Abstract
Few studies in amyotrophic lateral sclerosis (ALS) focus on the premotor interneurons synapsing onto motoneurons (MNs). We hypothesized inhibitory interneurons contribute to dysfunction, particularly if altered before MN neuropathology. We directly assessed excitability and morphology of ventral lumbar glycinergic interneurons from SOD1G93AGlyT2eGFP (SOD1) and wildtype GlyT2eGFP (WT) mice. SOD1 interneurons were smaller but density was unchanged. Patch clamp revealed dampened excitability in SOD1 interneurons, including depolarized PICs and voltage threshold. Renshaw cells (RCs; confirmed with immunohistochemistry) showed similar dampened excitability. Morphology and electrophysiology were used to create a “random forest” statistical model to predict RCs when histological verification was not possible. Predicted SOD1 RCs were less excitable (consistent with experimental results); predicted SOD1 non-RCs were more excitable. In summary, inhibitory interneurons show very early perturbations poised to impact MNs, modify motor output, and provide early biomarkers of ALS. Therapeutics like riluzole that universally reduce CNS excitability could exacerbate this dysfunction.
Introduction
Amyotrophic lateral sclerosis (ALS) is a rapidly evolving adult-onset neurological disease characterized by a progressive loss of corticospinal neurons and motoneurons (MNs). There has been considerable debate in the field over the role of hyperexcitability in neurodegenerative processes in ALS. While drug treatment for ALS based on nonspecific reduction of neuronal excitability (with riluzole, for example) (Bellingham 2011), does result in a modest increase in lifespan (Bensimon et al., 1994), similar treatment paradigms have been disappointing clinically. A more nuanced understanding of the excitability of vulnerable neurons could help in creating a more targeted and effective approach for treatment of ALS. For example, if inhibitory pathways are failing in ALS and neuronal excitability is universally reduced with riluzole, this could further exacerbate inhibitory dysfunction by reducing activity not only in vulnerable MNs but also in inhibitory interneurons presynaptic to MNs. Thus it is important to consider all aspects of neuronal excitability, including intrinsic excitability of vulnerable neurons, synaptic drive, and neuromodulation (Gunes et al., 2020). Intrinsic properties of MNs have been well studied but the same is not true for interneurons that are synaptically connected to them. In fact, no studies thus far have directly assessed activity of spinal premotor interneurons in an ALS model. Despite evidence that ALS patients have disrupted inhibition at spinal levels (Raynor and Shefner, 1994; Shefner and Logigian, 1998; Sangari et al., 2016; Howells et al., 2020), much is still unknown concerning the involvement of inhibitory circuitry in ALS. Morphological alterations in inhibitory circuits have been demonstrated in animal models of ALS. These include degeneration of spinal interneurons, fewer neurons expressing markers of inhibitory neurotransmitters including GlyT2 prior to loss of MNs (Martin et al., 2007; Hossaini et al., 2011), and loss of glycinergic boutons onto MNs prior to symptom onset (Chang and Martin, 2009a). Typically, fast motor neurons have greater numbers of inhibitory synaptic contacts, but these are largely lost in SOD1 mice beginning when motor unit atrophy is first observed (Pun et al., 2006; Hegedus et al., 2007; Allodi et al., 2020). Recurrent inhibitory circuits mediated by Renshaw cells (RCs) are impaired before symptom onset by both loss of MN collaterals which provide synaptic drive, and complex changes to RC-MN synaptic structures (Casas et al., 2013; Wootz et al., 2013). A few studies have suggested activity is decreased in inhibitory interneurons by indirect measurements. Quantification of synaptic inputs to MNs has shown that frequency of inhibitory postsynaptic potentials in spinal MNs is decreased embryonically in both SOD1 mouse and zebrafish models (McGown et al., 2013; Branchereau et al., 2019) and glycinergic inputs to MNs decay faster in SOD1 MNs (Medelin et al., 2016). On a larger scale, blocking the activity of V1 inhibitory interneurons during in vivo treadmill running was recently found to mimic the early locomotor deficits in SOD1 mice (Allodi et al., 2020), suggesting that inhibitory interneurons could be inactive or under-active. However, none of these studies directly examined electrophysiological activity in inhibitory interneurons.
We hypothesized that inhibitory spinal interneurons could contribute to the pathogenesis of ALS through a depression of MN inhibition. Decreased activity of inhibitory interneurons could result in synaptically-driven hyperexcitability of MNs, and other long-term changes in network function. In this study, we examined electrical and morphological properties of glycinergic interneurons in the spinal cord of SODG93A GlyT2-eGFP (SOD1) mice compared to GlyT2-eGFP (WT) using whole cell patch clamp and three-dimensional reconstructions, immunohistochemistry for the RC marker calbindin (CBD), and statistical modeling. We show here that significant dysfunction is present in SOD1 glycinergic interneurons, including RCs and other subclasses of inhibitory interneurons. Impairment in glycinergic interneurons should be explored as both a mechanism of vulnerability of MNs and a potential biomarker of early dysfunction of spinal circuits.
