Aaron J. Mackey, Ph.D. senior director HemoShea

Integrative analytics unmask dynamic human tissue systems.

A better understanding of the complex biology driving both human disease and drug response is critical to improving pharma productivity. Getting “back to human biology” enables better informed target and compound selection, produces more reliable investigative and predictive modeling, reduces uncertainty surrounding drug development risks, and allows more informative clinical studies. While immortalized cell lines and genetically tricked-out animal models provide a convenient proxy for human biology, they have not delivered reliable predictive accuracy to produce effective and safe drugs. The industry needs relevant human tissue systems that replicate in vivo human physiology and disease more accurately to better predict human response. However, with these more reliable systems comes a complexity of data interpretation that demands 1) broader thinking about how dynamic biological systems behave and 2) integrative computational methods to infer complex biological mechanisms from limited datasets.

Embracing Dynamicism

All living systems must employ complex networks of dynamic sensors, regulators, and effectors to ensure the maintenance of metabolic fluxes. These homeostatic systems never reach stasis, but are in constant compensation, governed by intricate feedback and feed-forward regulatory circuits, many of which exhibit oscillatory, temporal activity patterns. Oscillatory activity remains observable even in dissociated cell culture and is resynchronized upon each new perturbation (e.g., fresh addition of serum growth factors, new drug exposures, or any other external stimuli). Recent evidence demonstrates the vast rhythmic regulation of gene transcript expression across most mammalian tissues. In contrast, the majority of deductive biological research embraces a simplifying “constant effect” assumption, ignoring the influence of dynamic effects, e.g., testing simple “on-off” (or “up-down”) binary hypotheses of gene expression change in response to perturbation.

Human tissue systems are dynamic. They change over time, moving continuously from state-to-state across infinitesimally small time intervals. Between any two discrete snapshots in time, one can ask how the two states relate to each other—what changes, and what stays the same? One can attempt to build a causal model that predicts the state of the system at a future time, based on the state of the system at a previous time. A variety of methods exist for building such models, but all require input data to “learn” the models, and must be collected over multiple time points so that the temporal dynamics can be observed. The more time points that are collected and the greater the frequency of collection, the better the precision of the final model. Conversely, having observations at only a few time points, a dynamic system could be considered to be unstable or unpredictable, when in fact the system is exhibiting a biologically relevant, though complex, behavior that might be considered noise when observed at low resolution (Figure 1). The field must learn to embrace the idea of dysregulated genes that are neither up- nor downregulated in response to treatment, yet whose homeostasis has been nonetheless altered by drug perturbation.

Figure 1. Temporal dynamics of drug response. Illustrated are four of many possible dynamic responses in gene expression to drug perturbation. (A) The “typical” expectation that a drug causes a constant shift in an equilibrium baseline. (B) A variation of A where the equilibrium shift is achieved after a period of dynamical response. (C) Similar to A, a shift in the baseline but accompanied by a strong oscillatory element. (D) No shift in the baseline, but a change in magnitude of the oscillatory aspect. Arrows denote the fold changes that would be observed at particular sampling time points.

Integrative Computational Biology

As drug developers move away from conventional mono-cell cultures to translational tissue systems, unbiased genome-wide data interpretation must also evolve. Cells in living tissue are never at rest, but exist in perpetual dynamic response to hemodynamic shear forces, nutrient transport, and cell-cell communication. When drug perturbations are introduced to the system, responses must be measured and analyzed in the context of the baseline dynamics, with time-sensitive statistical models to test the score and magnitude of drug-induced dysregulation across time, rather than the magnitude of a single fold change seen at only one given time.

At HemoShear, we are using genome-wide transcriptomics as an unbiased metric of regulatory response to drug perturbation, coupled with a holistic, integrative method for biological pathway inference. For every transcriptomic drug response, measured at multiple time points, we deploy a variety of pathway enrichment search algorithms to explore an even wider variety of pathway databases. We have also built an ever-growing compendium of gene-gene co-expression modules, clustering together genes that respond in common cohort to a multitude of drugs. They also provide substantial opportunity to understand disease biology at a very deep level and identify novel therapeutic targets.

Use of statistical algorithms ensures that we neither miss drug-specific responses, nor include noise. For any given drug response measured across the tens of thousands of genes expressed in a given cell type, our various pattern search algorithms generate hundreds of pathways and modules. Despite the reduction in complexity achieved by collapsing gene- to pathway-level responses, many pathways and modules remain mechanistically redundant to one another, sharing the same key genes and regulatory connections. We employ the same network-clustering algorithms used for social networks, like Facebook and Twitter, to aggregate mechanistically-overlapping pathways into tens of cohesive functional “themes.” Themes can be thought of as organizing concepts that connect transcriptomic responses to functional endpoint assays used for tissue-level corroboration of the mechanistic hypotheses found at the molecular level (Figure 2).

As drug developers begin to leverage newly available translational human in vitro systems that more accurately reflect human in vivo biology, they must simultaneously embrace updated approaches to analytics. Instead of traditional benchmarks for experimental success that require artificial binary responses, understanding complex dynamic mechanisms will require a more holistic view and statistical approaches that look across both genes and mechanistic pathways. Embracing these challenges will yield an unprecedented ability to achieve a deeper and clearer understanding of the biology underlying human response and more confidence in target selection, validation and engagement with which the industry can expect to deliver safer and more effective drugs.

Functional themes as organizing concepts. For each transcriptomic study of drug response in HemoShear tissues, a suite of pathway enrichment tools are used to identify those biological processes and functions found significantly perturbed by the drug. Each pathway (represented as nodes in this graph) is connected to other significant pathways based on shared genes relevant to both pathways (represented as edges). Network community detection algorithms are used to identify strongly connected modules (identically colored nodes) that reflect excess commonality of drug perturbed gene responses. Each module can be summarized as a biological theme, and connected to functional endpoint assays that directly test mechanistic hypotheses evinced by each theme.

Aaron J. Mackey, Ph.D., ( [email protected])  is senior director of  computational biology at HemoShear.

Previous articleMost Popular Science Stories of 2014
Next articleAngelini to Commercialize Iclusig in Central, Eastern Europe