CHAPTER 1: Genome-wide Discovery of MicroRNA Biomarkers for Cancer Precision Medicine Free

Evidence-based medicine (EBM) has been considered the optimal practice for decision-making in health care for decades. Compared to traditional medicine, which relies more on clinical experience and pathophysiologic rationale, EBM emphasizes the use of best available evidence from well-designed and well-conducted research. Despite the wide adoption of EBM as the gold standard of clinical practice, it has a number of known limitations, such as limited usefulness when applied to individuals and potential bias in randomized controlled trials.^1–4  As a result, in clinical practice, patients suffering from the same disease are often provided with nearly the same therapeutic interventions (‘one-size-fits-all’ approach to therapy), which overlooks the difference between individuals. In fact, the therapeutic responses of individual patients may depart widely from the average treatment effect. For instance, patients with the same cancer type may have heterogeneous genetic makeup in the tumors associated with discrepant clinical outcomes, and therefore should be managed differently in the clinic.

Over the last decades, the rich knowledge about human diseases gained from extensive basic, preclinical, and clinical research and the rapid development of biotechnology have paved the way towards precision medicine. Different from EBM, precision medicine is a new paradigm of medicine empowered by comprehensive molecular and clinical profiling of patients.⁵  It tailors precise disease diagnosis, prognosis, and therapeutics to the individuals based on (epi-)genetic, phenotypic, and clinical characteristics and eventually aims to achieve the right treatment, to the right patient, at the right time.^5,6  Recent advances achieved in multiple related fields have collectively laid a strong foundation for the development of precision medicine in the coming decades. On the one hand, the rapid development of biotechnologies contributes significantly to reduction in the cost of high-throughput sequencing, enabling the generation of multi-omic profiles for patients on a large scale. On the other hand, the fast-growing computing power and development of analytical methods has made it possible to store, transfer, analyze, and interpret the ‘big’ biomedical data generated. More recently, artificial intelligence (AI) has demonstrated unprecedented performance in various preclinical and clinical studies,⁷  providing new opportunities to develop automated, intelligent systems in the clinic.

Among the broad spectrum of research areas in cancer precision medicine, biomarker development is a central mission. A biomarker is a biological substance that can be measured and evaluated as an indicator of normal or abnormal conditions or a sign of disease condition.⁸  It can be found in tissues, blood, or other body fluids. Molecules such as DNA, mRNA, miRNA, and proteins, as well as metabolites and microorganisms, can all be exploited as biomarkers. Numerous biomarkers have been developed for cancer screening, diagnosis, prognosis, and prediction of therapy response as well as disease monitoring. Importantly, tailored for specific clinical applications, molecular biomarkers can be customized based on a selection of different types of clinical specimens, diverse biomolecules to test, and various high- or low-throughput platforms for profiling.⁹ 

MicroRNAs (miRNAs) are small regulatory non-coding RNAs that can downregulate gene expression mainly via base pairing to 3′ untranslated regions of target mRNAs. These small molecules influence almost every cancer-related process involving cell proliferation, apoptosis, metastasis, and angiogenesis.^10–15  Compared to other molecules, miRNAs are relatively stable and abundant in various clinical specimens, presenting promising candidates for biomarker development. In the last decade, miRNAs have been exploited as valuable diagnostic, prognostic, and predictive biomarkers in various cancer types.¹⁶  Recently, the growing interest in ‘liquid biopsies’ has also put circulating miRNAs to the forefront of biomarker discovery and development.

In this chapter, we will first summarize miRNA-based biomarkers for various clinical applications based on a substantial literature review and discuss the potential limitations of traditional approaches. Recent years have seen large-scale cohort studies generating vast amounts of omics data, which will continue to grow exponentially in the following decades.¹⁷  These ‘big’ multi-omic datasets, together with histopathological reports and medical records, provide tremendous opportunities for biomarker discovery and in silico validations. Next, we will introduce a data-driven methodology for biomarker development and the major steps involved. Furthermore, we will summarize the major advantages of the data-driven methodology and discuss the main challenges limiting its implementation, possible solutions, and emerging opportunities.

Cancer-associated miRNAs were discovered in 2008, followed by hundreds of miRNA studies thereafter.¹⁸  Notably, miRNA is relatively stable in body fluids such as plasma and serum, and in exosomes and tissue samples, which increases their potential to be exploited as biomarkers.^19,20  Indeed, a number of studies have already demonstrated the versatile clinical value of miRNAs for diagnosis, prognosis, therapy response prediction, and disease monitoring, which will be reviewed in detail in this section.

Cancer diagnosis, especially at early stages, is crucial for more effective disease prevention. Early detection of cancer significantly increases the chances for successful treatment, leading to the reduction of mortality and improvement of long-term survival of patients. However, due to the general low sensitivity and specificity, early detection of cancer has been a bottleneck. Recent studies on miRNA-based biomarkers have shown a great potential in employing circulating miRNAs for early detection of cancers. For instance, miR-29a, miR-92, and their combined signatures achieved more than 80% sensitivity and 70% specificity in differentiating colorectal cancer (CRC) patients from healthy controls.^21,22  Circulating miR-210 could also well distinguish cancer patients from healthy controls in glioblastoma multiforme (GBM) and non-small-cell lung carcinoma (NSCLC).^23,24 

Tissue-specific miRNAs could also be employed to identify tumor origins, especially when it is not clear where a metastatic tumor originated.^22,25  Rosenfeld's group constructed a transparent classifier based on the expression levels of 48 tissue-specific miRNAs from 253 samples.²⁶  In an independent cohort of 83 samples, the classifier achieved a high-confidence accuracy of 89%. For metastatic tumors, the classification sensitivity also reached 85%, demonstrating a great potential to use the approach for tracing the tissue of origin. Another study succeeded in identifying 42 tumors from 509 samples, with 64 miRNAs measured by customized microarrays.²⁷ 

