Department Of Computer Science Western Michigan University, Kalamazoo, Michigan, Usa

Toward a Prefrontal Model of Short-term Synaptic Plasticity: Network Architecture and Working Memory Deficits Chris D.

1,3 Gore ,

Phillip J.

1 Gray ,

Julia

1 Kane ,

Mihály

1,2 Bányai

1Center

for Complex Systems Studies, Kalamazoo College, Kalamazoo, Michigan, USA, 2Department of Biophysics, KFKI Research Institute of Particle and Nuclear Physics of the Hungarian Academy of Sciences, Budapest, Hungary 3Department of Computer Science Western Michigan University, Kalamazoo, Michigan, USA Abstract

Methods

Working memory (WM) is an important cognitive function that refers to the maintenance and online manipulation of information for a short period of time. It has been suggested that short-term synaptic facilitation in recurrent connections of neocortical networks can sustain WM (Mongillo, Barak, &Tsodyks, 2008). We propose a prefrontal network model for relating short-term synaptic plasticity and network architecture to WM tasks such as the n-back task and a version of the Sternberg task that manipulates WM load. Our overall purpose is to model WM as realistically as possible in order to measure the effects of synaptic facilitation/depression and network architecture on WM performance.

The complexity of the brain, as a system of feedback loops and recurrent connections, and the number of neurons comprising it make it difficult to understand the totality of the processes relevant for schizophrenia and working memory solely by imagination. The methods of system science may be useful tools in understanding the physiology and dynamics of relevant neural systems. A computational approach to schizophrenia and short-term plasticity fosters an understanding of the nature and limits of the strategies neural systems employ in processing and transmitting information. One of the major problems in biological memory is how it is maintained and changed through multiple time scales. Computational tools can contribute greatly to our understanding of how synapses function and change in relation to external stimulus. By emphasizing descriptions of functionally and biologically realistic neural systems, a formal computational model captures the essential features of this prefrontal system at multiple spatial-temporal scales. Such a model can be used to test hypotheses—like the synaptic theory of working memory —that can be directly verified by current or future biological experiments. Models can help in understanding the loci of abnormal plasticity processes, and the consequences involved. It is generally accepted that schizophrenia is related to excessive pruning of cortical connections, and simple network studies have shown that this cortical pruning may lead to the formation of “pathological attractors". In our study we are developing a prefrontal network model of short-term plasticity giving consideration to differences in network architecture observed in healthy and schizophrenic brains.

Introduction Working memory (WM) is an important cognitive function that refers to the maintenance and online manipulation of information for a short period of time. WM is mediated by activity in recurrent networks in the prefrontal cortex (PFC). During WM tasks, a stimulus that is briefly presented to an animal often has to be remembered during a delay period that takes place between the presentation of a stimulus and the execution of a task. For instance, in computerized versions of Sternberg WM tasks, which are commonly used in research, a sequence of capitalized consonant letters (e.g., BGZXF) is presented on the screen and is followed by a delay period lasting 1-3 s during which the screen is blank. After the delay period, a letter-position probe (e.g., z=3) is presented and the subject is required to specify whether the probe is true or false. In versions of another commonly used WM task, the n-back task, a sequence of letters is presented one letter at a time and the subject is asked to hit a target key each time the letter on the screen matches the letter presented 1-, 2-, or 3-back (i.e., 1-, 2-, or 3-letters earlier) in the sequence or matches a predetermined target letter (0-back). Thus, every letter serves as a probe and the delay depends on the condition (0-, 1-, 2-, or 3-back). Early theories suggested that enhanced spiking activity in the form of persistent reverberations among neurons that code for a stimulus allow for the stimulus to be remembered during a delay period (Hempel et al., 2000). However, electrophysiological studies have shown that spiking activity may disappear during a delay period (Mongillo, Barak, &Tsodyks, 2008). ST facilitation reflects an increase in the probability of neurotransmitter release that can last up to 1 s (Abbot &Regehr, 2004). ST depression reflects a decrease in the probability of neurotransmitter release that last up to several seconds. A more recent synaptic theory of WM (Mongillo et al.) suggests that ST synaptic facilitation mediated by residual calcium levels at presynaptic terminals of pyramidal glutamatergic neurons in the PFC that code for a stimulus, allows a stimulus to be remembered for about 1 s without enhanced spiking activity (Altamura et al., 2007). In this case, a probe can act as a weak non-specific excitatory input that can reactivate the network and thus, the memory of the stimulus. ST depression is also taking place and plays a role. It should be noted that PFC activation has been shown to increase directly with WM load (e.g., the number of letters in a stimulus sequence) but not length of delay (Altamura et al., 2004). A possible explanation is that a larger network of neurons in the PFC is involved with a larger WM load. WM deficits are a core cognitive deficit in schizophrenic patients. Altered connectivity is the proposed mechanism. In the present research, we propose a prefrontal network model for relating ST synaptic plasticity and network architecture to WM tasks such as the n-back task and a version of the Sternberg task that manipulate WM load. Our overall purpose is to model WM as realistically as possible in order to measure the effects of synaptic facilitation/depression and network architecture on WM performance and in the process, to elucidate possible explanations for WM deficits in schizophrenic patients.

