Neural architecture of the electrosensory lateral line lobe: adaptations for coincidence detection, a sensory searchlight and frequency-dependent adaptive filtering

The electrosensory lateral line lobe (ELL) is the first and only site of termination of electrosensory afferents in teleost fish. The simple laminar structure of the gymnotiform ELL has led to a thorough analysis of its circuitry and in vivo and in vitro physiology. An earlier review (Carr and Maler, 1986) summarised the morphology of ELL cells and circuits in two species of gymnotiform fish, Eigenmannia viriscens and Apteronotus leptorhynchus (high-frequency wave species; see Bullock, 1969). Since then, these fish have continued to be the focus of numerous behavioral and electrophysiological studies. In this paper, we briefly update the ELL circuitry of these species and add information on the identification of the transmitters, receptors and second messengers of its neurons and afferents. We then focus on the physiology of an in vitro slice preparation of the ELL that has revealed important cellular adaptations for its sensory function.

The ELL consists of four segments: the medial (MS), centromedial (CMS), centrolateral (CLS) and lateral (LS) segments. The MS receives input from ampullary electroreceptors; tuberous electroreceptor afferents trifurcate and project to the CMS, CLS and LS (Heiligenberg and Dye, 1982). There are numerous differences between the three tuberous segments related to their putative role as spatial and temporal filters (Metzner and Juranek, 1997; Shumway, 1989a,b; Turner et al., 1996), but segmental differentiation will not be emphasised in this review. Electrosensory afferents are fasciculated and travel in stereotyped trajectories to terminate topographically in each of the ELL segments (Lannoo et al., 1989). Thus, although the ELL segments are of different sizes, each receives a complete topographic map of electroreceptors; the rostro-caudal body axis is mapped rostral-to-caudal in the ELL, while the dorso-ventral axis is mapped along the medial- to-lateral ELL axis.

All ELL segments have an identical laminar organization consisting of the following layers (see Fig. 1). (1) A deep fiber layer (DFL) consisting of electroreceptor afferents and axons of commissural neurons. (2) A deep neuropil layer (DNL), the site of termination of electrosensory afferents on the basal dendrites of ELL projection cells and interneurons. This layer also contains ovoid cells and spherical cells; the spherical cells are involved in time coding and are not discussed further (see Carr and Maler, 1986). (3) A granular layer (GCL) containing granular interneurons and a recently described deep basilar pyramidal cell (Bastian and Courtright, 1991). (4) The plexiform layer (PlL) containing mainly efferent axons of ELL pyramidal cells. (5) The pyramidal cell layer (PCL). (6) The stratum fibrosum (StF) containing axons of the direct feedback pathway (excitatory and inhibitory). (7) The ventral molecular layer (VML), the site of termination of StF excitatory fibers. (8) The dorsal molecular layer (DML), the site of termination of parallel fibers emanating from granule cells of the overlying cerebellum (eminentia granularis posterioris, EGp); this projection is described as the indirect feedback pathway. Vertical fibers in the DML, putatively emanating from Egp eurydendroid cells, are a second minor component of the indirect feedback pathway (Maler, 1979).

The most important features of ELL circuitry are summarised in Fig. 1. ELL pyramidal cells are of two basic types: basilar and non-basilar pyramidal cells, corresponding to the physiologically defined E and I cells (Saunders and Bastian, 1984). Basilar pyramidal cells have a thick basal dendrite ending in a dense arborization (see Fig. 2) that receives glutamatergic synaptic input from electrosensory afferents (Wang and Maler, 1994); α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid (AMPA) (and perhaps N-methyl-D-aspartate, NMDA) receptors mediate the response. Non-basilar cells lack a basal dendrite (Maler, 1979), but receive disynaptic electrotonic (gap junction) and inhibitory input from two types of ELL interneuron (see below). Both types of pyramidal cell have apical dendrites that ramify in the molecular layer, where they receive electrosensory feedback (VML and DML) and proprioceptive input (DML only, Bastian, 1995; Sas and Maler, 1987). Pyramidal cells can also be classified with respect to their dorso-ventral position within the PCL (Bastian and Courtright, 1991) and this is correlated with their intracellular Ca²⁺ stores (see below).

