Emerging Applications of Mass Spectrometry‐Based Metabolic Fingerprinting in Clinics

A number of recently developed mass spectrometry (MS)‐based metabolic fingerprinting analyses are promoting metabolic analysis systems into practical clinical applications, including but not limited to disease diagnosis. Herein, recent cutting‐edge research on mass spectrometry‐based metabolic fingerprinting analytical methodology and clinical applications are described. These developments resolve challenges regarding the practical clinical applications of MS‐based metabolic fingerprinting analysis systems, such as rapid signal readout, high throughput and direct MS analysis, and intelligent data mining of complex biological samples.

(CSF), and feces. Notably, sample preparation is highly dependent on the choice of MS techniques. Numerous MS techniques have been used for MS-based metabolomics applications, such as liquid/gas chromatography (LC/GC) MS, laser desorption ionization mass spectrometry (LDI) MS, and rapid evaporative ionization (REI) MS. Chromatography-MS (LC/GC MS) requires mazy sample pretreatment procedures (e.g., desalination, derivatization, and chromatography for metabolomics). [11] LDI is a soft ionization process, which generates mainly singly charged gaseous ions from solid or liquid analytes under the laser. [12] LDI MS is tolerant of salts, proteins, and contaminants, which makes it free of complex and time-consuming sample preparations when dealing with complex biological samples. REI produces gaseous ions by thermal evaporation, which is a by-product of the electrosurgical process (i.e., radiofrequency wave). Furthermore, REI is an ambient ionization technique, thus can achieve direct real-time analysis of biological fluids with almost no sample pretreatment. [13] After MS data acquisition, multivariate statistical methods are required to interpret the information obtained from MS datasets, [14] including unsupervised clustering and supervised machine learning (ML) model. Unsupervised methods, such as principal component analysis (PCA), [15] cluster samples based on the variation of metabolic fingerprints without the prior knowledge of the sample label. Supervised ML (e.g., partial least squares discriminant analysis [PLS-DA], [16] random forest [RF], [17] elastic net [EN] [18] ) supports the construction of predictive models based on the learning of the data matrix (X ) toward a matrix (Y ) that contains label information, such as disease or healthy control. Due to the ability to rapidly process complex and heterogeneous data, ML has been immensely used for statistical analysis of MS data toward disease diagnosis and biomarker discovery. Notably, a power analysis is required before data analysis to determine the minimum sample size required of statistically validated data. [19] 3. Clinical Application of Mass Spectrometry-Based Metabolic Fingerprinting MS-based metabolic fingerprinting have been conducted on the following diseases: cancers (including brain, [8b] lung, [20] pancreatic, [21] colorectal, [22] gastric, [8a,23] cervical, [10,24] and gynecological cancer [25] ), inflammatory or infectious disease, [9a,26] kidney disease, [2d] and cardiovascular/cerebrovascular diseases. [27] Compared with conventional diagnostic imaging and biopsy methods, MS-based metabolic fingerprinting affords desirable accuracy with minimum invasiveness and low costs.

LC/GC MS-Based Metabolic Fingerprinting
Coupled with chromatographic separation (e.g., GC, LC), MS has emerged as an essential instrument in the clinical laboratory for the qualitative or quantitative metabolic analysis of biological samples with high sensitivity and specificity. [11,28] A combination of untargeted and targeted LC coupled tandem mass spectrometry was utilized by Song et al., [9a] interpreting the plasma metabolome (404 polar metabolites) and lipidome (598 lipids) quantitated by 71 internal standards in coronavirus disease 2019 (COVID-19) group and healthy group (Figure 1a). A plasma biomarker panel of 10 metabolites (with the smallest p-value) presents the meaningful biological functions, which were suitable for COVID-19 diagnosis, achieving an AUC of 0.975 by a logistic regression model. To identify potential early-stage chronic kidney disease (CKD) metabolite biomarkers, Chen et al. [29] reported the high-throughput metabolomics of 2155 subjects by ultra-performance liquid chromatography coupled with mass spectrometry (UPLC MS) (Figure 1b). Five biomarkers were identified using an advanced bioinformatics approach (including machine learning and statistic feature reduction) and multistep validations. Notably, Chen et al. first identify 5-MTO as a critical biomarker for early-stage CKD detection. Yu et al. [23] conducted untargeted serum metabolomic analysis to discover potential biomarkers for gastric cancer diagnosis via UPLC MS. According to the filter criteria (p < 0.05 of t-test, fold change >1.5 or <2/3, and variable importance in the projection (VIP) >1 in PLS-DA), a panel of three metabolites was selected as diagnostic markers for gastric cancer, showing high potential diagnostic value with AUC of 0.986 by PLS-DA. Similarly, Ghosh et al. [30] performed GC MS-based serum metabolic fingerprinting and systemic inflammatory profiling of asthma COPD overlap (ACO), finding eleven metabolites significantly regulated in ACO compared with COPD or asthma only by OPLS-DA, including fatty acids, energy metabolites, and cholesterol. To investigate the metabolic diversity of different cancer, Li et al. [31] applied LC MS for metabolic profiling (225 metabolites) of 928 cell lines from more than 20 cancer types, confirming the existence of linage distinctions (148 out of 225 metabolites) in model cancer cell lines.
