Endometrial cancer (EC) is the most common gynaecological neoplasm in developed countries, and over 382,000 new cases were diagnosed worldwide in 2018 (1). In general, EC is diagnosed in postmenopausal women (85% of cases) and its incidence is rising due to longer life expectancy and life style associated risk factors. Women with BMI above 35 have an odds ratio of 5.7 for developing EC, with an increase of 1.1 odds ratio per BMI unit (2, 3). Exposure to unopposed estrogens or tamoxifen or genetic aberrations associated with Lynch syndrome confer a cumulative risk up to 70% (4, 5). Finally, endocrine disruptors and other environmental pollutants can also increase EC risk (6, 7). Therefore, an alarmingly high number of women in the general population is exposed to risk factors for developing EC.

In this context, screening programs would be extremely beneficial for these women, but, unfortunately, no minimally- or non-invasive diagnostic tool for EC exists today, and diagnosis relies on invasive endometrial biopsy and pathology investigation.

A second unmet clinical need in EC is the necessity to accurately stratify patients. EC is diagnosed at an early FIGO stage in 80% of the cases, and the five-year survival of FIGO stage 1a is around 95%. However, a proportion of women diagnosed with early-stage EC develop recurrent disease, which dramatically decreases survival rates (8). This represents a challenge as recent projections indicate that the worldwide EC mortality will increase by 70% by 2040 (Global Cancer Observatory, World Health Organisation - https://gco.iarc.fr).

Therefore, prognostic biomarkers to reliably predict patient prognosis are needed, both prior to any intervention - to decide on the most appropriate treatment and if needed optimally plan the surgical procedure - as well as post-operatively, to define the most appropriate adjuvant treatment, and avoid over-treatment and under-treatment. A number of prognostic markers like histological assessment of tumour type and grade, hormone receptor status, PTEN expression, mismatch repair proteins (MLH1, PMS2, MSH2, MSH6), POLE exon 3 mutation, CTTNB1 mutation, L1CAM overexpression, and TP53 aberrations allow stratification of patients according to their risk of recurrence (9–20). In particular, the recent introduction of The Cancer Genome Atlas (TCGA) molecular classification improved the risk stratification at the postsurgical (21, 22), but also improved the concordance between presurgical biopsy and pathology assessment at hysterectomy (23), which has been a problem in the past (24). This classification groups EC patients in four clusters with distinct prognosis and a number of studies demonstrated the reliability and the clinical applicability of this classification using surrogate analyses (i.e., IHC and POLE gene mutation analyses). Patients with POLE mutated tumours have the best prognosis, followed by mismatch repair deficient tumours and with the final groups having an intermediate and the worse prognosis (no specific molecular profile and p53 mutated, respectively) (14). Recently, also classification methods fully based on IHC, hence applicable also in centres with limited access to molecular infrastructures, showed robustness and reliability (19).

Nevertheless, these methods require invasive biopsies, and women consider the presurgical biopsy procedures discomforting and painful (25). Therefore, non- or minimally-invasive prognostic tools applicable presurgically are urgently needed.

Proteomic and metabolomic profiles are attractive approaches for identifying biomarkers that can be detected in tissues or body fluids obtained via non-, minimaly or semi--invasive procedures. The purpose of the present review is to explore the current research on the use of proteomics and metabolomics in the context of EC. This review provides a brief introduction to the wet-lab technologies, the computational pipelines for data analyses and a systematic review of all published studies aimed at using proteomics and/or metabolomics for diagnostic and prognostic biomarker discovery in EC. This is followed by conclusions and recommendations for future studies.

Proteomics and metabolomics represent fields that have grown significantly in the last decades, thanks to the important technological advances that allow accurate and sensitive analyses. Both approaches have been extensively used for biomarker discovery in various disorders (26–31).

Data from proteomic and metabolomic experiments can reveal molecules that can possibly serve as diagnostic or prognostic markers. However, even if well-designed and executed, experiments often result in noisy, biased and incomplete data due to a multitude of uncontrollable factors. Therefore, thorough data analysis needs to be performed to eliminate technical noise, while preserving genuine biological variation between samples. A set of computational methods used to analyse data are typically bundled together into one unified data analysis pipeline ( Figure 1 ), which treats raw data files as an input while providing the list of potential biomarkers as the output.