Results
Morphology of glycinergic interneurons
Ventrally located glycinergic interneurons in the lumbar enlargement of WT and SOD1 mice postnatal day (P)6-10 were reconstructed from image stacks. There was no change in soma size and density of ventral glycinergic interneurons between WT and SOD1 mice when interneurons throughout the ventral horn were assessed. However, when we narrowed our examination to only those glycinergic interneurons located adjacent to MNs within lamina IX (most likely premotor interneurons), SOD1 interneuron somata were smaller in several measures. Typical ventrolateral glycinergic neurons in the region of the MN pools in lamina IX are shown in the areas delimited with dotted lines in Figure 1A. Complete reconstructions of soma morphology were performed on 3902 GlyT2 interneurons (1918 WT and 1984 SOD1 interneurons) from fixed lumbar spinal cords (image stacks obtained using confocal microscopy) and 682 GlyT2 interneurons from only lamina IX (362 WT and 320 SOD1 interneurons). Reconstructions revealed that glycinergic interneurons in lamina IX were significantly smaller in every measurement of the soma (soma perimeter, largest cross-sectional area, soma surface area, and soma volume) (Fig 1B). Cell counts performed in the ventral lumbar spinal cord of 8 WT mice and 9 SOD1 mice showed no significant differences (Fig 1C). Dendritic morphology was not possible in these images due to the high number of GFP+ neurons and processes but was completed on a smaller subset of neurons that were imaged live after patch clamp electrophysiology with a 2-photon microscope.
Representative images of GFP interneurons in A, WT on the top and SOD1 on the bottom. Lamina IX delimited with dotted lines in images of ventral horns. (B) Morphological analysis showed no differences between glycinergic neurons in the ventral horn (top panels), but within lamina IX (bottom panels), SOD1 GlyT2 interneurons were smaller than WT in soma perimeter, soma largest cross-sectional area, soma surface and soma volume. (C) Density of GlyT2 interneurons were not significantly different between WT and SOD in the ventral horn or in lamina IX. Significance denoted with * indicates p ≤ 0.05. Scale bar in A: 100 μm.
In these experiments, 59 glycinergic neurons (34 WT and 25 SOD1) were recorded using visually guided patch clamp with Texas Red dye in the electrode as shown in Figure 2A. The location of patched neurons was distributed throughout the ventral horn, not limited to lamina IX (Fig 2B). Thus, similar to the larger dataset above, soma sizes of patched SOD1 interneurons were not significantly smaller than WT (Fig 2C). However, dendrites were larger in patched SOD1 GlyT2 interneurons, including greater dendritic surface area and volume (see Table 1 for complete results).
Images of typical interneurons in A, and locations of all patched interneurons in B (WT = blue, SOD1 = red), with dotted line indicating area of lamina IX. (C) In patched interneurons, soma surface area was unchanged, while dendrite surface area was larger in SOD1 interneurons than in WT. Significance denoted with * indicates p ≤ 0.05.
Reconstruction data of GlyT2-GFP+ interneurons in WT and SOD1 animals.
Electrophysiology of glycinergic interneurons
Whole cell patch clamp revealed very different intrinsic properties of ventral glycinergic interneurons. Recording was performed on 59 ventral interneurons from throughout the ventral horn (not limited to lamina IX) in transverse lumbar spinal cord slices from P6-10 mice. SOD1 inhibitory interneurons (n = 25) were found to have diminished intrinsic excitability compared to WT interneurons (n = 34).
Measurements of intrinsic excitability included depolarized onset and peak voltage of persistent inward currents (PICs) and depolarized threshold for action potential firing, as shown in Figure 3. In voltage clamp, slowly depolarizing voltage ramps were used to measure PICs (typical leak subtracted traces shown in Fig3A). Voltage sensitivity is determined by measuring voltage at PIC onset and peak. Both PIC onset and peak were significantly depolarized in SOD1 interneurons, indicating less intrinsic excitability. Amplitude of the PIC was unchanged. Depolarizing current ramps were used to measure the input-output relationship of the neurons in current clamp. Inhibitory interneurons from SOD1 mice were found to have significantly higher threshold voltage than control interneurons (Fig 3E-F). This shift is likely driven by the changes in PIC. The current at firing onset, or I-ON, was not significantly changed (p = 0.15), similar to the current at firing offset, or I-OFF (p = 0.07). After-spike afterhyperpolarization (AHP) was also significantly different in SOD1 interneurons. The duration (at half AHP amplitude) was shorter in SOD1 neurons, as shown in Fig 3H-I. Amplitude of the AHP was unchanged. While changes in the PIC and threshold indicate less excitability in SOD1 interneurons, the shortened AHP suggests an increased ability to fire at higher rates in SOD1 interneurons, a property that is associated with increased excitability. However, no changes in firing rates were detected in SOD1 interneurons. All other properties were found to be similar in SOD1 and WT glycinergic interneurons, including maximum firing rates, sag/rebound currents (IH), action potential parameters and membrane properties (see Table 2 for all electrical properties). See discussion for further interpretation of these findings.
