A biomarker can be a biomolecule that indicates a change in biological structures or functions and can be used to evaluate the state of a living organism (Silberring and Ciborowski, 2010). Since most clinical decisions are based on laboratory tests, biomarkers can influence clinical outcomes. Many studies aim to find specific biomarkers for certain diseases, and mass spectrometry is a common approach to search for them. In short, mass spectrometry instruments measure the mass-to-charge ratio ions in the gas phase of proteins, peptides, metabolites, lipids and other molecules from a given sample (i.e., cells, biofluids and tissues) to determine their molecular weight, structures, and quantities (Aksenov et al., 2017). Since this technique allows the identification and quantification of thousands of molecules, it is one of the most used technologies in “omics” studies. Proteomics is the science that aims to identify and quantify the complete set of proteins, their localization, interactions and post-translational modifications present in cells, tissues, biofluids, or other biological materials from an organism (Aslam et al., 2017). Several studies have applied proteomics to identify and quantify circulating and tissue biomarkers that could help in the clinical decision-making process (McArdle et al., 2021) and understand the mechanism of SARS-CoV-2 infection (Bojkova et al., 2020; Kim et al., 2023). Metabolomics is defined as the qualitative and quantitative study of the complete set of small molecules (metabolites) and their interactions within a biological sample (Hasan et al., 2021). Alterations in metabolic pathways was also demonstrated to be linked with many human diseases, including viral infection (Birungi et al., 2010; Chandler et al., 2016; Ussher et al., 2016; Uchiyama et al., 2017). Moreover, metabolomics has been used to identify and prioritize diagnostic and prognostic COVID-19 biomarkers (Shen et al., 2020; McCreath et al., 2021). Lipidomics is the study of the lipidome, which can be defined as the set of lipid species expressed in a biological system, such as biofluids, tissues or cells (Han and Gross, 2003; Yang and Hhan, 2016). Lipids are known to play a central role in viral infections, as they are required to compose the structures for both the virus and the cell (Abu-Farha et al., 2020). The proteome, metabolome, and lipidome are dynamic and modified based on biotic and abiotic insults such as viral infections. This review will focus on mass spectrometry-based omics analyses; however, genomics and transcriptomics are also considered when used in a multiomic approach.

Mass spectrometry-based omics approaches were used to characterize the SARS-CoV-2 infection using a wide range of biological samples, such as plasma, serum, urine, organ autopsies, airway mucus, bronchoalveolar lavage fluid, and saliva (Zeng et al., 2021; Liu Y. et al., 2022; Zhang Z. et al., 2022; Mansouri et al., 2022; Muñoz-Prieto et al., 2022). The molecular landscape of these studies contributed to elucidating the SARS-CoV-2 infection mechanisms and suggested many protein candidates as biomarkers (Shen et al., 2020; Danlos et al., 2021; Ren et al., 2021; Shi et al., 2021; Spick et al., 2021; Liu J. F. et al., 2022; Schuurman et al., 2022). Proteomics of serum samples obtained by LC-MS/MS determined a total of 93 proteins and 204 metabolites differentially expressed when compared between severe and non-severe COVID-19 patients (Shen et al., 2020); The urinary proteome mapped by LC-MS/MS identified 56 proteins differentially expressed when comparing mild and severe COVID-19 patients (Li et al., 2020). Many studies have focused on selecting individual biomarkers among the identified regulated proteins, mainly associated to as platelet degranulation, acute inflammatory response, and complement activation-related proteins (D’Alessandro et al., 2020; Li et al., 2020; Park et al., 2020; Shen et al., 2020; Liu et al., 2021; Mohammed et al., 2022). Due to the complexity of the disease, a single biomarker may be insufficient to indicate with accuracy the changes in the host system and if the disease is progressing to a severe case. Machine learning can be applied to search for specific biomarkers among hundreds of candidates and use sets of biomarkers to build robust classification models to classify patients or predict if a patient will progress to severe COVID-19. The combination of mass spectrometry-based omics and machine learning will be detailed below in the context of COVID-19, Table 1.

