic models for human diseases.

The response of a biological system to a perturbation (such as a knock-out, drug treatment or stress exposure) depends on many factors such as genetic background or environmental conditions. In order to make generalizable conclusions from protein expression changes, these factors need to be taken into account which requires the measurement of large sample sets consisting of many different conditions. (C) High-throughput proteomics allows new dimensions in gene function annotation: (1) Reverse proteome profiling for protein centric analysis: studying the behavior of proteins across many conditions (e.g., mutants, knock-outs) can provide functional information.
(2) Covariation analysis: s functionally related proteins are co-regulated, the covariation of proteins across many conditions can be used for functional assignment [18]. demands on sample numbers and measurement consistency over large sample sets or over multiple batches.
Recent developments have prompted a conceptual shift. When studying complex biological systems, many samples are required to reach sufficient statistical power ( Figure 1A). In clinical applications, samples are often heterogeneous and have high inter-individual variability [8][9][10][11]. Low sample size indeed might be one of the main reasons why biomarker studies often fail in validation [12]. Moreover, protein changes can be buffered, and changes can thus have low effect sizes [8]. Eventually, statistical tests on thousands of proteins at the same time require multiple testing corrections, many clustering methods struggle with multidimensional data, and machine learning models easily overfit, if the sample numbers are low in comparison to the number of variables the models are trained on. In these cases, large sample numbers with sufficient replicates, and precise measurements, help to alleviate these limitations.
There are other pressing needs which demand a scaling of biological experiments. Biological responses depend on a large number of genetic and environmental factors. Considering a limited space of conditions might lead to context specific results that are not generalizable ( Figure 1B). However, testing multiple conditions to the extent that it captures biological signals quickly scales the size of experiments, requiring high-throughput (HT) and cost-effective methods.
Finally, large-scale systematic analyses, that involve many conditions, could help to solve another major bottleneck of current biomedical research. Despite intense research efforts, a large fraction of genes remain without sufficient functional annotation, and research still overly concentrates on a subset of genes already known in the pre-genomic era [13][14][15]. This not only hampers a major goal of molecular biology-the explanation of phenotypes from genotypes-but also means that we miss potential targets for drug design and biotechnology. Studies in yeast have shown that combining genome-wide perturbation approaches with transcriptomics, proteomic, or metabolomic technologies constitutes a powerful tool for systematically capturing missing gene function [16][17][18].

HT PROTEOMICS EXPERIMENTS
Proteomics is increasingly facilitating biological experiments at scale.
The basic precondition is the necessity to achieve high sample throughput in a reasonable amount of time and at a low cost per sample. However, HT proteomics is not simply a question of running more samples,

WHAT MATTERS MORE, PROTEOMIC DEPTH, PRECISION, OR THROUGHPUT?
Often, large proteomics experiments require a compromise between proteomic depth on the one hand, and precision and throughput on the other hand. Obviously all three factors are important, but the individual importance of depth, precision, and throughput depends on the biological question. If one is ultimately interested in detecting certain low abundant proteins (e.g., several transcription factors, or extracellular signaling proteins such as cytokines are low abundant) depth is arguably a key variable in a proteomic experiment. However, the importance of depth decreases with sample size, because biological systems are organized in networks [22][23][24]. This means that biological responses involve concentration changes of many proteins [8,25,26] which not all need to be quantified to detect the response (Figure 2).
With the advent of multiparametric statistics and machine learning in biomedicine, combined with a better understanding of biological networks, this situation increasingly triggers a conceptual shift: the need for understanding biological systems that are organized in networks, makes it less crucial to quantify every single protein each time. Rather, it becomes more important to study many different conditions, and to learn how to interpret "signatures" that capture part of the response [7,27].

STUDY DESIGN
A HT proteomics experiment starts with an appropriate study design.

