An exciting aspect of neuroscience is the ability to peek into the brain, measuring the activity of neurons in behaving animals. Electrophysiology allows us to listen to the signals sent by individual neurons, as if intercepting a message in Morse Code over a radio. So far, neuroscientists have been able to discern the function of many types of neurons by presenting animals with stimuli, and observing which stimuli cause activation or inactivation of the neuron of interest. These types of measurements tell us the result of the computation performed by that neuron. A new and challenging puzzle will be to determine what kinds of signals are used in each calculation, and how they are combined.
Electrophysiological recordings in single brain regions typically result in a wide variety of activity recorded from individual neurons. This occurs because different types of neurons are intermingled in the brain. Therefore, to interpret the data, we have to classify neurons based on their molecular profiles and based on the connections that they make with other neurons. Dopamine neurons are a convenient model system for studying computation because they are found in only a few regions of the brain, and have a clear, targetable molecular profile.
In the ventral tegmental area (VTA) of the brain, dopamine neurons seem to have the uniform function of signaling when errors occur in the prediction of a reward (reward prediction error, RPE). We know that reward prediction error can theoretically be calculated with a simple subtraction: actual reward minus expected reward. This signal is important for guiding our future behaviors to maximize reward. If the actual reward is higher than the expectation (positive RPE), we may favor these actions, whereas if the actual reward is lower than our expectation (negative RPE), we may refrain from those actions in the future.
Although dopamine is important for our behaviors and dopamine RPE signals have been observed for 20 years, we still do not understand how dopamine neurons compute RPE. More precisely, although there have been many models which predict the mechanism of RPE computation in dopamine neurons, there was no direct method to test these ideas experimentally. One of main focuses of these models was, naturally, deciding which brain areas provide information about “actual reward” or “expectation” for dopamine neurons to calculate “actual reward minus expectation”.
We decided to tackle this question. Our first goal was to specifically label neurons that directly project onto dopamine neurons. To do this, we established a modified rabies-GFP virus system with mouse genetics to infect dopamine neurons and hop trans-synaptically exactly once, into presynaptic neurons. Using this technique, we mapped the monosynaptic inputs to dopamine neurons across the entire brain, 4 years ago (Watabe-Uchida et al., 2012). The next question was: what information do these inputs send to dopamine neurons? Which brain areas provide actual reward or expected reward information?
For this experiment, we used modified rabies virus carrying channelrhodopsin 2 (ChR2), a light-gated ion channel, so that only dopamine neurons and neurons that project to dopamine neurons express ChR2. We used this to identify neurons providing input to dopamine neurons across the brain, by seeing whether each neuron we recorded from was activated by blue light (with very low latency). We recorded the activity of neurons in input areas while mice behaved, and then shined blue light to determine which neurons were direct inputs to dopamine neurons. We recorded from 7 input-dense areas which have most often been suggested as important sources of signals in RPE models. With this data in hand, we were prepared to answer our initial question: which brain areas provide actual reward or expectation information for the computation? Which model is true?
The results turned out to be that each variable (actual reward and expectation) was distributed among inputs in all of the brain regions we recorded from. Furthermore, these variables were already mixed in many input neurons, such that they themselves could signal at least partial RPE information. Thus, it seems that our brain computes partial RPE at multiple nodes in the neuronal network and dopamine neurons gather this information together to compute a very precise RPE. In other words, computation in our brain can be distributed and redundant even for a simple arithmetic like the subtraction required for RPE. This type of redundancy likely contributes to the robustness of computations in our brain.
Overall, our data suggests that computations in the brain are different (and more complicated) than many proposed models and simple arithmetic is embedded in multi-layered neural circuits. When we study the brain, we often place too much focus on identifying cascades of brain areas and forget about one of the most exciting, though mysterious, aspects of our brain: integration of information. This study took a lot of work, more than 8 years, to prepare the systematic method to examine both connectivity and activity (Tian et al., 2016). We hope that we could contribute to the sense of wonder people feel when they think about computations in the brain, and demonstrate how unique and mysterious these neural computations can be compared to simple arithmetic. Finally, we hope that this study can help guide explorations of other computations, as the field slowly gathers a repertoire of model computations to look for themes across the brain.
Author: Mitsuko Watabe-Uchida