Electrosensory afferents also terminate on several types of interneuron including type 1 and type 2 granular interneurons (GC1, GC2), ovoid cells and polymorphic cells (they also terminate on multipolar cells, but this cell type has received little attention and will not be discussed further). Both GC1 and GC2 cells have basal dendrites in receipt of electrosensory input and axons terminating on the somata of pyramidal cells; the more numerous GC2 cell type also has an apical dendrite that ramifies in the molecular layer (Maler et al., 1981b; Maler and Mugnaini, 1994). The GC2 cell is γ-aminobutyric acid (GABA)-ergic and its synapses on pyramidal cells involve strictly GABA_A receptors (Berman and Maler, 1998a). The GC1 cell is also believed to be inhibitory, but its transmitter is unknown; its inhibition of pyramidal cells appears to be linked to K⁺ channels (Berman and Maler, 1998a). Earlier in vivo studies (Shumway and Maler, 1989) suggested that the GC2 cell is involved mainly in shaping the temporal response properties of pyramidal cells, while the GC1 cell generates the inhibitory surround of the pyramidal cell’s receptive field. Polymorphic cells are also GABAergic, but their main site of termination appears to be GC1 cells (Maler and Mugnaini, 1994). It has been proposed that they permit electrosensory feedback to regulate spatial frequency filtering of pyramidal cells (Maler, 1989; Shumway and Maler, 1989).

The axons of GABAergic ovoid cells project bilaterally and climb the length of the basal dendrites of basilar pyramidal cells (Bastian et al., 1993; Maler and Mugnaini, 1994), where they form multiple synapses that generate slow inhibitory postsynaptic potentials (IPSPs) (mainly GABA_B receptors: Berman and Maler, 1998a). Although these cells were hypothesised to be involved in the rejection of common-mode electrosensory inputs (Bastian et al., 1993), the role of slow inhibition in this function is unclear (Berman and Maler, 1998a).

ELL pyramidal cells project topographically to the midbrain (torus semicircularis) and the rhombencephalic nucleus praeminentialis dorsalis (Pd). The Pd is involved in two kinds of electrosensory feedback to the ELL: (1) direct feedback projection to the VML and (2) indirect feedback via the EGp projection to the DML.

Glutamatergic Pd stellate cells project via the StF, terminating in the VML on proximal apical dendrites of pyramidal cells and on GABAergic interneurons (Maler et al., 1981b, 1982); both AMPA and NMDA receptors are involved (Berman et al., 1997). The ELL–Pd connections are reciprocal and topographic (Maler et al., 1982). GABAergic Pd bipolar cells also project back to the ELL via the StF, but terminate on pyramidal cell somata where they act predominantly via GABA_B receptors (Berman and Maler, 1998b). This pathway is spatially diffuse (Maler and Mugnaini, 1994).

Several types of Pd cell project to the EGp (Sas and Maler, 1983), but with only a crude rostro-caudal topography. EGp granule cells project glutamatergic parallel fibers (Wang and Maler, 1994) in the medio-lateral axis across the DML of all ELL segments. These fibers terminate, using AMPA and NMDA receptors, on distal apical dendrites of pyramidal cells as well as on various interneurons. In addition, vertical fibers terminate in an orthogonal fashion to the parallel fibers; the vertical fibers probably emanate from EGp eurydendroid cells, are cholinergic and use muscarinic receptors in the ELL molecular layer (Maler, 1979; Maler et al., 1981a; Phan and Maler, 1983).

Recent studies have begun to elucidate the second-messenger cascades within neurons of the ELL. Electroreceptor afferents contain nicotinamide adenine dinucleotide phosphate diaphorase (NADPH-d; Turner and Moroz, 1995), a marker for nitric oxide synthase; the location of the target of nitric oxide (i.e. soluble guanylate cyclase; Vincent, 1994) are not known, nor has the function of this second messenger been explored in the ELL. Both protein kinase A (PKA) and protein kinase C (PKC) are present in the ELL and are greatly enriched in its molecular layer (Maler, 1999a,b); it is likely that these second messengers are localised both in afferent fibers of the VML and DML and in pyramidal cells (see Fig. 1). A Ca²⁺/calmodulin-dependent kinase (CaMK2) β-like subunit is found in ELL pyramidal cells and granular interneurons (Maler and Hincke, 1999). It has been proposed (Maler and Hincke, 1999) that PKA in feedback afferents and a CaMK2β-like subunit in pyramidal cells are together involved in the adaptive suppression of redundant input to the ELL (anti-Hebbian plasticity).