The most prominent difference in SOD1 interneurons was the shift in voltage dependence of PICs. (A) Representative, leak-subtracted current-voltage relationship of PICs from WT (P8) and SOD1 (P6) interneurons; arrows indicate voltage at PIC onset and voltage at PIC peak. (B-D) Mean onset, peak and amplitude of PICs in WT and SOD1 interneurons. (E-F) Threshold voltage (as measured by current ramps as shown) was depolarized in SOD1 interneurons. Starting potential for both interneurons was −65mV. The current at firing onset (I-ON) was not significantly different as shown in G. (H) The AHP in SOD1 interneurons was shorter, as shown by the smaller mean duration at half amplitude in a representative P8 WT and P7 SOD1 interneuron in I. Neither the AHP amplitude nor the AHP decay time constant, tau, were significantly altered in SOD1 glycinergic interneurons as shown in J-K. Vertical scale bars in E: top = 50pA, and bottom = 20mV. Horizontal scale bar in E = 0.5s and all scale bars apply to both left and right panels. Vertical scale bar in H = 10mV (APs were truncated from image), and horizontal scale bar = 50ms. Significance denoted with * indicates p ≤ 0.05.
Electrophysiological properties of GlyT2-GFP+ interneurons in WT and SOD1 animals.
Renshaw cells: electrophysiology and morphology
Glycinergic interneurons in the ventral spinal cord are not a homogenous population and we sought to determine more specifically what subtype of ventral inhibitory interneurons were morphologically and electrophysiologically altered in SOD1 mice. Since we recorded glycinergic interneurons in the region of lamina IX, we expected our population to include RCs. To distinguish them, we performed immunohistochemistry for the RC marker CBD on GlyT2 neurons labeled with dye during recording (see Figure 4A). We also examined density of RCs in lamina IX of the spinal cords, which was unchanged between WT (N = 8 mice) and SOD1 (N = 7 mice; Figure 4B). Similar to the results above for all GlyT2 interneurons, electrophysiological parameters of SOD1 RCs were very different compared to WT (7 WT and 6 SOD interneurons). The voltage dependence of the PIC (both onset and peak) was significantly depolarized in SOD1 RCs, indicating they are less excitable. Threshold voltage was also significantly higher in SOD1 RCs when measured from step protocols, though it was not significant when measured from current ramp protocols (Figure 4C). All RC electrophysiological properties are included in Table 3.
(A) Image of an interneuron which was patched and filled with Texas Red (left panel), and immunoreactive for CBD (middle panel); overlay provided in right panel. The RC is indicated by a white arrow in middle panel. (B) No differences were found in the density of CBD+ cells in lamina IX of WT and SOD1 mice. (C) Several electrophysiological parameters were measured from the cells. PIC onset, PIC peak and voltage threshold showed higher values in RCs from SOD1 animals, compared to WT RCs. There were no significant differences between groups in PIC amplitude, I-on and AHP duration. Significance denoted with * indicates p ≤ 0.05. Scale bar in A: 100 μm.
Electrophysiological properties of Renshaw cells, in both groups, WT and SOD1.
Predictive statistical modeling
Since it was not possible to identify all recorded GlyT2 neurons as RCs or non-RCs using immunohistochemistry, we performed predictive statistical modeling. We created a series of statistical models using random forest and principal components analysis (see methods for more details on creation of the models). All variables, including morphology and electrical properties, were grouped into 8 categories of related variables. For example, all variables pertaining to dendritic morphology were grouped into a single category. Importance of each variable was determined by the decrease in accuracy of the model when variables were removed. The variables that were used by the random forest model are listed in Table 4 in descending order of importance. Variables with large positive values were crucial for the model’s accuracy in predicting RC vs non-RCs, while variables that have negative values proved detrimental to the predictive value of the model. For comparison, the grouping that provided the best prediction of CBD-tested neurons using PCA were action potential features (LOOCV accuracy 70%, AIC =28, p=0.11 and 0.08, PC1 and PC2, respectively). Variables in this grouping that had effect sizes ≥ 0.5 were I-off, AP threshold, AP rise and AP fall. The most accurate random forest model correctly predicted the identity of the CBD-tested interneurons with 79% accuracy; thus we used this model to predict RCs and non-RCs in all recorded GlyT2+ interneurons in which RC identification using immunohistochemistry was not possible. The model predicted that 80% of our 59 patched interneurons were RCs. This percentage is greater than that measured from our immunohistochemistry results: 65% of our patched interneurons (15/23) were positive for CBD. Predicted RCs were then analyzed as a group for differences between WT and SOD interneurons, while predicted non-RCs were analyzed separately. We found that in predicted RCs, there was a depression of excitability very consistent with our experimental data. The predicted RC grouping had significant depolarization of the PIC and significantly depolarized threshold. PIC parameters of the predicted groups are shown in Figure 5. Interestingly, in predicted non-RCs excitability increased. Both the PIC amplitude and the PIC normalized to capacitance were larger in the group of predicted non-RCs from SOD1 mice. Further studies will be needed to investigate more fully the properties of different subtypes of non-RC glycinergic interneurons in this and other ALS models at this and other time points in the lead up to MN dysfunction and degeneration.
