Citations to this work should read:

Park, B. (1998, unpublished) A connectionist account of antidepressant action. Available from the Connectionist models of cognitive, affective, brain, and behavioral disorders website at www.sci.sdsu.edu/CAL/connectionist-models/

A connectionist account of antidepressant action

Brett Park

Summary

A neural net model of the action of antidepressants through the serotonergic system is presented. This model suggests that 5HT does not need to be immediately related to the coding of mood, and provides an explanation for antidepressant action that is related to learning rather than mood per se.

Key words

Depression, serotonin, antidepressants, neural net

Introduction

Despite considerable research effort, there is as yet only limited understanding of the role of the monoamine systems in their involvement in antidepressant action. A characteristic of the explanations suggested to date is that they make an immediate link between the monoamine and the clinical state eg the noradrenaline hypothesis of depression, the indolamine hypothesis etc. More complex formulations still directly link the action of a transmitter system to a given behavioural complex eg the serotonergic projection from the dorsal raphe nucleus restrains the response to unconditioned aversive stimulation (Deakin and Graeff 1991). It has become inceasingly clear that not only are transmitter systems involved in diverse disorders (eg abnormalities in serotonergic function occur in both anxiety disorders and depression ) but diverse transmitter systems are involved in the responses to relatively simple stimuli [eg noradrenaline, dopamine and serotonin are all involved in the response to stress (Anisman and Zacharko 1991) ] so that hypotheses making such direct links will at best only provide partial simplifications for the role of the transmitters in a given response. Anatomically monoamine systems are highly distributed and such hypotheses also provide little in the way of a role for the rest of the projection pathways. DRN output may well constrain the output of the periaqueductal grey (Deakin and Graeff 1991) but limiting its role to these effects says little about the fact that the same serotonergic projections are likely to be desynchronising the cortical mantle at the same time (Vanderwolf et al 1989).

Connectionist models (neural nets) may provide a useful tool in approaching these difficulties as they provide a means for examining the effects of distributed changes in highly interactive systems (Farah and McClelland 1993, Mesulam 1990). They have successfully been used at differing levels of sophistication to model the effects of transmission through different transmitter systems (Hasselmo and Bower 1993), receptor types (Traven et al 1993), and also to examine the effects of pharmacological manipulation (Servan-Schreiber et al 1990, Callaway et al 1994). They are increasingly being used within psychiatry (Park and Young 1994) with the aim of providing a bridge between neurophysiological and psychological phenomena and are showing a degree of predictive validity in addition to the understanding they can bring.

Neurophysiological basis for the model

The model is based on a single layer of idealised cortex (Fig 1). In light of the fact that in terms of monoaminergic innervation (vide infra) there may well be important distinctions between cell types, the model makes a clear separation between those units (cells) with excitatory outputs and those with inhibitory outputs. In addition, in keeping with what is known of cortical architecture, the inhibitory units only function within the local circuitry, whilst the units representing longer distance transmission are purely stimulatory (Douglas and Martin 1990).

In the light of the histological evidence, two types of inhibitory interneurone are modelled. There is increasing evidence that the median raphe projection may be to a subset of inhibitory interneurones (Hornung and Celio 1992, Miettinen and Freund 1992) which subserve feedforward inhibition in predominantly dendritic areas. The main receptor associated with this pathway is thought to be the 5HT1A receptor (Deakin 1990) and increasing transmission through this pathway will be modelled as a general decrease in the activity in these feed-forward inhibitory interneurones. This is compatible with neurophysiological evidence that stimulating the Median Raphe nucleus can increase the overall level of transmission through areas rich in 5HT1A receptors (Assaf and Miller 1978, Winson 1980), through a mechanism of disinhibition. There is similarly indirect evidence that dorsal raphe pathways have a stimulatory effect through 5HT2 receptors on a different set of inhibitory interneurones, located at a deeper cytoarchitectonic layer which may modulate feedback inhibition through GABA-A receptors (Blue et al 1988, Sheldon and Aghajanian 1990, Gellman and Aghajanian 1993, Morilak et al 1993). Increasing the output of the dorsal raphe projection was therefore modelled by increasing the level of the activity of the feedback inhibitory units within the model (Fig 2).

