Clinical lipidomics in the era of the big data

Applied analytical approaches in lipidomics

Historically, thin layer chromatography (TLC) and gas chromatography (GC) techniques were used for lipid analysis, but these usually required derivatization (which is time-consuming and chemically alters the original sample). Since the vast majority of lipidomics publications nowadays use the mass spectrometry (MS) technique for the analysis and since this technique is fully compatible with routine clinical applications (e.g., in neonatal screening or drug monitoring), it will be the main focus of our clinical lipidomics review. MS offers a sensitive, structurally detailed and high-coverage analysis of the lipidome of various biological samples, however, it also faces some shortcomings that have already been discussed in another review [3]. Additionally, for specific and key publications, alternative techniques beyond MS will also be mentioned. From an analytical point of view, clinical lipidomics is mainly performed on a MS instrument (with low or high resolution) with either direct injection or liquid chromatographic separation step. We can distinguish between two approaches applied in lipidomics: (a) untargeted and (b) targeted. The first one is mainly data-driven, and it is used for hypothesis generation. The analytical workflow (sample preparation protocol and analytical method) is optimised to detect as many molecules (lipids) as possible. Since there is no focus on a particular group of lipids, we are not eliminating the possibility of revealing the unanticipated changes, including unknown metabolites and pathways. The second approach is generally hypothesis-driven and quantitative, making it the perfect candidate for clinical (medical) practice application. It offers high sensitivity and specificity for the measurement of known lipids. With targeted lipidomics, we can establish a lipids panel that can be used for diagnostics.

Overview of general workflow and data acquisition

The general workflow in clinical lipidomics, which typically includes the design of experiment, sample collection, sample preparation, quality control, analysis, data processing, statistical evaluation, and validation, is summarized in Figure 1. The first step – the preanalytical phase/collection of the samples – cannot be underestimated. Many lipids (e.g. polyunsaturated fatty acids (PUFAs)) are prone to oxidative changes or other biological transformations/degradation. Thus, the storage and time between the sample collection and analyses are crucial. In terms of sample preparation, collection and storage, preceding the lipidomic analysis itself, many factors must be taken into account that may affect the obtained lipid profile and its reproducibility. Recommendations regarding these factors, the clinical trial design or reporting of results, have already been summarised and discussed in another review [4]. The non-polar compounds – lipids – are extracted from biological material (plasma, serum, tissues) by the simple addition of organic solvent, centrifuged and analysed by, in most cases, liquid chromatography coupled with mass spectrometry. The complex mixture of lipid molecules is then separated by either the length of their non-polar side chains and the number of double-bonds (reversed-phase separation) or the polar head group (hydrophilic interaction chromatography known as HILIC) [2]. We can even skip the separation part for high-throughput quantitative assays and use the flow-injection analyses coupled with high-resolution mass spectrometry (FIA-HRMS). Molecules such as free cholesterol (FC), cholesterol esters (CE) [5] or diacylglycerols (DG) and triacylglycerols (TG) can be analysed this way.

Figure 1:

Overview of the clinical lipidomic workflow. IQCs, internal quality controls; EQCs, external quality controls; ROC, receiver operating characteristics.

The analyses of large sample cohorts bring many challenging difficulties. The main issue is the stability of the analytical platform over a long period. Usually, data for extensive clinical studies need to be acquired over time in multiple batches. Analytical variations and batch effects are taking a severe impact on such data. Applying quality control (QC) and quality assessment (QA) samples is strongly encouraged to overcome this obstacle. The QC samples in the individual batches acquired across multiple weeks will differ in their signal intensities, although the samples are the same. Such an effect needs to be corrected either by the application of a machine learning algorithm (e.g., Systematic Error Removal using Random Forest (SERRF)) [6] or by piecewise polynomial interpolation (e.g., SPLINE). Often only corrections to stably labeled internal standards are used to eliminate the above trends, but this is not able to fully cover the variability in trends even within lipid classes.

After processing the data and obtaining the areas of the raw peaks, filtering and zero imputation are commonly applied and the identified lipids can be quantified at the level of absolute quantity or semiquantity using stable labeled standards. Lipid concentrations in biofluids (most commonly in plasma) are usually reported in molar concentration (nmol/L), but standardization to total lipids in the sample or percentage of lipids relative to lipid class is often used.

Data file size, mainly generated by the untargeted approach, largely depends on the length of chromatographic separation and the number of acquired data points (scanning speed of the instrument). Thus, using backup storage devices with multiple mirroring options across physical storage platforms is highly recommended. The generation of checksums files (e.g., MD5, SHA-256, etc.) should also be applied to ensure correct data transfer between the research groups.

A separate problem is the software dedicated to targeted or untargeted lipidomics. Due to the variety of analytical platforms utilized in lipidomics, many different software solutions for processing lipidomics data have been developed hand-in-hand. Currently, there are many commercial software solutions provided by vendors competing with rapidly evolving free-to-use software, which is driven by an active lipidomic community.

