© 2009 Nature America, Inc. All rights reserved.
news and views as being the major effectors that were affected by SHH-proteoglycan interactions. Consequently, these genes may be critical for SHH-controlled neuronal proliferation. Their expression was consistently detected in in vitro and in vivo models, reinforcing the strength of the finding. These data led the authors to propose that an SHH-dependent expression profile signature is associated with neural progenitor proliferation that is distinct from that involved in t issue patterning. This documentation of SHH proteoglycan–dependent gene expression is a landmark for the neurosciences. The Chan et al. study2 represents a major step toward a full understanding of SHH signaling complexity by assessing its role in the establishment and maintenance of a favorable microenvironment for neural p recursor p roliferation. Moreover, the authors demonstrate that proteoglycan-SHH interactions are essential for the signaling of proliferation versus differentiation. The results presented by Chan et al.2 also suggest that proteoglycans might contribute to the developmental restriction of SHHdriven proliferation in mitogenic niches,
albeit through very different interactions. An increasing number of endogenous inhibitory ligands have been identified that attenuate SHH mitogenicity during cerebellar postnatal development. Some of these also contain Cardin-Weintraub motifs and bind to heparan sulfate. For example, vitronectin12, protease nexin-1 (ref. 13) and FGF-2/FGFR1 (ref. 14) signaling can stop the expansion of granular precursors by specifically antagonizing the SHH pathway. It is tempting to speculate that these ligands might compete with SHH for proteoglycan binding sites. This occupancy would chase away SHH, thus depleting the proliferative niche and possibly triggering downstream pro-differentiating signaling. This hypothesis obviously requires further experimental evidence. Furthermore, both proteoglycans and SHH have been implicated in cancer. Tumor growth, angiogenesis and metastasis originating from tumorigenic cell lines are attenuated in mice lacking the heparan sulfate proteoglycan g lypican-1 (ref. 15). Thus, in addition to advancing our understanding of Shh signaling, Chan et al.’s2 approach of interfering with proteoglycan-modulated
signaling via the Cardin-Weintraub motif of proteoglycan-binding proteins may well inspire developmental and cell biologists and have an effect on studies investigating the role of SHH in cancer. 1. Kreuger, J., Spillmann, D., Li, J.P. & Lindahl, U. J. Cell Biol. 174, 323–327 (2006). 2. Chan et al. Nat. Neurosci. 12, 409–417 (2009). 3. Varjosalo, M. & Taipale, J. Genes Dev. 22, 2454–2472 (2008). 4. Rubin, J.B., Choi, Y. & Segal, R.A. Development 129, 2223–2232 (2002). 5. Ahn, S. & Joyner, A.L. Nature 437, 894–897 (2005). 6. Lewis, P.M., Gritli-Linde, A., Smeyne, R., Kottmann, A. & McMahon, A.P. Dev. Biol. 270, 393–410 (2004). 7. Vaillant, C. & Monard, D. Cerebellum published online, doi:10.1007/s12311-009-0094-8 (18 February 2009). 8. Ulloa, F. & Briscoe, J. Cell Cycle 6, 2640–2649 (2007). 9. Sasaki, H., Nishizaki, Y., Hui, C., Nakafuku, M. & Kondoh, H. Development 126, 3915–3924 (1999). 10. Aza-Blanc, P., Lin, H.Y., Ruiz i Altaba, A. & Kornberg, T.B. Development 127, 4293–4301 (2000). 11. Pan, Y., Bai, C.B., Joyner, A.L. & Wang, B. Mol. Cell. Biol. 26, 3365–3377 (2006). 12. Pons, S., Trejo, J.L., Martinez-Morales, J.R. & Marti, E. Development 128, 1481–1492 (2001). 13. Vaillant, C. et al. Development 134, 1745–1754 (2007). 14. Fogarty, M.P., Emmenegger, B.A., Grasfeder, L.L., Oliver, T.G. & Wechsler-Reya, R.J. Proc. Natl. Acad. Sci. USA 104, 2973–2978 (2007). 15. Aikawa, T. et al. J. Clin. Invest. 118, 89–99 (2008).