Machine learning is a technique that allows a computer to recognize patterns and thus “learn” from a given data. It is a heavily statistic-based technique that uses algorithms to find exploitable regularities in data (Laponogov et al., 2021). The most common applications are supervised and unsupervised learning (Rajoub, 2020); the former will be the focus of this work. In a given dataset, supervised learning assumes the existence of an unknown function y = f x or an unknown probabilistic distribution P y ∨ x that dictates the behavior of input features X and an output label y. Machine learning algorithms find this relationship by being trained using a set of samples where the values of the features are assigned to a label y i = f x i (Kelchtermans et al., 2014). For instance, proteomics and machine learning can be applied to develop a method to diagnose COVID-19 using plasma samples. In order to use machine learning for this purpose, it is needed to acquire the proteomic profile of each patient using a mass-spectrometry technique, being the proteins (features) identified together with their quantitative information. The features X are assigned to a label y (if this set of features are from a COVID-19 positive or negative patient). Additionally, specific machine learning algorithms, such as decision trees (Mann et al., 2021), are needed to identify the most important proteins for group discrimination from the features set, thus finding potential biomarkers for COVID-19 infection. The studies reported in Table 1 used supervised learning as a tool for the prognosis and diagnosis of COVID-19 and will be further discussed here.

A classification task can help discriminate groups, but training accurate models to perform such a task can be challenging. Sample collection, sample preparation, data acquisition, and preprocessing should be carefully planned to generate models that can generalize outside the training dataset. For instance, if a model is being trained for COVID-19 diagnosis, it is necessary to include patients with other infectious diseases in the control group to ensure that the model will be specific and will not misclassify other virus-infected patients as COVID-19 positive. Sample preparation and data acquisition are two variables that can be explored to maximize the classification performance of a model. Different protein fractions and preparation methods lead to different protein identifications, which can be more specific or less specific for the classification task. Data preprocessing is essential to ensure that the data will be comparable and avoid overfitting during model training (when the model “memorizes” the training dataset and performs poorly in the test dataset).

The most common approach described in the literature for COVID-19 prognosis is the use of omics data to train models for predicting if a patient has the disease or if they will develop severe COVID-19. One of the earliest studies involving this approach used serum samples of 18 non-severe and 13 severe COVID-19 patients to obtain a proteomic and metabolomic profile using LC-MS/MS and UPLC-MS/MS, respectively. Then, the identified proteins (894) and metabolites (941) were used to train a random forest algorithm for sample classification with a 10-fold cross-validation, a resampling technique used to train and test a model several times using the same dataset, which is helpful for small datasets. This resulted in a model with a high Receiver Operating characteristic curve area under the curve of 0.957 (high values indicates that the model is good at distinguishing between the classes); the authors proceeded to perform a validation step with two independent cohorts, one containing 10 patients, which 7 of them were correctly classified and the other containing 19 patients which 16 of them were correctly classified (Shen et al., 2020).

Usually, machine learning techniques work best with larger datasets since it is hard to accurately evaluate the model with a low error (Combrisson and Jerbi, 2015). It is best to use data obtained by the same method for model validation. The dataset size limitation is challenging to address, mainly if acquiring data is time-consuming and expensive but also when the samples are difficult to obtain. A low number of samples in an omics study usually leads to a substantially higher number of features than samples, which is called the curse of dimensionality, causing most machine learning methods vulnerable to overfitting (Mirza et al., 2019). Increasing the number of samples and reducing data dimensionality can reduce model overfitting, usually resulting in reduced training accuracy and improved performance in test sets. For instance, plasma proteomics of 18 non-severe and 33 severe COVID-19 patients using label-free LC-MS/MS identified 1200 proteins, of which 38 were regulated. The authors used the regulated proteins to train five machine learning algorithms for classification, achieving an accuracy of 88% using a support vector machine algorithm; however, they reduced the number of features (proteins) by performing a PLS-DA (Partial Least-Squares Discriminant) analysis and retrieving the 20 most important proteins. They trained the same models again with 20 features and achieved an accuracy of 84% (Suvarna et al., 2021). It is difficult to confirm if the overfitting was reduced with the dimensionality reduction without a test dataset, but the results obtained after the PLS-DA analysis indicate that the model performance before dimensionality reduction was slightly overestimated.