Furthermore, miRNA signatures may also be used for the identification of cancer molecular subtypes. Molecular subtyping studies classify cancer into multiple subgroups that are biologically homogeneous in relation to clinical outcomes. The immediate clinical relevance of cancer molecular subtypes has been implicated in their significant association with survival, risk of developing recurrence and metastasis as well as therapeutic response in a variety of cancers.^28–31  In breast cancer, miRNA expression signatures can be used to classify the status of HER2, ER, and PR receptors, which are widely adopted clinical biomarkers.^32,33  Youssef developed a classifier using a unique miRNA signature to classify renal cell carcinomas (RCCs) into three subgroups, clear cell, chromophobe, and papillary, with a cross-validation accuracy of 90%.³⁴  Another study on RCC molecular classification also identified miRNA signatures, which can not only distinguish between different RCC subtypes but also between benign oncocytomas and malignant chromophobe RCCs.³⁵ 

Traditionally, histopathological characteristics, such as tumor size, lymph node and distal metastasis, and differentiation grade, are used for prognosis. However, due to biological heterogeneity, patients with the same clinical characteristics may have very different outcomes, varying from a complete cure to death. MiRNAs could regulate the expression of genes involved in cancer-related biological processes such as epithelial to mesenchymal transition, and their dysregulations may lead to more invasiveness, metastasis to distant organs, and, eventually, affect patient survival.^36,37  Developing novel prognostic biomarkers may help with better patient management and optimized treatment. The potential utility of miRNAs as prognostic biomarkers has been widely studied in many cancers. For instance, pancreatic cancer patients are typically diagnosed with local advanced or metastatic disease at the time of presentation, but few effective therapeutic options are available for advanced patients.³⁸  Based on quantitative real-time polymerase chain reaction (qPCR) experiments for 107 pairs of pancreatic cancer and adjacent normal tissues, Li et al. found low expression of miR-218 was an independent predictor of poor prognosis.³⁹  As another example, in lung cancer, overexpression of let-7 could inhibit lung cancer cell growth in vitro, and its low expression showed significant prognostic value independent of clinical characteristics.⁴⁰ 

Cancer recurrence and metastasis are the leading causes of mortality, and therefore accurate prediction of recurrence and metastasis is a central task closely related to prognosis. Stratifying cancer patients according to the risk of recurrence and metastasis could help select high-risk patients for more effective treatment and avoid unnecessary overtreatment of low-risk patients. Taking CRC as an example, up to 50% of patients will develop metastases.⁴¹  Quite a few miRNAs such as miR-885-5p, miR-19, miR-8181a, miR-203 have been implicated for their functional roles in regulating cell migration, invasion, and metastasis, and have shown the potential to predict recurrence, regional and distal metastasis in CRC.^42–45  Furthermore, circulating miRNAs could also predict the risk of disease recurrence of early-stage colon cancer at the time of diagnosis, providing potential non-invasive biomarkers for early detection.⁴⁶ 

As another major cancer type extensively studied, about 5% of breast cancer patients suffer from incurable metastasis, and 10–15% more patients develop metastasis within 3 years of diagnosis.⁴⁷  Many miRNAs have been reported for their functional roles in suppressing (e.g., miR-31,^48,49  miR-126, and miR-335 ⁵⁰ ) or promoting (e.g., miR-374a⁵¹  and miR-200a⁵² ) metastasis in breast cancer. Some of these metastasis-associated miRNAs, such as miR-21 and miR-205, were found to be significantly associated with survival and could differentiate patients with higher risk.⁵³  Furthermore, circulating miRNAs such as miR-10b and miR-373 could also distinguish breast cancer patients with lymph node metastasis from non-metastatic patients.⁵⁴ 

Prediction of response to a specific treatment is of great clinical value. Accumulating evidence has shown that miRNAs played a significant role in regulating genes associated with drug resistance. For instance, miR-27a and miR-451 were found to be overexpressed in multidrug-resistant cancer cell lines and could upregulate ABCB1 and P-glycoprotein, whose products provoke cell resistance to many chemotherapies for cancer patients.⁵⁵  FOLFOX is a combination of chemotherapeutic regimens, made up of folinic acid (FOL), 5-fluorouracil (F), and oxaliplatin (OX)⁵⁶  and commonly used as the first-line chemotherapy in advanced CRC.⁵⁷  Chen et al. discovered that miR-19a was significantly associated with FOLFOX-resistance based microRNA microarray data analysis.⁵⁷  Based on further validation using qRT-PCR, they confirmed miR-19a as a potential predictive marker, with a sensitivity of 66.7% and a specificity of 63.9% to discriminate against the drug-resistant patients. Furthermore, miR-19a was also found to be related to gefitinib resistance in NSCLC cells.⁵⁸ 

Apart from chemotherapeutics, miRNAs may also confer resistance to radiotherapy and immunotherapy.^59,60  Radiotherapy uses ionizing radiations (IRs) to cure cancer by producing free radicals in the neoplastic cells. This process is closely related to the DNA damage response (DDR), which could also be regulated by miRNAs. For instance, overexpression of miR-21 might be related to repair of DNA damage and could induce radioresistance.⁶¹  Immunotherapy is a type of cancer treatment that activates the immune system to fight cancer,⁶²  and miRNAs are required for the development and function of the immune system.⁶³  For instance, dysregulated miRNAs can help tumor cells reduce the susceptibility to cytotoxic T-lymphocyte-mediated cytolysis,^63,64  or influence PD-L1 expression.⁶⁵  Furthermore, recent studies have found that circulating miRNA signatures can potentially predict immunotherapy response in NSCLC,^66–69  melanoma,^60,70  and other cancers.⁷¹ 