CHROMATOGRAPHY
Chromatography is an integral part of quantitative LC-MS-based workflows, and a key determinant of the throughput and longitudinal precision of a proteomics experiment. The gradient length as well as column wash and equilibration times define the total run time. In order to make short gradients applicable, chromatographic systems ideally operate with short overhead times (column wash and equilibration time). This has been achieved by increasing flow rates over the proteomics-typical nanoliter flow rate chromatography to capillary/micro-flow rates (5-50 µl/min) [49,60,[73][74][75][76] up to analytical flow rates (800 µl/min) [44,77]. Increased flow rates reduce dead times and solvent delay times and result in shorter wash and equilibration times. At the same time, higher flow rates result in increased robustness, spray stability, and less carry-over [44,60,73,75]. The main downside of high-flow chromatography is the higher sample dilution, requiring higher sample amounts per injection compared to lower flow rates. However, with instruments gaining in sensitivity, high flow rate chromatography is increasingly applicable to proteomics. For example, with the latest generation of qTOF instruments, we have recently shown that with just 1 µg injection amount, more than 4000 proteins can be quantified in a digested human cell lysate with 5-min high-flow chromatography (800 µl/min) [78]. High-flow chromatography is therefore attractive for many applications where robustness and short measurement times are required such as clinical applications (e.g., plasma, serum) [44,77,79] or screenings of cell lines or microbial samples [50,77].
Another innovative development is chromatography with preformed gradients, for instance, as implemented in the Evosep One system, which is designed for medium up to HT applications that require short overhead times [80]. This technology has been successfully applied in gradients of <6 min [81,82]. The platform runs with low flow rates, which addresses studies where only low sample amounts are available [80].
In general, short gradient chromatography is driven by advanced chromatographic systems and new column technologies that enable high peak capacities even in short separation times. The increased pressure limits of modern UHPLC systems and columns enable the use of small particles for fast and efficient separations. Further, new technologies such as micropillar array columns (PAC) [83], or core shell particles [84] increase the efficiency of peptide separations.

MS INSTRUMENTATION AND ACQUISITION METHODS FOR HT PROTEOMICS
The use of short chromatographic gradients necessitates the ability of the mass spectrometer to measure the eluting peptides at speed. As data-dependent acquisition (DDA) methods select individual precursor ions for MS/MS fragmentation, the proteomic depth they could achieve is limited at fast gradients by the MS/MS scan speed. Even though latest generation time-of-flight (TOF) instruments can measure more than 100 MS/MS scans per second [85], this factor still places DDA at a significant disadvantage when coupled to fast chromatography. Thus, the focus of HT proteomics has been placed on DIA, a technology that is rapidly becoming the new standard for proteome quantitation. In DIA methods such as Sequential Windowed Acquisition of all THeoretical mass spectra (SWATH-MS), the mass spectrometer cycles through a predefined set of wide Q1 quadrupole isolation windows, repeatedly isolating and fragmenting all the precursor ions within the mass range of interest [86]. HT DIA methods are driven by MS instrument developments that enable higher ion transmission efficiencies and shorter duty cycles. This is due to Q1 isolation widths being dependent on the applicable MS/MS duty cycle. Shorter duty cycles enable a higher number of narrow isolation windows which increases selectivity, that is the ability to separate signals from different peptides.
While early DIA methods often produced a lower proteomic depth compared to DDA methods, over time, DIA has been shown to increase proteomic depth, data completeness, robustness and quantification accuracy, specifically in proteomics experiments that use fast chromatographic gradients [65,71,[87][88][89]. Two independent studies addressed the reproducibility of DIA-based discovery workflows. The inter-lab reproducibility was assessed on the quantification of more than 4,000 proteins in a HEK293 cell line across 11 labs worldwide [90].
Moreover, in an intra-laboratory benchmark study, 1560 runs acquired along a 4 month period and across six different platforms demonstrated that longitudinal data acquired on multiple instruments can be integrated into the same dataset [91]. The benchmark studies indicate that parallelization is feasible, providing confidence for large-scale protein-quantification studies where samples are measured in several laboratories or on different instruments.
Stable and robust longitudinal data acquisition is prevented by signal drifts resulting from contaminations. Therefore, robust instrumentation that reduces the number of cleaning cycles is particularly important in large-scale experiments. Ion-mobility devices such as FAIMS or TIMS trap can prevent neutral or singly charged ions from entering the orifice of the mass spectrometer [81], keeping the instruments cleaner for longer periods. Further, instrument geometry such as the 90 degrees bent at the entrance of the TIMS device in the timsTOF pro increased robustness and prevented cleaning of the quadrupoles for more than 1.5 years [92]. Further, the instruments that use TIMS [82] or Zeno trap [93] increase the fragment ion signal, enabling decreased injection amounts in HT proteomics [78] and hence reduce contamination.