Both studies used as examples above use a mass spectrometry technique that can identify the proteins present in the samples, which is useful for biomarker discovery using machine learning. As we stated earlier, identifying the most important features for classification requires the use of less complex and interpretable models such as Bayesian, decision trees and linear models; for instance, using linear regression models allows us to access the weights that describe the coefficients of a curve, thus indicating which feature has more influence (Mann et al., 2021). This is useful in omics since it provides a way to search biomarkers among thousands of molecules without manually analyzing them. This is useful in omics since it provides a way to search biomarkers among thousands of molecules without manually analyzing them. This approach yielded a list of proteins, metabolites, and lipids that can be used as biomarkers for COVID-19 diagnosis and prognosis, including Serum Amyloid (SAA1 and SAA2), C-reactive Protein (CRP), SERPINA3, SERPING, FGG, Kynurenine, Ursodeoxycholic acid, Carnitine 3:0, Diglyceride 36:0, IL6, Gal-9, ITGB1BP2 and many other biomolecules (Shen et al., 2020; Suvarna et al., 2021; Yaşar et al., 2021; Ahern and COvid-19 Multi-omics Blood ATlas COMBAT Consortium, 2022; Castañé et al., 2022).

The use of proteomics and machine learning for the prognosis of COVID-19 has been demonstrated to be accurate compared to the use of clinical parameters. A study by Demichev et al. (2021) showed that a model trained with proteomic data achieved a similar performance of a model trained with clinical diagnostic parameters (ROC-AUC = 0.98 and ROC-AUC = 0.97 respectively); also, they demonstrated that the machine learning models significantly outperformed other predictive scores of COVID-19 risk factors such as age, BMI, CCI, and molecular predictors such as CRP and IL-6 Demichev et al. (2021). Another study also demonstrated that proteomic data could achieve similar performance of clinical parameters for prognosis classification. Sardar et al. (2021) achieved an accuracy of 89.47% in a model trained with clinical parameters and 89.01% in a model trained with proteomic data. This supports the relevance of proteomic data and machine learning as an important tool for COVID-19 Sardar et al. (2021).

Since mass-spectrometry can provide information on different types of “omics,” each one of them depicting the disease in different ways and explaining the system either by regulation of proteins, lipids or metabolites, the combination of different types of “omics” data can elucidate the causative changes that are responsible for the disease (Hasin et al., 2017). In this context multiple omics integration can provide more in-depth knowledge of what changes occur during COVID-19 infection, including omics data that are not obtained by mass spectrometry, such as transcriptomics and genomics. For instance, it is already reported that the addition of proteomic data to genomic and transcriptomic data help elucidate a disease, demonstrating that the information at protein level complemented the genomics information, leading to the identification of multiple pathways and processes taking place during disease pathogenesis (Subramanian et al., 2020). However, data integration between multiple omics approaches can be challenging, especially due to the heterogeneity of the different omics data and the amount of data generated by them (Subramanian et al., 2020). Machine learning can be applied to analyze multiomics data to address some of these limitations. Although machine learning is still new in multiomics data, it was already employed in several studies to understand several diseases (Reel et al., 2021).

For COVID-19 research, the integration of several omics data helped the elucidation of several pathways. A study performed by Overmyer et al. (2021) obtained the proteome, metabolome, lipidome and leukocyte mRNA expressions from 102 COVID-19 patients and 26 non-COVID-19 patients, where they found 219 biomolecules that were associated with COVID-19 severity. They demonstrated that the multiomics approach yielded a much more detailed picture of the system behavior during COVID-19 infection, showing the regulation of biological processes related to lipid transport, acute phase response, neutrophil degranulation and blood coagulation. Interestingly, the authors performed a correlation analysis between all omics data (which they called “cross-ome analysis”) to uncover the connections between different biomolecule classes. The multiomics data were also used to train machine learning models for COVID-19 prognosis. They demonstrate that the data combination yielded higher average precision and AUC-ROC (area under the receiver operating characteristic curve) than each dataset.

