Implications of Big Data for cell biology

Abstract

“Big Data” has surpassed “systems biology” and “omics” as the hottest buzzword in the biological sciences, but is there any substance behind the hype? Certainly, we have learned about various aspects of cell and molecular biology from the many individual high-throughput data sets that have been published in the past 15–20 years. These data, although useful as individual data sets, can provide much more knowledge when interrogated with Big Data approaches, such as applying integrative methods that leverage the heterogeneous data compendia in their entirety. Here we discuss the benefits and challenges of such Big Data approaches in biology and how cell and molecular biologists can best take advantage of them.

What is “Big Data,” and what, if anything, can it do for cell biologists? The definition of Big Data is changing as rapidly as genomics data are being generated. All biologists are faced with data growing at a rate that could not have been imagined just 20 years ago, when most labs were still running polyacrylamide gels to sequence individual genes over the course of a couple of days. Soon after, results from a single microarray were intimidating enough to most biologists to be considered Big Data. Now, it is routine to analyze the entire compendia of expression and protein–protein interaction data. A similar “data avalanche” is happening in DNA sequencing, in which thousands of genomes are being analyzed in concert, and in imaging, in which cellular and organismal phenotypes can be systematically assessed in high-throughput format. Rather than setting a size threshold to define it (lest we fall into a “620K is enough memory for anyone” trap), Big Data is a moving bar that is set just beyond what we can, at a particular time, routinely annotate, analyze, and visualize—that is, Big Data is positioned where the challenges are in interpreting the wealth (and noisiness) of data now readily available. In other words, Big Data can be characterized by the three Vs: volume, variety, and velocity. The key question here is whether these virtual mountains of expression, sequence, proteomics, imaging, and other data can be transformed into biological knowledge in such way that it is both trusted and useful to cell biologists.

It is interesting that the two most-cited articles in this journal are among the very first Big Data papers (Spellman et al., 1998; Gasch et al., 2000). David Botstein showed, in a retrospective, that indeed it was the data that were important: roughly half of the citations for these articles came from computational biologists and statisticians (Botstein, 2010). Thus, these articles not only defined for the first time a genome-wide set of genes regulated by the cell cycle and stress response, respectively, but they also provided data for follow-up analyses, both experimental and computational, that enabled systems-level understanding of these processes and how they work in concert with other pathways. For example, a subsequent article used data from these two studies combined with growth rate under different limiting conditions to characterize the coordination of cell cycle, stress response, and growth rate in Saccharomyces cerevisiae (Brauer et al., 2008). Since those articles, Big Data has grown, not simply in size—more than 1.3 million samples are now available from the Gene Expression Omnibus database alone—but also in diversity. Nowadays, no area of molecular biology or genetics is insulated from high-throughput data—whether it is exploring genomic diversity in the context of evolution or human disease, considering epigenetic changes in development, or understanding transcriptional regulation of genes or posttranslational protein modifications. These are produced not only by individual laboratories, but also by large consortia, such as the Encyclopedia of DNA Elements (ENCODE) project to identify all the functional elements in the human genome (www.encodeproject.org/), the GTeX project to generate expression data across different human tissues (www.gtexportal.org/), and the LINCS program to produce a large-scale set of cellular signatures in response to different molecular or genetic perturbations (www.lincsproject.org/).

These data sets can already be helpful to the everyday bench biologist, where one might find some new information about particular genes of interest that are discussed in the text or pop up in a reanalysis of an individual data set. However, these data, when analyzed in concert both within and across data types have the potential to provide substantial biological knowledge, generating hypotheses that cannot be readily extracted from either the literature or any individual, even genome-scale, data set. For example, combining information in the Gasch et al. (2000) yeast stress response data set and the Spellman et al. (1998) yeast cell cycle data set can enable us to identify cell cycle–regulated genes that respond to DNA damage. This can then be extended with additional data to include identification of regulatory networks (e.g., by including chromatin immunoprecipitation followed by high-throughput sequencing [ChIP-seq] and nucleosome occupancy data). One such method, BETA, integrates ChIP-seq data sets that provide information about transcription factor binding with gene expression data informative of differential expression to predict direct target genes of transcription factors (Wang et al., 2013). Emerging Big Data analysis methods and prediction platforms can generate hypotheses about the function, interactions, and regulation of proteins, RNAs, and other biomolecules, their behavior in biological pathways, and relationships between various molecular entities and phenotypes. For example, Inferelator 2.0 uses a combination of Markov chain Monte Carlo and ordinary differential equations to learn both the topology and dynamics of regulatory networks in Halobacterium (Madar et al., 2009). Furthermore, many of these approaches can, through integrative analysis of Big Data collections, provide insight into biological contexts (such as specific tissues, developmental stages, and perturbations) that are challenging to directly assay experimentally. For example, in humans, the Genome-scale Integrated Analysis of gene Networks in Tissues (GIANT) webserver predicts functional maps of protein–protein relationships for 144 tissues through Bayesian integration of thousands of expression, sequence, and protein–protein interaction experiments (Greene et al., 2015); experimental biologists can explore these functional maps to better understand cell-specific processes in human disease.