Cancer is an evolving process in nature, posing a significant challenge to accurate diagnosis and treatment. Timely and accurate monitoring of disease status is essential to improve the therapy quality and detect recurrence and metastasis earlier.⁷²  Due to the close relationship between miRNAs and disease progress, they could provide time-course information to disease monitoring.⁷³  Circulating miRNA biomarkers could serve as a prospective indicator to monitor the dynamic disease status, treatment response, and adverse events, and are likely to surpass conventional protein biomarkers, which generally have lower sensitivities and specificities.⁷⁴ 

As an example, Li's group surveyed 749 miRNAs in the plasma of 20 pre- and post-operative stage II/III colorectal patients.⁷⁵  They examined miRNA expression profiles in the plasma to monitor CRC dynamics and used differential miRNAs to predict patient survival. A screening model was constructed by the training set (20 pre-/post-operation/controls), followed by internal validation (65 pre-/post-operation and 40 controls) and independent external validation (90 pre-/post-operation and 70 controls). Their results showed that miR-145, miR-106a, and miR-17-3p were significantly differentially expressed between pre-operation and post-operation. In addition, they found that high expression of miR-17-3p and miR-106a was a powerful indicator for shorter disease-free survival (DFS). Another study in lung carcinoma found that the expression levels of miR-21 and miR-24 were decreased in the sera of post-operative patients compared to that of pre-operative patients.⁷⁶  These studies suggested that miRNA expression biomarkers have the potential to be used for disease status monitoring, which is important in optimizing treatment and improving outcomes.

The vast majority of pre-existing biomarkers were developed based on classical, knowledge-based approaches, mostly relying on literature review or research experience. The intrinsically biased method of biomarker identification employed in empirical methods has known limitations, such as limited predictive power, specificity, and robustness. These limitations greatly affect the reproducibility of biomarkers and eventually hamper their clinical translation.⁷⁷  Recent years have seen an emerging data-driven methodology transforming the way biomarkers are traditionally developed. Different from the empirical methods, the novel methodology is based on unbiased data mining from high-throughput miRNA expression profiles, followed by in silico and clinical validations. The schematic diagram of this methodology is shown in Figure 1.1. Representative biomarker studies are summarized in Table 1.1.

The data-driven methodology provides an unbiased, systematic, and cost-efficient paradigm for efficient mining of cancer biomarkers. In the following sections, we will introduce in more detail the major steps involved in a typical workflow for biomarker development.

MiRNAs are abundant and relatively stable small molecules in tumor tissues and various types of body fluids. Fresh frozen (FF) and formalin-fixed paraffin-embedded (FFPE) tissue specimens are commonly used as a gold standard for the diagnosis of most cancers. ‘Liquid biopsy’, which can be obtained not only from blood but also from urine, saliva, stool, and the like, has demonstrated a great potential for early detection of cancers in recent years. These different sources of clinical specimens have pros and cons, and the selection of appropriate samples should be based on the detailed clinical question, accessibility, sample degradation, and quality control, time, and cost for sample processing as well as invasiveness.

Fresh frozen tissues preserve cells and molecules well in a native state, providing an ideal source of miRNAs for biomarker studies. However, tissue samples require rapid freezing as soon as possible after collection to avoid ice crystals and should be stored at extremely low temperatures. Liquid nitrogen is usually used for snap-freezing, and the biobank center with liquid nitrogen barrels must be very close to the operating room. The problem is that liquid nitrogen evaporates rapidly, particularly in small containers, which have to be consequently refilled at intervals. The high cost and time-consuming and laborious procedures mean that liquid nitrogen is not always available.⁷⁸ 

The FFPE method is also time-consuming and laborious, but FFPE samples can be kept for a long time at regular room temperature while preserving the structure of cell and tissue. These advantages make FFPE a convenient and cost-effective way of collecting and storing patient samples in hospitals or research institutions. The FFPE method is also a standard procedure for pathological diagnosis. With the clinical follow-up information, FFPE samples are easily accessible and apply well to large-scale retrospective studies. Although intact mRNA is not available from FFPE materials, miRNAs are more stable and less affected by formalin fixation and temperature, due to their small size, stability, and close association with several large protein complexes.⁷⁹  Studies have shown fairly good consistency between FFPE and FF samples,^19,80–82  suggesting the applicability of using FFPE samples as appropriate resources for miRNA analyses. Nonetheless, it remains a concern that miRNA levels may decrease when FFPE samples are stored for a long time, and the degradation level might be related to their Guanine-cytosine content.^83,84 

Tissue specimens are invasive and more suitable for molecular subtyping, cancer prognosis, and prediction of therapy response than early detection and disease monitoring. Liquid biopsy as a non-invasive tool for biomarker discovery is gaining increasing attention.⁸⁵  It provides a promising strategy for early diagnosis and continuous monitoring of disease status that is more affordable and more acceptable to the patients. The first miRNA to be described in biological fluids was miR-21, detected in the serum of B-cell lymphoma patients.¹⁸  As a biomarker, miR-21 expression was higher in the sera of patients than that of healthy participants and was associated with relapse-free survival. Pieces of evidence have shown that miRNAs are relatively stable in other sources like urine and saliva,^86,87  suggesting the potential to use miRNAs as biomarkers for non-invasive detection of urologic cancers.⁸⁸  Furthermore, fecal miRNA biomarkers have also been developed in several studies,^89–94  though the technology is still challenged by gut microbiome influences.⁹⁵ 

In addition to cell-free miRNAs, miRNA molecules are also abundant in extracellular vesicles (EVs) released by tumor cells. EV are membrane particles, including exosomes, ectosomes, microparticles, microvesicles, and apoptotic bodies, with similar constituents, including miRNAs as their parental cells or tissues.⁹⁶  EVs like exosomes can be taken up by neighboring cells or distant cells and play an essential role in cellular regulation and information exchange. Dysregulated exosomal miRNAs, which might be more stable due to lipid membrane protection, could also serve as biomarkers in cancer.⁹⁷  The studies of miRNAs in circulating tumor cells (CTCs) as biomarkers are still in progress due to the challenge of reliable detection and isolation of CTCs.⁹⁸  Combining with single-cell sequencing technology, the study of CTC miRNAs may pave a new way for clinical application.