Proteomic and metabolomic raw data is first processed by background correction, signal transformation, outlier detection, and normalisation. Pre-processing is essential to minimise unwanted technical bias and enable comparisons of samples. Further, integration of clinical information enables comparisons of average metabolite and protein signals between phenotypic groups of interest (48–50).

Background correction addresses different effects in proteomics and metabolomics. In protein microarrays, it is challenging to correctly quantify the fluorescent signal produced by the biological reaction avoiding local residuary background (51–53). In metabolomics, background (or baseline) correction is used to eliminate low frequency artefacts and differences generated by the measuring instrument (54). Data log-transformation is common practice, as this renders fold changes symmetric around zero, reduces potential skew in the data and provides a good approximation for the normal distribution, which is a prerequisite for most computational methods (52, 55), especially for linear models. Following background correction, outlier detection is performed by the three standard deviations technique and subsequently removed or replaced.

In large-scale studies based on MS proteomics or metabolomics, samples will be distributed into several analytical batches, which may introduce instrumental variabilities into the data set. Such batch effects can be very destructive as they render comparison between phenotypic groups ineffective. Normalisation strategies for metabolomic and proteomic experiments make use of control samples and control molecules (52, 56, 57), which are usually assumed to exhibit constant signal levels. Any differences in signal values are considered to be technogenic and thus, corrected for. The most popular normalisation strategies are global scaling, quantile normalisation, cyclic loess and the ones involving linear models (56).

After appropriate pre-processing, the data is used for statistical analysis, where a large number of techniques are available. Characteristics of data and the research question determine the choice of the statistical method. In biomarker discovery studies, molecules that can reliably distinguish between two (or more) groups, like disease versus controls, are referred to as significantly differential and can be used as biomarker candidates. The process of identifying such molecules is termed differential analysis (52, 58). Classical univariate statistics such as Student’s t-test (requiring normal data distribution) or Mann-Whitney U test (non-parametric test that does not rely on parameterized data distribution) can be used for differential analyses. Differential analyses have a low probability (usually less than 5%) to deliver significant results by mistake; however, if repeated multiple times, as for omics studies, can result in the generation of false significant hits. Therefore, the number of tests performed needs to be taken into account. The simplest and one of the most popular methods for multiple testing correction is ‘Bonferroni correction’ (59), which adjusts the p-value threshold by dividing it by the number of tests. This is a conservative approach that may result in a high number of potential biomarkers being ignored, hence, less stringent methods can be considered (like Benjamini and Hochberg False Discovery Rate correction) that keeps the number of falsely significant results at a predefined level (e.g., 5%).

While classical statistical methods analyse the significance of each molecule of interest independently (60), machine learning algorithms are able to efficiently assess the predictive performance of multiple proteins, metabolites, features and even their combinations. Machine learning is a field of computer science that studies algorithms capable of learning valuable relationships from data without being explicitly programmed. Myriads of machine learning algorithms have been developed over the past years (61) and are frequently used in biology to discover biomarkers for various diseases (49, 62). The most popular machine learning methods are decision tree (63), support vector machine, random forest (64) and gradient boosting machines (65).

It can be challenging to build reliable machine learning models, because most model algorithms can learn random patterns that can only explain data these models were exposed to. This phenomenon is known as overfitting and might cause models to report completely irrelevant biomarkers and thus, render the entire study obsolete. In order to account for potential overfitting and keep its influence at minimum, various strategies have been proposed (66). One of the most important techniques is k-fold cross validation. By using only one part of the data to build a model (training set) and the remaining part to assess its performance (test set), researchers can be confident that the biomarkers identified by the model are not random fluctuations in the training data.

With this systematic review we aimed to respond to the following question: Can proteomics and metabolomics contribute to identification of biomarkers for diagnostics and prognostics in EC? The review was conducted according to the PRISMA guidelines (67) and is registered at the ‘International prospective register of systematic reviews’ (PROSPERO, Registration number CRD42022245880).