Rather than having separate parts of the model for neocortex, hippocampus and striatum and thalamus only a single layer of cortex was modelled with the aim of looking for an explanation, that can link the role of 5HT in these diverse areas. A separate justification for doing this arises from the parallels that exist in the serotonergic innervation of these diverse areas. Median raphe projections appear to be dominant in the earlier areas of information processing in a given site [ hilar region of the dentate in the hippocampus (Lidov et al 1980) and shell region of the ventral striatum (Van Bockstaele and Pickel 1993) ] versus the dorsal raphe predominance in dorsal hippocampus and core region of the ventral striatum. In this model this is taken to be equivalent to the denser thick varicose innervation and 5HT1A receptor binding in layers I - III of neocortex (Kosofsky and Molliver 1987) versus the denser innervation from the dorsal raphe (Blue et al 1988) and binding of 5-HT2 recognition sites (Hoyer et al 1986) in deeper (III - V) cortical layers.

The action of antidepressants was modelled as an increase in transmission through the 5HT1A receptor (Blier et al 1987) coupled with a decrease in the transmission through the 5HT2 receptor, analagous to a combination of 5HT2 blockade and 5HT2 receptor down regulation (Stahl and Palazidou 1986) for classical tricyclics (TCA's) and an increase in transmission throught his pathway to model the effects of selective serotonin reuptake inhibitors (SSRI's) (Cadogan et al 1993, Fuller 1993) . Dietary tryptophan depletion which has been shown to acutely reverse the effects of antidepressant treatment with SSRI's (Delgado et al 1989) and which is though to lead to an acute reduction in central serotonergic function (Biggio et al 1974, Gartside et al 1992) was modelled by decreasing transmission through both receptor types.

Neural net architecture and implementation.

The model is implemented using a seven layer fan-in, fan-out feedforward backpropagation network (Figs 2). The pyramidal cells are each modelled using three units. The first represents the dendritic region to which the inputs and feedforward inhibitory unit connects. The second and third units represent the cell soma at two time points with the feedback excitation and inhibition occurring between these two units. Each dendritic unit only connects to its own pair of soma units, with the weight of the connection linked. The biases are linked in all three units.

The initial training set includes three "reward" inputs which map to two "reward" outputs, and an equivalent set of "punishment" responsive units (Fig 3). The initial training set is dichotomous in that there are no inputs presented which combine both stimulus types. After 100 epochs of training, the network is then presented with compound stimuli, which have a mixture of reward and punishment type inputs but evaluate in a dichotomous fashion. Patterns with rewarding output are presented at twice the rate of punishing ones, though with compound stimuli the bias is towards perception of punishing stimuli as an approximation to the Schwartz and Garamoni States of Mind model (Schwartz and Garamoni 1986,1989). (Though the network can be trained on the latter set alone, this method reduces problems with local minima.) In this manner affective responses are coded into the glutamatergic/ GABAergic type units with no involvement of serotonergic function at this stage.

To achieve a "depressed" state within the network, once it had been trained to criterion on the original set of patterns, the patterns are changed to a set where there are no longer any inputs presented that cause rewarding outputs and one of the previously rewarding inputs is changed to signal punishment. This is intended to reproduce the effect of combined loss and a stressful environment.This network is then tested on its responses to the remaining rewarding stimuli as a way of examining the extent to which it has become anhedonic. To simulate the recovery process, the network is trained on the full set of reward and punishment stimuli. The total error is used as a measure of the extent to which the network is "depressed", as it is predominantly made up of the errors occurring on the reward patterns which have been returned to the training set. The effects of altered serotonergic transmission are modelled by manually changing the biases of the inhibitory units.

Results

As there are more units than are required for "solving" the coding of the input/output relations, there were a number of different networks created. Significant activation of both inhibitory units by patterns of both valences occurred in different solutions. After training on the "depressogenic" set, the network became "anhedonic", with a reduction in reward responses to previously rewarding inputs and indeed single reward inputs were responded to as punishing (Fig 6). Simulating a decreased MRN input during training with the depressogenic set, increased the length of time for spontaneous recovery from the simulated depressed state.