A - D Predicted RCs have depolarized PIC onset, similar to overall experimental results. This is accompanied by depolarized threshold (data not shown). E - H Predicted non-RC glycinergic interneurons have larger PIC amplitudes, including larger normalized PIC amplitudes. * Significance denoted with * indicates p ≤ 0.05.
Variables used to create the random forest predictive model. Variables are arranged in descending order of importance for prediction of RCs and non-RCs, based on the mean decrease in accuracy test.
Discussion
This study shows for the first time that RC are depressed in excitability and glycinergic interneurons may be contributing to ALS pathology. Alterations in both morphology and excitability of glycinergic interneurons occurs at a very early stage. RCs are depressed in activity and lamina IX interneurons are smaller in SOD1 mice. Based on our statistical modeling, we predict the remaining non-RC inhibitory interneurons have increased excitability, including larger PICs. We speculate that these shifts in interneuron excitability could alter the excitability of postsynaptic MNs, modify motor output, and provide an early biomarker of ALS.
Intrinsic and synaptic hyperexcitability
A core issue this study brings to light is potential long-term disruption to the balance of excitatory and inhibitory synaptic transmission in ALS. Spinal cord and cortex both show imbalance between excitatory and inhibitory synaptic transmission in SODG93A mice (Avossa et al., 2006; Saba et al., 2016). Clinically, ALS patients show evidence of altered synaptic activity as well, including a reduction in inhibition. A reduction of glycinergic receptor binding in the anterior gray matter of the spinal cord (Hayashi et al., 1981; Whitehouse et al., 1983) has been demonstrated along with abnormal glycine and gamma amino butyric acid (GABA) levels in blood serum (Malessa et al., 1991; Niebroj-Dobosz and Janik, 1999). Electrophysiology also suggests disruption in spinal inhibitory circuits in ALS patients (Raynor and Shefner, 1994; Shefner and Logigian, 1998; Sangari et al., 2016). An interesting question is whether the neurons are altered in excitability only within the motor circuit (i.e. corticospinal neurons and the synaptically-connected spinal inhibitory and cholinergic interneurons, and MNs), or if disturbances are more widespread throughout the nervous system. In the spinal cord, interneurons that serve as conduits between vulnerable corticospinal and spinal MNs include cholinergic interneurons, Ia inhibitory interneurons and RCs. There is abundant evidence of disruption in these cholinergic, GABAergic and glycinergic synapses from animal models (Martin et al., 2007; Chang and Martin, 2009b; Pullen and Athanasiou, 2009; Hossaini et al., 2011; Herron and Miles, 2012; Casas et al., 2013; Wootz et al., 2013; Saxena et al., 2013; Milan et al., 2015; Dukkipati et al., 2016; Medelin et al., 2016). Similarly, in the cortex, neurons that are not themselves vulnerable to neurodegeneration in ALS also exhibit altered excitability and morphology in a parallel time course to the vulnerable corticospinal neurons (Clark et al., 2017, 2018; Kim et al., 2017). The changes in excitability of other neuronal populations must contribute to the imbalance of both excitatory and inhibitory neurotransmission that has been clearly demonstrated in ALS patients and animal models.