Another study by Byeon et al. (2022) also reported an increase in the performance of predictive models when merging multiple omics datasets. They quantified 1463 cytokines and circulatory proteins, 902 lipids, and 1018 metabolites from 455 COVID-19 patients belonging to three groups depending on disease severity. They found that the models trained with all omics data outperformed each model in the held-out test dataset. The same was observed in another study where machine learning models were trained to predict patient survival, Richard et al. (2022) obtained the proteomic and metabolomic profile of 40 COVID-19 patients, they further trained models for patient survival prediction; the model trained using the proteomics and metabolomics data outperformed both datasets individually.

Although multiomics approach can provide more in-depth knowledge of the disease and generate models with higher performance, the data analysis still imposes some challenges. As pointed out in a review made by Reel et al. (2021), problems such as data heterogeneity, especially when merging dataset that uses different types of normalization and scaling (i.e., proteomics and transcriptomics), class imbalance, and high data dimensionality (when the number of features is much higher than the number of observations) should be taken into consideration to develop appropriate pipelines for model training. The use of deep learning techniques for multiomics has increased in popularity in the past years, especially in analyzing large-scale multiomics datasets (Reel et al., 2021). Deep learning has also demonstrated to have superior performances when dealing with non-linear problems, having great advances in cancer survival research (Chai et al., 2021). Deep learning is a subfield of machine learning with a structure composed of multiple layers of non-linear processing units for feature extraction and processing (Shinde and Shah, 2018). For multiomics applications, deep learning is more appropriate, since it can analyze and extract patterns from large amounts of data obtained from different sources (Shinde and Shah, 2018). In omics, Deep Neural Networks, Recurrent Neural Networks and Convolutional Neural Networks have been applied to predicting DNA and RNA sequence structure, drug design, protein structure/function and protein interactions (Zhang et al., 2019). Moreover, deep learning was also used to develop predictive models for the diagnosis and prognosis of many types of cancers (Chai et al., 2021) (Chaudhary et al., 2018; Xie et al., 2019; Tong et al., 2020; Rong et al., 2021).

For COVID-19, multiomics integration with deep learning is still in its early stages. However, the technique’s potential has already been demonstrated. A study developed by Rahnavard et al. (2022) obtained the proteomic and metabolomic profile of 28 patients with severe COVID-19 and compared it to a control group composed of 28 healthy patients, 25 non-COVID-19 patients with similar clinical symptoms and 25 non-severe COVID-19 patients. They evaluated the performance of three machine learning models - KNN (k-nearest neighbors), RF (random forest) and LR (linear regression) and a deep neural network (DNN) for COVID-19 prognosis. Their results demonstrated that DNN outperformed the other models’ overall accuracy and was better at classifying the severe COVID-19 patients as well.

In summary, some works have used machine learning models for multiomics data (Sun et al., 2021; Richard et al., 2022; Liu et al., 2023), but the number of features used in these studies were relatively low (less than 300), which could justify the use of machine learning. Another point to consider is the necessity for sample processing, choosing deep learning can alleviate this need, since steps such as feature reduction could be omitted, while choosing machine learning models would require extensive data treatment to reduce the features. The integration of multiomics data for COVID-19 holds great promise.

One of the main reasons to search for biomarkers is to use these molecules as a surrogate of the disease and its progression to develop diagnostic and prognostic methods.