Antidepressant medication had both a prophylactic effect and increased the rate of recovery from the depression (Fig 4). The most important part of the simulated antidepressant action was the increased transmission through the simulated postsynaptic 5HT1A receptor as evidenced by antidepressant efficacy of both antidepressant types. The simulated antidepressant treatments typically showed an initial "worsening" and a dose response effect was also apparent in that small doses generally led to recovery little different to the untreated rate.The model supports Deakin's prediction (Deakin et al 1993) that a postsynaptic 5HT1A antagonist would delay the recovery compared with no intervention (Fig 4).

The SSRI treated networks showed a differential vulnerability to the tryptophan depletion compared to TCA treatment though the effect was not marked and a return to the original depressed state of the network did not occur. The effect was more apparent if the tryptophan depletion was limited to a decrease in the bias of the feedback inhibitory unit. In figure 5 scores are calculated for each output by subtracting punishment unit activations from reward unit activations and then multiplying by 100. The valence change caused by trptophan depletion is then calculated by subtracting equivalent pattern scores between the two conditions.

Discussion

Mechanism of simulated antidepressant action

The mechanism whereby, antidepressant medication exerts its effects within this simulation can be understood in terms of the solution the neural network comes to in coding the input/output pattern relations within the weights between units. The identification of which stimulus type an input unit belongs to occurs in the connections between the first three layers of the network, whilst the higher level task of balancing the response types occurs to a greater extent in the later layers of the network. When the network is trained on the depressogenic patterns, in the absence of rewarding patterns the network as a whole adapts to the change in valence of part of the input. The input unit which has changed valence shows some change in its connections towards those of the original punishment type inputs, however changes also occur in the later stages of the network to further bias its responding with punishment type outputs. A selective decrease in the likelihood of the inhibitory units being on during recovery, from the perspective of the backpropagation algorithm increases the extent to which the excitatory units contribute to the error and hence increases the rate at which they relearn to respond to the appropriate stimuli. In the case of the simulated antidepressant action, this leads to a selective increase in learning in that part of the network which is responsible for identifying the input type. The value of this increased learning can also be seen in the simulation of the prophylactic effect of antidepressants in that the increased rate at which the network learns to change the identification of the input unit, leads to a decreased requirement to adjust the overall bias of responding in the later stages of the network and reward responses are preserved. The architectural features of the network necessary for this antidepressant effect is the fan in at the input stage and the smaller number of inhibitory units compared with excitatory, both of which are recognised features of cortical organisation.

The initial worsening that is seen with the simulation of antidepressant medication again points to an antidepressant effect that is not dependant on an immediate link between 5-HT and mood. From a clinical perspective this is a recognised side effect of treatment but is a potential difficulty for the simulation in that the initial worsening was marked and occurred on all runs of the simulation whereas it is certainly not an invariable accompaniment of antidepressant treatment. This however may be where receptor changes are of importance. The simulation is of an immediate increase in serotonergic transmission, whereas it is likely that with antidepressant treatment the increase is more gradual, initially being significantly attenuated by the effect of the stimulation of the 5HT cell body and terminal autoreceptors. The down regulation of the cell body and terminal autoreceptors, then leads to a gradual increase in serotonergic transmission.

Experimental support for mechanism

The mechanism of antidepressant action is compatible with the effects of Median raphe stimulation on the induction of long-term potentiation in the dentate gyrus (Klancnik and Phillips 1991). Further neurophysiological evidence compatible with this model, is the enhancement in LTP induction that is seen with general reductions in inhibitory neurotransmission (Davies et al 1991,Tomasulo et al 1993, Kaibara and Leung 1993, Yasui et al 1993). There are some studies however which suggest that serotonergic transmission has an inhibitory effect on LTP formation (Villani and Johnston 1993) and that at least part of this inhibitory effect is mediated by 5-HT1A receptors located on pyramidal cells (Corradetti et al 1992). Possible explanations for these discrepancies are that these studies used bath application of 5-HT to hippocampal slices which may well have led to a different pattern of delivery of 5-HT to the receptors and target neurons than either MRN stimulation or antidepressant treatment in an intact brain. In addition the study by Coradetti and colleagues demonstrated that 5HT in the presence of a 5-HT3 receptor blocker [postulated to block a stimulatory serotonergic input to inhibitory interneurones (Ropert and Guy 1991)] led to an enhanced and persisting response to stimulation, which remains compatible with the simulation.