A case for hyperexcitability and glutamate excitotoxicity involvement in ALS pathogenesis has been more controversial. Altered intrinsic excitability has been consistently reported in both corticospinal projection neurons and spinal MNs, but these perturbations appear to fluctuate based on age/disease progression. Corticospinal projection neurons have increased firing and other signs of hyperexcitability during the first week of postnatal development (Pieri et al., 2009; Saba et al., 2016, 2019; Kim et al., 2017), but excitability is normal in young and adult pre-symptomatic ages (Kim et al., 2017; Saba et al., 2019), and may become hypoexcitable at the age of symptom onset (Saba et al., 2019). Similarly, spinal MNs show intrinsic hyperexcitability at very early (embryonic) stages (Kuo et al., 2004, 2005; Martin et al., 2013). At a postnatal age similar to the inhibitory interneurons in this study, spinal MNs do not have an altered threshold, rheobase or frequency-current relationship, though other abnormalities are present in PICs, action potential duration, AHP and dendritic Ca2+ entry (Pambo-Pambo et al., 2009; Quinlan et al., 2011, 2015; Leroy et al., 2014). Thus, at the time point studied here, inhibitory interneurons are more disrupted in excitability than the vulnerable spinal MNs. At symptom onset, MNs may finally succumb to hypo excitability, though not all studies agree on this point (Delestrée et al., 2014; Martinez-Silva et al., 2018; Jensen et al., 2020). Notably, MNs derived from ALS patients’ induced pluripotent stem cells also show initial hyperexcitability followed by hypoexcitability (Wainger et al., 2014; Devlin et al., 2015), and increasing excitability of MNs enhanced neuroprotection rather than neurodegeneration (Saxena et al., 2013). A unifying theme throughout the findings is that vulnerable MNs show excessive homeostatic gain in response to perturbation (Kuo et al., 2020), and not only excitability but other cellular properties wildly fluctuate over the lifespan of the animal (Irvin et al., 2015). In light of that, the present findings that inhibitory interneurons are less excitable should be further explored to fully characterize their contribution to neurodegenerative processes.
Changes in neuron morphology
In ALS models, altered morphology in spinal MNs occurs so early that it could be viewed as a deficit in normal development. Embryonically, SOD1 MNs have shorter projections but this is reversed during postnatal development. At 1-2 weeks of age, MNs begin to show more dendritic branching and larger soma sizes (Amendola and Durand, 2008; Filipchuk and Durand, 2012; Martin et al., 2013). Spinal MNs remain larger through adulthood (Dukkipati et al., 2018). In contrast to the spinal MNs at this age, we found that inhibitory interneurons in lamina IX were smaller and patched interneurons had expanded dendritic surface area and volume. This is reminiscent of changes in corticospinal neurons: presymptomatic corticospinal neurons also have increased arborization but with smaller soma diameters (Ozdinler et al., 2011; Saba et al., 2016). In both sporadic and familial ALS patients, postmortem tissue shows degeneration of apical dendrites in corticospinal neurons and smaller soma size (Genç et al., 2017). Changes in neuronal size over the life span could greatly increase the impact of other stressors (e.g., proteostasis, metabolic deficits, intracellular transport) on vulnerable neuron populations.
After spike afterhyperpolarization
The AHP can contribute to a neuron’s firing rate such that a neuron with a large, long-lasting AHP will typically fire at a slower rate (typical of slow/type I MNs) than a neuron with small, fast-decaying AHP (typical of fast/type II MNs). In ALS patients, AHP is shortened in MNs of patients with little force deficits and later elongated in patients with large force deficits (Piotrkiewicz and Hausmanowa-Petrusewicz, 2011), which could reflect early changes in physiology of vulnerable fast MNs and later, the remaining, less-vulnerable population of slow MNs. Similarly, SOD1 mouse MNs show shorter AHPs early, and longer AHPs around symptom onset (Quinlan et al., 2011; Jensen et al., 2020). Here we show that AHP is shorter in duration in SOD1 interneurons very early, however we did not find changes in any measure of the minimum or maximum firing rates in GlyT2 interneurons or RCs, so the physiological importance of this finding is not clear.
Inhibitory interneurons: Subtype specific effects
The interneurons in this study were mostly composed of RCs, based on both our positivity rate with immunohistochemistry for CBD (65% positive) and our statistical modeling (80% RC / 20% non-RC). Interneurons located in lamina IX likely reflect a mix of RCs and Ia inhibitory interneurons. Based on highly significant results in our RC group and the prediction of higher excitability in non-RC inhibitory interneurons, further studies are warranted to examine how inhibitory interneurons may be differently modulated in a subtype-specific manner in animal models of ALS.
Functional implications for impaired inhibitory circuits
Outward signs of improperly functioning inhibitory interneurons may be apparent from subtle changes in motor patterning and could potentially be used as a biomarker of early ALS. Locomotor disturbances have been demonstrated in vivo in presymptomatic SOD1 mice during walking/running (Vinsant et al., 2013a, 2013b; Akay, 2014; Quinlan et al., 2017; Allodi et al., 2020). In the embryonic and neonatal spinal cord in vitro, increased duration of depolarizing events in MNs and a slower period of locomotor-related bursting has been described (Medelin et al., 2016; Branchereau et al., 2019), which indicate that even during development the functioning of locomotor circuits is already impaired. In presymptomatic adult mice capable of running on a treadmill, subtle locomotor differences are observed including advanced intermuscular phasing and a slower speed (Quinlan et al., 2017; Allodi et al., 2020) which indicate that before any large scale neurodegeneration occurs, there are changes in patterns of activity in MNs and interneurons that could be exploited for use as a biomarker of ALS. At initial symptom onset, the first failure of the tibialis anterior manifests as abnormality in the swing phase of the ankle (Akay, 2014). Interestingly, silencing different interneuron populations (including inhibitory interneurons and cholinergic interneurons) reveals the contribution of each population: silencing cholinergic interneurons increases locomotor deficits in SOD1 mice (Landoni et al., 2019), while silencing inhibitory interneurons has no effect suggesting these interneurons are already impaired (Allodi et al., 2020).
