Big Data: Challenges and opportunities for clinical pharmacology

An eminent physicist has remarked that the future truths of Physical Science are to be looked for in the sixth place of decimals. Scientists are trained to recognize that correlation is not causation. … Petabytes allow us to say: ‘Correlation is enough’. We intend, in this editorial, to highlight some key elements constituting Big Data, and to reflect on associated challenges and opportunities for clinical pharmacology. Two of the initial challenges are evidence of perceptual incongruity encumbering Big Data: (1) its definition and (2) the pretence of its emergence as a 21st century zeitgeist in science. Its definitions vary. Our working description of choice is by the Oxford Dictionaries 3, to wit: extremely large data sets that may be analyzed computationally to reveal patterns, trends and associations, especially relating to human behaviour and interactions. The ‘zeitgeist’ reference is to the school of thought that Big Data constitutes a philosophical revolution in science, but this seems a far-fetched embroidery of its capacity and, thus, its epistemological place in science. The ‘eminent physicist’ referred to by Michelson in the epigraph is thought to be Lord Kelvin and, perhaps the source of the quote misattributed to Kelvin: ‘There is nothing new to be discovered in physics now. All that remains is more and more precise measurement’. There is no doubt that both statements reflected the view of physics at the close of the 19th century, that there was no longer a need for theory, just data analysis. Enter Einstein, Heisenberg et al. ‘The tendency to conflate the undisputed usefulness of Big Data – which above all, is an informational tool – with its presumed ability to provide full scientific understanding, sometime leads Big Data specialists to overstate their claims’. ‘Big Data that are made available through data bases for future analysis turns out to represent highly selected phenomena, materials and contributions, to the exclusion of the majority of biological work. What is worse, this selection is not the result of scientific choices, which can therefore be taken into account when analyzing the data. Rather, it is the serendipitous result of social, political, economic and technical factors, which determines which data get to travel in ways that are non-transparent and hard to reconstruct by biologists at the receiving end’. This statement cautions that the vast sources of information can provide a great wealth of insight while simultaneously providing significant challenges to reality. Some of these concerns have been dismissed or considered as minimized. Yet they persist as substantial hurdles that must considered when exploiting the Big Data information pool. These challenges include the supposition of an n = all information capture (presumably obviating the need for statistics), validly linking information across disparate sources, whether to implement a supervised vs. unsupervised statistical learning approach in analyzing data sources, how to map known physiological or pharmacological interactions as a scaffold for these, assumption of a static system, the assertion that all necessary empirical observations have been made and that the data collection is complete and the assertion that correlation is sufficient and that causation is no longer necessary for valuable information to be extracted from Big Data datasets. ‘There are a lot of small data problems that occur in Big Data...They don't disappear because you've got lots of the stuff. They get worse’ 9 Collection of all available data does not necessarily ensure that all possible data points have been captured. Important sampling bias may still exist. For example, scientists at Rutgers University considered the possible scenario where government agencies might make decisions about disaster relief based on social media data. Their analysis of Twitter usage during Hurricane Sandy led to the conclusion that the borough of Manhattan suffered the worst damage and that partying peaked after the storm subsided. Of course, this was the entirely inaccurate result of biased data collection. The Big Data dataset contained relatively few tweets from the most severely affected coastal areas because of power outages, failing cell phone batteries and a general trend toward lower social media usage in those communities compared with Manhattan 10 When linking information across disparate sources, consideration must be given to the type of data, the data collection process, the level of detail of the data and the time epoch when the data were collected. All of these factors can potentially confound the analysis. The choice of a supervised vs. an unsupervised learning approach hinges on whether a target outcome that is desired to be predicted. A supervised learning approach requires having system output (target outcome) and then evaluation of possible predictors. These approaches range from LASSO, linear regression through to more complex approaches such as support vector machines and classification algorithms for the selection of optimal predictors 11. Many of these more sophisticated supervised learning approaches are automated, leveraging the power of machine learning to facilitate the exploration of the system. Unsupervised learning approaches, on the other hand, set out to evaluate the relationships among a set of variables in a dataset without attempting to predict anything in particular. Clustering algorithms or modularity analyses are those approaches that reveal interrelationships among variables in a dataset and can be useful in understanding how information groups together. However, these analyses are limited to correlational assessments and do not consider the underlying mechanistic constraints of the system. Mapping