Besides disease diagnosis and mechanism investigation, chromatography-MS has also been applied in the research of www.advancedsciencenews.com www.advintellsyst.com drug metabolism study. To investigate the attenuation mechanism of Caowu (traditional Chinese medicine) compatibility with Yunnan Baiyao, Ren et al. [32] adopted UPLC MS-based serum metabolomics of male rats (control group and treatments groups), identifying 63 endogenous metabolites related to the potential toxicity of Caowu according to p < 0.05 in t-test and VIP > 1 in PLS-DA. Further, four core attenuated metabolic pathways were identified, revealing the possible mechanism of inflammatory inhibition. Zimmermann et al. [33] applied LC MS for identification of bacteria-produced drug metabolites by evaluating the capability of 76 human gut bacteria to metabolize 271 oral medicines for mapping drug metabolism of the human microbiome.
Chromatography-MS has been widely utilized as a conventional MS tool for disease diagnosis, disease mechanism investigation, and drug metabolism study toward clinical application. However, sample cleanup (e.g., desalination) and chromatographic separation (e.g., LC/GC) procedures lead to laborand time-consuming sample preparation and large sample consumption, limiting the application of LC/GC MS for the point-of-care test.

LDI MS-Based Metabolic Fingerprinting
LDI MS has exhibited great potential for metabolic analysis in terms of the advantages of high throughput, fast analysis, and minimal sample preparation owing to the tolerance of salts, proteins, and contaminants toward large-scale clinical use. Noteworthy, the performance of LDI MS is highly dependent on the selection of matrix. The traditional organic matrices limit the metabolic analysis in low mass range (m/z < 700) by LDI MS because of the strong background noise. Recently, LDI MS-based metabolic fingerprinting has been applied for various diseases diagnosis, owing to the development of well-designed nanocomposites as matrices and adaptive application of the machine learning algorithms for data analysis.
In recent years, cancer studies using LDI MS-based metabolic fingerprinting have been performed by many research groups  [9a] Copyright 2020, Elsevier. b) Identification of 5 serum metabolites associating with CKD progression. Reproduced with permission. [29] Copyright 2019, Springer Nature.
www.advancedsciencenews.com www.advintellsyst.com (including our group). Cao et al. [8b] reported a high-performance brain tumor diagnosis by machine learning of LDI MS-based metabolic fingerprinting with Pd-Au synthetic alloys as matrix, achieving diagnostic AUC of 0.917 with accuracy of 89.9%, sensitivity of 94.0%, and specificity of 85.7% (Figure 2a). The optimized synthetic alloy enhanced the LDI efficacy and achieved direct detection of 100 nL of serum in seconds with high salt tolerance and protein endurance. Further, four biomarkers were selected with gradual changes to the healthy state during the radiotherapy process, indicating the potential of the platform for treatment response study. Huang et al. [20] conducted machine learning of serum metabolic fingerprinting for early-stage lung adenocarcinoma (LA) diagnosis. Ferric particle-assisted LDI MS achieved direct metabolic patterns extraction using only 50 nL of serum with diagnostic AUC of 0.921, sensitivity of 90%, and specificity of 93% by sparse regression machine learning, which is superior to traditional imaging and biopsy methods. A panel of 7 biomarkers (i.e., uracil, uric acid, 3-hydroxypicolinic acid) were confirmed beneficial for discriminating early-stage LA from healthy controls, achieving AUC of 0.894. Su et al. [8a] designed mesoporous PdPtAu alloys as matrix with enhanced optoelectric/ thermal effect and mesoporous surface morphology, yielding sensitivity of 92.0% and specificity of 92.0% for early diagnosis of gastric cancer. Pei et al. [25] constructed a FeOOH@Metal-Organic Framework-assisted LDI MS platform to diagnose gynecological cancers with AUC of 0.921-0.997, sensitivity of 97.2-98.7%, and specificity of 97.3-98.7% using orthogonal partial least squares discriminant analysis (OPLS-DA) (Figure 2b). The FeOOH@Metal-Organic Framework composites showed selective metabolic analysis and enhanced ionization efficiency Reproduced with permission. [25] Copyright 2020, John Wiley and Sons. c) COF@Au reveals specific serum metabolic fingerprints as the point of Crohn's disease diagnosis. Reproduced with permission. [26] Copyright 2021, John Wiley and Sons. d) Rapid deep learning-aided diagnosis of stroke by serum metabolic fingerprint based multi-modal recognition. Reproduced with permission. [27a] Copyright 2020, John Wiley and Sons.
