Modelling efficiency in insect olfactory information processing
BioSystems 89 (2007) 236–243
Modelling efficiency in insect olfactory information processing Yuqiao Gu a,b,c , Hans Liljenstr¨om a,b,∗ a c
Department of Biometry and Engineering, P.O. Box 7032, SLU, S-75007 Uppsala, Sweden b Agora for Biosystems, P.O. Box 57, S-19322 Sigtuna, Sweden School of Automation and Energy Engineering, Tianjin University of Technology, PR China Received 1 December 2005; accepted 6 April 2006
Abstract The olfactory system of insects is essential for the search of food and mates, and weak signals can be detected, amplified and discriminated in a fluctuating environment. The olfactory system also allows for learning and recall of odour memories. Based on anatomical, physiological, and behavioural data from the olfactory system of insects, we have developed a cross-scale dynamical neural network model to simulate the presentation, amplification and discrimination of host plant odours and sex pheromones. In particular, we model how the spatial and temporal patterns of the odour information emerging in the glomeruli of the antennal lobe (AL) rely on the glomerular morphology, the connectivity and the complex dynamics of the AL circuits. We study how weak signals can be amplified, how different odours can be discriminated, based on stochastic (resonance) dynamics and the connectivity of the network. We further investigate the spatial and temporal coding of sex pheromone components and plant volatile compounds, in relation to the glomerular structure, arborizing patterns of the projection neurons (PNs) and timing patterns of the neuronal spiking activity. © 2006 Elsevier Ireland Ltd. All rights reserved. Keywords: Olfaction; Antennal lobe; Oscillations; Noise; Amplification; Discrimination; Spatio-temporal coding
1. Introduction Natural odours are often mixtures of many molecules in relatively specific ratios (Laurent, 1999), and odorant signals transported by turbulent air or water plumes are usually very noisy (Getz and Lutz, 1999; Laurent, 2002). In the odorant signal space, the information can be characterized by quality (chemical compounds in a given odour), quantity (dose or concentration), spatial distribution, and temporal fluctuation. Odour quality ranges from highly specific, involving single molecules, ∗ Corresponding author at: Department of Biometry and Engineering, P.O. Box 7032, SLU, S-75007 Uppsala, Sweden. Tel.: +46 18 671728; fax: +46 18 673529. E-mail addresses: [email protected] (Y. Gu), [email protected] (H. Liljenstr¨om).
to more general, involving several types of molecules in a mixture (sometimes containing hundreds of volatile components (Laurent, 2002)). Therefore, odour information processing is a very complex, multidimensional, distributed and dynamical classification problem. Insects rely on their olfactory system to identify, classify, and memorize sex pheromonal odours and plant odours for mating and finding host plants, respectively. In the insect olfactory system, the odour information is processed through three known stages in the olfactory pathway: the antenna, the antennal lobe (AL), and the mushroom body (MB). The olfactory receptor neurons (ORNs) in the antenna are essential for the recognition of odour molecules. Some ORNs are extremely specific, and will only respond to a single odorant. Other ORNs display a broader response profile, and show a robust response
0303-2647/$ – see front matter © 2006 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.biosystems.2006.04.021
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
to e.g., a group of chemically related compounds. These observations have led to the designation of different coding schemes: specific pathways represent a so-called labelled-line code, while more generalist pathways contribute to an “across-fiber” pattern of coding (Hansson, 1999). The AL, containing intrinsic inhibitory local interneurons (LNs) and output excitatory projection neurons (PNs), is the primary odour information processing centre in the insect olfactory system. The AL neurons are organized in several glomeruli. In this structure, signal representation, amplification, suppression and discrimination of the odour stimuli is carried out through the complex circuit dynamics of the AL network. The circuitry of the AL is fairly well determined, and electrophysiological studies have revealed that it can generate oscillatory neural activity (Hansson, 1999). These oscillations seem to play an important role in odour information processing. When the antenna is presented with an odorant, many PNs can synchronously oscillate, while some LNs simultaneously respond with synchronized, subthreshold oscillations in their membrane potentials (Bazhenov et al., 2001a). The odour signal space is mapped to the AL network coding space in the form of ensemble dynamics by the way of slow, fast spiking activity and synchronous oscillations. Experimental data show that different odours evoke coherent oscillations in different, but usually overlapping, ensembles of neurons (Laurent and Davidowitz, 1994). Individual odours evoke complex temporal response patterns in many neurons. These patterns differ across odours for a given neuron and across neurons for a given odour. The response of each neuron to a stimulus differs from that of other neurons by its duration, its timing relative to the odour delivery, and its temporal structure (Laurent et al., 1996). Odours appear to be represented combinatorially by dynamical neural assemblies, defined partly by the transient, but stimulus-specific synchronization of their neuronal components (Laurent, 1996). Ca2+ -imaging studies demonstrate that different odorants are represented as unique combinations of activated glomeruli, and these glomerular representations vary with the dose change (Sadek et al., 2002; Carlsson et al., 2002; Carlsson and Hansson, 2003). So far, only a few neural network models of the AL have been developed. For example, Linster et al. (1993) have studied the detection of sexual pheromones, using a formal model composed of probabilistic neurons. Two classes of neurons, selective (to one odour component) and non-selective neurons are observed in their model. Getz and Lutz (1999), proposed a neural network model with logistic neurons for odour quality processing, based
237
on the glomerular structure of the AL, as well as on the labeled-line and across-fiber response features of the ORNs. Using Hodgkin–Huxley-type models of the PNs and LNs with measured ionic currents and with random network connections, Bazhenov et al. (2001a,b) demonstrated certain synchronous oscillation properties and temporal activity patterns induced by odour stimuli in the AL. Previously, we have developed neural network models with realistic anatomical circuits and physiological parameters of some subsystems of the nervous system for understanding the relationship between structure, dynamics and function. Using neural network models of the mammalian olfactory cortex, we have especially studied how olfactory memory develops and is stored and how the information process can be made more efficient with the help of non-linear dynamics, such as oscillations, noise and chaos (Liljenstr¨om, 1991,2003; Liljenstr¨om and Wu, 1995; Liljenstr¨om and Hasselmo, ˚ 1995; Liljenstr¨om and Arhem, 1997). We have also developed models of neocortex to simulate and analyze EEG-like signals, and investigated how the dynamics are affected by the internal local and global connection topology, parameters and different types of external stimuli and signals (Gu et al., 2004,2006). In the present work, we further develop and generalize our experiences and ideas from previous neural network modeling for insect olfaction. Based on morphological and physiological data from glomerular circuitry of insect AL, we have developed a cross-scale neurodynamical model of the AL, using Hodgkin–Huxley type models of the PNs and LNs, as proposed by Bazhenov et al. (2001a). We use this model to investigate effects of connectivity and complex dynamics in amplifying weak odour signals, signal discrimination and detection of similarity, difference and speciality. 2. The glomerulus model We base the structure of our neural network model on the basic anatomy of the AL. Fig. 1 shows the synaptic connection topology within a glomerular network in our model. The membrane potentials of the PNs and LNs are described by the following dynamical equations, with variables and parameters as given in Bazhenov et al. (2001a): Cm