Liquid biopsy, as an essential source of miRNAs, has several advantages in addition to non-invasiveness. MiRNAs extracted from liquid biopsies can be exploited as biomarkers to assess cancer risk at an early stage while other types of clinical specimens cannot. Because of the easy accessibility, miRNA biomarkers can also be developed using longitudinal samples to predict cancer development, treatment effect, surgical outcome, and cancer recurrence.^20,68,99,100  Despite the versatility, liquid biopsy suffers from several technological bottlenecks. First, fewer miRNAs can be extracted from liquid biopsy than from tissue samples, and therefore highly sensitive diagnostic technology is needed for biomarker development. Second, since miRNA levels in liquid biopsies are affected by various conditions, there is still no ideal endogenous control to normalize and estimate the abundance of miRNAs.¹⁰¹  Last but not least, the sensitivity and specificity remain unsatisfactory, especially for the detection of early-stage cancers, and improvement should be made before clinical translation.

High-quality data is the basis for genome-wide miRNA discovery. Two major technologies have been used for generating massive miRNA expression profiles on a genome-wide scale: miRNA microarrays and next-generation sequencing (NGS). Different protocols and reagents can be used for library preparation based on miRNAs or total RNAs extracted from clinical specimens. Microarrays or NGS are subsequently employed to transform the molecular information to analyzable miRNA expression data. With great success, both platforms have been powerful tools for discovering miRNA biomarkers for various clinical applications, such as diagnosis, prognosis, and therapy response prediction.

Microarray is the first method for genome-wide profiling of miRNA expression levels. A microarray includes thousands of miRNA oligonucleotide probes in duplicate, covering known miRNA sequences in the miRBase database.¹⁰²  Generally, miRNAs will be ligated with fluorophore-conjugated nucleotide or short oligonucleotide by ligase catalysis.¹⁰³  The abundance of miRNA can be estimated by the intensity of fluorescence after hybridization, washing, and scanning. The microarray platform produces miRNA expression data with high sensitivity and low cost. The first miRNA biomarker study identified a unique miRNA expression signature in chronic lymphocytic leukemia.¹⁰⁴ 

NGS has already become a pervasive and powerful tool for biomedical research. It offers a genome-wide solution for evaluation of the abundance of known miRNAs by massive short reads with high specificity, and discovery of previously uncharacterized miRNAs. Similar to the microarray platform, NGS also requires RNA size selection to reduce the influence of unrelated molecules. MiRNAs are sequenced parallelly after adapter ligation and cDNA library construction, which is more complicated compared to the microarray platform. Studies have shown a consistent correlation between NGS and microarray platform,^105,106  but potential bias due to PCR amplification and sequencing was often observed.^107,108 

Scanned images of microarrays are the original hybridization data and should be converted to spot intensity signal values by segmentation techniques.¹⁰⁹  The signal after background adjustment should be normalized by methods such as quantile normalization and spike-in controls or housekeeping genes.¹¹⁰  It should be noted that each normalization method has its specific assumption. For example, quantile normalization is a commonly used nonlinear normalization method which assumes that the intensities of all miRNAs across microarrays follow the same distribution.

After normalization, non-biological batch effects are a significant concern at the preprocessing stage. It refers to cumulative errors originating from variations in technical platforms, experimental protocols, laboratory conditions, and the same assay at different times or in different working places. Biological findings derived from a collection of data generated from different batches may be confounded by systematic errors, leading to spurious results and conclusions.¹¹¹  Generally, batch effects are not easy to avoid due to the nature of experimental design, especially for a large number of samples. For example, most of the existing microarray platforms can process 24, 48, or 96 samples at the same time. Still, hundreds or thousands of samples may be involved in a large-scale cohort study, which poses the challenge for a balanced experimental design. To detect potential batch effects, principal component analysis can be performed, and once confirmed, statistical methods such as ComBat may be used for adjusting batch effects based on an empirical Bayes method.¹¹² 

Unlike the design of microarrays, where probesets have clear target miRNAs, reads obtained from NGS are raw nucleotide sequences. The raw sequencing reads should be first examined for quality control using tools such as FastQC.¹¹³  For instance, reads of low quality should be removed, and adapter sequences should be trimmed off. Preprocessed reads are subsequently mapped to the human reference genome or miRNA sequence database, followed by the calculation of miRNA expression levels based on read count coverage. To reduce potential bias caused by varying library size, further normalization is needed, e.g., using counts per million mapped reads (CPM) rather than the raw counts.¹¹⁴ 

Over the last decades, various international, national, and regional projects/consortia have generated “big” omics data, which greatly facilitate cancer research and biomarker development. Based on these public data repositories, researchers can make biomarker discovery and in silico validations by integrative analysis of multiple datasets at a large scale for a specific clinical question.