In addition, Big Data, while naturally subject to signal-to-noise challenges, can compensate for the noisiness of each individual data set precisely because of its scale. Intuitively, signals that occur independently in multiple data sets are more likely to be “real”; for example, genes identified as cell-cycle regulated in multiple genome-scale studies are more likely to be truly cell-cycle regulated. Of course, simply identifying repeating signals can also zero in on common technical and biological artifacts or very broad (and thus often less interesting) biological signals, such as the general stress response that S. cerevisiae exhibit across essentially all treatments or broad growth regulators in human cell culture data. Sophisticated computational approaches based on Big Data collections, such as those described earlier, can specifically focus on the biologically informative signals relevant to a specific biological question, including those hard or impossible to detect by simple analyses.

Furthermore, as each individual experiment is inherently assaying only specific aspects of cellular complexity, combining data sets from different experimental conditions, platforms, and experimental approaches provides insights into molecular pathways that cannot be realized from individual studies done in isolation. Thus, analyses based on large collections of data hold the promise of systematic insights into how pathways function across mechanistic interaction and regulation modes, spanning, for example, transcriptional regulation by histone proteins and transcription factors (from ENCODE data), RNA stability effects, and protein transport, interactions, and posttranscriptional modifications from imaging and proteomics studies.

In the past decade, many systematic approaches have been developed for integrating across genomic Big Data collections to provide novel biological insights—for example, to identify biological processes in which genes with unknown function participate (Pena-Castillo et al., 2008) and to predict physical and regulatory protein interactions and posttranslational modification in a large-scale, automated way (Vaske et al., 2010; Zhong et al., 2014; Park et al., 2015). These approaches use statistical techniques that aim to isolate signal from noise in these diverse and heterogeneous data collections, often relying on examples of known biological associations (e.g., genes with previously discovered biological function) to identify informative signals and make new discoveries. For example, Integrative Multi-species Prediction (IMP; imp.princeton.edu) probabilistically combines a large collection of expression, sequence, and protein interaction data to provide functional network and function predictions for any protein in the human and major model organism genome or proteome. Although most of these methods analyze data in preset ways, a recent trend includes development of approaches that enable the user to focus analysis on a specific biological area or question, essentially directing the analysis of Big Data without having to do any programming. Many Big Data integrative methods now provide highly targeted analysis, such as in the tissue-specific functional networks provided by GIANT (giant.princeton.edu) or Th17-focused prediction of TH17 regulatory networks (Ciofani et al., 2012).

In fact, integrative analysis of functional genomics data coupled with computational modeling has effectively directed laboratory experiments and given rise to novel experimental discoveries in multiple model organisms and humans (Hess et al., 2009; Yan et al., 2010; Guan et al., 2012; Wong et al., 2012). For example, Doherty et al. (2012) showed that the BLM10-20S proteasome activator mediates DNA damage and other cellular stresses, in part by examining the predicted functional networks of the genes that were induced in blm10 mutants using the BioPIXIE tool (Myers et al., 2005). Similarly, Sanchez-Garcia et al. (2014) integrated expression data from primary breast tumors with data from RNA interference screens using their Helios tool to identify candidate cancer-driver genes. They went on to experimentally characterize one of their novel predictions, RSF-1, showing that when amplified, RSF-1 increased both tumorigenesis and metastasis in mouse models of breast cancer.