dvPN = −gL (VPN − EL ) − INa − IK − IA − gKL dt ×(VPN − EKL ) − IGABAA − Islow-inh − Istim ,
238
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
the role of specific connections and strengths. This is described in more detail in the next section. 3. Simulation results 3.1. Sensitivity and stochastic resonance
Fig. 1. Synaptic connection topology within a glomerular network, where each PN and LN receives input from the ORNs (dot dashed arrow lines), the LNs send inhibitory input to each other, and to the PNs (dashed arrow lines), the PNs send excitatory input to LNs and to mushroom body (MB) and lateral horn (LH) (double solid arrow lines). There are no PN–PN connections.
Cm
dVLN = −gL (VLN − EL ) − ICa − IK(Ca) − IK dt −gKL (VPN − EKL ) − IGABAA − InACh − Istim (1)
We connect several of these neurons in networks, more or less mimicking the circuitry of the AL, although we artificially modify the connectivity in order to determine
Using a glomerulus model with n PNs and m LNs, we investigate how noise affects the dynamics of the system, as well as its functional significance in signal detecting. To study the effect of noise in detecting weak signals in this model, we add a white noise current with zero mean to a subthreshold constant stimulus signal, which alone cannot be detected by the network. The white noise current can be either the external background input stimulus, or the intrinsic noise current of the network. The PNs respond with a random spiking activity. Fig. 2 shows the dynamical patterns of the membrane potentials of the first and the second PN of a glomerulus with 12 PNs and 4 LNs. When the signal amplitude is decreased further, the spiking activity of the PNs disappear. When the noise variance is increased, the PNs respond with a random spiking activity again. When the number of neurons is increased in the glomerulus, the threshold of the network is decreased. This implies that the sensitivity of the glomerulus increases with increasing number of neurons. Using a glomerulus with 21 PNs and 7 LNs, the input stimulus is now separated into two parts: one part is white background noise, and the other part is two odour signal pulses. When the mean of the background noise is subthreshold, with a small variance, the PNs show subthreshold and low amplitude oscillations. While two
Fig. 2. The response membrane potential, V, of PN1 and PN2 to a subthreshold constant stimulus signal and white noise with zero mean. The signal and the white noise are shown in the top panel, and the responses of the PN1 (left) and PN2 (right) are shown in the bottom panel.
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
239
pulse stimulus evoked high frequency oscillations of the membrane potentials of PNs in G1 and suppressed the activity of the PNs in the other glomeruli (G2–G4). Even when another odour pulse stimulates G2, with lower input amplitude than the first one, the PNs show the same activity as in Fig. 4, as if the second odour did not exist. This result, which we believe may be the first one demonstrating odour discrimination, as a result of glomerular morphology and network connectivity, corresponds to experimental data (Bill Hansson, personal communication). If PN–PN connections are added to the model network, the system cannot identify different odour components. 3.3. Spatial coding of sex pheromone components
Fig. 3. Two subthreshold odour pulses, presented at periods 200–400 and 1200–1400 ms, induce high frequency rhythmic oscillations during white background noise with subthreshold mean and small variance. The odour pulses and the white noise are shown in the top panel, while the simulated LFP is shown in the bottom panel.
weak subthreshold signal pulses are added to the background noise, the PNs show superthreshold rhythmic oscillations. The stimulation signal, the noise and the simulated local field potential (LFP), calculated as the mean membrane potential of all the 21 PNs, is shown in Fig. 3. The results described above, demonstrate a phenomenon similar to stochastic resonance, akin to what has been found in other neural systems (Moss et al., 2004). When the input white background noise level is increased, computer simulations show that the PNs and LNs oscillate slowly in a random way. When signal (odour) pulses are presented to the network, the neurons oscillate in a rhythmic way with higher frequency. 3.2. Network connectivity and discrimination In this section, the network is composed of four glomeruli, G1–G4, each with a pair of PN and LN (see Fig. 1). We use PNij to denote the ith PN in glomerulus j. One simulation result of G1 and G2 is shown in Fig. 4b and c. In this simulation, there is a constant background odour input to the network during the whole process. We use an odour pulse representing a specific odour to stimulate G1 during 200–500 ms. We found that the odour
To study the spatial coding of sex pheromone components, we connect four glomeruli to mimick the macroglomerular complex (MGC) in the AL of male moth (Carlsson et al., 2002). Because the pheromone sensitive PNs (or ORNs) do not always branch into a single glomerulus, as predicted by their physiological characteristics (Laurent, 1999), some PNs in our model arborize to a single glomerulus, while other PNs arborize also to neighbour glomeruli. In each simulation, we use two pulses as pheromone input signals to stimulate one of the four glomeruli. There is also a background noise input in each simulation. We use a four dimensional input vector (p1 , p2 , p3 , p4 ) to represent the input to the four glomeruli in MGC during each pulse. Neurons branching in glomerulus i receive input. Pi If a neuron arborizes in both glomerulus i and j, it receives input Pi + Pj If a neuron branches in glomerulus i, j and k, it receives input Pi + Pj + Pk In each simulation, only one glomerulus has the strongest response to the specific pheromone input pulses (data not shown). These results are close to experimental results (Carlsson et al., 2002). 3.4. Spatial and temporal coding of plant volatile compound Different ORNs have different response spectra to certain plant volatile compounds. Some ORNs have a strong response, some ORNs have a weak response, and others have no response at all. The same odorant often activates ORNs with different response spectra, and a certain receptor neuron can detect different molecules. Generally, the molecular receptive range of an ORN comprises closely related compounds, and a certain receptor is probably narrowly tuned to a molecular determinant that is common to many odorants (Carlsson et al., 2002).