Among various public databases, the three most popular repositories with miRNA expression profiles for large-scale cancer samples are The Cancer Genome Atlas (TCGA),¹¹⁵  Gene Expression Omnibus (GEO),¹¹⁶  and International Cancer Genome Consortium (ICGC).¹¹⁷  TCGA is a database with large-scale multi-omics data for over 20 000 primary tumor and tumor-adjacent normal samples across 33 cancer types. ICGC provides a much bigger online platform for international collaborative cancer research, which currently involves 86 major cancer projects, including TCGA and the Sanger Cancer Genome Project. GEO is a comprehensive database established by the National Center of Biotechnology Information (NCBI), representing one of the largest databases of gene expression microarrays in the world.¹¹⁸ 

To retrieve miRNA data from TCGA, ICGC, and GEO, one could either directly download from https://www.ncbi.nlm.nih.gov/gds/, https://icgc.org/ and https://www.ncbi.nlm.nih.gov/geo/, or via R packages such as UCSCXenaTools,¹¹⁹  GEOquery¹²⁰  and TCGAbiolinks¹²¹  with the specification of the cancer type or accession ID of interest. For instance, in GEO, the raw data (e.g., Affymetrix CEL files) and processed data may be accessible on the webpage of the project. For high-throughput sequencing data, the corresponding sequencing platform, protocol, and library information as well as raw data may be available in Sequence Read Archive (SRA).¹²²  Curated miRNA expression profiles are also available in other databases such as miRmine,¹²³  TissueAtlas,¹²⁴  and CGGA.¹²⁵  In addition to primary tumor-derived data, miRNA expression profiles are also available for many cancer cell lines in the CCLE database,¹²⁶  which are valuable for generating hypotheses for in vitro studies.

Having obtained miRNA expression profiles, either generated from in-house cohorts or curated from public datasets, potential miRNA biomarkers can be discovered based on genome-wide analysis, involving the following three major steps. First, miRNA candidates are prioritized based on statistical evaluation of the predictive power to differentiate between different groups depending on the clinical question, e.g., cancer patients vs. healthy controls, patients with recurrence/metastases vs. those without, and drug-responsive vs. drug-resistant. Second, statistical and machine learning methods are employed to select further the miRNA candidates to build a risk-scoring model. Last but not least, in silico validation should be performed in other independent datasets to confirm the predictive performance of the model.

Prioritizing miRNA candidates can be considered a feature selection in machine learning. The most straightforward methods are statistical tests to identify the most differential miRNAs between cases and controls. For instance, Student's t-tests and Wilcoxon signed-rank tests are the most basic parametric and non-parametric statistical tests, which can be used to select the most statistically significant miRNAs among thousands of miRNAs. Model-based approaches such as the empirical Bayes model in the ‘limma’ package¹²⁷  and negative binomial model in the ‘DESeq2’ package¹²⁸  were tailored explicitly for differential expression analysis based on miRNA microarrays and sequencing data, respectively. Importantly, since multiple hypothesis tests are performed in the genome-wide analysis of miRNA expression profiles, adjustment for multiple comparisons (e.g., using the Benjamini-Hochberg method) is needed to control false positives.

To prioritize miRNA candidates, several other further filtering steps are often required. First, differential miRNAs are not necessarily discriminative, that is, miRNAs may not be able to distinguish between cases and controls even if they are significantly differentially expressed. Therefore, very often, statistical analyses such as univariate logistic/Cox regression and area under the receiver operating characteristic (ROC) curves are employed to prioritize the most predictive miRNA candidates further. Second, in clinical implementation, low-throughput assays such as qRT-PCR are used to measure the expression of miRNAs, which usually have a much lower sensitivity. Therefore, lowly expressed miRNAs need to be filtered out for more robust detection of biomarkers in the clinical practice. Last but not least, too few miRNAs left may run into the risk of insufficient biomarkers after clinical validation, while having too many candidates will increase the cost for clinical validation and application. In order to determine the number of miRNA candidates prioritized, one needs to strike a balance between the predictive performance and the cost-efficiency.

Based on the prioritized miRNA candidates, a risk-scoring model can be established to assess the joint predictive value of the entire panel. Depending on the clinical question, different statistical models may be trained. Multivariate logistic regression provides a straightforward linear model, which is suitable for classification problems such as distinguishing cancer patients from healthy people. For instance, logistic regression was used to build a multivariate linear model for early detection of pancreatic cancer using saliva samples.¹²⁹  Cox regression analysis, instead, can be used for prognostic risk scoring. As an example, an 8-miRNA expression signature was employed to train a Cox regression model to predict recurrence-free survival in stage II and III CRC, demonstrating superior performance to pre-existing biomarkers.¹³⁰ 

The linear regression models are advantageous in their easy interpretability and applicability in the clinic. However, the most significant limitation of these models lies in the well-known collinearity issue, i.e., highly correlated miRNAs may lose their predictivity, leading to a confusing and unstable model. To overcome the limitation, other machine learning approaches such as decision trees, random forests,¹³¹  support vector machines (SVMs),¹³²  and artificial neural networks (ANNs) may be considered. For instance, an miRNA-based decision tree classifier was used to identify the tissue origin of cancers of unknown primary origin, with an accuracy of 89% on the validation set.²⁶  In another study, random forests and SVM algorithms were used to classify breast cancer patients and health participants.¹³³  Moreover, Ali et al. trained a recurrent neural network, based on the expression levels of 35 miRNAs, to classify kidney cancer subtype with a high overall accuracy of ∼95%.¹³⁴ 

Feature selection and modeling may be simultaneously performed. The most frequently used approach may be Least Absolute Shrinkage and Selection Operator (LASSO) regression, which can generate a simple, sparse linear model when used in multivariate regression analysis.^135,136  Prediction Analysis for Microarrays (PAM) is also a popular classification method with the selection of the optimal miRNA signature based on shrinkage centroid regularized discriminant analysis.¹³⁷  Other popular machine learning approaches, such as linear discriminant analysis (LDA) and neighborhood component analysis (NCA), have also been employed for miRNA-based biomarker development.^134,138 

Validation is crucial to ensure the performance, robustness, and reproducibility of the miRNA panel. Before clinical validation, in silico (or analytical) validation using independent public datasets is a powerful and cost-free strategy, representing a unique advantage of the data-driven methodology.