NEW ACQUISITION TECHNIQUES FOR FAST DIA EXPERIMENTS THAT ACHIEVE HIGH PROTEOMIC DEPTH
A challenge of DIA in comparison to DDA methods has been the lack of precursor mass assignment to the MS2 traces. In SONAR [94,95], and Scanning SWATH [77] a Q1 quadrupole continuously slides across the mass range of interest, as opposed to "discrete" cycling through predefined isolation windows in standard DIA. Relating the continuously recorded MS/MS signal to the position of the sliding Q1 quadrupole, allows to pinpoint the precursor mass corresponding to an observed fragment trace with a precision defined by the transmission profile of the quadrupole. This provides an extra "Q1" dimension of information, that promotes computational deconvolution of the data. Scanning SWATH is thus particularly beneficial for fast chromatography methods, gaining up to 70% in precursor identifications at 5-min gradients [77], quantifying over 5000 proteins from a k562 whole-cell tryptic digest using the DIA-NN software [96].
Ion-mobility devices can decrease spectral complexity and add another separation dimension to LC-MS platforms [5]. In dia-PASEF technology [82,97], trapped ion mobility separation (TIMS) [98] is explored to separate charged precursor ions based on their flight times through neutral gas, determined by the ion collisional cross section.
The additional separation improves selectivity, but it also increases sensitivity via its ability to carry out MS/MS scans while trapping the incoming ions for future scan cycles. Further, single charged ions can be excluded from fragmentation, reducing the background noise. The dia-PASEF technology likewise shows performance with fast methods, quantifying over 6000 proteins in 11 min [99] and more than 5000 proteins in 4.8 min [100]. Further, DIA-PASEF has been applied to quantify 12,000 phosphopeptides in 15 min measurements [101].
The effect of reducing the background signals is also explored by the FAIMS technology [102,103], which also utilizes ion mobility for extra separation of peptides. Unlike the trapped ion mobility concept employed in PASEF, FAIMS does not operate as a trap, but filters the incoming ions. FAIMS-DIA has likewise been shown to perform well with short gradient methods [81]. The FAIMS interface has been further used in combination with direct infusion and DIA, quantifying more than 500 proteins within minutes of measurement time, enabling fast proteomic measurements without liquid chromatographic separation [104].
Another recent technology termed ZenoSWATH uses an ion trap before the TOF analyzer to sequentially release the ions from high to low m/z [93,105]. This centers the ions in the TOF accelerator for nearly 100% duty cycles, achieving a sensitivity increase by the factor of 5-15 [105], depending on the fragment m/z [78]. This enables fast scan speeds, allowing for acquisition schemes with larger numbers of narrow isolation windows, producing cleaner and less convoluted spectra. We have recently demonstrated that the ZenoSWATH approach allows quantification of about 5000 proteins in 5 min gradients from a whole-cell tryptic digest standard (K562) [78].

DATA PROCESSING FOR HT PROTEOMICS
Before the advent of HT proteomics, the data processing software was optimized for handling long-gradient data and failed to confidently identify peptides from data acquired using 30-minute or faster gradients [95]. Shorter gradient runs, characterized by lower peak capacity, result in highly complex DIA data, requiring software tools that are capable of deconvoluting signals and removing interferences.
The first DIA software specifically developed to support fast proteomic experiments, DIA-NN, uses neural networks to filter out false precursor-spectrum matches and thus enable confident identification and quantification of peptides and proteins in short gradients [95].
Further, the commercial software Spectronaut [87] has been successfully used for the analysis of short gradient DIA runs [81,99,101], demonstrating its applicability for large-scale and HT discovery studies.
For the downstream analysis, several methods have been developed for large-scale datasets, including specific normalization strategies, batch correction, missing-value replacement, and statistical testing [91,106,107]. In addition, tools for quality control and management have been developed [108][109][110] that are particularly relevant for large-scale projects to ensure long-term stability and avoid instrumentspecific batch effects.