www.advancedsciencenews.com www.advintellsyst.com owing to size-exclusion effect and good light response, thereby achieving direct metabolic fingerprinting extraction using 1 μL of serum by LDI MS without any pretreatment. Besides cancer diagnosis, LDI MS-based metabolic fingerprinting has also been widely used in the research of other diseases, such as Crohn's disease and cardiovascular/cerebrovascular diseases. Yang et al. [26] designed a covalent organic framework@Au (COF-V@Au) for serum metabolic fingerprinting of Crohn's disease patients and healthy controls, achieving of AUC of 0.984 by OPLS-DA (Figure 2c). The COF-V@Au matrix possesses a large specific surface area and nano-porous structure to provide abundant sites for metabolites adsorption, achieving great ionization efficiency. Xu et al. [27a] conducted deep learning to integrate serum metabolic fingerprinting extracted by nanoassisted LDI MS and clinical index of stroke patients, achieving enhanced AUC of 0.845 for stroke screening compared with only serum metabolic fingerprinting or clinical index (Figure 2d). This work highlighted the importance of multi-modal data integration. A novel deep learning model-based feature selection approach was developed, addressing the limitation of deep learning for feature selection. Zhang et al. [27b] developed a deep stabilizer to map native LDI MS results to high-quality results, focusing on the data reproducibility for one accurate and stabilized performance. The established protocol afforded an AUC of 0.86 and sensitivity of 0.98 and CV < 10% at the 74th percentile specificity for the diagnosis of coronary heart disease.
So far, various nanocomposites as matrices for LDI MS detection and machine learning algorithms for data analysis have been developed to resolve multifarious biological fluids of different types of diseases. LDI MS-based metabolic fingerprinting shows high potential for large-scale clinical application of various diseases, owing to the high throughput, rapid analysis, and high diagnostic performance.

REIMS-Based Metabolic Fingerprinting
Ambient ionization allows direct MS analysis of sample surface under atmospheric pressure with almost no sample pretreatment, including but not limited to REIMS (rapid evaporative ionization mass spectrometry).
REIMS utilizes real-time metabolic analysis of the aerosol produced from various biological materials without sample preparation, including biological fluids, cell culture lines, human tissue, and microorganisms, [13a] indicating the capability of real-time diagnosis and especially intraoperative detection.
Innovative REIMS technologies such as intelligent knife (iKnife) have been developed for rapid and simple biological tissue-type identification, [34] focusing on solving the limitation of general histological analysis, such as time-consuming procedure and high pathological-to-pathological variance. Koundouros et al. [9b] conducted metabolic fingerprinting of aerosolized tissue material using iKnife, achieving an accurate diagnosis of PIK3CA MUT breast cancers (accuracy of 90%) with random forest as the classifier. Further, metabolic phenotyping by REIMS was adopted for molecular markers prediction. The results turned out that oncogenic PIK3CA led to an upregulation of arachidonic acid and overproduction of derived eicosanoids (Figure 3a). Taking advantage of iKnife, Tzafetas et al. [10] achieved accurate machine learning-aided discrimination of healthy from premalignant and invasive cervical lesions (accuracy of 100%), leading to improved postoperative oncological outcomes and fertility preservation of women with cervical cancer (Figure 3b). Further, significant MS peaks contributing to cancer-to-normal separation were revealed by univariate analysis. REIMS/iKnife can make nearly real-time (1-2 s) predictions of relevant tumor characteristics based on metabolic fingerprinting, thus providing a new approach for cancer diagnosis and precision treatment.