The miRNA candidates prioritized based on the discovery dataset should be validated first. Inconsistent miRNAs showing contradictory expression trends in the validation sets should be removed from the panel. It should be noted that due to different platforms or annotations used in the validation datasets, some miRNAs may be missing and should be excluded from the panel as well. Using the new panel with consistent miRNAs, a new risk-scoring model should be established and calibrated.

The performance of the risk-scoring model should also be rigorously evaluated by calculating statistics such as areas under the ROC curve (AUC) and hazard ratios with confidence intervals, sensitivities and specificities. The performance of prognostic biomarkers can be evaluated by Kaplan-Meier analysis with log-rank tests and Cox regression analysis with Wald tests. Furthermore, chi-squared tests and Fisher's exact tests may be used to assess the performance of diagnostic and predictive biomarkers.

As an example, in our previous study, we developed a novel miRNA expression signature for recurrence prediction in stage II and III CRC.¹³⁰  Using the TCGA-COAD cohort with 158 samples, eight miRNA candidates were prioritized to train a Cox regression model for recurrence risk prediction. The model was subsequently validated in another two public datasets with 107 and 40 samples, respectively.¹³⁰  Based on the successful in silico validation, we were able to develop a robust qRT-PCR-based miRNA-recurrence classifier, demonstrating superior performance to standard clinicopathological features in the identification of high-risk CRC patients.¹³⁰ 

Analytical validation can significantly increase the confidence to identify consistent miRNAs and develop reproducible risk-scoring models. Further clinical validation based on retrospective or prospective cohort studies is pivotal to confirm the clinical utility in real practice. Importantly, high-throughput assays such as miRNA microarrays and RNA-seq are ideal for biomarker discovery, but not appropriate for clinical implementation due to inadequate accessibility, and much higher turnaround time and cost. Therefore, in a typical clinical setting, low-throughput assays for measuring miRNA expression levels are more practical for clinical translation.

Multiple types of low-throughput assays can be considered for developing and validating miRNA biomarkers. Northern blotting, one of the classical methods of miRNA detection, used to be the gold standard in early miRNA studies.¹³⁹  Due to the simplicity and low cost, northern blotting is still widely used. It involves the separation of total RNA by polyacrylamide gel electrophoresis and transferring separated RNA to a nitrocellulose membrane, followed by detection using a miRNA-specific labeled probe. However, northern blot is time-consuming and requires a large amount of total RNA. Therefore, in real clinical applications, it is not applicable when only a limited amount of samples are available.¹⁴⁰ 

A more frequently used method for miRNA biomarker development is qRT-PCR. While traditional qPCR cannot be used to detect mature miRNAs due to the small size, more specialized probe designs have been proposed.^141,142  Although qPCR is quantitative and highly sensitive compared to other techniques, only a few miRNAs can be quantified at a time. Therefore, for a large panel of miRNAs, molecular testing using qPCR could be very laborious. Furthermore, the miRNA template is so short that it can be difficult to distinguish between miRNAs with differences of only one or two nucleotides.¹⁴³  Droplet digital PCR (ddPCR), based on sample partitioning into thousands of micro-reactions in a water-oil emulsion droplet, has more precise quantitation, higher sensitivity, and robustness than qPCR.¹⁴⁴ 

Compared to techniques quantifying miRNA expression from clinical samples of mixed cell types, miRNA in situ hybridization enables visualization of the expression patterns at the cellular level by localizing the pre-labeled probes hybridized with target miRNAs.^145–147  Further improvement has been made by multicolor miRNA in situ hybridization (MMISH),¹⁴⁸  which combines in situ hybridization with immunofluorescent stainings, thus creating a multicolor image for miRNA visualization in cells.¹⁴⁹  MiRNA ISH and related methods have demonstrated excellent performance in distinguishing between various tumor types.^149,150 

Several new technologies can be tailored to a targeted panel of miRNAs of interest, providing a more scalable solution compared to conventional high-throughput microarrays. For instance, customized miRNA microarrays can be efficiently applied to a small miRNA panel in a large-scale cohort study, and have demonstrated the validity in various clinical applications.^151–153  NanoString nCounter uses synthetic RNA segments labeled with different fluorochromes to hybrid to single-strand DNAs, and the absolute abundance can be subsequently estimated by counting fluorescent barcodes.¹⁵⁴  Using this platform, miRNA biomarkers have been successfully identified.^155–158  This platform could also be used in clinical applications with large-scale samples more cost efficiently.¹⁵⁹ 

Using data generated from low-throughput assays, it is still necessary to filter out inconsistent candidates and finalize the miRNA panel. However, the distributions of high-throughput data such as miRNA microarray and NGS data are very different from low-throughput data such as qRT-PCR-derived cycle threshold (Ct) values. Therefore, risk-scoring models developed based on high-throughput data cannot be directly applied to data generated by low-throughout assays. A new model needs to be trained for the low-throughput assay, and the performance needs to be statistically evaluated in independent retrospective or prospective cohorts.

To summarize, compared to traditional methods, the data-driven methodology has the following advantages:

• (1) More unbiased biomarker discovery. Typically, a large-scale dataset is used for prioritization of biomarker candidates, which usually are differentially expressed and discriminative miRNAs between cases (e.g., tumor tissues, sera specimens) and controls (e.g., tumor-adjacent normal samples, sera of healthy people). Therefore, the identified miRNA biomarkers are more unbiased with higher specificity than empirical methods.

• (2) Increased statistical power. The data-driven methodology employs a large-scale dataset used for biomarker discovery, which can significantly increase the statistical power to identify the most promising candidates.