NORMALISATION, BATCH EFFECTS AND MISSING VALUES
Large-scale studies are often run on multiple instruments in parallel and across a long period of time [18,91,111], with time-dependent signal drifts and cleaning cycles [112]. Further, non-standardized protocols, reagent lots or different lab personnel can introduce additional biases [112]. Mitigating such artifacts requires appropriate normalization, batch effect correction and imputation strategies.
Proteomics datasets are usually normalized to adjust for different injection amounts and/or instrument dependent changes in signal intensities. Therefore, the distributions of all peptides/proteins (e.g., medians) are aligned between samples [35]. Such normalizations, however, are based on the following assumptions: (1) all proteins have a linear response and (2) the total protein amount remains constant.
Both assumptions need to be carefully considered, if samples are very heterogeneous or total intensities vary a lot between samples.
In addition to normalizations, batch effect corrections can be applied to correct discrete batch effects or drifts. In contrast to normalization, batch effect corrections are applied on feature level (protein/peptide). Drifts can be corrected by fitting and subtracting a linear or a non-algorithm (e.g., LOESS functions). For the correction of discrete batch effects several algorithms exist such as median centering, ComBat or RUV-III-C [35,91,113]. The selection of the right batch correction method depends on the batch sizes, sample type (e.g.,

DIA PROTEOMICS WITH MULTIPLEXING
Multiplexed proteomics allows for simultaneous analysis of multiple samples in a single MS acquisition. Thus, multiplexing via chemical labeling of peptides has the potential to provide higher throughput in mid-size experiments. DDA-based multiplexing with isobaric tags such as TMT [114], or iTRAQ [115] or non-isobaric labeling with mTRAQ [116] or dimethyl labeling [117] have found wide applications in proteomics. Recently, the introduction of a 16-plex label (i.e., TMTpro) has facilitated 16 measurements in parallel [45,118,119]. It has been demonstrated that combining the 16-plex TMT with FAIMS Pro and real time search enables the measurement of 16 samples in a 3-h DDA run [45].
Until recently multiplexed DIA methods were hampered by the complexity of the data processing. Recently, a multiplexing module in the DIA-NN software has been specifically tuned to take advantage of the co-elution of peptides which are differentially isotopically labeled.
This resulted in the development of the plexDIA technology [120], which offers three-fold increased throughput of DIA proteomics, using mTRAQ tagging, while maintaining comparable data quality. Most importantly, plexDIA is applicable to any chromatographic setup, and allows for HT proteomics at low flow rates, thus combining speed and sensitivity. For example, plexDIA enables measuring about 1000 proteins in single U-937 monocytes on a timsTOF SCP instrument (Bruker Daltonics), using a 15-min gradient, that is, 5-min per cell considering the 3-plex setup [120].