Besides tissue-type identification, REIMS has also been applied to biological fluid-based disease diagnosis toward clinical application. Van Meulebroek et al. [35] established a laser-assisted (LA) REIMS platform for rapid metabolic fingerprinting of feces (<10 s without sample pretreatment), possessing the capability of metabolic diseases diagnosis and treatment evaluation (Figure 3c). This platform achieved accurate discriminative fingerprinting (>90%) for type 2 diabetes and controls using OPLS-DA (Q 2 of 0.734) and successfully cross-validated by the time-consuming and redundant LC-MS, confirming the reliability of the LA-REIMS for fecal metabolic fingerprinting. Wijnant et al. [36] optimized LA-REIMS-based salivary metabolic fingerprinting for metabolic discrimination between adolescents with normal weight and overweight/obese, achieving the accuracy of 97.1% (sensitivity of 100% and specificity of 94.3%) using OPLS-DA (Q 2 of 0.808). Cameron et al. [37] utilized pulsatile carbon dioxide (CO 2 ) laser for thermal desorption of analytes, achieving REIMS data acquisition with high signal-to-noise ratios toward metabolic biofluid phenotyping (accuracy of 94%). Paraskevaidi et al. [24] employed LA-ESIMS in cervical cancer screening by metabolic fingerprinting of liquid-based cytology (LBC) sample, achieving AUC of 0.916 with sensitivity of 94% and specificity of 83% by random forest algorithm. This platform has also been proved useful for providing simultaneous information of disease severity, affording AUC of 0.867 with the sensitivity of 91% and specificity of 73% in discriminating high-grade pre-invasive from normal precancerous cells.
In summary, REIMS-based metabolic fingerprinting allowed the construction of real-time predictive models reflecting the metabolic perturbations of various diseases without sample pretreatment, resulting in a powerful candidate for intraoperative detection and large-scale screening and diagnosis.

Conclusion
Mass spectrometry is a powerful tool for metabolic fingerprinting of biological fluids or tissues for clinical application owing to its high sensitivity and resolution. This minireview intends to provide an overview of the research progress on MS-based metabolic fingerprinting in recent three years, focusing on three typical MS techniques (LC/GC MS, LDI MS, and REIMS) and their clinical applications.
LC/GC MS is a dominant tool for metabolic analysis toward disease diagnosis, disease mechanism investigation, and drug metabolism study. However, sample cleanup (e.g., desalination) and chromatographic separation (e.g., LC/GC) procedures lead to labor-and time-consuming sample preparation and large sample consumption, limiting the application of LC/GC MS for www.advancedsciencenews.com www.advintellsyst.com point-of-care test. By contrast, LDI MS requires minimal sample preparation due to the selective and sensitive ionization with the assistance of the recently developed nanocomposites as matrix. LDI MS-based metabolic fingerprinting shows high potential for large-scale clinical application of various diseases, owing to the high throughput, rapid analysis, and high diagnostic performance. REIMS requires almost no sample pretreatment owing to the ambient ionization process. REIMS-based metabolic fingerprinting achieves real-time prediction of the metabolic perturbations of various diseases, making it a powerful candidate for large-scale screening and diagnosis, especially intraoperative tissue-type identification. Although numerous progresses have been made by MS-based metabolic fingerprinting toward clinical application, there are still some challenges to be addressed. First, most of the recently developed applications by MS-based metabolic fingerprinting are focusing on a satisfactory diagnostic performance toward disease diagnosis. Expanding the application for therapeutic response assessment and prognosis is a promising direction for precision medicine, especially for LDI MS and REIMS. Second, it remains a great challenge for the identification and biological interpretation of metabolic biomarkers. With the selected signals as differential features, it is necessary to understand the biological mechanism of how the biomarkers interact with the biological systems. Furthermore, rather than the single metabolic fingerprinting analysis, the integration of multi-omics datasets (such as metabolomics, proteomics, and genomics) would be beneficial to interpret the complex biological systems for clinical application. Developing efficient multi-omics data acquisition platforms and intelligent data interpretation algorithms (such as machine learning) are efficient methods to address this issue, which will be a breakthrough point for MS-based metabolic fingerprinting in clinical use.  [9b] Copyright 2020, Elsevier. b) The iKnife and its intraoperative diagnostic advantage for the treatment of cervical disease. Reproduced with permission. [10] Copyright 2020, National Academy of Sciences. c) Metabolic phenotyping of feces for type 2 diabetes diagnosis. Reproduced with permission. [35] Copyright 2020, American Chemical Society.