The CaMK2α subunit is present in the excitatory direct feedback pathway only (StF and VML); it has been proposed that this second messenger is important for non-linear facilitation of a sensory searchlight (Wang and Maler, 1998; Maler and Hincke, 1999; see below).

Ryanodine receptors are found in pyramidal cells (soma and dendrites) within the PCL but are lacking in deep basilar pyramidal cells (Zupanc et al., 1992). The inositol 1,4,5-trisphosphate (InsP₃) receptor is also found in the ELL, but is confined to the somata and proximal dendrites of only the most superficial pyramidal cells within the PCL (Berman et al., 1995). This correlates with the morphology and degree of ‘phasicness’ of pyramidal cells (Bastian and Courtright, 1991): superficial pyramidal cells (InsP₃ + ryanodine receptors) have extensive apical dendrites and are highly phasic; intermediate pyramidal cells (ryanodine receptors only) have modest apical dendrites and are phasic-tonic; deep basilar pyramidal cells (which have no intracellular Ca²⁺ stores) have very small apical dendrites and give tonic responses. The exact connection between the Ca²⁺ stores and temporal processing in pyramidal cells is unknown.

On the basis of the response of electroreceptors to moving objects, Bastian (1981) predicted that convergence of primary afferent inputs and detection of coincident spikes by a higher-order neuron would be required to achieve the behavioral sensitivity of the electrosensory system. In apteronotids, the P type tuberous electroreceptor afferents (P units) fire in phase with the electric organ discharge (EOD) cycles (Hopkins, 1976); there is a range of probabilities of firing per EOD cycle that depends on the amplitude of the EOD (Nelson et al., 1997). This arrangement would permit coincidence detection with the resolution of the EOD (approximately 1 ms), which is consistent with the brief duration of electroreceptor-afferent-evoked excitatory postsynaptic potentials (EPSPs) (see below). Recent quantitative analysis has again suggested the need for coincidence detection in the active electrosense of Apteronotus leptorhynchus (Nelson et al., 1997). In recordings of EPSPs in vitro (Berman and Maler, 1998a–c), a striking feature of those generated in the basilar pyramidal cell dendrite arborization is their rapid dynamics compared with EPSPs generated by feedback pathways. Electroreceptor-afferent-evoked EPSPs peak at 1–2 ms and decay within 10 ms. In contrast, EPSPs evoked from StF or parallel fibers peak at 4–6 ms and last more than 10 ms. Such rapid electroreceptor afferent postsynaptic potential (PSP) dynamics may be essential for the conservation of timing information for the detection of coincident electrosensory inputs; this may be similar to the role of fast EPSPs and coincidence detection in the auditory system (for a review, see Trussel, 1997).

The electrosensory afferent EPSPs are generated far from the cell body on the basal arborization (Fig. 2A). Theoretically, the cable properties of the basilar dendrite would temporally smear the EPSP and reduce its amplitude by the time it reached the soma, thus drastically low-pass-filtering the electrosensory signal and removing much of its high-frequency information. Two mechanisms may operate to alleviate this problem to allow fine-grained resolution of timing information in the afferent stream: inhibitory control of EPSP duration and voltage-gated dendritic channels. We have found evidence for the first mechanism (Berman and Maler, 1998a). Tuberous-afferent-evoked EPSPs in basal pyramidal cells are appreciably shortened by GABA_A IPSPs, although they do not affect its amplitude; these IPSPs are due to the type 2 granular interneuron (Maler, 1979; Maler and Mugnaini, 1994) and prevent temporal smearing of the phase-locked P cell input.

Not only will the IPSP prevent the effects of a single presynaptic spike from lingering too long, but it will also depress the effects of a presynaptic spike that occurs within the IPSP window. For instance, if a presynaptic spike arrived 2–8 EOD cycles after the EPSP peak in Fig. 2B, its effect would be inhibited by the IPSP. The IPSP therefore acts as a high-pass filter or coincidence detector; spikes arriving on the same or the next EOD cycle will summate, whereas a spike arriving later will be inhibited. A stylised impression of this mechanism of coincidence detection is shown in Fig. 2. Without inhibition, the spike streams from two primary afferents summate and drive the cell (BP output) at a rate proportional to the rate of the incoming signal. With inhibition, only coincident spikes (within 1–2 EOD cycles) summate; IPSP inhibition and the faster EPSP relaxation prevent non-coincident spikes from summating. Interestingly, GABA-mediated inhibition has already been shown to enhance coincidence detection in this way in the auditory brainstem of the chick (Funabiki et al., 1998). Note that the cut-off frequency will be determined by the dynamics of the IPSP of the granule cell, which may itself be controlled (perhaps by GABAergic polymorphic cells, Maler and Mugnaini, 1994, see above).