A fast, low cost and accurate technique for proteomic data acquisition is essential to maximizing the potential of machine learning for diagnosis and prognosis. Matrix-assisted laser desorption/ionization-time of flight mass spectrometry (MALDI-TOF MS) is a technique that can analyze biomolecules without prior chromatographic separation; it is easy to use, requires a low sample amount, and sample preparation is simplified (Hajduk et al., 2016). However, MALDI-TOF MS does not provide the details and scale of protein identifications as other mass spectrometry techniques. Instead of protein/peptide identifications, the MALDI-TOF MS spectra mainly provide a series of mass-to-charge ratio peaks corresponding to the biomolecules analyzed and their intensities. The set of identified peaks composes a profile with distinctive features depending on the analyzed sample; for instance, a serum sample from a severe COVID-19 patient may have a different MALDI-TOF MS profile than a mild COVID-19 patient. While it may not be possible to visually recognize patterns in these spectra for diagnosis/prognosis purposes, a machine learning approach may be sufficient to perform this task with high precision. Thus, acquiring proteomic spectra using MALDI-TOF MS combined with training machine learning models is a tool to detect the patterns of each group of interest. However, this approach comes at some costs, such as a lower identification rate (a low number of peaks can be assigned to a specific biomolecule), and low precision in quantifying and sampling a subset of molecules due to ionization suppression (Hajduk et al., 2016).

The first work that performed this analysis for COVID-19 used 362 nasal mucous secretion samples (nasal swabs) for diagnostic purposes using the proteomic profile acquired by MALDI-TOF MS (Nachtigall et al., 2020). A total of 211 COVID-19-positive and 151 negative samples were processed, and the spectra obtained were used to train six different machine-learning algorithms. In this case, instead of proteins, the data consisted in m/z peaks that represent unidentified proteins. The authors found 88 protein peaks that were used to train the models. Their findings indicate that a Support Vector Machine Radial was the best algorithm for COVID-19 diagnosis, with an accuracy, specificity, and sensitivity of 93.9%, 94.7%%, and 92.6%, respectively (Nachtigall et al., 2020). The number of samples is higher than in most works that use other mass spectrometry techniques. This is facilitated by MALDI-TOF MS experiments requiring less resources than a LC-MS/MS, for example. The absence of protein identification did not impede training the classification models. This aspect highlights the advantages of using MALDI-TOF as a data acquisition method. However, the authors did not validate the findings using a test set. Another similar study performed nasal swabs MALDI-TOF MS proteomics of 226 samples split into 82 for training/validation and 117 for the test (assessing the model’s true performance) (Tran et al., 2021). The authors used an automated platform denominated Machine Intelligence Learning Optimizer (MILO) to train and test the models (Jen et al., 2021). This platform generated 379,269 models, achieving an accuracy of 96.6%. A study with serum samples also achieved high performance on the training set for COVID-19 diagnosis, reaching a value of 99% accuracy, 98% sensitivity, and 100% specificity in a dataset of 298 samples (146 COVID-19 positives and 152 controls) (Yan et al., 2021).

Although MALDI-TOF MS allowed the analysis of large samples, some aspects still need to be optimized, such as data preprocessing.

Optimizing a pipeline involving mass spectrometry and machine learning for biomarker discovery in the context of diagnosis and prognosis is essential for developing high-performance models that can be generalized outside the training data. Although the studies mentioned in this review are promising, the combinations and impact of different analytical sample preparation, data acquisition, and machine learning methods are still poorly explored. Several protein fractions could be analyzed for COVID-19, such as glycoproteins, phosphoproteins, protein in vesicles, and protein aggregates, which were already described to influence COVID-19 and other diseases such as cancer (Harsha and Pandey, 2010; Nagaraj et al., 2010; Drake and Kislinger, 2014; Pan et al., 2016; Basile et al., 2022; Pongracz et al., 2022). This variety of biomolecules can be accessed through different sample preparation methods such as solid phase extraction (SPE), magnetic beads, silica beads, ultrafiltration, dialysis, hydrophilic interaction chromatography (HILIC) and metal oxide affinity chromatography (MOAC) for example, (Hajduk et al., 2016). Also, samples such as plasma and serum could benefit from depletion methods to remove the highly abundant proteins, since the top 20 most abundant proteins compose roughly 97% of the total protein mass in plasma samples, for example, (Anderson and Anderson, 2002). The depletion of these proteins results in increased sensitivity of the mass spectrometer to detect less abundant proteins (Qian et al., 2008).