Data handling and statistical methods in big data lipidomics

Due to the complex nature of lipid data containing hundreds and thousands of compounds, data handling and statistical analysis are key steps providing the basis for the subsequent interpretation of results. This overview in Figure 2 was created based on the use of statistical tools in the publications listed in Table 1. At the beginning of each project, a power of study should be calculated to define the relationships between sample numbers and the expected statistical significance of each test. This should be followed by an evaluation of the nature and quality of the data with respect to normality, skewness, outliers, presence of missing values and breadth of scale. This step goes hand in hand with descriptive statistics and logarithmic transformation, which provides a Gaussian distribution for subsequent use of parametric tests. Most statistical analyses then involve univariate methods, usually based on hypothesis testing or correlation analyses, in either parametric or non-parametric settings. In clinical big data projects, regression models are often applied, allowing a combination of ordinal and categorical data types. More recently, multivariate methods based on different variants of supervised discriminant analysis have emerged allowing the creation of statistical models description and classification of samples. The systematic names of lipids contain information about their structure (number of double bonds and number of carbons), these chemical and physical ontologies can be used to find and visualize trends at many different levels. The list presented in Figure 2 does not aim to be a comprehensive overview of all statistical methods but rather represents the trends and options that are currently used in data evaluation. In the near future, we can expect to see greater use of Bayesian statistics and the deployment of artificial intelligence and machine learning, which may bring new advanced tools for data evaluation.

Figure 2:

Statistical evaluation of big data clinical lipidomics. UV scaling, unit-variance scaling; kNN, K-nearest neighbours algorithm; LOD, limit of detection; SD, standard deviation; QR, quartile range; ANOVA, analysis of variance; ROC, receiver operating characteristics; PLS, partial least squares discriminant analysis; OPLS, orthogonal projections to latent structures discriminant analysis; LDA, linear discriminant analysis; AUC, area under the curve.

Harmonization of lipidomics

Mass spectrometry-based workflows have become more technically robust and user-friendly for their easier implementation in routine clinical laboratories and thus for an expansion of the number of lipids measured beyond only cholesterol and triacylglycerols. A wide range of analytical platforms are used in studies for mass spectrometry-based lipidomics, but simplified and streamlined workflows which are compatible with routine clinical laboratory settings are in demand. The same issues apply to sample collection and storage, for example, the lack of community-approved analytical protocols brings a challenge in the translation of lipidomics results (i.e. biomarker panels) into the clinical setting [7].

These challenges can be overcome by the efforts of the “Clinical Lipidomics Interest Group” (CLIG) within the International Lipidomics Society (ILS), which was established in 2019. The CLIG is focused on managing the inter-laboratory ring studies that were initiated in 2022. Shared materials (e.g. human plasma reference material) are being analysed by all participating laboratories according to well-described mass spectrometry-based workflows (using validated analytical methods according to FDA and EMA guidelines) [8]. After finding out major gaps in current knowledge, a constructive dialogue between representatives from all disciplines: lipidomics and biomedical researchers, analytical chemists, representatives from regulatory agencies (i.e. FDA, EMA, ILS), and data scientists will be needed to propose solutions to these pitfalls. It is the lack of harmonization and standard reference measurements that is the current challenge in lipidomics. A similar situation occurs with other body fluids such as urine, tears, saliva, or sweat. The standardization of the measurement of these alternative body fluids is another problem, due to uncontrolled procedures for collection, transport, storage, and measurement. Therefore, to further ensure that lipidomics studies are reproducible, a reporting checklist has recently been published, which will help to standardize the description of lipidomic studies [9].

Cardiovascular diseases

The number of patients with cardiovascular disease (CVD) is increasing due to an ageing population and improving survival rate after CVD events due to better healthcare. However, at the individual level or in specific or comorbid subpopulations, the risk of an event can vary widely, requiring risk estimation tools to improve patient management. Effective risk stratification should identify at-risk individuals requiring more intensive or different treatment. Conversely, patients with favourable prognosis should be identified to avoid the overuse of medications and associated adverse events. For all these reasons, clinical lipidomics studies have been carried out to search for lipids with prognostic value as markers of CVD-related events and death. Different plasma ceramide ratios have been identified as significant predictors of CVD-related death in patients with stable coronary artery disease and acute coronary syndrome [25]. Additionally, polyunsaturated phosphatidylcholines (PC) in plasma, in combination with ceramides, can be used to further enhance the predictive ability and accuracy of prediction models [22]. Based on these and other studies, the CERT2 score for cardiovascular risk and event prediction using plasma ceramides (with possible improvement by including polyunsaturated PC) has been developed and its usefulness has already been summarized in previous reviews [42, 43]. It has also been shown that the CERT2 score is strongly associated with CVD, coronary artery disease, and stroke mortality in the elderly, while the association between these events and conventionally measured lipids (HDL-C, LDL-C, TG, total cholesterol) was weak [28]. Moreover, several mono-, di- and trihexosylceramides, alkyl- and alkenyl- phosphatidylcholines and other lipids improved current lifestyle and clinical risk factors for prediction of CVD events or death in patients with type 2 diabetes mellitus (T2DM) [23]. Multiple large cohort studies have provided similar results with regard to Mediterranean dietary intervention [24], heart failure risk [26] or all-cause mortality [27].

