Figure 1. Re-iterative algorithm for data smoothing. Click to enlarge.

Scientists occasionally have to invent things to get their work done. In research areas involving elaborate experiments or heavy computation--neuroscience being an example of both--technical innovation can be a major component of research. One of the things that drew me to research was the chance to create new devices, techniques and computational algorithms.

As a first-year graduate student I studied lipid metabolism in cultured adipocytes (fat cells in a dish). Existing assays were far too insensitive for measuring glycerol and glucose levels in the extracellular fluid, so I developed a fundamentally different type of assay, based on radioactive tracers rather than the usual spectrophotometric techniques. The new assay is precise and linear down to about 50 pico-moles/liter--about 100 million times smaller than the typical plasma glucose concentration.

Figure 2. Relational database for electrophysiology experiments. Click to enlarge.

In later metabolic studies, I became concerned that the methods we used to smooth plasma metabolite time courses might introduce a bias, since the experimenter has to decide just how smooth he wants the time course to be. I developed a statistical/computational algorithm that not only chooses the smoothing level automatically, but in fact chooses that level in such a way as to strip away all but the meaningful component of the data.

When I set up a neuroscience laboratory at the University of Chicago, I wanted to avoid the "data nightmare" that is common in large labs. The problem is that every lab member saves his data in his own way, so it is nearly impossible for anyone else to analyze the data. We created a system of data storage that was standardized for all experiments, and in which the design of each experiment was completely evident from the data itself; that is, no external information (notes, files) was required. The core is an SQL relational database, designed for all electrophysiological experiments (not just ours) and extensively optimized and tested to bring it into second normal form. We wrote C++ and ODBC code to pump data in real-time during experiments into the dbase, and C++, ODBC and Matlab code to extract and analyze data. Thus far groups at Harvard and Stanford have implemented or expressed interest in implementing this database in their labs.

Figure 3. Modified microdrive for bundle technique.

Figure 4. Dispersion of electrodes in a bundle.

Much of the work in my Chicago lab involved electrophysiological recording from cortical neurons. Historically this was done with a single electrode, but since we were interested in ensemble coding, we needed to study multiple cells simultaneously. No existing technique was suitable for our needs. The method we designed uses existing technology--motorized advance, electrode fabrication, surgical methods, amplification--thus reducing the development time and maximizing the availability of the system to other labs. Instead, we created a bundle of 3-15 electrodes that was placed where an ordinary single electrode would be. The idea was simple but it took about 2 years to modify the electrodes, connectors and layout so that the bundle fit in a conventional micro-drive and the electrodes were optimally spaced to record from multiple neurons.