physiological or pharmacological information as a scaffold can be a very powerful means of organizing information in datasets. In particular, it can be a framework related to cellular functioning or, perhaps, to a pharmacologic pathway where data can be applied in a structured sense to organize the information gleaned on the basis of what is known about the system. These models also assume the system observed is static. By creating a large network of correlations, there is little information on mechanism or the fundamentals of the system that would allow for it to be utilized for prediction. This makes Big Data predictions and models vulnerable to changes in the system. One may develop a predictive model in a particular set of data that arose under very specific conditions. If those conditions change, or if the system changes dynamically, the model will no longer be able to predict outcomes adequately. This is what was observed with the Google Flu predictor where search terms were analyzed to determine and compared with flu outbreaks 12. It was a robust predictor at the outset. Publicizing the predictor was followed by a subsequent two-fold over-prediction of influenza-like along with complete misses of flu outbreaks (including the non-seasonal A-H1N1 pandemic in 2009). Plausibly, knowledge of the predictor changed what individuals were searching for and therefore confused the classifier/predictive algorithm. Lazer et al. point out that ‘quantity of data does not mean that one can ignore foundational issues of measurement and construct validity and reliability and dependencies among data’ attributing this issue as part of the problem with the Google Flu predictor 12. This is not to distract from the fact that Big Data usher in an information age that is and will continue to be will no doubt be transformative by the power it affords the scientific enterprise. For example: Lang Li at the Center for Computational Biology and Bioinformatics, Indiana University School of Medicine focuses on identification of drug interactions in the context of personalized medicine. The first step in his work was to establish an ontology and corpus of information using text-mining techniques that identify likely drug–drug interactions by searching abstracts of publications in PUBMED. These hits were assessed for both the intensity of the interaction reported as well as the significance level. Once pruned, the most interesting interactions predicted using these Big Data text mining approaches were investigated further 13. In the case of a loratadine and simvastatin interaction identified using this technique, evaluation of interaction at various CYP enzymes and transporters was undertaken. When these assessments did not result in a mechanistic explanation of the possible interaction, Han et al. went on to examine interactions between these compounds at the level of the muscle cell 14. What was identified was a pharmacodynamic interaction at the level of the muscle tubule that otherwise would not have been known. This work contributed to knowledge of mechanisms that may help to understand the high rates of myopathy observed with statin treatment. Another example is the studies by Bernhard Palsson and coworkers at UCSD. They apply systems frameworks to -omics data sources. Their work has facilitated an understanding of cellular dynamics with such detail that they have predicted across scales using this approach ‘from bacteria to human’. This was accomplished using non-linear optimization in conjunction with the -omics datasets that provide the constraints necessary to focus these large quantities of information into useful scaffolds that further increase the insight that can be discovered. Recent accomplishments include publications reporting ‘systems biology guided identification of synthetic lethal gene pairs and its potential use to discover antibiotic combinations’ 15 and the determination of a metabolic signature that facilitates the evaluation of quality of a unit of red blood cells 16. The red blood cell work utilized an approach that combined endothelial markers and clinical outcomes to develop an algorithm that determines the red blood cell unit quality. These are significant challenges, the opportunity to identify novel antibiotic combinations that may have activity in resistant organisms and the means of evaluating the likely utility of a unit of red blood cells with respect to clinical outcomes, and highlight how Big Data can be leveraged to address critical public health problems. Both of these examples illustrate the utility of Big Data techniques for generating hypotheses or identifying gaps in current scientific knowledge, but they also demonstrate the importance of going beyond signal detection and exploring potential explanatory and mechanistic underpinnings for their findings. These findings are often extended with specific experiments to refine findings and provide new insights that make important contributions to their respective fields. Big Data provide enormous opportunities for clinical pharmacology as a nexus to help guide and provide frameworks that leverage this resource, but with full cognizance of limitations. Accordingly, while ‘cheerleading’ from the popular press is one thing, we find similarly conveyed acclamation from the scientific community to be disquieting. Even the ‘BD2K’ label the NIH attached to its data science initiative has this flavour. We feel compelled to caution: ‘Be careful, there are impressionable students listening. Do you really want to introduce them to the end of theory?’ The enthusiasm for the remarkable instrumental utility Big Data is justified. Nonetheless, the logic of scientific discovery, manifest within hypothesis and theory, lives, waiting, perhaps, for another century when its demise is again predicted.