Peter Dayan of Theoretical Neuroscience fame gave a talk today at Harvard University that I was fortunate enough to attend. One of the great things about Boston/Cambridge area is the sheer pull of our collective universities in bringing top neuroscientists to the area. Anyway, some of his newer work that he was presenting today was entitled, “Norepinepherine and Neural Interrupts.”
There are about four main neuromodulatory systems that each use a particular neurotransmitter as its chemical of action. They seem to have wide ranging connections that span large neural surfaces, and they include dopamine (DA), acetylcholine (ACh), serotonin (5-HT), and norepinepherine (NE). As much as brain chemistry is certainly important, the effects of these particular neuromodulators on the systems that they innervate is incredibly complex, as they make millions of synapses on a what seems to be a wide variety of neurons. Dayan was presenting an interesting viewpoint on the role of NE involved in uncertainty states. The essential idea is that uncertainty must be a top-down mediated process, which starts either with an expectation or no expectation at all. Both of these states are uncertain states, which can be restated as either an expected uncertainty or an unexpected uncertainty. In either case, the brain’s environment must be favorable to a learning condition.
Consider your walk through your school or office. Let’s say that you know that they have been remodeling your wing of the office, so as soon as you turn the last corner, you have an expectation that things will look different. You know already that uncertainty exists. This situation may be neurobiologically different from if you had no idea that they painted the walls pink with green dots, so when you turn the corner, there is (certainly!) an unexpected uncertainty.
Presumably, ACh mediates the former – that is, the expected uncertainty, where NE may modulate the unexpected uncertainty. The experimental evidence seems to come primary from non-human primate studies that measure levels of NE modulation during a target-response visual task that shows a spike in NE pathway activation after the salient target is presented, and no change from baseline levels of activation with the presence of a distractor.
Dayan’s model was a simple state dependent modulation between either the target and distractor that included probabilities of NE activation due to being presented with the two stimuli. He had a pretty simple error predictor that was pointed out did not account for motor type errors, which presumably should be averaged and removed across task difficulty – that is, as task difficulty increases, the number of motor specific mistakes may stay relatively constant. I think this is a reasonable assumption for present purposes.
However good the model was at reproducing the qualitative features of the experimental evidence – and it was, more or less – it was very difficult not to call into question several possible shortcomings. The model set an arbitrary threshold for activation, at 95% probability required for activation, but there did not seem to be any physiological reason for such a threshold. There were also some unexplained qualitative microfeatures of the model’s output that also seemed curious. Additionally, the model included some features such as the random resetting after a certain number of trials that did not seem to have an immediate physiological basis, though I may well just not understand the particulars of the system.
For being a model that simply takes into account the states of NE pathway activation based on the evidence of target presentation, there seemed to emerge some interesting features. However, I think that the next step might be to suggest a mechanism by which this system is acting.
To address this particular question, Dayan suggested a top-down neural interrupt hypothesis presumably governed by the prefrontal cortex and the anterior cingulate cortex. Thus control over the NE neuromodulatory pathway via locus coeruleus may have the ability to globally set the system up for learning in the cases of unexpected uncertainty.
Clearly, many questions remain concerning this proposal, especially considering the lack of a biophysical mechanism, as well as the basic understanding of cortical specificity.
However the idea is certainly interesting nonetheless, in the sense that yet another neuromodulatory system may have a tangible role in a complex behavior such as uncertainty mediation.
I heard a pretty great talk today by Jennifer Mann from Baylor College of Medicine. Her talk was on a natural phenomenon of the long DNA strands called knotting. It is well known that DNA must be compacted in order to avoid taking up an unreasonable amount of space within the nucleus of a cell. However one can envision that this compaction process is not without its intricacies.
Imagine taking a long string of some sort, not unlike your headphone cord. When you hastily throw your headphones into your pocket, there’s a good chance that when you pull them out later on, they are in a headache of a knot. It’s somewhat of a mystery how the cord ended up in such a tangled configuration, and every once in awhile, this happens in the compacting process of DNA, which is a long strand of nucleic acids strung together that eventually give rise to the amino acids that make proteins, which are largely the stuff of life. (Forgive my really rudimentary biological descriptions.)
Several interesting properties arise, and one can describe the topology of the resultant knots, if they do form, within the DNA, as pictured here. Though there are an infinite number of minor changes within a given conformational structure, the topology can still be quite similar and thus create tractable problems. This was termed “knot invariant.”
Well, without getting into what amounted to some heavy an fanciful biology, one can imagine a multitude of ways to untangle the knot of DNA, which is quite a problem for a cell to undertake naturally. Essentially, Mann was providing evidence that the cell’s mechanism to untangle the DNA knot was via the easiest (read: lowest energy) change that would then allow the entire molecule to relax.
The computational approaches that such a system presents are vast. For one, there is undoubtedly a conformational strain on the natural configuration of the twisted, knotted DNA strand, and those changes are likely to have some energy stored in them, not unlike a spring, that is eager to be released into a lower energy “rest” configuration.
From a dynamical systems perspective there might be many such stable states of relaxation, and small perturbations from the knotted conformation (i.e. untangling) might allow one to predict – based on optimized bond length and energy parameters – to what stable fixed point the molecule might end up.
Among the several other interesting questions that exist, minimization of energy is among my favorite. But from a biological point of view, it would be nice to be able to selectively target a site at which just one small change (a cut in the loop, twist, and rejoin) can greatly reduce or eliminate the knotting.
For more information, see Z. Liu, J. Mann, E. Zechiedrich, and H. Chan. Topological Information Embodied in Local Juxtaposition Geometry Provides a Statistical Mechanical Basis for Unknotting by Type-2 DNA Topoisomerases. J Mol Biol, 2006.