Clinical-translational implications
Evidence of MN hyperexcitability in ALS patients is present alongside signs of dysfunction in inhibition (Rothstein et al., 1992; Plaitakis and Constantakakis, 1993; Raynor and Shefner, 1994; Shefner and Logigian, 1998; Andreadou et al., 2008; Bae et al., 2013). Perhaps reducing excitability (with riluzole, for example) only results in a modest increase in lifespan (Bensimon et al., 1994) because it could be helpful to MNs while exacerbating inhibitory dysfunction by further depressing excitability of inhibitory interneurons like RCs. Thus, a more targeted approach to reduce MN excitability could be beneficial early in disease pathology. Stimulation of inhibitory nerves or certain protocols of transcranial stimulation could be explored, along with augmentation of inhibitory neurotransmission.
Materials and Methods
Ethics Statement
Experiments were performed in accordance with the United States National Institutes of Health Guide for Care and Use of Laboratory Animals. Approval of the Northwestern University’s Animal Care and Use Committee was obtained for all experiments performed in this study. All efforts were made to minimize animal suffering and to reduce the number of animals used.
Animals and Tissue harvest
Transgenic B6SJL mice overexpressing the human SOD1G93A gene (strain 002726, Jackson Labs, Bar Harbor, ME, USA) and their wild type littermates were used (nontransgenic for the human SOD1G93A gene). Transgenic animals were identified using standard PCR techniques by Transnetyx (Cordova, TN, USA) and were bred with GlyT2-eGFP mice (Zeilhofer et al., 2005) generating SOD1G93A GlyT2-eGFP mice, here called SOD1. Inhibitory glycinergic interneurons express the Na+ and Cl- coupled glycine transporter 2, or GlyT2. GlyT2-eGFP expression is driven by the GlyT2 promotor. Specifically, GlyT2-eGFP females were bred with SOD1G93A GlyT2-eGFP males, and progeny were used for experiments without knowing if they were SOD1G93A GlyT2-eGFP or GlyT2-eGFP. Genotyping was performed post-experimentally. For the following studies juvenile mouse pups were used between postnatal day (P) 6 - 10. Mice were deeply anesthetized with isoflurane (Henry Schein Animal Health, Dublin, OH, USA), decapitated and eviscerated. The lumbar spinal cord from L1 - L6 was removed and embedded in 2.5% w/v agar (No. A-7002, Sigma-Aldrich, St Louis, MO, USA). The agar block was then superglued with Loctite 401 (Henkel Corporation, Rocky Hill, CN, USA) to a stainless steel slicing chuck and 350 μm transverse slices were made using the Leica 1000 vibratome (Leica Microsystems, Buffalo Grove, IL, USA) as described previously (Quinlan et al., 2011). During both spinal cord isolation and slicing, the spinal cord was immersed in 1–4°C high osmolarity dissecting solution containing (mM): sucrose 234.0, KCl 2.5, CaCl2 · 2H2O 0.1, MgSO4 · 7H2O 4.0, HEPES 15.0, glucose 11.0, and Na2PO4 1.0. The pH was adjusted to 7.35 when bubbled with 95% O2/5% CO2 using 1 M KOH (Fluka Analytical, Sigma-Aldrich). After cutting, the slices were incubated for >1 h at 30°C in incubating solution containing (mM): NaCl 126.0, KCl 2.5, CaCl2 · 2H2O 2.0, MgCl2 · 6H2O 2.0, NaHCO3 26.0, glucose 10.0, pH 7.4 when bubbled with 95% O2/5% CO2 (all reagents for solutions were purchased from Sigma-Aldrich).