APPLICATIONS OF HT PROTEOMICS IN BIOMEDICAL STUDIES
One promising application of HT proteomics is the analysis of clinical specimens such as biofluids and tissue samples for exploring clinical phenotypes, revealing molecular dysregulation, discovering new biomarkers, and eventually, to create prognostic and predictive models for diseases.
Blood is a rich source of biomarkers, easily accessible and routinely collected [121]. Therefore, many large-scale studies focused on plasma and serum samples. The effect of weight-loss on the proteome has been studied in 52 obese participants, along 8 weeks (1294 plasma samples).
In total, 93 proteins changed with weight, including apolipoproteins and other proteins involved in lipid metabolism, and it was found that weight loss is associated with reduced low-grade inflammation [42].
Another weight-loss study generated an even larger dataset of 1508 samples. With a throughput of 31 samples per day, this study identified on average 408 proteins and quantified those with quantification precision (CVs) of 10.9% [49]. These two independent weight-loss studies both identified the same panel of apolipoproteins as the most significantly changing proteins, despite employing different technologies (DDA and DIA). Recently, alcohol-related liver disease biomarkers were discovered in a cohort with 569 individuals, measured in just 3 weeks. The biomarker panel enabled prediction of fibrosis as well as mild inflammation, outperforming conventional clinical assays [122].
Large-scale plasma proteomics was further applied to dissect protein variability into genetic, environmental, and longitudinal factors by analyzing a twin study (232 samples) [123], and to identify aging markers by measuring proteomes (among other molecular data) of 106 individuals over 48 months [124]. Further, the progress in MS based technology has enabled deep plasma profiling even in large-scale studies as demonstrated in a cancer cohort of 180 plasma samples, identifying 2732 proteins and reporting several biomarker candidates [71].
While plasma and serum are the most frequently analyzed body fluids, large-scale studies have also been conducted for cerebrospinal fluid (CSF) [51,125] and urine samples [126] to find proteins associated with Alzheimer's disease [51] and Parkinson's disease [125,126].
A large-scale study has also been performed of bone marrow biopsies, measuring quantitative proteomes of 252 patients with acute myeloid leukemia and identifying different disease subtypes [127].
The Clinical Proteomic Tumor Analysis Consortium (CPTAC) has reported a range of tumor-tissue studies that are not necessarily HT but have sample sizes larger than 100 [128][129][130][131][132][133][134][135][136] and are thus worth mentioning in this review. These studies focus on proteogenomic analysis, identifying cancer-relevant pathways or potential therapeutic targets. The data are available on the CPTAC portal [137], making these datasets a rich and comprehensive resource for studying cancer biology. The prediction of drug responses to platinum in 130 ovarian serous carcinoma patients [138] as well as the identification of disease subtypes in 204 FFPE prostate cancer samples [60] and 110 early-stage hepatocellular carcinoma [139] indicate the potential of proteomics for patient stratification and personalized therapies.

CLINICAL PROTEOMICS EXEMPLIFIED BY THE COVID-19 HOST RESPONSE
The COVID-19 pandemics has highlighted the value of HT DIA proteomics to rapidly foster a molecular understanding of a disease, and to identify biomarker panels with diagnostic, predictive, and prognostic value.
To this end, within months after the outbreak of SARS-CoV2, plasma proteomics successfully provided a molecular map of the patient response to the virus, highlighting the impact of the innate immune system, blood coagulation, complement, and acute phase response proteins [44,140,141]. Next, we and others showed that identified plasma protein profiles can successfully classify COVID-19 severity, and are prognostic of disease outcome, outperforming established clinical scores [44,78,100,142,143]. This demonstrated that plasma proteomics could be a valuable tool to assist clinical decision making and drug development.
Further, it became evident that a set of peptides was sufficient for outcome prognosis [144] or severity stratification [44,79]. This was crucial for the next step in translating such peptide signatures into a clinical assay, which required multiplexing capabilities, absolute quantification, and a conversion to the analytical instrumentation available to routine diagnostics labs, that is, high-flow chromatography coupled to triple-quadrupole mass spectrometers. Facilitated by the use of high-flow LC coupled with DIA proteomics, we then converted a panel of peptides selected by discovery proteomics for COVID-19 to a multiplexed, targeted multiple-reaction-monitoring (MRM) method combined with the use of isotope-labeled protein standards for absolute quantification [145]. The assay, which operates on triple-quadrupole instruments, successfully stratified severity and was prognostic about outcome of COVID-19 disease in a multi-cohort clinical study [145].
The outlined workflow of using HT DIA for the discovery of biomarker panels that are translated to MRM based absolute quantification assays highlights a potential route for MS-based proteomics to enter routine diagnostic labs and regulated environments. This might be crucial to drive precision medicine and the development of a new generation of targeted treatments.

