Identification of Mouse Claustral Neuron Types Based on Their Intrinsic Electrical Properties

Abstract Although its dense connections with other brain areas suggests that the claustrum is involved in higher-order brain functions, little is known about the properties of claustrum neurons. Using whole-cell patch clamp recordings in acute brain slices of mice, we characterized the intrinsic electrical properties of more than 300 claustral neurons and used unsupervised clustering of these properties to define distinct cell types. Differences in intrinsic properties permitted separation of interneurons (INs) from projection neurons (PNs). Five subtypes of PNs could be further identified by differences in their adaptation of action potential (AP) frequency and amplitude, as well as their AP firing variability. Injection of retrogradely transported fluorescent beads revealed that PN subtypes differed in their projection targets: one projected solely to subcortical areas while three out of the remaining four targeted cortical areas. INs expressing parvalbumin (PV), somatostatin (SST), or vasoactive intestinal peptide (VIP) formed a heterogenous group. PV-INs were readily distinguishable from VIP-INs and SST-INs, while the latter two were clustered together. To distinguish IN subtypes, an artificial neural network was trained to distinguish the properties of PV-INs, SST-INs, and VIP-INs, as independently identified through their expression of marker proteins. A user-friendly, machine-learning tool that uses intrinsic electrical properties to distinguish these eight different types of claustral cells was developed to facilitate implementation of our classification scheme. Systematic classification of claustrum neurons lays the foundation for future determinations of claustrum circuit function, which will advance our understanding of the role of the claustrum in brain function.

While the claustrum may be involved in higher cognitive functions, it remains entirely unclear how input is processed by different types of claustral neurons and their local circuits. To better understand claustral signal processing, it is fundamentally important to identify the types of neurons that exist within the claustrum. Previous work has described a few different types of claustral neurons in mice and rats (Shibuya and Yamamoto, 1998;Kim et al., 2016;Chia et al., 2017; but did not comprehensively characterize claustrum neurons, because only a limited number of neurons and a limited set of characteristics were considered in each case. In particular, analysis of claustrum interneurons (INs) thus far has been limited to INs expressing the calcium-binding protein, parvalbumin (PV), and it is likely that several other types of INs also are present (Kowia nski et al., 2001(Kowia nski et al., , 2009Real et al., 2003;Kim et al., 2016;. Here, we develop a comprehensive scheme for classifying claustral neurons. For this purpose, we measured the intrinsic electrical properties, anatomical projections and immunohistochemical properties of hundreds of claustrum neurons. We show that claustral projection neurons (PNs) can be clearly distinguished from INs based on their intrinsic electrical properties, and that both groups can be further divided into subtypes. PN were divided into five subtypes, with one of the groups exclusively projecting to the subcortex and several other subtypes projecting to the cortex. Claustral INs were also distinguished according to their expression of PV, somatostatin (SST), or vasoactive intestinal peptide (VIP). This classification scheme provides a firm cellular foundation for future studies that examine higher-order functions of the claustrum. We also developed software, based on artificial neuronal network algorithms, that identifies claustral neuron types based on their intrinsic electrical properties. This software provides an easy-touse tool that will standardize claustral cell classification.

Retrobead labeling of PNs
Retrograde transport of fluorescent latex microspheres (retrobeads) was used to identify PNs (Katz et al., 1984). To inject retrobeads into their brains, mice were anesthetized with isoflurane while kept on a heating blanket. A small craniotomy was performed using a dental drill. Hamilton syringes with a 33 gauge needle were used to inject either red (1:4 dilution in distilled water) or green (undiluted) retrobeads (Lumafluor) using the injection coordinates, angle, and volumes that are stated in Table 1. Injection speed was set to 10-15 nl/min. After the final volume was injected, the needles were kept in place for 5-10 min and then slowly retracted (1 mm/ min). The wound was sutured, and acute brain slices were prepared from injected mice two to three weeks after surgery.

Identifying the claustrum in brain slices
Because the claustrum is a small and thin structure, it was important to have a reliable means of identifying it during brain slice recordings. The claustrum core is located ventro-laterally to an inflection point of the external capsule (Fig. 1A, arrow), which served as a reliable structural landmark. Further, PV, SST, and VIP were expressed in patterns that defined the location of the claustrum. PV was enriched in the claustral core (Fig. 1A,D), as reported previously (Real et al., 2003). VIP-INs formed a dense neurite plexus that partially overlapped with PV in the dorsal parts of the claustrum while ventral parts of the plexus were largely adjacent to the PV core (Fig. 1B,D). SST was enriched in the area directly adjacent to the PV core (Fig. 1C,D). This is the first demonstration that the position of VIP-IN or SST-IN can be used to locate the claustrum.
These IN-specific expression patterns served as the basis for locating claustral neurons during patch clamping in brain slices. Because our experiments used transgenic mouse lines with selective expression of membranebound eGFP in one of these three IN types, it was possible to guide electrode placement by using a two-photon microscope (Olympus FV-1000; 950-nm excitation wavelength) to visualize eGFP fluorescence in live slices. eGFP expression in PV-IN marked the claustrum core, while eGFP expression in VIP-IN or SST-IN could also be used to locate the claustrum because of the consistent distribution of the processes of these two subtypes of IN. This is illustrated for representative images of individual slices in Figure 1E-G and for images averaged across many slices in Figure 1H-J. VIP neurites formed a dense plexus at the height of the inflection point that corresponds to the most dorsal parts of the PV-enriched claustral core and were less prominent in the central core area of the claustrum (Fig. 1F,I). The distribution of SST neurons was complementary to that of the PV-enriched core, with a lower neurite density in an area ventral to the inflection point of the external capsule and within the claustrum core (Fig.  1G,J). Neurons were therefore considered to be within the claustrum if they were located within 50 mm of the PV-enriched or SST-deprived claustral core or within the dorsal part of the VIP-enriched neurite plexus or 200 mm ventral to this structure. Finally, in a majority of experiments, brain slices were fixed after completing recordings and then stained for PV expression and neurobiotin labeling to confirm the location of patched cells within the PV-enriched core of the claustrum. Taken together, we are confident that all results presented in this paper are based on recordings made from claustrum neurons.