Such data-driven approaches provide an important complement to the highly curated, aggregate databases that provide access to valuable information such as comprehensively curated physical and genetic interaction data (e.g., BioGRID; Chatr-Aryamontri et al., 2015) or phenotype information for model organisms through the model organism databases (e.g., Engel et al., 2010; Bult et al., 2013; Deans et al., 2015). Because they are based on high-throughput data, not literature-based curation or collections of specific experiments, the Big Data–based resources tend to be less biased toward prior knowledge and are able to make predictions even in areas in which prior knowledge may be very sparse or nonexistent. The price for this is of course the higher potential for errors due to noise levels in the data, although these can be mitigated by careful analysis, making genomic data collection a great source of hypotheses that can drive traditional experiments. Together the Big Data–driven methods and the curated databases are powerful tools for the cell biologist. The curated data within the databases can serve as important gold standards for evaluating computational predictions, and the predictions can be used to guide and refine the annotation provided by the curated databases.

Big Data also has the potential of revolutionizing our use of model organisms, enabling accurate, less-biased, molecular-level identification of the most informative model for genes and diseases in the least expensive and most tractable experimental system. The key advantage is the ability to go beyond sequence-based orthology to systematically assess functional conservation, promising a functional mapping of proteins, pathways, and phenotypes across organisms. For example, biologists can use a method based on probabilistically mapping protein networks from a large compendium of high-throughput expression data across organisms to systematically predict which genes are most likely to participate in the same biological process and thus have analogous function in different organisms (Singh et al., 2008; Chikina and Troyanskaya, 2011; Park et al., 2013). Such approaches can succeed where sequence-based methods often fail, such as resolving paralogues based on tissue expression and correctly identifying functional divergence when orthology is predicted based on sequence and evolutionary relationships. The growing Big Data compendia in model organisms and humans, combined with sophisticated computational approaches, are bringing in an era in which we will be able to quantitatively and systematically identify the best experimental model (or models) for a given disease or process, pinpoint specific aspects of relevant biology that are or are not conserved across organisms, and generally be able to effectively and accurately integrate our knowledge across organisms.

What does this mean for a cell biologist? This means that all biologists contemplating their next study should consider using Big Data—based tools to inform their hypotheses, whether to identify additional proteins that may be relevant to the process they are studying, examine predicted molecular functions of a protein of interest, or consider pathways that may be relevant to the experimental treatment or genetic modification they are considering. If they are interested in a specific cell type or developmental stage, they can use Big Data–based resources to identify proteins expressed in this cell type, tissue-specific interactions and functions, and perhaps even cell-type specific predictions of perturbation effects on phenotypes. Many of the prediction systems and algorithms necessary for these analyses are available publicly, often in a user-friendly form aimed at biomedical researchers with no or limited computational training (Table 1). A more in-depth analysis, especially involving unpublished data from a cell biologist's laboratory, can be undertaken either in their own lab (if they have the requisite computational skills or through the newly emergent systems that enable sophisticated computational analysis by nonspecialists; e.g., Greene and Troyanskaya, 2011) or with a computational collaborator. Most institutions now include “card-carrying” computational biologists or bioinformaticians, as well as many experimentalists with substantial computational expertise, although looking for the best collaborative fit may require crossing departmental boundaries into computational biology, computer science, or similarly focused departments.

TABLE 1:

Examples of user-friendly systems for Big Data analysis in biology.

galaxyproject.org	Platform for genome-scale biomedical research
imp.princeton.edu	Functional networks in model organisms and humans
giant.princeton.edu	Tissue-specific networks and genome-wide association studies in humans
thebiogrid.org	Database of protein and genetic interactions
seek.princeton.edu	Cross-platform search engine for expression data
genomespace.org	Framework for integrative genomics analysis
cbioportal.org	Visualization and analysis of cancer genomic data

Does this mean that experimental cell biologists should look for alternative careers? Absolutely not! Computational approaches based on Big Data generate hypotheses, not experimentally verified biological knowledge. In addition, the broader, less-biased, Big Data–driven information can be a powerful guide for cell biology studies. In the past, cell biologists would read articles and look at their last gel to inform the next set of experiments. As we continue to make progress in harnessing Big Data, cell biologists can obtain a new, valuable tool from its the broader, less-biased information. Cell biologists who do not use Big Data to inform their experiments are squandering a valuable resource. It is analogous to a biologist doing DNA amplification manually in water baths when PCR machines are available. The wealth that Big Data brings will enable cell biologists to better design and focus their experimental programs with the expectation that biological insights will come faster and more efficiently. We are not even close to replacing individual experiments (and the cell biologists who do them!) with computers, but instead are in the midst of an exciting time when we are just beginning to tap the major effect of Big Data on the world of cell biology.

Abbreviations used:

GTeX

Genotype-Tissue Expression

LINCS

Library of Integrated Network-based Cellular Signatures.