In the past two decades, the advancement of high-throughput technologies has led to the discovery of genomic, transcriptomic, and epigenomic modalities involved in cancer initiation, progression, and treatment response. Multiple groups have started to effectively utilize molecular data produced by high-throughput oncology experiments to identify biomarkers of progression and therapeutic response in cancer patients (Sorlie et al., 2001; Zhang et al., 2001; van’t Veer et al., 2002; Zhan et al., 2002, 2006; Sotiriou et al., 2003; Ayers et al., 2004; Allen et al., 2006; Jain et al., 2009; Lim et al., 2009; Petty et al., 2009; Zhao et al., 2009; Carro et al., 2010; Lefebvre et al., 2010; Shaughnessy et al., 2011; Bae et al., 2013; Aytes et al., 2014, 2018; Mitrofanova et al., 2015; Robinson et al., 2015; Wang et al., 2016; Giulietti et al., 2017; Heng et al., 2017; Hoadley et al., 2018; Abida et al., 2019; Epsi et al., 2019; Arriaga et al., 2020; Panja et al., 2020; Rahem et al., 2020). Yet, our understanding of the mechanisms involving these modalities, their upstream regulation, and effective therapeutic targeting remains incomplete.

A biomarker is an objective measure (e.g., classically a genomic/transcriptomic/epigenomic alteration, gene, protein, metabolite, or their groups), typically used to predict the incidence of disease, its progression, or treatment outcome (Strimbu and Tavel, 2010; McDermott et al., 2013). In the context of oncology, biomarkers are classically used for cancer risk assessment and screening, tumor staging, disease recurrence, selection of initial therapy, alternative therapy choices, and monitoring for therapeutic toxicities (Ludwig and Weinstein, 2005). While employed in clinical use, the existing biomarkers are still sparse and suffer from issues of reproducibility and heterogeneity, alongside a lack of understanding of their underlying regulatory mechanisms (Ludwig and Weinstein, 2005; Boutros, 2015).

One of the reasons for such a knowledge gap is the fact that the majority of biomarkers are identified from gene-centric approaches (we will refer to gene/protein/metabolite etc.,-centric approaches as gene-centric approaches for simplicity), where either a specific gene is investigated (based on previous biological assumptions) or a gene(s) is selected based on differential behavior without connection to the upstream and downstream molecular mechanisms. Gene-centric findings are often limited in mechanistic interpretability and connectivity to other molecular processes, positioning such biomarkers as passengers, rather than drivers, of the biological process and thus are often dataset specific (Michiels et al., 2005; Chng et al., 2016).

In classical gene-centric approaches, genes (without their connections to one another or underlying mechanisms) are utilized as inputs into white- and black-box statistical and machine learning models, which have been successfully applied to identify gene-centric markers in breast cancer (van’t Veer et al., 2002; Wang et al., 2005; Zhang et al., 2013), lung cancer (Beer et al., 2002), multiple myeloma (Shaughnessy et al., 2007; Kuiper et al., 2012), colon cancer (Zhang et al., 2001; Yan et al., 2012), and prostate cancer (Garzotto et al., 2005; Erho et al., 2013), among many others. It is important to note that in white-box models (e.g., linear regression and decision trees) the relationship between input variables (i.e., genes) and output variables (i.e., disease outcomes) is understandable/explainable as they often identify linear or monotonic relationships (Zhang et al., 2001; Garzotto et al., 2005; Rosenfeld et al., 2008; Huo et al., 2017; Panja et al., 2018). On the other hand, black-box models (e.g., neural networks, gradient boosting, or ensemble models such as random forest) are able to capture non-linear/non-monotonic relationships, yet often suffer from model interpretability and subsequent limited clinical adoption (Wang et al., 2009; Ayer et al., 2010; Zhang et al., 2013). Even though both white- and black-box learning are excellent tools for predictive modeling, they mostly capture associative relationships when applied as gene-centric approaches and often miss the complexity of mechanisms inherent in biological systems, especially in the context of cancer.

Several groups have addressed this problem by developing biomarker discovery methods based on mechanism-centric approaches, which are not focused on single genes and take into account complex mechanisms implicated in cancer initiation, progression, and treatment response. In this review, we will discuss the mechanism-centric approaches based on construction and mining of co-expression networks (Freeman, 1977; Zhang and Horvath, 2005; Zhang and Huang, 2014; Han et al., 2016), regulatory networks (Basso et al., 2005; Lefebvre et al., 2010; Alvarez et al., 2016; Dhingra et al., 2017), protein–protein interaction (PPI) networks (Chuang et al., 2007), and molecular pathways (Epsi et al., 2019; Rahem et al., 2020; Figure 1). Through an in-depth understanding of upstream and downstream molecular mechanisms, such techniques open a door for the discovery of functionally interpretable molecular drivers (rather than passengers) and potential targets for precision therapeutics.