240
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
Fig. 4. The oscillatory response patterns of two glomeruli, G1 and G2 for a constant background input to the entire AL during 0–700 ms, and a specific odour pulse stimulus to G1 during the period 200–500 ms. Panel (a) shows the background input and the specific odour pulse. Panels (b) and (c) show the response membrane potential of the PNs in G1 and G2, respectively.
To investigate how a plant volatile compound is encoded spatially by the response distribution among glomeruli, and temporally by the response timing patterns of PNs in the AL, we connect eight model glomeruli in a circle, which will represent the plant odour recognition system in the AL (Carlsson et al., 2002). Each glomerulus is here composed of six PNs and two LNs. Each LN connects to all glomeruli. We use PNij to denote the ith PN in glomerulus j. In an arbitrary glomerulus j, PN1j and PN2j arborize to four neighbour glomeruli, PN3j and PN4j arborize to two nearest neighbour glomeruli, whereas PN5j and PN6j only connect to glomerulus j. We use an eight dimensional input vector,(Q1 , Q2 , . . ., Q8 ), to represent a plant volatile compound stimulus, relayed from the ORNs to the eight AL glomeruli. There is also a background noise stimulus. Neurons branching into glomerulus i, Gi, receive input Qi . If a neuron branches into glomerulus i and j, Gi and Gj, it receives
input Qi + Qj . If a neuron branches into Gi, Gj, Gk and Gl, it receives input Qi + Qj + Qk + Ql . In the first simulation, we suppose that G8 receives the strongest inputs from the ORNs, G7 receives inputs from ORNs weaker than those of G8, but a little bit stronger than those of G1. The other glomeruli do not receive any inputs at all. Thus, the input vector is (200, 0, 0, 0, 0, 0, 300, 500). The simulation result of the local field potential (LFP) of each glomeruli is shown in Fig. 5, which demonstrates that G8 has the strongest response to the input compound, the response of G7 is weaker than that of G8, but stronger than that of G1. Some PNs (PN12 , PN22 , PN16 and PN26 ) of G2 and G6 also respond to this compound. This result demonstrates that a plant odour compound is spatially encoded by several glomeruli with different response strengths. In the second simulation, we use (200, 0, 0, 0, 0, 300, 0, 500) as input vector, which is similar to the input vector in the first simulation, as input vector
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
241
Fig. 5. Spatial representation of a plant odour compound in eight glomeruli, G1–G8, which are arranged in a circle, with each glomerulus having six PNs. In an arbitrary glomerulus j, PN1j and PN2j arborize to four neighbour glomeruli, PN3j and PN4j arborize to two nearest neighbour glomeruli, PN5j and PN6j only innervate glomerulus j. We assume that G8 receives strongest inputs from ORNs, G7 receive inputs from ORNs weaker than those of G8, but (a little) stronger than those of G1. The other glomeruli do not receive inputs. The figure shows the LFP of each glomerulus.
Fig. 6. Dynamical activity patterns of the membrane potentials of individual PNs show discrimination of two similar odour inputs presented to the network. Response of the PNs in (a) G6 with input pattern (200, 0, 0, 0, 0, 0, 300, 500), (b) G7 with input pattern (200, 0, 0, 0, 0, 0, 300, 500), (c) G6 with input pattern (200, 0, 0, 0, 0, 300, 0, 500), and (d) G7 with input pattern (200, 0, 0, 0, 0, 300, 0, 500).
242
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
to the eight glomeruli. Simulation results of the LFP show similar spatial distributions as in Fig. 5 (data not shown). However, the temporal oscillatory patterns of the membrane potentials of PNs in G6 and G7 are quite different. In particular, the PNs (PN56 , PN66 , PN57 and PN67 ), which innervate a single glomerulus show completely different dynamics in response to the two similar stimuli. Responses of individual PNs in G6 and G7 in the first and second simulation are shown in Fig. 6. This result demonstrates temporal coding in our model, in addition to the spatial coding. The results are also in agreement with the experimental finding that similar plant odours are encoded by the same group of glomeruli, while the differences between similar odours are encoded by different dynamical patterns (Lei et al., 2004). Here, we demonstrate that the PNs, which innervate a single glomerulus, play an important role for differentiating similar odours by temporal coding. In a third simulation, we use a different input vector, (0, 0, 200, 500, 100, 0, 0, 0) as odour stimulation to the eight glomeruli. In this case, the spatial distribution is quite different from what is shown in Fig. 5. In these simulations, we also found that the PNs that arborize to neighbouring glomeruli respond to both of the compounds detected by several type of ORNs, and compounds detected by only one type of ORNs. However, the PNs that arborize to a single glomerulus show slower oscillation, or no response to compounds detected by several types of ORNs. They show faster oscillations in response to compounds detected by one type of ORNs. Therefore, we suggest that PNs arborizing to neighbour glomeruli detect similarity, whereas PNs arborizing to a single glomerulus detect difference and speciality.