Neuronal communication: a detailed balancing act Emilio Salinas What controls the functional connections between sending and receiving neurons? A new model suggests that each receiver circuit has a local switch that is controlled by the balance between excitation and inhibition. Learning and experience can modify the physical connections between neurons in the mammalian brain over time scales of minutes to days. However, the flexibility of everyday behavior provides abundant evidence that the functional connections between sensory and motor systems can change virtually instantaneously. If, for example, you are typing a document, I can ask you to press the H key with your right index finger or with any other finger, and vice versa, I can ask you to use the right index finger to press any key. Thus, the map between letters and finger presses can change at any moment and any possible combination can be executed on demand. At the neural level, this means that the sensory afferents conveying visual information about
Department of Neurobiology and Anatomy, Wake Forest University School of Medicine, Winston-Salem, North Carolina, USA. e-mail:
[email protected]
372
the keyboard must be able to communicate with all the efferent motor circuits that control the fingers. Therefore, a mechanism must exist to gate the underlying connections according to the ongoing behavioral task so that only the appropriate communication channel is switched on at any moment. Understanding how this switchboard works is a fundamental, but difficult, problem in systems and computational neuroscience. In this issue, Vogels and Abbott1 describe a biophysically plausible account of how this switchboard may work. According to their new model, each excitatory input to a downstream group of receiver neurons is carefully cancelled by local inhibition, except when those receiver neurons need to replicate one of the inputs; in that case, that one signal is ‘gated on’ by upsetting the detailed balance between excitation and inhibition. Thus, their model provides a specific mechanism for how neural signals may route through different circuits.
This study1 is the result of advances in understanding two phenomena in cortical neurophysiology: the balance between excitation and inhibition in local circuits and the transmission of information from one circuit to another. The first issue has been thoroughly studied by both theoreticians and experimentalists. Modeling studies have shown that the simultaneous action of excitation and inhibition generates large fluctuations in membrane potential, which are critical for explaining the high variability that characterizes cortical neuron spike trains2. Modeling has also revealed that the dynamics of both single neurons3,4 and whole networks5 are extremely different in the balanced (fluctuation driven) and unbalanced (excitation driven) regimes. Correspondingly, experiments have shown6,7 that feedback mechanisms in local microcircuits ensure that excitation and inhibition stay notably proportional, maintaining a delicate balance throughout a large range of activation levels.
volume 12 | number 4 | april 2009 nature neuroscience
© 2009 Nature America, Inc. All rights reserved.
news and views a
b
Gates off
100
Sender 1
Sender 2
Receiver
Firing rate (spikes per s)
The second issue, relating to information transmission, is a computational problem that, on the surface, seems to be rather innocuous. Suppose that the neurons in layer L1 of a network respond to a stimulus and that this signal goes from L1 to L2, from L2 to L3, and so on. Under what conditions will the signal be faithfully transmitted? Or, how many layers can the original signal reach w ithout it being corrupted? This problem dates back to a study that proposed8 that faithful t ransmission may be achieved if the neurons that are active in a given layer fire volleys of synchronous spikes that excite the next layer, and so on. This model was developed further9 and an alternative proposal10 later emerged suggesting that s ynchronicity is not critical and that asynchronous changes in firing rate can also be transmitted in a stable manner. However, in either case, there are two severe complications. First, real neuronal networks are not purely feedforward (L1 → L2 → ... → LN), but instead have extensive feedback connections; in fact, cortical areas that communicate with each other are t ypically interconnected bidirectionally11. Second, neurons do not simply relay information through waves of excitation; instead, when a neuron is activated, it normally receives large quantities of both excitation and inhibition. Therefore, the stable propagation of activity needs to coexist with the irregular f luctuations that characterize b alanced networks. Reliable signal t ransmission is tricky, because the signal may either fade away and die amidst the sea of fluctuations or else grow uncontrollably into a sort of shock wave. Nevertheless, recent simulation studies 12,13 have found conditions for reliable signal propagation under realistic constraints, both for synchronous volleys of activity and for v ariations in firing rate. Vogels and Abbott developed one of those models13, which is the starting point for their current work. The network constructed by Vogels and Abbott1 contains about 20,000 neurons that produce spikes and interact through model synaptic conductances. It represents a sparse description of a large swath of cortex spanning many areas. Without any input, the recurrent connectivity of the net produces irregular, asynchronous activity with low firing rates around 8 spikes per s. Embedded in this structure, there are smaller subnetworks representing distinct cortical areas, designated as ‘senders’ and ‘receivers’ (Fig. 1a). The crucial part of the model’s design is that a sender group excites both excitatory and inhibitory neurons in a
Gate 1 on
Gate 2 on
0 100
0 100
0 100
0
0
Time (ms)
900
Figure 1 Gating neuronal inputs through detailed balance of excitation and inhibition. (a) Schematic of the model by Vogels and Abbott1. Receiver neurons get excitatory input from two groups of sender neurons (red and purple). Senders contact both excitatory (orange) and inhibitory receiver neurons (dark and light blue). (b) Responses as functions of time. The excitatory receiver neurons (orange trace) may either not respond (gates off), follow the signal from sender 1 (gate 1 on) or follow the signal from sender 2 (gate 2 on), depending on the gain of the local inhibitory neurons (blue traces).