HT PROTEOMICS IN SYSTEMS BIOLOGY
Protein expression levels depend on many factors such as localization, genetic background, or environment. Understanding those dependencies can guide protein function annotations, disease models and the discovery of new drug targets. With the increased throughput, proteomics has matured to a technology for screening 100s to 1000s of different conditions. Therefore, several groups have used proteomics in recent years to systematically characterize different cell types, organisms, and tissues as well as systematically mapped different genotypes, or perturbations.
Several studies functionally characterized cancer cell lines with proteomics [113,146,147]. The largest and most comprehensive cancer cell line map includes 949 cancer cell lines across 28 tissue types and represents a major resource for target identification and for studying the relation between phenotypes and proteomes [110]. Our group characterized the proteome diversity of 1011 natural yeast isolates and enabled the systematic analyses of aneuploidies, revealing chromosome-wide dosage compensation at the proteome level which was associated with increased degradation [50].
With higher throughput of proteomics technologies, systematic screens of loss of function mutants became a promising tool for gene annotation and for understanding genotype-phenotype relationships.
For example, large-scale multi-omic profiling addressed hundreds of mitochondrial knock-outs in HAP1 cell lines [148] as well as yeast [149] to functionally annotate mitochondrial proteins as well as for the understanding of mitochondrial dysfunctions [148,149]. A further resource for inferring protein functions was generated by measuring protein abundances as well as thermal stability in response to 121 genetic perturbations in E. coli. Among others, this dataset explains growth phenotypes and demonstrates how functionally related proteins are coregulated across genetic perturbations [150]. Our lab measured proteomes for the Saccharomyces cerevisiae protein kinase knock-outs and used this data to generate a machine learning approach to predict the metabolome. We demonstrated the application of machine-learning approaches to describe the complex multifactorial relationships of enzymes and metabolites [74]. More recently, we measured proteome profiles for ∼4500 knock-out yeast strains and created, to our knowledge, one of the largest and most systematic proteomic datasets with almost nine million measured protein quantities. Moreover, this study has revealed a key advantage of com- Proteins are targets of most drugs and various proteomics and chemoproteomics methods are increasingly used for preclinical drug discovery [152,153]. HT discovery proteomics is particularly promising as a screening tool for the systematic and simultaneous analysis of cellular drug responses. It has been shown that drugs with similar mechanisms of action cluster together, identifying targets and offtarget effects and identifying resistance [78,154,155]. For example, this approach discovered mitochondrial function and autophagy modulation by kinase inhibitors by measuring 53 compounds in 5 cell lines [154]. Such proteomics screens can hence combine the advantages of target based and phenotypic drug screens. With proteomics increasing in speed and decreasing in price, systematic proteomics screens have thus the potential to provide unique and comprehensive readouts in early-stage drug discovery.
One major factor required to functionalize the human genome is an understanding of the association of genetic variation and protein abundances. Linking large scale proteomics with genomics has enabled the identification of genetic loci and alleles that are associated with changes of protein levels (protein quantitative trait loci, pQTL) [8, 156,157]. Whereas genomic quantitative trait loci (QTL) simply link genetic variants with phenotypes and diseases, pQTLs enable insights into the molecular mechanisms underpinning these associations, for example, by pinpointing the affected proteins and disentangling effects on protein dosage from protein structure [158]. Partly due to their accessibility, the analysis of human biofluids such as plasma have proven promising, for example, yielding a comprehensive plasma proteo-genomic map of human health and disease [156,159,160]. Knowledge that a genetic variant links to a disease, and is associated with changing protein levels is particularly valuable to establish target-indication pairs, and thus for the mining of drug targets [161]. and distant (trans) pQTLs that drive protein abundance changes, many of which were not detected in the transcriptome [162].

CONCLUSION AND OUTLOOK
There are several equally important reasons why the next genera-

CONFLICT OF INTEREST
The authors declare no conflict of interest.

DATA AVAILABILITY STATEMENT
Data sharing not applicable -no new data generated.