Network Architecture Experiments

Artificial Neural Networks A biological neuron, like those found in the brain.

We simulated the Sternberg delay by storing sets of 3, 5, and 8 consonants in networks of 15, 25, and 40 neurons respectively. Each letter is represented in binary; for example:

CMPLX

3 13 16 12 24

00011 01101 10000 01100 11000

In a recurrent network, all the synapses fire simultaneously, each time bringing the values closer to a known pattern. To queue the network to remember the pattern, we give it one of the letters and positions. For the above example, we could ask it if it recognizes the M with the queue 00000 01101 00000 00000 00000. We ran 100 simulations for each set size with all patterns chosen at random. It was able to closely replicate actual results from human subjects. Results from the traditional Sternberg delay are on the left, while our simulation results are on the right.

A simple artificial neuron that fires a synapse if the sum of its incoming synapses is above a threshold.

4

Initial Analysis of Synaptic Theory of Working Memory For our initial study in working memory, we took some steps with Synaptic Theory of Working Memory (Mongillo et al. Science 2008): §Exploration of model’s parameter space for facilitation and depression §Test simulations with/without network connections §Test simulations with/without read out signal

After exploring this model for working memory, we found that we were unable to evaluate its effectiveness due to its focus on realistic biology rather than neural computation. Thus, we decided to temporarily leave this model to build our own with a real-world task. The task we chose is a well-known psychological test known as the Sternberg delay.

Here we can see a scale-free Barabasi-Albert (BA) network along side a Highly Diluted Hopfield Network (HDHN). These were two of the networks used to store 8 letters (40 neurons because each letter requires 5 neurons). Although they have the same number of neurons and connections, the scale-free network is much more effective and much closer to a biological neural network.

Discussion

Selected References

From the simulation data it is evident that as network connectivity decreases, network performance also decreases. The random graph network was slower and much less accurate than the scale free graph network in task performance. This is consistent with previous findings and highlights the importance of network architecture in working memory performance. Schizophrenia has been characterized by excessive pruning of cortical connections and the loss of frontal network hubs. These abnormalities in network connectivity may indeed be what leads to such prevalent working memory deficits observed in schizophrenia.

Abbot, L.F., &Regehr, W.G. (2004). Synaptic computation. Nature, 431, 796-803.

Future Directions We would like to convert our model to a more biophysically accurate network model of short-term plasticity in the PFC by considering the parameters present in biophysically realistic single-synapse models such as the model proposed by Hennig, Postlethwaite, Forsythe, and Graham (2006).

Altamura, M., Elvevag, B., Blasi, G., Bertolino, A., Callicot, J.H., Wienberger, D.R., Mattay, V.S., & Goldberg, T.E. (2007). Dissociating the effects of Sternberg working memory demands in prefrontal cortex. Psychiatry Research, 154, 103-114. Basset, D.S., Bullmore, E., Verchinski, B.A., Mattay, V.S., Weinberger, D.R., & MeyerLindenberg, A. (2008). Hierarchical organization of human cortical networks in health and schizophrenia. Journal of Neuroscience, 28, 9239-9248. Hempel, C.M., Hartman, K.H., Wang, X.-J., Turrigiano, G.G., & Nelson, S.B. (2000). Multiple forms of short-term plasticity at excitatory synapses in rat medial prefrontal cortex. Neurophysiology, 83, 3031-3041. Henneg, M.H., Postlethwaite, M., Forsythe, I.D., & Graham, B.P. (2007). A biophysical model of short-term plasticity at the calyx of Held. Neurocomputing, 70, 1626-1629. Mongillo, G., Barak, O., &Tsodyks, M. (2008). Synaptic theory of working memory. Science, 319, 1543-1546.

Project Supervisor: Dr. Péter Érdi1,2