We performed a systematic search of the literature in the PubMed^® and OVID^® Embase databases on July 20, 2022 using the search terms listed in Supplementary Table S1 . We focused on proteomic and metabolomic studies performed in physiological fluids and tissue samples. There was no restriction on publication date. Reports were retrieved, and titles/abstracts were screened according to the inclusion and exclusion criteria ( Table 1 ) independently by authors AR and TLR. Disagreements were discussed and consensus was reached. The search strategy is provided in Table 2 .

The selected reports were read in detail and the following relevant data was extracted (when applicable): author and year of publication, country, fundings; sample: tissue (kind), plasma, serum or other body-fluid; study design; methods: omics approach, targeted/nontargeted proteomics/metabolomics; analytical methods; patient selection: case/control, stratification of patients according to reference test; patient characteristics: number of patients, characteristics of the enrolled patients with EC (e.g., mean age, body mass index [BMI], type of EC, histological differentiation, FIGO stage, menopausal status) and control patients or healthy women (e.g., mean age, BMI, diagnosis, menopausal status); study phase and statistical methods: discovery, validation phase, machine learning approaches used; differentially abundant proteins and metabolites in the study groups; diagnostic characteristics (e.g., sensitivity, specificity, area under the curve [AUC], positive predictive value [PPV], negative predictive value [NPV]) or prognostic characteristics (overall survival [OS], disease-free survival [DFS]), and hazard ratios [HR]; diagnostic or prognostic models; disclosures: affiliations with industry, industrial funds, patents.

Reporting was performed under the guidance of the PRISMA diagnostic test accuracy checklist (67). The risk of bias and quality of individual diagnostic accuracy studies were assessed following the QUADOMICS tool, an adaptation of QUADAS (68) that was designed specifically for omics studies (69) ( Supplementary Table S2 ). This tool focuses on study design, patient selection, index test, reference standards, flow of timing, pre-analytical and analytical procedures, and statistical analysis and nine questions per study were specifically answered ( Supplementary Tables S2–S4 ). Additional potential financial, commercial and conflict of interest biases were further examined ( Supplementary Tables S5 ).

The quality of the studies included was assessed systematically according to the QUADOMICS tool ( Supplementary Tables S2–S4 ; Figure 3 ). We evaluated study design and pre-analytical, analytical, and post-analytical bias of all included studies. The majority of the studies described the criteria for patient selection (question 1) in appropriate detail. Approximately 50% of metabolomic and 20% of proteomic studies did not reflect the real clinical setting, because they compared patients with healthy women, who are not likely to need a diagnostic test (see discussion - question 2). The assessment of pre-analytical bias (questions 3A and 3B) revealed that only a fraction of all studies reported appropriate descriptions of the samples, including the procedures for sample collection and processing (e.g., centrifugation time, type of blood tube). Furthermore, the majority of the studies did not report any information about the time of sample collection, the time between blood draw and centrifugation, the time between sample acquisition and storage, and the number of freeze/thaw cycles. In 75% of the metabolomic and 51% of the proteomic studies sufficient information on the clinical and physiological factors that can affect -omics data was not provided (question 4; e.g., BMI, menopausal status, menstrual phase cycle, fasting status). Approximately 70% of the included studies reported detailed descriptions on sample storage and metabolite extraction (question 5). Almost all samples were stored at -80°C or in liquid nitrogen, but several studies failed to report this information. The time between the reference standard and the index test (metabolomics or proteomics) was not clear for 31% of the included metabolomic and 21% of the proteomic studies (question 6), while in 52% of the metabolomic and 22% of the proteomic studies, the verification by reference test was not performed in all patients (question 7). With respect to analytical biases, we observed that only 21% of all studies provided a detailed description of the metabolomic analysis (question 8). Most studies did not provide information on sample randomisation for MS-based metabolomics, for the use and type of quality control samples, and occasionally, important MS parameters were not given. In proteomics, 78% of all included studies described the index test in sufficient detail. Regarding post-analytical biases, we observed that 24% of the metabolomic studies described the statistical analysis in sufficient detail, while 42% of studies provided incomplete description. These studies failed to report information on missing value treatment, sample-to-sample normalization, data transformation and scaling and in one case also on model calculation and cross-validation. In proteomics, 91% of all included studies reported the statistical analysis in sufficient detail.