Electrophysiology
Whole cell patch clamp was performed on MNs from the lumbar segments using 2–4 MΩ glass electrodes pulled from glass capillary tubes (Item #TW150F-4, World Precision Instruments, Sarasota, FL, USA) with a Flaming-Brown P-97 (Sutter Instrument Company, Novato, CA, USA). Electrodes were positioned using a Sutter Instrument MP-285 motorized micromanipulator (Sutter Instrument Company). Whole-cell patch clamp measurements were performed at room temperature using the Multiclamp700B amplifier (Molecular Devices, Burlingame, CA, USA) and Winfluor software (University of Strathclyde, Glasgow, Scotland). Briefly, slices were perfused with a modified Ringer’s solution containing (in mM): 111 NaCl, 3.09 KCl, 25.0 NaHCO3, 1.10 KH2PO4, 1.26 MgSO4, 2.52 CaCl2, and 11.1 glucose. The solution was oxygenated with 95% O2/5% CO2 and the perfusion rate was 2.5 – 3.0 ml/min. Patch electrodes contained (in mM) 138 K-gluconate, 10 HEPES, 5 ATP-Mg, 0.3 GTP-Li and Texas Red dextran (150 μM, 3000 MW, from Invitrogen, Life Technologies, Grand Island, NY, USA). In voltage-clamp mode, fast and slow capacitance transients, as well as whole-cell capacitance, were compensated using the automatic capacitance compensation on the Multiclamp. Holding potential was set at −90 mV, and neurons were subjected to slow, depolarizing voltage ramps of 22.5 mV s-1, bringing the cell to 0 mV in 4 s, and then back to the holding potential in the following 4 s. In current clamp, neurons were subjected to depolarizing current ramps for testing I-on (the current level at firing onset), I-off (the current level at cessation of firing), and the slope of the frequency–current relationship. Negative current was often necessary to prevent action potential (AP) firing.
Neuron Selection
Glycinergic interneurons were visually selected for recording based on 1) expression of GFP, and 2) location near the MN pools. Neurons that did not repetitively fire or did not maintain a resting membrane potential below −35 mV were excluded from electrophysiological analysis. Neurons that were not clearly stained with Texas Red after patch clamp and/or could not be imaged clearly were not included in anatomical analysis.
Electrophysiological analysis
Persistent inward current (PIC) parameters were measured after subtraction of leak current. Onset was defined as the voltage at which the current began to deviate from the horizontal, leak-subtracted trace. PIC peak was the voltage at which the PIC reached peak amplitude. The linear portion of leak current (usually between −90 and −75 mV) was used to calculate whole cell input resistance. The first action potential in the train evoked by a depolarizing current ramp in current clamp mode was used to measure all parameters relating to action potentials. Threshold voltage was defined as the voltage at which the slope exceeded 10 V/s. Threshold for action potential firing was tested in two ways. The first was to use the voltage at firing onset from the current ramps up to 130 pA/s (in current clamp). The second was the voltage at which a single action potential could be evoked with a current step. Action potential overshoot is the voltage past 0 mV the first spike reaches. Duration of the action potential is measured at half of action potential height (height is defined as overshoot – threshold voltage). Rates of rise and fall are defined as the peak and the trough of the first derivative of the action potential profile.
Immunohistochemistry
Immediately after electrophysiological recording, slices were fixed in (4%) paraformaldehyde for 24 h. As described by Kupferschmidt and colleagues (Kupferschmidt et al., 2015), sections were then incubated in sodium borohydride (NaBH4; 5mg/mL in diH2O) to quench aldehyde-induced catecholamine fluorescence and reduce free aldehyde groups, followed by Sudan Black B solution (0.2% in 70% ethanol). Slices were washed twice for 30 min in room temperature PBS to extinguish the immunohistochemistry reaction. Slices were blocked using 2% normal horse serum in PBST (0.2% Triton-X100 in PBS) for 4 h, and incubated in PBST with rabbit anti-Texas Red antibody (polyclonal IgG, 1:2000, ThermoFisher Scientific®, A-6399); mouse anti-Calbindin-D-28K (monoclonal; 1:3000, Sigma, C9848) for 72 h at 4°C. Following four washes in PBST over a total of 16–24 h, slices were incubated in the appropriate fluorescent dye-conjugated secondary antibodies (goat anti-rabbit IgG, Texas Red-conjugated, and goat anti-mouse IgG, Cyanine5-conjugated, both ThermoFisher Scientific®, 1:500, T-6391, and 1:500, A10524, respectively) for 48 h. Then they were washed again in 0.1M PBS and DAPI treated for 5 min, washed and covered using Fluoromount® and glass coverslips.
Morphological analysis
Neuron morphology was assessed from two types of images: 1) 2 photon image stacks of live neurons taken immediately after obtaining electrophysiological parameters, and 2) confocal image stacks taken after fixation and tissue processing. Images collected using these two methods were analyzed separately. Measurements from 2-photon images included measurements from both soma and dendrites (soma volume, maximum soma cross-sectional area, soma perimeter in the largest plane, soma surface area, number of primary dendrites, and average dendrite length, dendrite surface area and dendrite volume). Measurements from confocal images only included soma measurements.