Ageing, obesity, diabetes and lifestyle factors

Changes in the lipidome due to ageing, lifestyle, obesity, and T2DM have been the focus of many lipidomic studies. Lipidomic fingerprints have been shown to be related to age with sex-specific differences (also influenced by the postmenopausal stage of life in females) in multiple studies [12, 13]. Some studies point to the interconnection of physiological factors (age, sex, BMI) to the alterations of lipidome [15]. Ratios between lipids can be used as potential biomarkers of age, as these sometimes describe the lipid metabolism dynamics better than sole lipid abundances [14]. Obesity is caused by excessive accumulation of fats, i.e. various lipids, which have also historically been determined as biochemical parameters in routine clinical tests. However, a detailed view using lipidomics can help understand how obesity relates to increased mortality or risk of developing T2DM or CVD. By using lipidomic analysis of human plasma, it was possible to predict BMI, total body fat percentage, waist circumference and waist-hip ratio and its associations with sphingomyelin differing by 1 double bond (suggesting involvement of specific desaturases) have been shown [16]. A mother-offspring study has pointed to the change in lysophospholipids in relation to birth weight and multiple other lipid classes were altered with respect to obesity risk [17]. Several other studies have shown that lipidomics can identify a subset of individuals at high risk for T2DM (even in combination with CVD) years before the onset of the disease, and thus the potential utilization of a lipid panel to identify these at-risk individuals should be considered [19], [20], [21, 23].

Cancer

Alterations in the lipidome have been demonstrated in almost all types of cancer. Not just in cancer cell lines in vitro, but also in patients at the level of the tumour, surrounding tissues and in peripheral biofluids [44]. However, only a limited number of studies have demonstrated the diagnostical potential of lipidomics in routine analysis or screening. Based on changes in serum/plasma lipids, it was possible to distinguish a group of colorectal and gastric cancer patients from controls [31] or lung cancer, including benign lung disease [30, 33, 34]. Another recent study showed increased levels of lactosylsulfatides and decreased levels of sulfatides with hydroxylated fatty acyls in patients with renal cell carcinoma compared to controls not only at the tumour tissue level but also in plasma and urine [32]. One of the most promising studies, which results are supported by several parallel methods carried out in multiple laboratories, showed dysregulation of very long-chain sphingomyelins, ceramides and (lyso)phosphatidylcholines in the plasma of pancreatic cancer patients [29]. The sensitivity and specificity of these markers exceed 90%, outperforming even conventional CA 19-9, especially at an early stage, and the accuracy was comparable to established diagnostic imaging methods [29]. Additionally, we would like to mention that another direction in the clinical lipidomics of cancer points to a significant impact on the metabolism of fatty acids, such as sapienic acid [45] and others, whereas the results and potential of these studies have already been summarized previously in another review [46].

Neurodegenerative disorders

To date, not many large-scale lipidomics studies focused on neurodegenerative diseases such as Alzheimer’s or Parkinson’s disease have been carried out. One study discovered lipid patterns associated with Alzheimer’s disease resulting from changes in ether lipids, sphingolipids (especially GM3 gangliosides), phosphatidylethanolamines, and triacylglycerols. Constructed subsequent multivariate lipid models improved the classification of disease [35]. Additionally, metabolites and lipids identified in cerebrospinal fluid combined into an elastic net regression model were able to discriminate between AD, PD and healthy controls [37]. In another study, the non-invasively collected sebum from people with Parkinson’s disease revealed a difference in its lipid composition compared to controls, independent of dopaminergic medication [36].

Non-alcoholic fatty liver disease

The mechanisms by which non-alcoholic fatty liver disease (NAFLD) progresses from simple steatosis to steatohepatitis (NASH) is not well understood. It is known that obesity is a major cause of NAFLD, but it is not known why fatty liver develops in lean individuals nor why NASH develops only in a small part of NAFLD patients. A study focusing on serum lipidomic analysis of patients (with histologically confirmed NAFLD based on liver biopsy) showed that a BMI-dependent lipidomic profile can be used to reliably distinguish patients with NASH from steatosis, and highlighted new targets for therapeutic intervention [38]. Analysis of serum lipidome of patients with NAFLD and MAT1A-KO mice with steatohepatitis revealed three subtypes of NAFLD (M, non-M and indeterminate subtype) and identified markers that differentiate steatosis from NASH in each subtype [39]. It has been shown that individuals with NAFLD had increased serum levels of triacylglycerols with low carbon number and double-bond content while lysophosphatidylcholines and ether phospholipids were decreased compared to healthy controls [40]. By lipidomic analysis of the patient’s liver tissues, it was possible to describe the transition from a healthy liver to NAFLD to NASH. Changes were observed in the concentrations of unsaturated cholesteryl esters, di- and triacylglycerols and sphingomyelin ratios [41].