We also evaluated whether studies clearly stated the financial support, disclosed any potential conflict of interest, whether authors were affiliated to industry and whether the studies complied with the open science policy and deposited their data on public repositories ( Supplementary Table S5 ). Although older studies tend not to report any information on the financial support or the presence of any conflict of interest, more recent studies provide this information. The source of funding was declared in 88% of the studies (19 out of 23 proteomic and 27 out of 29 metabolomic studies) and the presence of any potential conflict was declared in 77% of the studies (14 out of 23 proteomic and 26 out of 29 metabolomic studies). Three studies (6% of those declaring the source of funding) received industrial supports and four studies included authors affiliated to companies. Only 25% of the studies deposited the data in public repository (1 out of 23 proteomic and 12 out of 29 metabolomic studies) and four proteomic studies declare that data are available upon request.

From the systematic literature search, 72 research papers were selected based on title and abstract. From these, 19 papers were excluded because focusing on basic cell mechanisms of carcinogenesis with no further investigation on the diagnostic or prognostic potential (70–80), response to metformin (81), side effect to radiotherapy (82), racial disparities (83, 84), drug resistance (85–87), premenopausal endometrial physiology (88). Three studies were in silico analyses (89–91). Additionally, 11 papers (22, 92–101) were based on multi-omics approaches or focused on pan-cancer biomarkers and will be discussed in paragraph 3.4, thus resulting in 39 papers for review.

Papers that included less than 10 subjects per study group are not further discussed here (n=16; see Supplementary Table S6 for details). This resulted in a total of 23 papers focusing on proteomic biomarkers for diagnosis, prognosis, risk stratification or classification ( Table 3 and Supplementary Table S7 ). Ten studies used blood (serum or plasma), two studies used uterine aspirate whereas eight studies used fresh frozen tissues and three used formalin-fixed-paraffin-embedded (FFPE) tissues. Various technologies were used, with the most common being 2D-DIGE/MS based methods. Although studies on tissue proteomics preceded chronologically those in body fluid, since this review focus of diagnostic/prognostic biomarkers where body fluids represent the most suitable biomaterial, we will start in the next paragraphs describing studies using body fluids for biomarker discovery.

Our literature search identified 29 studies using metabolomics in EC ( Table 4 and Supplementary Table S8 ), with the majority evaluating the metabolic profiles in blood samples (10 serum, seven plasma, one serum and plasma, one dried blood). Seven studies focused on endometrial tissue samples, one on cervical lavage, one on endometrial brushing and one study on urine samples. Most studies aimed to identify diagnostic and/or prognostic biomarkers, better understanding the mechanisms of carcinogenesis (studies in tissue samples), and also to determine associations between metabolic profiles and EC (131, 132). Plasma, serum and urine represent appropriate sources for discovery of diagnostic/prognostic biomarkers (155). However, also metabolic profiles of cervical lavage, brushing endometrial samples, and tissue samples (if obtained as pre-surgical biopsy) may be of clinical relevance. Non-targeted metabolomics was more commonly applied (20 studies) as compared to targeted metabolomics (10 studies). Only six studies used NMR analysis, and there was one study that combined NMR with the most commonly used LC-MS/MS (138). The majority of the targeted metabolomic studies focused on lipids and amino acids.

There was only one study that employed a combined omics approach (101) and performed nontargeted metabolomics on 24 cancer and 20 control tissue samples ( Supplementary Table S7 ) and also nontargeted proteomics on a subset of 12 cancer and 9 control tissue samples by LC-MS/MS. The authors identified 1,445 proteins significantly up- or down-regulated in the EC group compared with the control group (adj. p<0.05, FC<1.5). To further characterise any relation between the metabolic and proteomic profiles, the authors performed network analysis that showed 28 metabolites and 135 proteins with 212 connections. Glutamine, dopamine, noradrenaline, adenosine-5-monophospate, and guanosine-5’-monophosphate were the major centres of sub-networks showing differences in amino acid and nucleotide metabolism.

A number of studies explored the proteome in patients with EC, but such analyses were part of a larger multi-omics approach, or part of systems-biology/pan-cancer approaches to study human disease or to establish databases and repositories. Most of these studies did not use proteomics for biomarker discovery, but as tools to understand the pathophysiological processes.