Two photon imaging
An Olympus BX-51WIF microscope fitted with an Olympus 40x/0.8NA water-dipping objective lens was used. Two-photon excitation fluorescence microscopy was performed with a galvanometer-based Coherent Chameleon Ultra II laser (Coherent, Santa Clara, CA, USA) tuned to 900 nm. A red Bio-Rad 2100MPD photomultiplier tube (Bio-Rad, Hercules, CA, USA) (570 – 650 nm) was used to collect emission in the red channel. Z-stacks were obtained for each interneuron at 1024 x 1024 pixels (308 x 308 μm) resolution and roughly 100 μm in depth (step size 1 μm). From these Z-stacks, Texas Red® -filled neurons were three-dimensionally reconstructed using Neurolucida software (MBF Bioscience, Williston, VT, USA).
Confocal imaging
Sections were observed with a confocal laser microscope (LSM510, Zeiss, Oberkochen, Darmstadt, Germany) with argon (488 nm) and helium-neon (543 nm) lasers and a 20X objective. Each confocal image obtained in triple staining experiments was split into three channels. Co-localization analysis of CBD, Texas Red and GlyT2 positive cells included visual inspection of labeling, size, and shape of the cell through the stack. From these Z-stacks, GlyT2 positive interneurons were three-dimensionally reconstructed using Neurolucida software (MBF Bioscience, Williston, VT, USA). CBD positive cells were counted using ImageJ (National Institutes of Health, Bethesda, MD).
Statistical analysis
The assumptions of homogeneity of variances and the normality of the distribution of values for each measured characteristic were evaluated with Levene’s and Shapiro-Wilk tests, respectively. Group comparisons (WT vs SOD1) were performed using ANOVA on characteristics that satisfied both conditions and using the Kruskal-Wallis test for the rest of the characteristics. Results are presented as means ± SEM. Significance level was set at p≤ 0.05.
Statistical modeling: Random Forest and Principal Component Analysis
As it was not possible to confirm or rule out RC identity in the majority of our neurons using immunohistochemistry, we created statistical models based on those we had tested to predict RC identity among remaining untested neurons. To do this, we employed a random forest method of statistical learning that generates a large number of decision trees, each constructed using a different, randomly selected (with replacement) subset of the training data set. Thus, all variables in the dataset were first split into 8 subgroups of related variables. Groups included membrane properties, action potential measurements, PIC properties, firing frequency measurements, AHP parameters, IH characteristics, soma measurements and dendritic properties. For example, all variables pertaining to PIC characteristics (PIC onset, peak voltage, PIC amplitude and normalized amplitude) were grouped. Next, a random forest model (Breiman, 2001) was created for each subgroup of variables varying the number of measurements to include in each tree. This method determined the number of variables that yielded the most accurate model. The reason this was performed (instead of including all measured parameters) is that models with larger numbers of variables tend to overfit the training dataset and fail to classify accurately unseen/test cases. Therefore, it was essential to distill down our parameters to only include the most important variables for classifying neurons as RC or non-RC. The decision trees are then used to classify an observation by selecting the most common label among multiple trees. Random forest allows varying two parameters: the number of variables to consider when building each tree and the number of generated trees. One can choose a set of these two parameters that yields the most optimal model. The random forest model with the lowest error was then used to evaluate the most important variables in each subgroup for distinguishing between RCs and non-RCs. Here we used Mean Decrease in Accuracy Index to evaluate the importance of the various measured characteristics of the neurons in predicting RCs vs non-RCs. After the important variables in each subgroup were identified, a dataset was created with only these variables and the above process was repeated using these important variables. Model accuracy was evaluated by using a leave-one-out cross validation (LOOCV). The accuracy for the best random forest model created from the previously determined important variables was 79%. The best random forest model was then used to predict the RC/non-RC classification for untested cells in the dataset. Once the RC/non-RC classification was determined, normality and homogeneity tests were run on each group followed by ANOVA and Kruskal Wallis in order to determine group differences between WT RCs and SOD1 RCs as well as between WT non-RCs and SOD1 non-RCs. Principal components analysis (PCA) (Hotelling, 1933) of each previously created subgroup of variables was also evaluated for predictive value. The scores from the first two principal components in each subgroup were used in a logistic regression model to predict the known CBD-tested cells. Using PCA, the most important variables were identified as the same as those identified by the random forest method. The model accuracy was also tested using LOOCV. The best PCA model accuracy was 70%, therefore, the random forest model (with 79% accuracy based on LOOCV) was selected to predict classification of the remaining untested neurons rather than this method.
Acknowledgments
This project was funded by a springboard fellowship from Target ALS to KAQ and a University of Rhode Island Graduate School Tuition Scholarship to LMM. The authors thank Drs Emily Reedich and Claudia Fallini for comments on the initial draft of the manuscript.