BAM and the Molecular Ticker Tape
In my little corner of cyberspace BAM is now generating no little controversy, so let’s review some of major points. The Brain Activity Map (BAM) project is a proposal to fund a massive collaborative effort between neuroscientists and others to characterize the dynamic activity of complete neural circuits, first in “simpler” animals, then mammals, perhaps primates, and finally humans. The proposal may have been behind president Obama’s praise for neuroscience in the state of the union address, but the outline of a proposal (that does seem to be in some flux) appeared in the New York Times last week. There are apparently many scientists involved in the project, but Harvard molecular biologist George Church and Columbia neuroscientist Rafael Yuste seem to be the ones taking calls from the media and bending the ears of presidential advisors. I recently talked to a scientist who was a post-doc in the Yuste Lab who helped develop some of the high throughput neural recording techniques that are at the center of the BAM proposal. These imaging and electrophysiology methods are already capable of monitoring the activity of about 1,500 cells in brain slices, and somewhat fewer in behaving animals. One of the main questions of the feasibility of the BAM project is whether these techniques can be scaled up to the proposed goals, which could include neural systems with 1,000,000 cells (still far below the size of a whole human brain). As Ian Stevenson has pointed out, cell-count advances in multicellular recording have proceeded at a very sub-Moore’s Law pace, doubling only every 7.4 years. There is good reason to believe that even with substantially better funding these techniques may quickly run into basic physical limits before given us the high throughput data we want from neural systems (see the figure for pace of progress curves for neural recording, computing and molecular biological techniques).
One scientist who is not willing to wait for the kind of complete data set he wants to understand the brain is Konrad Kording, head of the Bayesian Behavior Lab at the Rehabilitation Institute of Chicago (and co-author of the Stevenson paper). He has a surprisingly eclectic curiosity and a sharp sense of humor. He is senior author on a commentary paper in Nature (with Daniel Acuna and Stefano Allesina, 2012) which introduces a model for predicting a metric of a scientist’s future influence from his/her current productivity. I’m sure the web-based version they set up get many more hits than any of us would like to admit (What? Scientists egotistical? I’m shocked!). He’s also managed to get a LOLCats figure into one of his papers; clearly he’s not one for taking the conventional path. Since he calculated that he would be long dead before getting his hands on the data he needed by conventional methods (he was senior author on the Stevenson paper mentioned above), he started to think about how he might be able to use the rapidly advancing techniques of molecular biology to radically increase the throughput of neural recording.
I heard this story when I met with Kording in his 14th floor lab (with a great view of Lake Michigan). He recalled that the idea of using molecular techniques was sort of in the air when he began to think about it in 2011, but most scientists were skeptical that a practical approach could be found. Kording, however, presented a possible strategy in a whimsically titled paper, Of Toasters and Molecular Ticker Tapes (open access, by the way), where he suggests it could be possible to have individual cells (in genetically accessible animals such as mice) record their own activity in strands of DNA. This would, he explains, work by inserting template DNA strands along with the special versions of DNA polymerase into the cells. DNA polymerases are the molecular machines that are responsible for copying DNA (typically, when cells divide). This process is generally pretty reliable under normal conditions in the nucleus of cells, but with certain versions of the molecule in certain conditions, the copying process can be susceptible to error, and the error rate can be proportional to the amount of calcium ions in the cell. Calcium ions are one of the main types of electrically charged particles used by neurons process and transmit signals, and in many neurons the flow of calcium ions into the cell tends to be greatest when that cell is actively communicating with other neurons. In fact, as I mentioned last time, it is exactly this property that allows neuroscientists like Jason MacLean to read neural activity from calcium imaging. So if you could compare each strand of DNA created by DNA polymerase with the template strand, you could tell how active the neuron was by how many mistakes were made in the copy. Inside the neurons in the animal’s brain, many molecules of DNA polymerase would be busy copying the target DNA strand while the neuron goes about its normal business receiving, processing and passing on electrical signals. As Kording explains it, the scientist would need to provide a synchronizing signal of some sort, say with a prominent sensory signal or by inducing a seizure, then after the animal is put through its experimental paces, the DNA could be recovered from the brain tissue and sequenced using modern high throughput techniques. The error rate of each strand in comparison with the template strand would provide a readout of each cell’s activity, with multiple strands improving the precision of the measurement. By carefully processing the brain tissue in a fine 3D grid pattern, an activity map of the animal’s entire brain could be recreated.
According to Kording, there would be some limitations to the technique. DNA polymerase from E. coli can work up to speeds of 1,000 bases a second, which is comparable to current electronic recording methods. Still the dynamics of cellular calcium are slow compared to the finest temporal scales of neural signaling, so the exact timing of neural activity with this method would be a little blurry. Nonetheless, in a multi-institution collaboration, Kording (with his collegues Keith Tyo and Ed Boyden) helped researchers from George Church’s lab (Zamft, 2012, also open) to establish the basic feasibility of using DNA replication as a calcium sensor (at least in a test tube), but the polymerases are not yet sensitive enough to detect the levels of calcium you would find in a normally functioning neuron.
Moving one step further into the hypothetical, Kording is particularly excited about the possibility of combining the molecular ticker tape method with another transformative approach recently proposed by Anthony Zador (et al.). They suggest that some of the genetic machinery used by the immune system to generate combinatorial complexity could be used to create a unique genetic barcode for each neuron in an animal’s brain, and that a virus could be engineered to allow these barcodes to fused together in a way that would indicate which cells were linked into connected neural circuits. In principal, this would allow the connections between most of the cells in an animal’s brain to be identified. This connectome sequencing method could be combined with the molecular ticker tape method to produce a nearly comprehensive representation of the connectivity and activity of a single animal’s brain. Kording argues that such technologies are close enough to practical application that a coordinated research effort could produce dramatic returns in the proposed time frame of the BAM proposal.
In a previous post, I argued that analytical methods might not be up to the task of handling such data sets, but Kording disagrees. He points out that including connectivity data would actually make understanding the activity network feasible because it would limit the number of possible explanations for any given data set. Circuit models could be analyzed on a cell-by-cell basis in terms of the activity of only directly connected partners, reducing the complexity of the task dramatically. Without going too far into the math, this simplification is the difference between problems requiring planetary scale computers from ones that can be solved with contemporary clusters.
Correction (2/28/2013): In an earlier version, I mistakenly stated that the Zamft 2012 paper established the feasibility of DNA polymerase calcium sensors in cell culture. In fact, it was in vitro. Thanks to co-author Adam Marblestone (see comment below) for alerting me to the error.