receiver group such that, by default, when the senders are highly active, the receiving excitatory neurons remain rather quiet, firing at low rates. This is because the receiving inhibitory neurons act locally to cancel the senders’ excitation. The authors refer to this local equilibrium as ‘detailed balance’. Now, a particular receiver subnetwork can be connected to many sender subnetworks, but as long as each sender’s signal is balanced by a number of local inhibitory neurons, nothing much happens to the excitatory neurons at the receiving end. However, if the detailed balance of a particular incoming signal is disrupted (for example, if the gain of the corresponding local interneurons is decreased) then that signal is gated on and drives the receiver excitatory neurons, which reproduce it quite faithfully (Fig. 1b). What is interesting about this circuit organization is that by selectively altering the local balance of the receiver subnetwork any of the sender signals can be gated on. Thus, the excitatory receiver neurons may respond to any one of a large number of sender inputs without scrambling their signals. There are three aspects of this work that are notable. First, the sender and receiver populations are part of a larger network with realistic dynamics (irregular, asynchronous and low firing) that does not require an external source of noise and the activation of these subnetworks does not disrupt the activity of the rest of the model neurons. Thus, the overall dynamical regime is quite accurate. Second, the idea that gating functions can be implemented by tweaking the balance between excitation and inhibition is an old one, but the authors, without invoking any
nature neuroscience volume 12 | number 4 | april 2009
exotic assumptions, have achieved a specific and reasonable account of how this may happen. In particular, they performed a beautiful analysis of the number of input signals that can be cancelled and gated by a local population of inhibitory neurons. They demonstrate that if a receiver signal is composed of N sender signals (firing rates), there is a fundamental limitation in the accuracy with which a single component signal can be extracted that depends on N and is independent of the extraction method. This in itself is a remarkable result about neural coding. What it means for the detailed-balance gating mechanism is that as long as no more than ~5 sender inputs are active simultaneously, 200 local inhibitory neurons are capable of gating on, with high accuracy, any one of 600 possible sender inputs (see Fig. 7 in Vogels and Abbott1). Third, the model makes a specific and testable prediction: at least some inhibitory neurons should be strongly active when their nearby excitatory targets are quiet (gating off). The converse pattern should also be seen, albeit not always; when those excitatory targets are strongly active (gating on), the inhibitory responses should decrease sometimes, but not always, because although the excitatory neurons should be able to react to any sender, each inhibitory neuron should correlate only with some of them. Relatively few studies14,15 have investigated mechanisms for switching information from one cortical circuit to another. Notably, however, one of them14 works in the opposite way, modulating the gain of the sender neurons. In contrast, in Vogels and Abbott’s model1, routes already exist from the senders
373
© 2009 Nature America, Inc. All rights reserved.
news and views to the receiver, but they are shut down by default; the relevant connection is activated locally, at the receiver end, without the need to alter the sender’s activity. It is possible that real circuits exploit both strategies, depending on the current behavior, the circuit’s function or its connectivity constraints. These mechanisms could be tested by recording from suspected sender and receiver neurons in tasks in which the relationship between multiple sensory signals and motor effectors can change, as in our keyboard example. In addition, the mechanism put forth by Vogels and Abbott1 should have a distinct microanatomical substrate reflecting the convergence of excitation and inhibition driven by the same input. The model by Vogels and Abbott1 may solve part of the neuronal switchboard puzzle, but many questions still remain. How is the detailed balanced between excitation and inhibition maintained? How are the connection pathways established and kept
separate? Perhaps the most pressing issue is the nature of the control signal that disrupts the local balance of excitation and inhibition to open each gate. The authors showed that it may work by acting on very few interneurons (
