We know that the passage of time at multiple time scales can influence neural activity. But what brain mechanisms are responsible for the encoding of time itself? Is there a specialised mechanism or module that measures the elapsed time and informs other brain areas or neuronal populations about the current position on the time axis? Or is time encoded in a distributed way, with – in principle – any part of the cortex involved in processing some information also intrinsically containing information about time?
In a recent article in Neuron, Goel & Buonomano argue for the latter, showing that neurons in a cortical slice in vitro – by definition isolated from any external circuit dedicated to measuring time – can be trained to represent time elapsed (at a subsecond scale) between an electric stimulus and optogenetic activation. The experiments were conducted in slices of cortical tissue implanted with electrodes used to deliver electric shocks which evoked neural activity. The slices were also transfected with a virus expressing channelrhodopsin, thanks to which cells were activated by optical stimulation (blue light shone above the tissue). First, the slices were subject to prolonged “training”, in which optical stimulation was delivered after the electric shock and a brief, constant time interval (100, 250, or 500 ms). After training, slices were “tested” – the electric shocks were delivered without being followed by the light. However, most superficial pyramidal cells in the slices spiked not only in response to shock, but also later, several hundreds of milliseconds after the initial peak. Crucially, the interval between the two waves of activity depended on previous training: the second wave of activity appeared around 100 ms if the slices had been trained to receive light 100 ms after the electric shock, and progressively later if the training had involved longer intervals.
While this finding shows that a generic part of the cortex can be trained to represent temporal regularities in its activity – without receiving inputs from another area serving as an external “clock” – these results might be explained either by single cells adapting their time constants in an experience-dependent way, or by the whole network of neurons encoding time in a dynamically evolving pattern of its activity. To disambiguate between these two possibilities, another experiment was conducted. This time the slices were implanted with two electrodes. In the training phase, one electrode was paired with an optical stimulus as before, delivered 100 ms after the shock. The other electrode was used to stimulate the cortical slice without an associated light stimulus (“unpaired” pathway). In the test phase, after a spike evoked by the electric shock, many of the neurons that had been classified as lying on either the paired or the unpaired pathway spiked again later. Crucially, this second wave of activity in neurons on the paired pathway was more pronounced and occurred markedly closer to the 100 ms mark than the secondary activation of neurons on the unpaired pathway. These results can be interpreted as pathway-sensitive learning, where the paired pathway is associated with more polysynaptic activity, and the timing of this activity – while not always coinciding with the expected interval (see below) – is closer to the expected interval.
To better understand the time-specific secondary response of the network in a more dynamic and mechanistic way, the authors performed two further experiments. First, they interleaved the training phase with test phases after 1 and 2 hours of pairing electric shocks with light. The results of this experiment showed that after 1 hour, the activity along the paired pathway increased (compared to the unpaired pathway) but not in a time-specific way. However, after 2 hours of training, the secondary peak of activity also depended on the trained interval, with the neurons trained using the 100 ms interval firing earlier than those trained at 500 ms.
Interestingly, the authors also observed that the probability of a light-evoked spike changed over the course of training. In cells which typically responded to light alone, just after pairing the light stimulus with an electric shock, the probability of a spike evoked by the light decreased dramatically to only 25%, and recovered during training to around 70%. This suggests that, during the initial phase of the training, the electric shock induces strong inhibition in the network which prevents the optical stimulus from activating the light-sensitive neurons. Later, this inhibition might decrease, leading to more neurons firing in response to light. The authors tested this hypothesis by separately measuring the excitatory and inhibitory currents evoked by the electric shock. It turned out that the ratio of excitation and inhibition differed between the paired and unpaired pathways only around the trained time interval, but not earlier or later. This result demonstrates that the network might have learned to dynamically shift the balance of excitation and inhibition in a training-dependent way.
Across the experiments, the exact timing of the secondary activity did not always coincide with the trained interval. For example, looking at slices trained at the 100 ms interval, experiment 1 – with only the paired pathway – resulted in secondary activity after approximately 100 ms, but further experiments – with both the paired and unpaired pathways – showed secondary activity peaking closer to 200 ms or even 250 ms. While the authors do not address these discrepancies, one could argue that the paired and unpaired pathway will strongly overlap in a densely interconnected slice, with neurons being indirectly depolarised by inputs from both pathways.
The time-specific balancing of excitation and inhibition suggests that, when looking at dynamic modulations of neural excitability, subtle measures of network activity might be a better choice than overall firing rates. Similar arguments have recently been raised in cognitive neuroscience, where non-invasive measurements often show dynamic changes of neural states without accompanying persistent activity.
It is important to note that the experiments were confined to a subsecond time scale. Different mechanisms might be at play at longer time scales. Crucially, temporal encoding in the order of seconds might more likely rely on cortico-hippocampal loops rather than local cortical networks.
Source article: Goel A, Buonomano D (2016) Temporal interval learning in cortical cultures is encoded in intrinsic network dynamics. Neuron 91(2):320-327. doi: http://dx.doi.org/10.1016/j.neuron.2016.05.042