• (3) Increased reproducibility. The identified miRNA candidates are further validated in multiple independent public datasets. Inconsistent miRNA candidates between different public datasets can be filtered out, and the remaining candidates will be further validated experimentally, which can dramatically increase the reproducibility and robustness.

• (4) More cost-efficient. Biomarker development could be costly, especially when large-scale high-throughput data is generated at the discovery stage. The data-driven methodology employs public datasets for prioritization of potential candidates and therefore does not incur any experimental cost during the stage of biomarker discovery.

• (5) Scaling up the study. The data-driven methodology efficiently integrates multiple public datasets that are mostly generated by different research groups, enabling a scaled-up multi-center cohort study.

• (6) Providing a more comprehensive functional landscape. Since high-throughput and even genome-wide miRNA expression profiles are used for biomarker discovery, itis more efficient for prediction with the possible mechanism as support by identifying downstream or upstream changes.^160,161 

Although the data-driven methodology has demonstrated superiority over classical methods, it can only be applied under certain situations when high-quality “big data” is available. As shown in Figure 1.2, several major limitations and challenges may limit the implementation in practice.

• (1) Availability of high-quality data. The data-driven methodology is more suitable for common cancer types such as breast cancer, colon cancer, and gastric cancers, for which multiple public miRNA expression datasets for large-scale tumor samples are available. For rare cancer types with low incidence rates, however, large-scale datasets are still lacking, either due to the difficulty of collecting clinical specimens (e.g., pancreatic cancer) or insufficient numbers of samples, possibly leading to protectionism (e.g., nasopharyngeal carcinomas). These limit further research and hinder the development in related fields of cancer precision medicine.

Even if public datasets are available, there are still challenges for integrative analysis. First, public data may come from all over the world, with diverse genetic backgrounds. The identification of miRNA candidates may be confounded by mixed ethnicity or other complex environmental factors. Second, independent public datasets may be generated from different platforms at different times and locations under different library preparation protocols. Direct integration is likely to ignore data bias and non-biological batch effects, but statistical methods are not always available and suitable for reliable data correction. Third, sample imbalance is another big challenge. The TCGA project significantly improved our understanding of cancer genetics and provided a valuable database for biomarker development. However, there are far more tumor samples than tumor-adjacent normal samples in the TCGA database, and therefore, using it for biomarker discovery and validation may induce bias leading to unreliable candidates. Last but not least, most of the published data lack corresponding clinical information, much limiting the power of the dataset and more in-depth analysis.

• (2) Liquid biopsy biomarkers. There is no doubt that liquid biopsies provide a valuable, promising source to develop circulating biomarkers for non-invasive cancer diagnosis. Challenges remain to be addressed with regard to the reproducibility, robustness, sensitivity, and specificity of circulating miRNA biomarkers. Notably, it is hard to distinguish whether the circulating miRNAs originated from plasma or tumor cells. Previous studies showed that the majority of circulating miRNAs actually originated from blood cells, which were regarded as contaminants and may mask the true tumor-related miRNA origins and activities.^162,163  MiRNA expression alterations sometimes only reflect the inflammatory or immune responses of the human body rather than specific tumors.¹⁶⁴  Furthermore, biomarkers identified from tissue samples may not be consistent in liquid biopsy, posing a significant challenge to tumor-specific biomarker development. Therefore, biomarkers identified from tissue specimens should be further confirmed and validated in liquid biopsies.

Also importantly, the processes of liquid biopsy sample collection, processing, storage, and expression quantification have not been standardized.¹⁶⁵  Standardization is essential when combining data from different locations or experimental batches. All clinical samples should be of high quality and should be normalized before moving to further analysis in case of the non-biological variations between different patients and labs.^165,166  Due to the lack of a consensus endogenous normalizer, the quantification of miRNAs by qRT-PCR is often not accurate. Current normalization procedures mainly use synthetic spiked-in controls and endogenous internal controls (housekeeping miRNAs), but how to normalize across multiple independent cohorts remains challenging.

• (3) Cancer heterogeneity. Cancer development is a very complex evolutionary process with substantial heterogeneity within and between tumors. However, the vast majority of biomarkers developed previously ignored the heterogeneity nature of tumors. Cancer patients of different molecular subtypes may have diverse clinical outcomes, and therefore developing cancer subtype-specific biomarkers may lead to more precise clinical decision making.¹⁶⁷  To dissect intra-tumoral heterogeneity, single-cell sequencing has become a promising strategy that also allows miRNA expression profiling at the single-cell level.^168,169  However, the clinical usage of the new technology, especially how to develop cancer biomarkers, remains mostly unexplored. Nonetheless, identification of circulating tumor cells coupled with single-cell sequencing may greatly promote the development of the entire field. With more advanced analytical methods developed and more large-scale single-cell sequencing datasets generated, the data-driven methodology will be compelling for more precise biomarker development in the future.

On the other hand, integrating multiple types of information, such as mutations, mRNA, miRNA, and protein expression alterations, as well as other epigenetic modifications, may help develop more specific biomarkers. Two strategies may be considered: integrating new biomarkers to existing biomarkers, or employing de novo analysis by integrative analysis of multi-omic data to identify putative biomarkers.¹⁷⁰  However, multi-omic data integration remains challenging computationally due to the intrinsic differences between the types and scales of data generated from different platforms.¹⁷⁰  Even if multi-omic biomarkers can be prioritized from high-throughput data, how to develop a reliable low-throughput assay that is translatable in the clinic would be a big challenge.

• (4) The big data challenge. Data lays the foundation for biomarker discovery, but also raises a considerable challenge to storage, transfer, and analysis. Biomedical data is continuously expanding. To meet the demand for massive data storage, computing infrastructure should be well designed and maintained. A high-performance computing center may be established to guarantee computational capacity and security. Ultrafast internet connection is also needed to ensure efficient data download from public databases and data transfer between local and remote servers.¹⁰¹  With the fast development of cloud computing, more efficient handling of big data and standardized analytical workflow may be established in the cloud server.

