Every week or two one of the post-doc’s in our lab passes around a paper that is as much philosophy of science as experimental result. One that recently made the rounds is Can a Biologist Fix a Radio?, by Yuri Lazebnik, a spirited and humorous dissection of the flaws of the standard methods of experimentation and analysis in cell and systems biology. The primary contention is that most biologists are not sufficiently quantitative in their descriptions of the interactions of components in their system, which leads to an abundance of confusing, conflicting experimental results. If we hope to understand complex biological systems, we must describe the components in sufficient detail, with a standardized language, to allow accurate models built from numerous components. A radio can be perfectly described by its circuit diagram due to the quantitative knowledge of the component properties. Using a simplistic biological pathway diagram provides little insight.
Recently, a similar paper, On the Precarious Path of Reverse Neuro-Engineering, in Frontiers in Computations Neuroscience, highlighted a logical flaw in how much neuroscience research is done. When we find patterns or correlations between an external signal to neuronal activity, the tendency is to proclaim “Aha! This is how the circuit represents information and what it uses to make decisions.” But correlation does not equal causation. In this paper the authors build a physical Braitenberg Vehicle II, which is a component of a classic thought experiment in neuroscience. It navigates an environment by acquiring optical input, processing it through a network of cultured neurons, and then controlling a motor output through a simple rule about the firing of the neurons. Recording the activity of the neurons leads to the discovery of a ‘population rate code’ in the activity of the neurons, which describes the behavior of the robot fairly accurately. An experimental biologist would be tempted to call this finding the ‘neural code’ of the system. But, the navigation rule that robot actually follows is not based spike rates at all, but rather the order of first spikes in a handful of the neurons in the network. The correlation between population rate activity and behavior is real, but it is not what is relevant to the circuit and the behavior. This example cuts deep, and hits my unease with much current brain circuit analysis. There are huge numbers of papers that show particular stimulus features can be extracted from the firing patterns of neurons in a particular brain region, but rarely does a study conclusively demonstrate that the circuit actually uses these correlations in the computational task. Demonstrating causality is difficult, but essential if we are to gain an understanding of how brain circuits really work.