Deep Science

The decade of the brain is maturing into the century of the mind. New bioengineering techniques can resolve and perturb brain activity with unprecedented specificity and scope (including neural control with optogenetics, circuit visualization with fiber photometry, receptor manipulation with DREADDs, gene sculpting with CRISPR/Cas9, and whole brain mapping with CLARITY). These technical advances have captured well-deserved media coverage and inspired support for brain mapping initiatives. But conceptual advances are also needed. More rapid progress might occur by complementing existing “broad science” initiatives with “deep science” approaches capable of bridging the chasms that separate different levels of analysis. Thus, some of the most interesting neuroscientific news on the horizon might highlight not only new scientific content (for example, tools and findings) but also new scientific approaches (for example, deep versus broad science approaches).

What is “deep science”? Deep science approaches seek first to identify critical nodes (or units) within different levels of analysis, and then to determine whether they share a link (or connection) across those levels of analysis. If such a connection exists, that might imply that perturbing the lower-level node could causally influence the higher-level node. Some examples of deep science approaches might include using optogenetic stimulation to alter behavior, or using FMRI activity to predict psychiatric symptoms. Because deep science first seeks to bridge different levels of analysis, it often requires collaboration of at least two experts at different levels of analysis.

The goals of deep science stand in contrast to those of broad science. Broad science approaches first seek to map all nodes within a level of analysis as well as links between them (for example, all neurons and their connections in a model organism such as a worm). Comprehensive characterization represents a necessary step towards mapping the landscapes of new data produced by novel techniques. Examples of broad science approaches include connectomic attempts to characterize all brain cells in a circuit, or computational efforts to digitally model all circuit components. Broad science initiatives implicitly assume that by fully characterizing a single level of analysis, a better understanding of higher-order functions will emerge. Thus, a single expert at one level of analysis can advance through persistent application of relevant methods.

Due to more variables, methods, and collaborators, deep science approaches pose greater coordination challenges than broad science approaches. Which nodes to target or levels to link might not be obvious at the outset, and might require many rounds of research. Although neuroscientists have long distinguished different levels of analysis (for instance, Marr’s descending goal, process, and hardware levels of analysis), they have often emphasized one level of analysis to the exclusion of others, or assumed that links across levels were arbitrary and thus not worthy of study. New techniques, however, have raised possibilities for testing links across levels. Thus, one deep science strategy might involve targeting links that causally connect ascending levels of analysis. For instance, recent evidence indicates that optogenetic stimulation of midbrain dopamine neurons (the hardware level) increases FMRI activity in the striatum (the process level), which predicts approach behavior (the goal level) in rats and humans.

While deep science findings are not yet news, I predict they soon will be. Deep science and broad science are necessary complements, but broad science approaches currently dominate. By linking levels of analysis, however, deep science approaches may more rapidly translate basic neuroscience knowledge into behavioral applications and healing interventions—which should be good news for all.