Biological and biomedical data in nature are multi-source, multi-modal, high-dimensional, and much more heterogeneous than other big data fields such as physics.¹⁷¹  As mentioned earlier, machine learning algorithms such as regression models, random forests, and SVMs have been frequently employed for biomarker discovery.^172–174  In recent years, AI has become a new powerful driving force in biomedical research and clinical applications. AI can learn imperceptible human knowledge from extensive datasets and is more potent in handling multi-modal data compared to traditional machine learning methods.¹⁷⁵  In a study of miRNA biomarkers for cancer classification, convolutional neural networks were demonstrated to outperform classical machine learning methods.¹⁷⁶  In another study, a variational autoencoder was employed to predict potential disease-associated miRNAs, which also obtained better performance than other methods.¹⁷⁷  It is believed that in the near future, AI will be more prevalent in the field of biomarker discovery, especially for large-scale datasets with more complicated data structures.

• (5) MiRNA molecular mechanisms. The data-driven methodology can facilitate researchers to identify the most predictive miRNA biomarkers. However, the vast majority of biomarker discovery studies focus only on clinical utility, leaving the biological functions of miRNA biomarkers unexplained. As summarized earlier, one advantage of the data-driven methodology lies in its unbiased, unsupervised genome-wide analysis, which provides a landscape of functional overview of miRNAs. Higher-level bioinformatic analysis, such as network inference, can be employed to pinpoint the critical upstream regulators further. For instance, our previous miRNA-mRNA regulatory network inference coupled with master regulator analyses identified miR-200 family and miR-508-3p as the determinants of the mesenchymal identities in CRC and ovarian cancer, respectively.^178,179 

According to the functions of miRNAs in cancer, the role of miRNAs can be generally classified into either tumor-suppressive or oncogenic. Therefore, the corresponding functional study strategies of miRNAs in oncology mainly involve two pathways.¹⁸⁰  The first pathway aims to inhibit oncogenic miRNAs through the utilization of miRNA antagonists. This can be achieved by employing chemically modified single-stranded oligonucleotides, with their sequences complementary to miRNAs. Alternatively, mRNAs with multiple target sites for a given miRNA can be used to sequester target miRNAs.^181–183  The second pathway, miRNA replacement, focuses on the loss of miRNA function. To restore the levels of miRNA, double-stranded miRNA mimics and DNA structures coding for specific miRNAs are developed, as a substitute for defective miRNAs.^181,182  With the advances in miRNA functional and mechanistic studies, miRNA expression perturbation might be a novel therapeutic approach in the future.

• (6) Towards clinical translation. Translating miRNA biomarkers into clinical practice is not only the most significant challenge but also the most exciting opportunity. Even though the US Food and Drug Administration has approved nucleic acid-based diagnostic assays in many diseases such as HER2 evaluation kits in breast cancer, the relatively new miRNA-based biomarkers are still in clinical trials. Putative miRNA biomarker panels are just beginning to accrue. In addition to experimental challenges discussed earlier, there are still some challenges specific to developing clinically accessible assays to assist decision making. First of all, high-level evidence is required to support the clinical value of miRNA biomarkers. Rigorous clinical trials should be performed before biomarkers can be approved for clinical practice. In order to generate high-level evidence, the performance of the biomarker should be statistically evaluated in a multi-center cohort study, especially a prospective cohort study. Benchmarking the performance against the current gold standard in the clinic is also needed to demonstrate superiority. Second, considering the cost-effectiveness and turnaround time, a small panel of miRNAs is usually preferred, so that the frontline users can prepare reagents more efficiently and timely. Third, it is recommended to use easily accessible and mature platforms such as qPCR for miRNA testing during the development of biomarker assays. Besides, simple and standard operating procedures are also preferred in real clinical practice. Last but not least, to facilitate the implementation of the biomarker, clinical specimens should be easy to collect. Despite the abovementioned limitations, liquid biopsy has unparalleled advantages compared to tissue specimens. We foresee that many more cost-efficient, easily accessible non-invasive miRNA diagnostic assays will come out and be broadly used in clinical tests in the near future.

AI:

Artificial intelligence

ANN:

Artificial neural networks

AUC:

Areas under the ROC curve

CPM:

Counts per million mapped reads

CRC:

Colorectal cancer

Ct:

Cycle threshold

CTC:

Circulating tumor cell

ddPCR:

Droplet digital PCR

DDR:

DNA damage response

DFS:

Disease-free survival

EBM:

Evidence-based medicine

EV:

Extracellular vesicle

FF:

Fresh frozen

FFPE:

Formalin-fixed paraffin-embedded

GBM:

Glioblastoma multiforme

GEO:

Gene expression omnibus

ICGC:

International Cancer Genome Consortium

IR:

Ionizing radiations

ISH:

In situ hybridization

LASSO:

Least absolute shrinkage and selection operator

LDA:

Linear discriminant analysis

miRNA:

MicroRNA

MMISH:

Multicolor miRNA in situ hybridization

NCA:

Neighborhood component analysis

NCBI:

National Center of Biotechnology Information

NGS:

Next-generation sequencing

NSCLC:

Non-small-cell lung carcinoma

OncomiR:

Oncogenic miRNA

PAM:

Prediction analysis for microarrays

PD-L1:

Programmed death-ligand 1

qRT-PCR:

Quantitative real-time polymerase chain reaction

RCC:

Renal cell carcinomas

ROC:

Receiver operating characteristic

SRA:

Sequence read archive

SVM:

Support vector machine

TCGA:

The cancer genome atlas