In the ELL, this coincidence detection mechanism may be enhanced by the presence of active Na⁺ channel activity on the dendrites that would amplify the larger EPSPs non-linearly and boost their rise time (Haag and Borst, 1996; Magee and Johnston, 1995). Voltage-dependent channels have been shown to boost the high-frequency component of synaptic inputs in the visual system of the fly (Haag and Borst, 1996) and in the electrosensory system of fish (midbrain: Fortune and Rose, 1997a,b). There is morphological evidence for Na⁺ channels on basal dendrites (Turner et al., 1994), suggesting the possibility of active boosting of high-frequency components of the EPSP, although direct evidence for this mechanism is not yet available.

As first suggested by Bratton and Bastian (1990), the direct feedback pathway to the ELL has many of the features initially proposed by Crick (1984) to underlie a cortico-thalamic sensory ‘searchlight’: (1) a subset of ELL–Pd connections is reciprocal, topographic and excitatory, thus forming a positive feedback loop (Sas and Maler, 1983, 1987); (2) the Pd stellate cells that project to the VML have receptive fields larger than those of ELL pyramidal cells, with a high gain appropriate to this role (Bratton and Bastian, 1990); (3) the inhibitory direct feedback projection emanating from Pd bipolar cells is topographically diffuse (Maler and Mugnaini, 1994), again consistent with Crick’s view of the role of the thalamic reticular nucleus. A searchlight mechanism may also require a non-linear element to amplify signals with respect to noise (Crick, 1984). We present below recent evidence for four types of non-linearity in the direct feedback pathway to the ELL: (1) voltage-dependent EPSPs; (2) dendritic spike bursts; (3) voltage-dependent inhibition (see Fig. 3); and (4) frequency-dependent synaptic facilitation (see Fig. 4).

Stimulation of the StF in an ELL slice demonstrated that VML input to pyramidal cells has both AMPA and NMDA receptor components. The NMDA receptor contribution has a relatively fast onset so that it contributes to the peak of the EPSP. In comparison with primary-afferent-evoked EPSPs (see above), the StF-evoked EPSP has a slow time course (peak approximately 6 ms, duration more than 10 ms; Berman and Maler, 1998b; Berman et al., 1997). The StF-evoked EPSP is highly voltage-dependent both because of its NMDA component (Fig. 3C) and because of the activation of a prominent somatic persistent Na⁺ current (Fig. 3D, I_NaP; Berman et al., 1997; Mathieson and Maler, 1988; Turner et al., 1994). This led Berman et al. (1997) to propose that temporal coincidence of electrosensory and direct excitatory feedback input to pyramidal cells would cause a voltage-dependent increase in feedback EPSP amplitude and thus non-linearly augment the response to the electrosensory input (Fig. 3E).

We have previously shown that Na⁺ channels are found on the proximal apical dendrites of pyramidal cells and can support active propagation of action potentials (Fig. 3A; Turner et al., 1994). Dendritic spikes are passively propagated back to the soma, where they produce depolarising afterpotentials (DAPs); depending on the activity of K⁺ channels, DAPs can trigger action potentials resulting in high-frequency spike bursts (Turner et al., 1994). Gabbiani et al. (1996; see also Metzner et al., 1998) have suggested that these spike bursts are a thresholding mechanism that allows pyramidal cells to detect temporal features of electrosensory input (Fig. 3E). The voltage-dependence of the direct feedback EPSPs and the dendritic spike burst mechanism probably form a closely linked cascade, but a better understanding will require more rigorous computational and experimental analysis.