Hence, we believe our model has captured some of the major features important for insect olfactory information processing. Our simulation results also indicate how the network neurodynamics, including oscillations and noise, can make this information processing efficient. The spatio-temporal activity patterns are clearly related to the network circuitry. In future work, we intend to investigate dose discrimination in odour blends, and encoding time signal of pheromone filaments. In particular, we are interested in understanding the spatio-temporal encoding mechanisms in AL for odour blends and discrete puffs of odours, as well as the mechanisms that decode the spatiotemporal patterns in the AL to spatial patterns, forming spatial memories in the Kenyon cell array in the MB. We also intend to investigate how plasticity and neuromodulation increase the sensitivity and discrimination of behaviourally relevant odours. We will extend our model to include olfactory receptor neurons and several types of neurons in higher levels, as well as special structures important for control of excitability and plasticity. We believe that this kind of computer modeling, combined with and based on experimental studies, will have significance for behavioural ecology and biological technology, such as biological control and electronic noses. Clearly, some of the results achieved so far would have been difficult to achieve by experimental methods alone. Acknowledgements This work has been made possible by a generous grant from Vinnova, the Swedish Agency for Innovation Systems. We thank Bill Hansson for fruitful discussions. References
4. Discussion and conclusions In this work, using a neural network model of the insect antennal lobe, with an anatomically inspired circuit and biophysical model neurons, we have investigated relationships between structure, dynamics, and function. Our model demonstrated phenomena akin to stochastic resonance in detecting weak signals, and we suggest that this phenomenon should be looked for in the insect olfactory system. In a model network with four glomeruli, the network can easily identify specific odour stimuli to one glomerulus, while suppress the response of other glomeruli to the constant background input. In particular, we found that the PNs that innervate different numbers of glomeruli play separate roles in coding similarity, difference and speciality by spatial and temporal coding.
Bazhenov, M., Stopfer, M., Rabinovich, M., Huerta, R., Abarbanel, H.D.I., Sejnowski, T.J., Laurent, G., 2001a. Model of transient oscillatory synchronization in the locust antennal lobe. Neuron 30, 553–567. Bazhenov, M., Stopfer, M., Rabinovich, M., Huerta, R., Abarbanel, H.D.I., Sejnowski, T.J., Laurent, G., 2001b. Model of cellular and network mechanisms for odour-evoked temporal patterning in the locust antennal lobe. Neuron 30, 569–581. Carlsson, M.A., Galizia, C.G., Hansson, B.S., 2002. Spatial representation of odours in the antennal lobe of the moth Spodoptera littoralis. Chem. Senses 27, 231–244. Carlsson, M.A., Hansson, B.S., 2003. Dose-response characteristics of glomerular activity in the moth antennal lobe. Chem. Senses 28, 269–278. Getz, W.M., Lutz, A., 1999. A neural network model of general olfactory coding in the insect antennal lobe. Chem. Senses 24, 351–372. Gu, Y., Halnes, G., Liljenstr¨om, H., Wahlund, B., 2004. A cortical network model for clinical EEG data analysis. Neurocomputing 58–60, 1187–1196.
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243 Gu Y., Halnes G., Liljenstr¨om H., von Rosen D., Wahlund B., Liang H., 2006. Modelling ECT effects by connectivity changes in cortical neural networks, Neurocomputing, 69, 1341-1347. Hansson, B.S., 1999. Insect Olfaction. Springer. Laurent, G., 1996. Dynamical representation of odours by oscillating and evolving neural assemblies. TINS 19, 489. Laurent, G., 1999. A system perspective on early olfactory coding. Science 286, 723. Laurent, G., 2002. Olfactory networks dynamics and the coding of multidimensional signals. Nat. Rev.: Neurosci. 3, 884. Laurent, G., Davidowitz, H., 1994. Encoding of olfactory information with oscillating neural assemblies. Science 265, 1872. Laurent, G., Wehr, M., Davidowitz, H., 1996. Temporal representations of odours in an olfactory network. J. Neurosci. 16, 3837–3847. Lei, H., Christensen, T.A., Hildebrand, J.G., 2004. Spatial and temporal organization of ensemble reresentations for different odour classes in the moth antennal lobe. J. Neurosci. 24, 11108–11119. ˚ Liljenstr¨om, H., Arhem, P., 1997. Investigating amplifying and controlling mechanisms for random events in neural systems. In: Bower, J.M. (Ed.), Computational Neuroscience. Plenum Press, New York.
243
Liljenstr¨om, H., Wu, X., 1995. Noise-enhanced performance in a cortical associative memory model. Int. J. Neural Syst. 6, 19– 29. Liljenstr¨om, H., 1991. Modeling the dynamics of olfactory cortex using simplified network units and realistic architecture. Int. J. Neural Syst. 2, 1–15. Liljenstr¨om, H., 2003. Neural stability and flexibility—a computational approach. J. Neuropsychopharmacol. 28, S64–S73. Liljenstr¨om, H., Hasselmo, M.E., 1995. Cholinergic modulation of cortical oscillatory dynamics. J. Neurophysiol. 74, 288–297. Linster, C., Masson, C., Kerszberg, M., Personnaz, L., Dreyfus, G., 1993. Computational diversity in a formal model of the insect olfactory macroglomerulus. Neural Comput. 5, 239–252. Moss, F., Ward, L.M., Sannita, W.G., 2004. Stochastic resonance and sensory information processing: a tutorial and review of application. Clin. Neurophysiol. 115, 267–281. Sadek, M.M., Hansson, B.S., Rospars, J.P., Anton, S., 2002. Glomerular representation of plant volatiles and sex pheromone components in the antennal lobe of the female Spodoptera littoralis. J. Exp. Biol. 205, 1363–1376.