Although sample preparation is important, the patient cohort should be carefully chosen before starting the experiments. Patients with a certain disease must be paired with individuals that do not have that disease but present similar symptoms to avoid bias. For instance, severe COVID-19 patients should be paired with patients suffering from a severe condition caused by other respiratory viruses to ensure that the analysis will yield specific markers for severe COVID-19. Selecting the appropriate number of samples is also important since it improves the accuracy of machine learning algorithms. Sometimes it is not possible to have a large patients cohort thus, it is important to evaluate the epidemiological and analytical factors involving the studied disease and the mass spectrometry platform used to acquire the data, in order ensure that the protein expression variability will be enough to capture differences or the number of samples should be larger (Nakayasu et al., 2021). A power analysis could also be performed to find a minimal sample size (Cohen, 1992).

Sample storage is also important to ensure the integrity of the experiments. It is recommended to aliquote serum samples for single use since multiple cycles of freezing and thawing can result in protein degradation, resulting in low-quality samples that will fail or limit the identification of potential biomarkers (Valo et al., 2022). For urine, it is recommended to store the samples immediately after removal of cells or debris at −20°C and subsequently at −80°C (Thongboonkerd, 2007). Implementing appropriate pre-analytical factors such as biofluid collection methods and storage conditions, can be based on guidelines reported in the literature.

MALDI-TOF MS target plate preparation is another step that requires optimization. A process known as the analyte suppression effect occurs in complex biological samples, which is the competition between co-existing components of the sample for desorption and ionization (Lou et al., 2015). This means that the ionization efficiency will vary with the presence of ions that can suppress the peak intensity of other ions; also, low reproducibility of peak intensity can be caused by matrix amount and crystallization process variability (Hajduk et al., 2016). Therefore, it is advisable to perform a standardization step prior to sample processing. In this standardization step, the target molecules that will be measured in the experiment should be evaluated since the MALDI matrices will impact this step. For instance, the alpha-cyano-4-hydroxycinnamic acid matrix is suited for peptides, while sinapinic acid is suitable for proteins larger than 3 kDa (O’Rourke et al., 2016). However, if the targets are low molecular weight compounds, some organic matrices may present signals in the low mass-to-charge range, which will interfere with the analyte signals; a review performed by Calvano et al. (2018) discusses the matrices for low molecular weight compounds and provides a comprehensive overview on MALDI matrices and their applications.

Data processing can also be addressed to achieve more stable results. Usually, the procedures for MALDI data processing involve quality control, normalization, transformation, smoothing, baseline correction, peak detection, and binning (alignment) (Hajduk et al., 2015; Hajduk et al., 2016). There are many method variations to perform each one of these tasks, and each type of data will respond better to a certain method. Thus, evaluating them to find the optimized data processing pipeline is necessary. Also, sample processing on different days or with the same equipment in other laboratories can cause peak mass shift (Rossel et al., 2021); subsequently, data shift will impair the classification model’s performance. Internal standards could be implemented to solve this issue.

While multiomics holds great promise in improving the understanding of a disease or in the generation of more robust machine learning models, data integration is considered a major bottleneck to multiomics studies (Pinu et al., 2019). If choosing for an interpretable model using a machine learning approach, the high dimensionality of the dataset will impose a great challenge for the analysis. Therefore, reducing features is a necessary step. This can be done by performing a feature selection step, which will select a smaller set of features that holds enough information to characterize the groups, or perform a feature extraction method to combine the features and transform them in another set of features (such as a Principal Component Analysis) that also holds enough information of the entire dataset (Picard et al., 2021). If interpretability is not an issue (meaning that it will not be possible to understand which features are used for learning and used to find biomarkers), a Deep Learning approach could be used in concatenated multi-omics data (Picard et al., 2021). The handling of missing data, which frequently arises during multi-omics data integration, should be considered since many statistical approaches do not process missing values. Data imputation often introduces less bias than the complete removal of the features with missing values (Liebal et al., 2020).