There is some evidence that antidepressant treatment with SSRI's is associated with enhanced learning (Weingartner et al 1983, Flood and Cherkin 1987). The model is also consistent with work carried out in our unit which suggests that decreasing serotonergic function by tryptophan depletion can selectively impair aspects of learning in normal volunteers (Park et al 1994). In addition it is also compatible with the effects of 5HT3 antagonists on learning (Crook 1991) though the mechanism suggested by the model is not necessarily through an interaction with cholinergic systems as is typically suggested (Barnes et al 1989, 1990), rather that it arises through the blockade of 5HT3 receptors stimulating GABAergic interneurones (Ropert and Guy 1991).

Implications of the model

From the perspective of this model it is the transmission through the 5-HT1A receptors of the median raphe projection that is crucial for the antidepressant effect which supports Deakin's (Deakin and Graeff 1991) and Blier's (Blier et al 1987) hypotheses. Simulations of isolated increases and decreases in transmission through the DRN projection/ 5HT2 receptor were not associated with the same antidepressant effect, and the simulations of TCA's and SSRI's were largely equipotent. The mechanism also provides a model for Deakin's disconnection hypothesis of Median Raphe action (Deakin 1983) although without the limit of any assumptions regarding the direction of change in valence of the stimulus.

The model suggests a way in which mood and affective responding does not need to be directly linked to monoaminergic transmission, yet altering the neuromodulatory input mimics observed clinical phenomena. Models which imply that affective responses are coded in terms of monoamines that predominantly depend on slower transmission of information through G protein linked receptors, are less able to give an account of the fact that threatening stimuli can be recognised as such immediately, and the appropriate responses can also be engaged within a rapid time course. Additionally, given that there are a relatively small number of monoaminergic neurons there may well be a limit to the amount of information that they could carry compared with what may be necessary for the appropriate identification and responding required for complex emotional processing. At a different level of observation, the relationship between mood and serotonergic function has also proved very difficult to pin down in empirical terms. In normal subjects serotonergic manipulations and antidepressant medication do not appear to have any consistent effect on mood. The relationship is also problematic for the explanations of the delay in antidepressant action. Certainly the side effects noted with the administration of SSRI's are typical of an immediate increase in serotonergic function, yet the effect to elevate mood is delayed. Furthermore the separation of the effects of the serotonergic innervation from the valence of the stimuli, can more easily allow for modulatory effects on both positively (Wogar et al 1991, 1993) and negatively reinforced operant behaviour.

Unlike receptor (Sulser 1978, Blier et al 1987) and dysregulation theories (Siever and Davis 1985) the explanation for the mechanism of action of antidepressants provided by this model, is not restricted to adaptive or functional changes at particular receptor types but puts them in the context of the effect of serotonergic transmission on its target neurones. Though these adaptive changes may clearly be of importance in determining the level of transmission through a particular part of the monoaminergic system, such explanations can readily be seen as simplifications compared with the level of explanation offered by this model which allows for a much more general spectrum of action for serotonergic transmission.

A further aspect of this simulation is that it suggests, that impaired serotonergic transmission would predispose to the development of a depressed mood state but is not necessary for its development. This would relate to the findings from neuroendocrine challenge studies, that show that the differences in serotonergic function between depressed patients and controls are statistical rather than absolute.

The findings from the tryptophan depletion simulation, though they do not fully replicate the findings of the Delgado study are nevertheless interesting. In the course of the treatment with SSRI's, the network uses the increased feedback inhibition to reduce overly punishing type responses. If this increased activity is then removed these responses return in strength; this is compatible with the specificity of symptom return reported in that study and the theories of serotonergic function related to restraint (eg Spoont 1993).

Limitations of the model

In discussing the ability of this simulation to replicate the effects of serotonergic manipulation and antidepressant action, it is necessary to look at the effects in the light of the numerous assumptions and simplifications made. The use of backpropagation of error as the training method is problematic in that it is typically seen as a curve fitting procedure with little neurobiological validity. However for this simulation, it is the component which links learning to the activation of the units which is crucial for the simulated antidepressant effect and this is an integral part of the mechanisms involved in LTP induction within the brain.