Modelling efficiency in insect olfactory information processing Yuqiao Gu a,b,c , Hans Liljenstr¨om a,b,∗ a c
Department of Biometry and Engineering, P.O. Box 7032, SLU, S-75007 Uppsala, Sweden b Agora for Biosystems, P.O. Box 57, S-19322 Sigtuna, Sweden School of Automation and Energy Engineering, Tianjin University of Technology, PR China Received 1 December 2005; accepted 6 April 2006
Abstract The olfactory system of insects is essential for the search of food and mates, and weak signals can be detected, amplified and discriminated in a fluctuating environment. The olfactory system also allows for learning and recall of odour memories. Based on anatomical, physiological, and behavioural data from the olfactory system of insects, we have developed a cross-scale dynamical neural network model to simulate the presentation, amplification and discrimination of host plant odours and sex pheromones. In particular, we model how the spatial and temporal patterns of the odour information emerging in the glomeruli of the antennal lobe (AL) rely on the glomerular morphology, the connectivity and the complex dynamics of the AL circuits. We study how weak signals can be amplified, how different odours can be discriminated, based on stochastic (resonance) dynamics and the connectivity of the network. We further investigate the spatial and temporal coding of sex pheromone components and plant volatile compounds, in relation to the glomerular structure, arborizing patterns of the projection neurons (PNs) and timing patterns of the neuronal spiking activity. © 2006 Elsevier Ireland Ltd. All rights reserved. Keywords: Olfaction; Antennal lobe; Oscillations; Noise; Amplification; Discrimination; Spatio-temporal coding
1. Introduction Natural odours are often mixtures of many molecules in relatively specific ratios (Laurent, 1999), and odorant signals transported by turbulent air or water plumes are usually very noisy (Getz and Lutz, 1999; Laurent, 2002). In the odorant signal space, the information can be characterized by quality (chemical compounds in a given odour), quantity (dose or concentration), spatial distribution, and temporal fluctuation. Odour quality ranges from highly specific, involving single molecules, ∗ Corresponding author at: Department of Biometry and Engineering, P.O. Box 7032, SLU, S-75007 Uppsala, Sweden. Tel.: +46 18 671728; fax: +46 18 673529. E-mail addresses: [email protected] (Y. Gu), [email protected] (H. Liljenstr¨om).
to more general, involving several types of molecules in a mixture (sometimes containing hundreds of volatile components (Laurent, 2002)). Therefore, odour information processing is a very complex, multidimensional, distributed and dynamical classification problem. Insects rely on their olfactory system to identify, classify, and memorize sex pheromonal odours and plant odours for mating and finding host plants, respectively. In the insect olfactory system, the odour information is processed through three known stages in the olfactory pathway: the antenna, the antennal lobe (AL), and the mushroom body (MB). The olfactory receptor neurons (ORNs) in the antenna are essential for the recognition of odour molecules. Some ORNs are extremely specific, and will only respond to a single odorant. Other ORNs display a broader response profile, and show a robust response
0303-2647/$ – see front matter © 2006 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.biosystems.2006.04.021
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
to e.g., a group of chemically related compounds. These observations have led to the designation of different coding schemes: specific pathways represent a so-called labelled-line code, while more generalist pathways contribute to an “across-fiber” pattern of coding (Hansson, 1999). The AL, containing intrinsic inhibitory local interneurons (LNs) and output excitatory projection neurons (PNs), is the primary odour information processing centre in the insect olfactory system. The AL neurons are organized in several glomeruli. In this structure, signal representation, amplification, suppression and discrimination of the odour stimuli is carried out through the complex circuit dynamics of the AL network. The circuitry of the AL is fairly well determined, and electrophysiological studies have revealed that it can generate oscillatory neural activity (Hansson, 1999). These oscillations seem to play an important role in odour information processing. When the antenna is presented with an odorant, many PNs can synchronously oscillate, while some LNs simultaneously respond with synchronized, subthreshold oscillations in their membrane potentials (Bazhenov et al., 2001a). The odour signal space is mapped to the AL network coding space in the form of ensemble dynamics by the way of slow, fast spiking activity and synchronous oscillations. Experimental data show that different odours evoke coherent oscillations in different, but usually overlapping, ensembles of neurons (Laurent and Davidowitz, 1994). Individual odours evoke complex temporal response patterns in many neurons. These patterns differ across odours for a given neuron and across neurons for a given odour. The response of each neuron to a stimulus differs from that of other neurons by its duration, its timing relative to the odour delivery, and its temporal structure (Laurent et al., 1996). Odours appear to be represented combinatorially by dynamical neural assemblies, defined partly by the transient, but stimulus-specific synchronization of their neuronal components (Laurent, 1996). Ca2+ -imaging studies demonstrate that different odorants are represented as unique combinations of activated glomeruli, and these glomerular representations vary with the dose change (Sadek et al., 2002; Carlsson et al., 2002; Carlsson and Hansson, 2003). So far, only a few neural network models of the AL have been developed. For example, Linster et al. (1993) have studied the detection of sexual pheromones, using a formal model composed of probabilistic neurons. Two classes of neurons, selective (to one odour component) and non-selective neurons are observed in their model. Getz and Lutz (1999), proposed a neural network model with logistic neurons for odour quality processing, based
237
on the glomerular structure of the AL, as well as on the labeled-line and across-fiber response features of the ORNs. Using Hodgkin–Huxley-type models of the PNs and LNs with measured ionic currents and with random network connections, Bazhenov et al. (2001a,b) demonstrated certain synchronous oscillation properties and temporal activity patterns induced by odour stimuli in the AL. Previously, we have developed neural network models with realistic anatomical circuits and physiological parameters of some subsystems of the nervous system for understanding the relationship between structure, dynamics and function. Using neural network models of the mammalian olfactory cortex, we have especially studied how olfactory memory develops and is stored and how the information process can be made more efficient with the help of non-linear dynamics, such as oscillations, noise and chaos (Liljenstr¨om, 1991,2003; Liljenstr¨om and Wu, 1995; Liljenstr¨om and Hasselmo, ˚ 1995; Liljenstr¨om and Arhem, 1997). We have also developed models of neocortex to simulate and analyze EEG-like signals, and investigated how the dynamics are affected by the internal local and global connection topology, parameters and different types of external stimuli and signals (Gu et al., 2004,2006). In the present work, we further develop and generalize our experiences and ideas from previous neural network modeling for insect olfaction. Based on morphological and physiological data from glomerular circuitry of insect AL, we have developed a cross-scale neurodynamical model of the AL, using Hodgkin–Huxley type models of the PNs and LNs, as proposed by Bazhenov et al. (2001a). We use this model to investigate effects of connectivity and complex dynamics in amplifying weak odour signals, signal discrimination and detection of similarity, difference and speciality. 2. The glomerulus model We base the structure of our neural network model on the basic anatomy of the AL. Fig. 1 shows the synaptic connection topology within a glomerular network in our model. The membrane potentials of the PNs and LNs are described by the following dynamical equations, with variables and parameters as given in Bazhenov et al. (2001a): Cm