The direct feedback inhibitory input acts primarily via GABA_B receptors (IPSP reversal potential −90 mV, duration more than 800 ms; Berman and Maler, 1998c). This inhibition is also highly voltage- and time-dependent: it is far more effective at suppressing the spike response to weak than to strong depolarisations. The underlying mechanism appears to be that the large hyperpolarisation produced by stimulating Pd bipolar cell axons in the StF removes inactivation from an I_A-like K⁺ channel; the subsequent activation of this K⁺ channel by depolarisation (EPSPs) slows the depolarisation, preventing or delaying spiking (conditional inhibition, Fig. 3B). Strong depolarisations rapidly inactivate the I_A channel and permit spiking; steady-state depolarisation of the pyramidal cell also causes inactivation of I_A. This non-linearity will act in concert with those described above to filter out weak electrosensory inputs but allow strong inputs to be detected with high gain.

The direct excitatory feedback pathway is characterised by prominent post-tetanic potentiation (PTP; Bastian, 1996b; Wang and Maler, 1997, 1998) that can increase the evoked EPSP by more than 200 %. The VML is highly enriched in CaMK2α, and this kinase is essential for PTP at these synapses (but not in the DML; Wang and Maler, 1998). The ability of CaMK2α to enhance its activity by autophosphorylation (De Konnick and Schulman, 1998; Hanson et al., 1994; Hanson and Schulman, 1992) causes this kinase to show a non-linear dependence on frequency: a switch-like behavior (Hanson et al., 1994; Miller and Kennedy, 1986). We have shown that PTP of StF fibers is dependent on the frequency of stimulation (PTP occurs with 100 Hz, but not with 50 Hz stimulation; Wang and Maler, 1997), and this led to the hypothesis (Wang and Maler, 1998) that CaMK2α was responsible for a frequency-dependent switch that turned on potentiation of the direct feedback pathway. This switch might operate during the scanning movements used by these fish during foraging (Fig. 4): if the first pass across the food object (Fig. 4, scan 1) caused strong activation of StF fibers (>100 Hz), they would potentiate and therefore optimise the detection of the object on the second scan (Fig. 4, scan 2) via the first two non-linear mechanisms described above.

There has been a resurgence of interest in cortico-thalamic pathways, and several recent papers have proposed that this feedback also subserves stimulus-specific enhancement (Ergenzinger et al., 1998; Rauschecker, 1998; Sillito et al., 1994; Yan and Suga, 1996, 1998) in a manner similar to Crick’s original searchlight formulation. It will be interesting to compare the underlying cellular mechanisms of the putative sensory searchlight in these very different neural systems.

The crude topography and diffuse projection characteristics of the indirect feedback pathway (Sas and Maler, 1987), together with the poor response of Pd multipolar cells to moving electrosensory stimuli (they respond tonically to direct current EOD changes; Bastian and Bratton, 1990), have suggested an alternative role for this pathway compared with the direct projection’s putative searchlight function. Bastian (1986a,b) and Shumway and Maler (1989) found that the pathway is likely to mediate an adaptive gain control phenomenon whereby ELL pyramidal cells respond with a change in firing rate that is relatively insensitive to global changes in the EOD amplitude. Removal of the EGp input to the ELL (Bastian, 1986a,b) or antagonism of GABA_A receptors in the ELL (Shumway and Maler, 1989) disrupted the response constancy, suggesting that control of molecular layer inhibition by the EGp input was responsible for gain control.

More recently, it has been shown that the PF pyramidal cell dendritic input displays anti-Hebbian plasticity; this provides an adaptive mechanism whereby pyramidal cell responses remain within an optimum range despite naturally recurring background changes in EOD amplitude (Bastian, 1996a,b, 1998).

The PF pyramidal cell EPSP consists of an AMPA and an NMDA receptor component. While both AMPA and NMDA components contribute to the peak of the EPSP (peaks at 4–6 ms), the NMDA component is longer-lasting and contributes to the slow decay of the EPSP (see Berman and Maler, 1998c; N. J. Berman and L. Maler, unpublished observations). There is also a substantial somatic/proximal dendrite I_NaP contribution (approximately 50 %) to the EPSP. These inward currents are closely controlled by PF-activated inhibition.