Five studies aiming at improved patient classification were performed within the TCGA consortium and used RPPA as proteomic method (22, 92, 94, 95, 99). Two studies demonstrated the utility of functional proteomics based on RPPA next to genomics and transcriptomics (92) and further created the bioinformatic resource ‘The Cancer Proteome Atlas’ (TCPA; (94)). The study of Kandoth and co-workers (13) explored a cohort of 373 EC patients, including 307 endometrioid and 53 serous or mixed histology cases to assess somatic mutations, copy number alterations, RNA expression, protein expression, DNA methylation and micro-RNA expression. With regard to proteomics, 293 samples were analysed by RPPA and several differentially expressed proteins were associated with other specific molecular tumour features (22). One of the latter two TCGA studies did not use proteomics (99), whereas the other study (95) explored 57 carcinosarcomas and -also in this case-, protein analyses were used to confirm other features identified with the molecular analyses (like EMT transitions, PI3K/AKT pathway activation, low steroid hormone receptor signalling). Similar to the TCPA (initiated within TCGA), the Clinical Proteomic Tumor Analysis Consortium (CTPAC) generated large proteomic datasets across various tumour types, and further characterised the proteogenomic landscape in EC (100). A pan-cancer study (98) aimed to characterise the actionable mutations across different solid tumours (including EC) and used RPPA to demonstrate PI3K and MAPK signalling pathways, whereas one study was focussed on human diseases other than cancer (97).

The limitations reported above related to the papers retrieved and reviewed are reflected also in the present study, i.e., small study cohorts, potenial pre-analytical and analytical bias, potential bias due to ethnic background, lifestyle; lack of validations; no comparison with a reference or gold standard. This, in combination with the heterogeneity in study designs and in the technologies adopted precluded us making any meta-analyses of identified candidate biomarkers, whose potential can be assessed at a qualitative level only. This implies the impossibility to make any clinically relevant recommendation or conclusion at this moment. Nevertheless, the strength of the present study is that, due to the rigorous systematic approach we adopted, it offers a balanced and realistic view of the potential of these technologies for the future. It sets the milestones in proteomic and metabolomic biomarker discovery research, and indicates the path to follow in the future (see paragraph 4.2. Recommendations for further biomarker discovery).

It should be also noted that recently systematic reviews focused on metabolomic and/or proteomic biomarkers for diagnosis of EC (160), and liquid biomarkers for diagnosis of EC (161). However, our systematic review has additional unique strengths: we did not limit our analyses on one biospecimen only, but focused on metabolomics and proteomics in different biospecimens; we rigorously assessed the study quality using QUADOMICS and analysed additional potential conflicts of interest. The study quality was also assessed by Karkia and co-workers (161), however, these authors included only studies published in the two years prior to the publication, whereas we did not set a time limit for publication. Additionally, we used strict inclusion and exclusion criteria, which were defined and deposited in the PROSPERO repository prior to the start of our work. We finally provide comprehensive tables with all available diagnostic accuracy data (AUC, sensitivity, specificity).

As thoroughly discussed, body fluids and liquid biopsies represent the most suitable material for diagnostic and prognostic biomarkers (162, 163), as they capture the disease heterogeneity better than a biopsy and they are non- or minimally invasive, thus create less anxiety in patients. In the context of the most appropriate body material, plasma preparation is less prone to technical (pre-analytical) biases than serum. Therefore, plasma represents the preferred source at least for the discovery phase of non-invasive diagnostic/prognostic biomarkers. Urine also represents an important clinical sample for non-invasive diagnostics (155), calling for further biomarker discovery studies. However, urine poses a challenge for biobanking, as large sample volumes are needed for analyses (this applies not only if 24-hour urine needs to be collected, but also for morning urine, which is common in biomarker discovery studies).

In terms of methodology, statistical and experimental validation in independent cohorts should be an intrinsic part of biomarker discovery studies, and inclusion of study populations with distinct lifestyles, geographical regions of origin and ethnic backgrounds is essential to identify candidate biomarkers truly associated with diseases. Strict standard operation procedures for sample collection and processing should be prepared by experts and rigorously followed (164). The importance of adhering to quality standards is also emphasized in a recent narrative review on biomarker discovery in EC (165).