dvPN = −gL (VPN − EL ) − INa − IK − IA − gKL dt ×(VPN − EKL ) − IGABAA − Islow-inh − Istim ,
238
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
the role of specific connections and strengths. This is described in more detail in the next section. 3. Simulation results 3.1. Sensitivity and stochastic resonance
Fig. 1. Synaptic connection topology within a glomerular network, where each PN and LN receives input from the ORNs (dot dashed arrow lines), the LNs send inhibitory input to each other, and to the PNs (dashed arrow lines), the PNs send excitatory input to LNs and to mushroom body (MB) and lateral horn (LH) (double solid arrow lines). There are no PN–PN connections.
Cm
dVLN = −gL (VLN − EL ) − ICa − IK(Ca) − IK dt −gKL (VPN − EKL ) − IGABAA − InACh − Istim (1)
We connect several of these neurons in networks, more or less mimicking the circuitry of the AL, although we artificially modify the connectivity in order to determine
Using a glomerulus model with n PNs and m LNs, we investigate how noise affects the dynamics of the system, as well as its functional significance in signal detecting. To study the effect of noise in detecting weak signals in this model, we add a white noise current with zero mean to a subthreshold constant stimulus signal, which alone cannot be detected by the network. The white noise current can be either the external background input stimulus, or the intrinsic noise current of the network. The PNs respond with a random spiking activity. Fig. 2 shows the dynamical patterns of the membrane potentials of the first and the second PN of a glomerulus with 12 PNs and 4 LNs. When the signal amplitude is decreased further, the spiking activity of the PNs disappear. When the noise variance is increased, the PNs respond with a random spiking activity again. When the number of neurons is increased in the glomerulus, the threshold of the network is decreased. This implies that the sensitivity of the glomerulus increases with increasing number of neurons. Using a glomerulus with 21 PNs and 7 LNs, the input stimulus is now separated into two parts: one part is white background noise, and the other part is two odour signal pulses. When the mean of the background noise is subthreshold, with a small variance, the PNs show subthreshold and low amplitude oscillations. While two
Fig. 2. The response membrane potential, V, of PN1 and PN2 to a subthreshold constant stimulus signal and white noise with zero mean. The signal and the white noise are shown in the top panel, and the responses of the PN1 (left) and PN2 (right) are shown in the bottom panel.
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
239
pulse stimulus evoked high frequency oscillations of the membrane potentials of PNs in G1 and suppressed the activity of the PNs in the other glomeruli (G2–G4). Even when another odour pulse stimulates G2, with lower input amplitude than the first one, the PNs show the same activity as in Fig. 4, as if the second odour did not exist. This result, which we believe may be the first one demonstrating odour discrimination, as a result of glomerular morphology and network connectivity, corresponds to experimental data (Bill Hansson, personal communication). If PN–PN connections are added to the model network, the system cannot identify different odour components. 3.3. Spatial coding of sex pheromone components
Fig. 3. Two subthreshold odour pulses, presented at periods 200–400 and 1200–1400 ms, induce high frequency rhythmic oscillations during white background noise with subthreshold mean and small variance. The odour pulses and the white noise are shown in the top panel, while the simulated LFP is shown in the bottom panel.
weak subthreshold signal pulses are added to the background noise, the PNs show superthreshold rhythmic oscillations. The stimulation signal, the noise and the simulated local field potential (LFP), calculated as the mean membrane potential of all the 21 PNs, is shown in Fig. 3. The results described above, demonstrate a phenomenon similar to stochastic resonance, akin to what has been found in other neural systems (Moss et al., 2004). When the input white background noise level is increased, computer simulations show that the PNs and LNs oscillate slowly in a random way. When signal (odour) pulses are presented to the network, the neurons oscillate in a rhythmic way with higher frequency. 3.2. Network connectivity and discrimination In this section, the network is composed of four glomeruli, G1–G4, each with a pair of PN and LN (see Fig. 1). We use PNij to denote the ith PN in glomerulus j. One simulation result of G1 and G2 is shown in Fig. 4b and c. In this simulation, there is a constant background odour input to the network during the whole process. We use an odour pulse representing a specific odour to stimulate G1 during 200–500 ms. We found that the odour
To study the spatial coding of sex pheromone components, we connect four glomeruli to mimick the macroglomerular complex (MGC) in the AL of male moth (Carlsson et al., 2002). Because the pheromone sensitive PNs (or ORNs) do not always branch into a single glomerulus, as predicted by their physiological characteristics (Laurent, 1999), some PNs in our model arborize to a single glomerulus, while other PNs arborize also to neighbour glomeruli. In each simulation, we use two pulses as pheromone input signals to stimulate one of the four glomeruli. There is also a background noise input in each simulation. We use a four dimensional input vector (p1 , p2 , p3 , p4 ) to represent the input to the four glomeruli in MGC during each pulse. Neurons branching in glomerulus i receive input. Pi If a neuron arborizes in both glomerulus i and j, it receives input Pi + Pj If a neuron branches in glomerulus i, j and k, it receives input Pi + Pj + Pk In each simulation, only one glomerulus has the strongest response to the specific pheromone input pulses (data not shown). These results are close to experimental results (Carlsson et al., 2002). 3.4. Spatial and temporal coding of plant volatile compound Different ORNs have different response spectra to certain plant volatile compounds. Some ORNs have a strong response, some ORNs have a weak response, and others have no response at all. The same odorant often activates ORNs with different response spectra, and a certain receptor neuron can detect different molecules. Generally, the molecular receptive range of an ORN comprises closely related compounds, and a certain receptor is probably narrowly tuned to a molecular determinant that is common to many odorants (Carlsson et al., 2002).