Mechanism-centric approaches provide a powerful solution for informed biomarker discovery, yet common challenges that these methods need to account for include sufficient cohort sizes, data variability and scaling, comprehension of existing knowledge bases, and tissue-specificity (Table 1).

As many of these methods utilize association-based analyses (i.e., correlation, mutual information, regression, etc.), a sufficient cohort size is required to be able to accurately estimate relationships between variables. One of the direct solutions to this problem includes combining analyses in multiple datasets; however, batch effects among different acquisition methods, profiling platforms, and even institutions where datasets were collected might hamper such implementation.

In addition to a sufficient cohort size, substantial variability of expression profiles is also required to be able to accurately predict associations between variables. This task is feasible, yet it requires careful consideration, meticulous initial experimental design, and in-depth investigation of the amount of final variability necessary for successful analysis. Another challenge is the need for well-defined phenotypes, as they often require a substantially large number of samples inside each phenotype group while also demanding intra-sample homogeneity, as in the eigengene approach, MARINa, PPI network-based method by Chuang et al., pathCHEMO, etc.

At the same time, methods that rely on single-patient/sample mining (e.g., VIPER, the PPI network-based method by Chuang et al., and pathER) rely on dataset scaling to define its single-sample signatures (defined by comparing each gene to the average of its expression in the dataset of interest) making interpretation of any findings from such analyses dataset-specific.

Another known challenge is tissue-specificity, commonly faced in PPI network-based and pathway-based approaches, though some tissue- and cell-specific interaction databases are now available such as TissueNet (Basha et al., 2017), the Integrated Interactions Database (Kotlyar et al., 2019), and HumanBase (Greene et al., 2015). Tissue-specificity in these methods is usually achieved by overlaying gene expression data onto the PPI networks or molecular pathways, such as in Chuang et al., pathCHEMO, and pathER.

Furthermore, limitations of mechanism-centric approaches that utilize knowledge bases (e.g., RegNetDriver, PPI network-based approach, pathCHEMO, and pathER) lie in their reliance on known biological relationships among groups of genes/proteins/other functional units contained within a database. Various annotation, pathway, and PPI databases depend on existing information and do not include functional units that have not been previously studied, thus limiting de novo discoveries.

The wide availability of large-scale data produced by high-throughput technologies has created a wealth of information for biomarker discovery. A vast majority of these biomarkers have been identified using gene-centric methods, yet their interpretability and clinical utility have been limited as they do not account for the relationships among genes. Utilizing methods that consider biological underpinnings of the data (i.e., mechanism-centric methods) can vastly improve interpretable biomarker discovery, clinical applicability and targeting, and reproducibility of results.

In particular, advantages of mechanism-centric over gene-centric approaches can be illustrated through their ability to (i) identify a tightly connected, cooperative group of genes unified by the same function, as opposed to individual genes (which might not be related); (ii) provide a mechanism-level view, which enhances the understanding of the biological mechanisms implicated in a phenotype (e.g., therapeutic resistance, cancer metastasis); (iii) look at alterations in biological structures, which enhances the likelihood of identifying functionally relevant targets; (iv) identify driver as opposed to passenger markers, which allows for their effective therapeutic targeting; (v) focus on molecular structures, rather than individual genes, which decreases the chance of detecting results due to experimental noise present in biological experiments (i.e., robustness of results); and finally (vi) identify biomarkers that are more accurate and more reproducible between different cohorts.

From a computational point of view, mechanism-centric approaches can be used for interpretable feature engineering and selection (i.e., reduction), subsequently reducing the number of hypotheses to be tested. This is clearly demonstrated by gene co-expression networks, regulatory networks, PPI networks, and pathway-based methods, where cooperative groups of genes, instead of a long list of singular genes, are assessed for their association with clinical outcomes.