The PF disynaptically inhibits pyramidal cells via stellate and vml cells. GABAergic IPSPs activated by the indirect feedback pathway differ in their temporal dynamics according to their target site (Fig. 5A,B). Selective application of GABA_A antagonists (see Berman and Maler, 1998c) revealed that PF-recruited inhibition of distal dendrites (i.e. from nearby stellate cells) controlled mainly the peak amplitude of PF EPSPs (Fig. 5A–C), whereas inhibition that terminated on or near the soma (i.e. from vml cells) controlled mainly the slow decay phase of the EPSP. When PFs are stimulated at high rates (100 Hz activation of the PFs; Berman and Maler, 1998c), stellate cells modulate the amplitude of individual EPSPs, whereas vml cells modulate the low-frequency component of the inputs (the summating depolarising wave). We hypothesise that these timing differences determine the frequency-dependent gain control of these inhibitory pathways (Fig. 5C); stellate cells act as low-pass filters (control the gain of high-frequency inputs), while vml cells act as high-pass filters (control the gain of low-frequency inputs). Interestingly, these timing differences are the opposite of those seen in cortical cells, where distal inhibition lasts longer than proximal inhibition (Pearce, 1993). In ELL pyramidal cells, the longer proximal inhibition may be tuned to control slow inward currents activated by StF fibers, i.e. NMDA EPSPs and I_NaP currents, or slower summed EPSPs from distal inputs. Not only will descending feedback exert complex controls on the tuning properties of pyramidal cells via these interneurons, but plasticity in the descending input to molecular layer interneurons (Bastian, 1998) may further influence the dynamics of feedback sensory filtering properties.

Stimulation of the PF pathway in vitro evoked inhibition even with very low stimulus intensities. The inhibition was revealed by an increase in EPSP amplitude following GABA_A antagonism. This means that stellate cells can be driven past their firing threshold by the same intensity of stimulus that evokes a barely discernible EPSP in the pyramidal cell. While stellate cells may have a steeper activation function than pyramidal cells (as suggested by Maler and Mugnaini, 1994), another explanation based on morphological evidence considers the dendritic span of stellate and pyramidal cells. Stellate cells have small compact dendritic trees, whereas pyramidal cells have large apical dendritic trees that span the full extent of the molecular layer. The artificial nature of in vitro stimulation would tend to activate a beam of parallel fibers in the vicinity of the stimulating electrode. Any stellate cell that was encompassed by this beam would be strongly influenced by its activity. Conversely, a pyramidal cell in the field of termination would have only a fraction of its full dendritic tree sampling the beam’s activity. Therefore, on purely morphological grounds, we would expect stellate cells to have a lower threshold of activation than pyramidal cells.

Although there is no evidence for beams of fiber activity in vivo, the hypothesis may be relevant to the studies of Bastian using natural stimuli. Bastian (1986a) found that removal of PF input (EGp lesions or anesthetic block) increased stimulus-driven responses, suggesting that inhibitory cells are strongly driven under natural conditions. However, he also found that the spontaneous activity of pyramidal cells decreased significantly after EGp lesions. These apparently contradictory results could be reconciled if stimulus-driven activity were spatially coherent (i.e. beams) but spontaneous activity were spatially diffuse. In the latter case, stellate cells would be largely unaffected by the experimental removal of this diffuse fiber activity, whereas the pyramidal cell apical dendritic tree, which integrates PF activity over a large area, would experience a substantial decrease in excitatory input (Fig. 5). This would lead to a net excitatory effect during non-stimulus-driven conditions, and a net inhibitory effect during stimulus-driven conditions.

The intensive study of an in vitro preparation of the ELL has provided some cellular correlates of certain functions of this structure already suggested by in vivo experiments, notably the adaptations of synaptic transmission of electrosensory afferents to coincidence detection and temporal processing. These new discoveries have, however, also demonstrated unexpected subtleties in the cellular physiology of ELL neurons, the consequences of which will require both computational analyses and more refined in vivo studies. The ELL is the first and only target of electroreceptors but its major input is from massive feedback pathways. Further, these pathways are associated with multifunctional serine/threonine kinases and are capable of several forms of plasticity on time scales ranging from milliseconds to hours. Clearly, the computations of the ELL cannot be envisaged as static but rather as adapting so as to optimize, in real time, the extraction of electrosensory information from a continuously changing environment. Although the plasticity of ELL function will make it far more difficult to study, it is also likely to lead to insights into sensory processing that may be readily generalized to other sensory systems.

N.J.B. and L.M. were supported by the Canadian Medical Research Council.