240
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
Fig. 4. The oscillatory response patterns of two glomeruli, G1 and G2 for a constant background input to the entire AL during 0–700 ms, and a specific odour pulse stimulus to G1 during the period 200–500 ms. Panel (a) shows the background input and the specific odour pulse. Panels (b) and (c) show the response membrane potential of the PNs in G1 and G2, respectively.
To investigate how a plant volatile compound is encoded spatially by the response distribution among glomeruli, and temporally by the response timing patterns of PNs in the AL, we connect eight model glomeruli in a circle, which will represent the plant odour recognition system in the AL (Carlsson et al., 2002). Each glomerulus is here composed of six PNs and two LNs. Each LN connects to all glomeruli. We use PNij to denote the ith PN in glomerulus j. In an arbitrary glomerulus j, PN1j and PN2j arborize to four neighbour glomeruli, PN3j and PN4j arborize to two nearest neighbour glomeruli, whereas PN5j and PN6j only connect to glomerulus j. We use an eight dimensional input vector,(Q1 , Q2 , . . ., Q8 ), to represent a plant volatile compound stimulus, relayed from the ORNs to the eight AL glomeruli. There is also a background noise stimulus. Neurons branching into glomerulus i, Gi, receive input Qi . If a neuron branches into glomerulus i and j, Gi and Gj, it receives
input Qi + Qj . If a neuron branches into Gi, Gj, Gk and Gl, it receives input Qi + Qj + Qk + Ql . In the first simulation, we suppose that G8 receives the strongest inputs from the ORNs, G7 receives inputs from ORNs weaker than those of G8, but a little bit stronger than those of G1. The other glomeruli do not receive any inputs at all. Thus, the input vector is (200, 0, 0, 0, 0, 0, 300, 500). The simulation result of the local field potential (LFP) of each glomeruli is shown in Fig. 5, which demonstrates that G8 has the strongest response to the input compound, the response of G7 is weaker than that of G8, but stronger than that of G1. Some PNs (PN12 , PN22 , PN16 and PN26 ) of G2 and G6 also respond to this compound. This result demonstrates that a plant odour compound is spatially encoded by several glomeruli with different response strengths. In the second simulation, we use (200, 0, 0, 0, 0, 300, 0, 500) as input vector, which is similar to the input vector in the first simulation, as input vector
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
241
Fig. 5. Spatial representation of a plant odour compound in eight glomeruli, G1–G8, which are arranged in a circle, with each glomerulus having six PNs. In an arbitrary glomerulus j, PN1j and PN2j arborize to four neighbour glomeruli, PN3j and PN4j arborize to two nearest neighbour glomeruli, PN5j and PN6j only innervate glomerulus j. We assume that G8 receives strongest inputs from ORNs, G7 receive inputs from ORNs weaker than those of G8, but (a little) stronger than those of G1. The other glomeruli do not receive inputs. The figure shows the LFP of each glomerulus.
Fig. 6. Dynamical activity patterns of the membrane potentials of individual PNs show discrimination of two similar odour inputs presented to the network. Response of the PNs in (a) G6 with input pattern (200, 0, 0, 0, 0, 0, 300, 500), (b) G7 with input pattern (200, 0, 0, 0, 0, 0, 300, 500), (c) G6 with input pattern (200, 0, 0, 0, 0, 300, 0, 500), and (d) G7 with input pattern (200, 0, 0, 0, 0, 300, 0, 500).
242
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243
to the eight glomeruli. Simulation results of the LFP show similar spatial distributions as in Fig. 5 (data not shown). However, the temporal oscillatory patterns of the membrane potentials of PNs in G6 and G7 are quite different. In particular, the PNs (PN56 , PN66 , PN57 and PN67 ), which innervate a single glomerulus show completely different dynamics in response to the two similar stimuli. Responses of individual PNs in G6 and G7 in the first and second simulation are shown in Fig. 6. This result demonstrates temporal coding in our model, in addition to the spatial coding. The results are also in agreement with the experimental finding that similar plant odours are encoded by the same group of glomeruli, while the differences between similar odours are encoded by different dynamical patterns (Lei et al., 2004). Here, we demonstrate that the PNs, which innervate a single glomerulus, play an important role for differentiating similar odours by temporal coding. In a third simulation, we use a different input vector, (0, 0, 200, 500, 100, 0, 0, 0) as odour stimulation to the eight glomeruli. In this case, the spatial distribution is quite different from what is shown in Fig. 5. In these simulations, we also found that the PNs that arborize to neighbouring glomeruli respond to both of the compounds detected by several type of ORNs, and compounds detected by only one type of ORNs. However, the PNs that arborize to a single glomerulus show slower oscillation, or no response to compounds detected by several types of ORNs. They show faster oscillations in response to compounds detected by one type of ORNs. Therefore, we suggest that PNs arborizing to neighbour glomeruli detect similarity, whereas PNs arborizing to a single glomerulus detect difference and speciality.