In addition there are multiple absences from this model which include other serotonin receptor types, noradrenergic and dopaminergic innervations and the interactions between different parts of the brain, which have been implicated in the changes in brain function related to depression and antidepressant action. Against this considerable problem, however, is the fact that simplifications are necessary in modelling a system as complex as the brain. Additionally, despite such simplifications, it is still possible to capture key computational aspects of complex systems (eg Lockery and Sejnowski 1993).

Limiting the simulation to a single, idealised layer of cortex, leads to problems in terms of relating it to the integrated action of different brain areas. However it has the advantage that not only can the model represent the evaluation of external stimuli, but it can also function with respect to internal stimuli. If the inputs to the network are conceived as positive and negative perceptions of the self, and the output representing an integrated evaluation then the network can also be taken as a model of the effect of antidepressants on lowered self esteem within depression.

Conclusion

Implementing a degree of neuroanatomical and neuropharmacological detail within this simulation in an attempt to bridge the considerable gap between neural networks and the real nervous system (as noted by several authors eg. Hasselmo and Bower 1993, Crick and Asanuma 1986, Shepherd 1990) has allowed a possible bridge to be modelled between the pharmacological and psychological effects of antidepressant drugs. The most significant differences between the explanation of antidepressant action provided by this model and others is that the action of 5HT occurs through a selective effect on learning rather than through a direct effect on mood and that the delay in appearance of treatment effect is related to the learning process rather than to adaptive receptor changes as is typically proposed (Artigas 1993, Briley and Moret 1993). The changes in 5HT transmission brought about by antidepressants allow a network to more quickly learn and adapt to a change in contingencies.

Refs

Anisman H and Zacharko R M (1991) Multiple Neurochemical and behavioural consequences of stressors: Implications for Depression. In Psychopharmacology of Anxiolytics and Antidepressants. Ed S E File. Pergamon Press, New York, 57-82.

Artigas F (1993) 5-HT1A receptor antagonists in combination therapy. 5-HT and antidepressants: new views from microdialysis studies. Trends in Pharmacological Sciences , 14, 262.

Assaf S Y, and Miller J J (1978) Neuronal transmission in the dentate gyrus: role of inhibitory mechanisms. Brain Research, 151, 587-592.

Barnes J M, Costall B, Coughlan J, Domeney A M, Gerrard P A, Kelly M E, Naylor R J, Onaivi E S, Tomlins D M and Tyers M B (1990) The effects of ondansetron, a 5-HT3 receptor antagonist, on cognition in rodents and primates. Pharmacol Biochem Behav. 35, 955-962.

Barnes J M, Barnes N M, Costall B, Naylor R J and Tyers M B (1989) 5-HT3 receptors mediate inhibition of acetylcholine release in cortical tissue. Nature, 338, 762-763.

Biggio G, Fadda F, Fanni P, Tagliamonte A, Gessa GL (1974) Rapid depletion of serum tryptophan, brain tryptophan, serotonin and 5-hydroxyindoleacetic acid by a tryptophan free diet. Life Sciences, 14, 1321-1329.

Blier P, de Montigny C, and Chaput Y (1987) Modifications of the serotonin system by antidepressant treatments: implications for the therapeutic response in major depression. J Clin Psychopharmacol , 7, 245-255.

Blue M E, Yagaloff K A, Mamounas L A, Hartig P R, and Molliver M E (1988) Correspondence between 5-HT2 receptors and serotonergic axons in rat neocortex. Brain Research, 453, 315-328.

Briley M and Moret C (1993) 5-HT1D autoreceptor antagonists as potential antidepressants. 5-HT and antidepressants: in vitro and in vivo release studies. Trends in Pharmacological Sciences , 14, 396-397.

Cadogan A K, Marsden C A, Tulloch I and Kendall D A (1993) Evidence that chronic administration of paroxetine or fluoxetine enhances 5-HT2 receptor function in the brain of the guinea pig. Neuropharmacology, 32, 249-256.

Callaway E, Halliday R, Naylor H, Yano L, Herzig K (1994) Drugs and human information processing. Neuropsychopharmacology, 10, 9-19.

Corradetti R, Ballerini L, Pugliese A M, and Pepeu G (1992) Serotonin blocks the long-term potentiation induced by primed burst stimulation in the CA1 region of rat hippocampal slices. Neuroscience, 46, 511-518.