Mechanism-centric methods can both (i) provide interpretable inputs to white- or black-box approaches or (ii) contribute to inner model interpretability (i.e., such as in visible machine learning). First, results from mechanism-centric methods can be utilized as inputs into learning models to significantly improve predictive performance (over gene-centric inputs). One such example was demonstrated in Rahem et al., where pathway-based markers were utilized as inputs into Cox proportional hazards regression modeling and outperformed gene-centric markers for tamoxifen resistance in ER-positive breast cancer (Rahem et al., 2020). Similarly, Chuang et al. showed that markers identified by their PPI network-based method could be effectively used as inputs into a regression model and outperformed gene-centric markers in classification of metastatic breast cancer (Chuang et al., 2007). Though not in cancer, several methods have also suggested utilizing hierarchical structures (such as those inherent in Gene Ontology) as inputs for predictive models (Carvunis and Ideker, 2014; Yu et al., 2016). Second, mechanism-centric methods can potentially be incorporated into model building, such as in “visible learning,” where the relationships between inputs and outputs can be interpreted (Yu et al., 2018). One such (outside of cancer) neural network method, DCell, was proposed by Ma et al., where the hierarchy of molecular relationships determined from prior knowledge (Gene Ontology and CliXO) was built into the model itself (i.e., hierarchies were utilized by nodes of the neural network) (Ma et al., 2018). Recently, Kuenzi et al. developed an extension of DCell, called DrugCell, which utilized chemical drug structures as a part of the neural network learning model to predict drug response in cancer cells (Kuenzi et al., 2020). This interpretable deep learning model was shown to be able to predict cell sensitivity/resistance to specific drugs, synergistic drug mechanisms, and effective drug combinations for treatment.

Further improvements in the interpretability of biological processes that inform discovery of mechanism-centric biomarkers can be made through multi-level data and method integration. For example, several groups have combined co-expression WGCNA modules with PPI networks to uncover hubs with functional connections as biomarkers in endometrial cancer (Liu et al., 2019) and bladder cancer (Wang Y. et al., 2019). Wang et al. constructed an Active Protein-Gene network model using transcriptional regulatory and PPI networks to quantify TF activity and elucidate both upstream and downstream regulations (Wang et al., 2013). Even though this study was done in diabetes, it could be applicable to mechanism-centric biomarker discovery in cancer. Ahsen et al. embedded VIPER within a new framework (NeTFactor) to identify TFs that most likely regulate a gene-centric biomarker signature (Ahsen et al., 2019). While this method was applied to asthma and peanut allergy, it could easily be extended to cancer studies. At the same time, multi-omic integration in RegNetDriver improved the interpretability of the proposed model to explain the impact of mutations, structural variants, and DNA methylation on TF activity in prostate cancer (Dhingra et al., 2017). A recent study by Broyde et al. constructed a multi-omic lung adenocarcinoma tissue-specific oncoprotein interaction network using information obtained from ARACNe, CINDy (an algorithm identifying post-translational modulators), VIPER, and PPI predictions (Broyde et al., 2021), which depicted a complex network of interactions for KRAS and could potentially be utilized for mechanism-centric biomarker discovery. Such multi-level approaches in conjunction with mechanism-centric methods promise to uncover a deeper understanding of mechanisms involved in gene regulation and post-translational modifications in biomarker discovery.

Finally, recent technological advances, such as those seen in single-cell studies, promise to improve our understanding of intra-tumor heterogeneity, clonal evolution, and the role of microenvironment in cancer progression and therapeutic response. Single-cell gene expression offers a granular view of active pathways in a cell type-specific manner and potentially allows for the construction of cell type-specific networks. In fact, the rapid advances of single-cell sequencing technology have already allowed network analysis methods to be applied directly to data from single-cell RNA-sequencing (scRNA-seq) (Crow et al., 2016; Aibar et al., 2017; Chan et al., 2017; Fiers et al., 2018; Papili Gao et al., 2018; van Dijk et al., 2018; Lamere and Li, 2019; Jackson et al., 2020; Sekula et al., 2020; Ye et al., 2020) with integration of other data modalities for improved network inference (Aibar et al., 2017; Chan et al., 2017; Papili Gao et al., 2018; van Dijk et al., 2018; Jackson et al., 2020; Pratapa et al., 2020). Furthermore, matching single-cell and bulk patient samples could provide an invaluable resource for single-cell driven network investigations that can be compared to and related back to bulk tissues. As more single-cell data become available (e.g., RNA sequencing, targeted DNA sequencing, ATAC-seq, etc.), we foresee advances in single-cell technologies and data analysis to be central to understanding precise, clone-specific biomarkers, unveiling trajectories of tumor evolution and providing accurate ground for informed time-cautious precision therapeutics.

In summary, mechanism-centric approaches (based on gene co-expression networks, regulatory networks, PPI networks, and molecular pathways) identify biomarkers that are biologically meaningful, interpretable, reproducible, have higher translational potential, and provide greater predictive power over biomarkers identified by gene-centric methods. Thus, mechanism-centric approaches are the future of clinically relevant rational biomarker discovery, personalized treatment planning, and precision therapeutics in cancer.

CY and AM conceived and wrote the manuscript. Both the authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.