Hence, we believe our model has captured some of the major features important for insect olfactory information processing. Our simulation results also indicate how the network neurodynamics, including oscillations and noise, can make this information processing efficient. The spatio-temporal activity patterns are clearly related to the network circuitry. In future work, we intend to investigate dose discrimination in odour blends, and encoding time signal of pheromone filaments. In particular, we are interested in understanding the spatio-temporal encoding mechanisms in AL for odour blends and discrete puffs of odours, as well as the mechanisms that decode the spatiotemporal patterns in the AL to spatial patterns, forming spatial memories in the Kenyon cell array in the MB. We also intend to investigate how plasticity and neuromodulation increase the sensitivity and discrimination of behaviourally relevant odours. We will extend our model to include olfactory receptor neurons and several types of neurons in higher levels, as well as special structures important for control of excitability and plasticity. We believe that this kind of computer modeling, combined with and based on experimental studies, will have significance for behavioural ecology and biological technology, such as biological control and electronic noses. Clearly, some of the results achieved so far would have been difficult to achieve by experimental methods alone. Acknowledgements This work has been made possible by a generous grant from Vinnova, the Swedish Agency for Innovation Systems. We thank Bill Hansson for fruitful discussions. References
4. Discussion and conclusions In this work, using a neural network model of the insect antennal lobe, with an anatomically inspired circuit and biophysical model neurons, we have investigated relationships between structure, dynamics, and function. Our model demonstrated phenomena akin to stochastic resonance in detecting weak signals, and we suggest that this phenomenon should be looked for in the insect olfactory system. In a model network with four glomeruli, the network can easily identify specific odour stimuli to one glomerulus, while suppress the response of other glomeruli to the constant background input. In particular, we found that the PNs that innervate different numbers of glomeruli play separate roles in coding similarity, difference and speciality by spatial and temporal coding.
Bazhenov, M., Stopfer, M., Rabinovich, M., Huerta, R., Abarbanel, H.D.I., Sejnowski, T.J., Laurent, G., 2001a. Model of transient oscillatory synchronization in the locust antennal lobe. Neuron 30, 553–567. Bazhenov, M., Stopfer, M., Rabinovich, M., Huerta, R., Abarbanel, H.D.I., Sejnowski, T.J., Laurent, G., 2001b. Model of cellular and network mechanisms for odour-evoked temporal patterning in the locust antennal lobe. Neuron 30, 569–581. Carlsson, M.A., Galizia, C.G., Hansson, B.S., 2002. Spatial representation of odours in the antennal lobe of the moth Spodoptera littoralis. Chem. Senses 27, 231–244. Carlsson, M.A., Hansson, B.S., 2003. Dose-response characteristics of glomerular activity in the moth antennal lobe. Chem. Senses 28, 269–278. Getz, W.M., Lutz, A., 1999. A neural network model of general olfactory coding in the insect antennal lobe. Chem. Senses 24, 351–372. Gu, Y., Halnes, G., Liljenstr¨om, H., Wahlund, B., 2004. A cortical network model for clinical EEG data analysis. Neurocomputing 58–60, 1187–1196.
Y. Gu, H. Liljenstr¨om / BioSystems 89 (2007) 236–243 Gu Y., Halnes G., Liljenstr¨om H., von Rosen D., Wahlund B., Liang H., 2006. Modelling ECT effects by connectivity changes in cortical neural networks, Neurocomputing, 69, 1341-1347. Hansson, B.S., 1999. Insect Olfaction. Springer. Laurent, G., 1996. Dynamical representation of odours by oscillating and evolving neural assemblies. TINS 19, 489. Laurent, G., 1999. A system perspective on early olfactory coding. Science 286, 723. Laurent, G., 2002. Olfactory networks dynamics and the coding of multidimensional signals. Nat. Rev.: Neurosci. 3, 884. Laurent, G., Davidowitz, H., 1994. Encoding of olfactory information with oscillating neural assemblies. Science 265, 1872. Laurent, G., Wehr, M., Davidowitz, H., 1996. Temporal representations of odours in an olfactory network. J. Neurosci. 16, 3837–3847. Lei, H., Christensen, T.A., Hildebrand, J.G., 2004. Spatial and temporal organization of ensemble reresentations for different odour classes in the moth antennal lobe. J. Neurosci. 24, 11108–11119. ˚ Liljenstr¨om, H., Arhem, P., 1997. Investigating amplifying and controlling mechanisms for random events in neural systems. In: Bower, J.M. (Ed.), Computational Neuroscience. Plenum Press, New York.
243
Liljenstr¨om, H., Wu, X., 1995. Noise-enhanced performance in a cortical associative memory model. Int. J. Neural Syst. 6, 19– 29. Liljenstr¨om, H., 1991. Modeling the dynamics of olfactory cortex using simplified network units and realistic architecture. Int. J. Neural Syst. 2, 1–15. Liljenstr¨om, H., 2003. Neural stability and flexibility—a computational approach. J. Neuropsychopharmacol. 28, S64–S73. Liljenstr¨om, H., Hasselmo, M.E., 1995. Cholinergic modulation of cortical oscillatory dynamics. J. Neurophysiol. 74, 288–297. Linster, C., Masson, C., Kerszberg, M., Personnaz, L., Dreyfus, G., 1993. Computational diversity in a formal model of the insect olfactory macroglomerulus. Neural Comput. 5, 239–252. Moss, F., Ward, L.M., Sannita, W.G., 2004. Stochastic resonance and sensory information processing: a tutorial and review of application. Clin. Neurophysiol. 115, 267–281. Sadek, M.M., Hansson, B.S., Rospars, J.P., Anton, S., 2002. Glomerular representation of plant volatiles and sex pheromone components in the antennal lobe of the female Spodoptera littoralis. J. Exp. Biol. 205, 1363–1376.