Crick F H C and Asanuma C (1986) Certain Aspects of he Anatomy and Physiology of the Cerebral Cortex. In McClelland J L, Rumelhart D E and the PDP Research Group, Parallel Distributed Processing: Explorations in the microstructure of cognition, Vol 2: Psychological and Biological Models, 333-371.

Crook T H (1991) Effects of ondansetron in age-associated meory impairment. 5th World Congress of Biological Psychiatry, Satellite Symposium, The role of ondansetron, a novel 5-HT3 antagonist, in the treatment of psychiatric disorders.

Davies C H, Starkey S J, Pozza M F and Collingridge G L (1991) GABA(B) autoreceptors regulate the induction of LTP. Nature, 349, 609-611.

Deakin J F W, Graeff F G, and Guimaraes F S (1993) Altered 5-HT1A receptor mediated neurotransmission. Clinical implication of microdialysis findings. Trends in Pharmacological Sciences, 14, 263.

Deakin J F W and Graeff F G (1991) 5-HT and mechanisms of defence. Journal of Psychopharmacology, 5, 305-316.

Deakin J F W (1983) Roles of serotonergic systems in escape, avoidance and other behaviours. In Cooper S J (Ed) Theory in Psychopharmacology. Vol 2. Academic Press, New York 179-204.

Farah M J and McClelland J L (1992) Neural Network Models and Cognitive Neuropsychology. Psychiatric Annals, 22, 148-153.

Flood J F and Cherkin A (1987) Fluoxetine enhances memory processing in mice. Psychopharmacology, 93, 36-43.

Gartside S E, Cowen P J and Sharp T (1992) Effect of amino-acid loads on hippocampal 5-HT release in vivo evoked by electrical stimulation of the dorsal raphe nucleus and d-fenfluramine administration. British Journal of Pharmacology, 107, 448P

Gellman R L and Aghajanian G K (1993) Pyramidal cells in piriform cortex receive a convergence of inputs from monoamine actviated GABAergic interneurones. Brain Research, 600, 63-73.

Hasselmo M E and Bower J M (1993) Acetyl choline and memory. Trends in Neurosciences, 16, 218-222.

Hornung J P and Celio (1992) The selective innervation by serotonergic axons of calbindin-containing interneurons in the neocortex and hippocampus of the marmoset. The Journal of Comparative Neurology, 320, 457-467.

Hoyer D, Pazos A, Probst A, Palacios J M (1986) Serotonin receptors in the human Brain. II. Characterization and Autoradiographic Localization of 5-HT1C and 5-HT2 Recognition Sites. Brain Research, 376, 97-107.

Kaibara T and Leung L S (1993) Basal versus apical dendritic long-term potentiation of commisural afferents to hippocampal CA1: A current-source desity study. Journal of Neuroscience, 13, 2391-2404.

Kosofsky B E and Molliver M E (1987) The serotoninergic innervation of cerebral cortex: different classes of axon terminals arise from dorsal and median raphe nucleus. Synapse ,1, 153-168.

Klancnik J M and Philips A G (1991) Modulation of synaptic plasticity in the dentate gyrus of the rat by electrical stimulation of the median raphe nucleus. Brain Research, 557, 236-240

Lidov H G W, Grzanna R, and Molliver M (1980) The serotonin innervation of the cerbral cortex in the rat - an immunohistochemical analysis. Neuroscience, 5, 207-227.

Lockery S R and Sejnowski T J (1993) The computational leech. Trends in Neurosciences, 16, 283-290.

Mesulam M-M (1990) Large-Scale Neurocognitive Networks and Distributed Processing for Attention, Language and Memory. Annals of Neurology, 28, 597-613.

Miettinen R, and Freund T F (1992) Convergence and segregation of septal and median raphe inputs onto different subsets of hippocampal inhibitory interneurons. Brain Research, 594, 263-272

Morilak D A, Garlow S J, and Ciaranello R D (1993) Immunocytochemical localization and description of neurons expressing serotonin2 receptors in rat brain. Neuroscience, 54, 701-717.

Park S B G and Young A H (1994) Connectionism and Psychiatry. Philosophy, Psychiatry and Psychology (in press)

Park S B, Coull J T, McShane R H, Young A H, Sahakian B J, Robbins T W and Cowen P J (1994) Tryptophan depletion in Normal Volunteers Produces Selective Impairments in Learning and Memory. Neuropharmacology (in press)

Ropert N and Guy N (1991) Serotonin facilitates serotonergic transmission in the CA1 region of rat hippocampus in vitro. J Physiology , 441, 121-136.

Schwartz R M and Garamoni G L (1986) A structural model of positive and negative states of mind: Asymmetry in the internal dialogue. In Kendall P C (Ed.) Advances in cognitive-behavioural resarch and therapy Vol 5, Academic Press, New York, 1-62.

Schwartz R M and Garamoni G L (1989) Cognitive balance and psychopathology: Evaluation of an information processing model of positive and negative states of mind. Clinical Psychology Review, 9, 271-294.

Servan-Schreiber D, Printz H W and Cohen J D (1990) A network model of catecholamine effects: Gain, signal to noise ratio, and behaviour. Science, 249, 892-895.

Sheldon P W and Aghajanian G K (1990) Serotonin (5HT) induces IPSP's in pyramidal layer cells of rat piriform cortex: evidence for the involvement of a 5-HT2- activated interneurone. Brain Research, 506, 62-69.

Sheldon P W and Aghajanian G K (1991) Excitatory responses to serotonin (5-HT) in neurons of the rat piriform cortex: Evidence for mediation by 5-HT1C receptors in pyramidal cells and 5-HT2 receptors in interneurons. Synapse , 9, 208-218.

Shepherd G M (1990) The significance of Real Neuron Architectures for Neural Network Simulations. In Schwartz E L (Ed) Computational Neuroscience, MIT Press, Cambridge Massachusetts, 82-96.

Siever L J, and Davis K L (1985) Overview: Toward a dysregulation hypothesis of depression. American Journal of Psychiatry, 142, 1017-1031.

Stahl S M and Palazidou L (1986) The pharmacology of depression, Trends in Pharmacological Sciences, 7, 349-354.

Sulser F (1978) New perspectives on the mode of action of antidepressant drugs. Trends in Pharmacological Sciences, 1, 92-94.

Traven H G C, Brodin L, Lansner A, Ekeberg O, Wallen P and Grillner S (1993) Computer simulations of NMDA and non-NMDA Receptor-mediated Synaptic Drive: Sensory and Supraspinal modulation of Neurons and Small Networks. Journal of Neurophysiology, 70, 695-709.

Tomasulo R A, Ramirez J J and Steward O (1993) Synaptic inhibition regulates associative interactions between afferents during induction of long-term potentiation and depression. Proceedings of the National Academy of Sciences of the United States of America, 90, 11578-11582.

Van Bockstaele E J, and Pickel V M (1993) Ultrastructure of Serotonin-Immunoreactive Terminals in the Core and Shell of the Rat Nucleus Accumbens: Cellular Substrates for Interactions with Catecholamine Afferents. Journal of Comparative Neurology, 334, 603-617.

Vanderwolf C H, Leung L -W S, Baker G B (1989) The role of serotonin in the control of cerebral activity: studies with intracerebral 5,7-dihydroxytryptamine. Brain Research, 504, 181-191.

Villani F and Johnston D (1993) Serotonin inhibits induction of long-term potentiation at commissural synapses in hippocampus. Brain Research, 606, 304-308.

Weingartner H, Buchsbaum M S and Linnoila M (1983) Zimelidine effects on memory impairment produced by ethanol. Life Sciences, 22, 2159-2163.

Winson J (1980) Influence of raphe nuclei on neuronal transmission from perforant pathway through dentate gyrus. J Neurophysiology, 44, 937-950.

Wogar M A, Bradshaw C M, and Szabadi E (1993) Effect of lesions of the ascending 5-hydroxytryptaminergic pathways on choice between delayed reinforcers. Psychopharmacology, 111, 239-243.

Wogar M A, Bradshaw C M, and Szabadi E (1991) Evidence for an involvement of 5-hydroxytryptaminergic neurones in the maintenance of operant behaviour by positive reiforcement. Psychopharmacology, 105, 119-124.

Yasui M, Kawasaki K, Matsushita A and Satoh M (1993) Benzodiazepine inverse agonists augment long-term potentiation in CA1 and CA3 of guinea pig hippocampal slices. Neuropharmacology, 32, 127-131.