Patch clamp recordings of intrinsic electrical properties
The electrical properties of claustral neurons were measured with borosilicate glass pipettes (5-9 MX) filled with internal solution containing the following: 130 mM Kgluconate, 10 mM KOH, 2.5 mM MgCl 2 , 10 mM HEPES, 4 mM Na 2 ATP, 0.4 mM Na 3 GTP, 5 mM EGTA, 5 mM Na 2 phosphocreatinin, and 0.2% neurobiotin (290-295 mOsm, pH 7.4). All whole-cell patch clamp recordings were performed at 24°C (except for a few recordings also done at 30°C) with a Multiclamp 700B amplifier (Molecular Devices) and a Digidata 1440 interface (Molecular Devices). Signals were acquired at 50 kHz and filtered at 10 kHz. The mean access resistance (R a ) was 18.9 6 0.4 MX (SEM) and remained non-compensated. Voltage values stated here also do not account for the liquid-liquid junction potential of À11.7 mV. Recordings were excluded from analysis if R a exceeded 30 MX, the ratio of R a /R m (R m = membrane resistance) exceeded 20% and/or the resting membrane potential (RMP) was unstable or less negative than À50 mV. Action potentials (APs) also had to overshoot 0 mV to be included in our analysis. A few INs whose AP firing frequency did not saturate at the highest current intensity applied were also excluded because their maximum AP firing frequency and adaptation values could not be determined accurately.
Electrical signals were analyzed with a variety of software: Clampfit 10.2 (Molecular Devices), Origin (OriginLab), Excel (Microsoft), Orange3 (Demsar et al., 2013), and custom-written MATLAB or R routines (MathWorks, R2015; R-Core Team). Up to 63 intrinsic electrical properties of claustrum neurons were measured in response to 1-s-long current pulses. RMP and R m were derived from the slope of a linear fit of plots of the relationship between membrane potential and applied current, derived from a series of current steps ranging from À60 pA to the largest subthreshold depolarizing current. The current threshold (ct) was defined as the minimum current level required to elicit AP firing, with "ctp2" corresponding to twice the threshold current level and "max activity" corresponding to the depolarization level that evoked the highest AP frequency.
AP threshold was defined as the membrane potential at which the first derivative of the membrane potential was .10 mV/ms. The threshold level was then used as a reference point to determine AP amplitude, half-width, threshold/AP amplitude, AP rise rate/half-width, and afterhyperpolarization (AHP) amplitude. AP waveform was described by the maximum AP rise and decay time as well as the ratio of maximum rise/decay time. Some parameters were the ratios of individual properties and were included to improve IN separation. The presence of an Figure 1. Identification of the claustrum in coronal brain slices. A, The claustrum core (cla) is enriched in PV-positive neuropil and is positioned ventro-laterally relative to an inflection point of the external capsule (ec; arrow), between the caudate putamen (CPu) and insula (ins). The distribution of VIP-IN and SST-IN can also be used to identify the claustrum: both IN types differ in their neurite distribution between the PV-enriched claustrum core and its surrounding shell. B, VIP-IN surround the claustrum core and their neurites form a mesh of neuropil around the PV-core. The neuropil density in insular cortex and endopiriform nucleus (end) is lower than in the claustral shell region. SST neurons (C) are enriched at the fringes of the claustrum core and avoid the PV-rich core. D, Merger of A-C showing the PV-rich claustral core, embedded into the VIP-enriched neuropil of the claustral shell and an enrichment of SST neurons in the fringe area of the claustral core. E-G, Visualization of IN-specific expression of eGFP in live brain slices from transgenic mice expressing eGFP in different populations of claustrum IN. Arrows indicate inflection points of the external capsule. E, PV-promoter-driven eGFP expression is enriched in neuropil in the core of the claustrum. F, VIP-promoter-driven eGFP expression is enriched in neuropil surrounding the claustrum core. G, SST-promoter-driven eGFP expression shows that SST-positive neuropil has a complementary distribution to PV and is lowest in density in the claustrum core. H-J, Images of mean intensity (gray scale) of eGFP fluorescence from 22-28 claustral image z-stacks from mice with eGFP expression in PV-IN (H), VIP-IN (I), and SST-IN (J). Images were aligned relative to the external capsule (dashed line), with the external capsule inflection point positioned at the upper left. Differences in fluorescence between the claustrum core (red outline) and its surrounds permit identification of the claustrum in live slices from all three mouse lines. The low fluorescence in the lower right corners of E-J is caused by glass recording pipettes.
Research Article: New Research afterdepolarization (ADP; see Fig. 2D below) was measured at the ct level. If the AHP was followed by a local maximum, this indicated an ADP. To calculate integrated ADP amplitude, a linear fit of the membrane potential between the AHP and the local minimum after the ADP was subtracted from the trace and the resulting positive values were averaged. The local minimum was defined as the minimum value after the AHP that preceded the change from a negative to a positive value of the low-pass filtered (eight-pole Bessel, 50 Hz cutoff frequency) membrane potential slope.
The 63 intrinsic electrical properties also included many measurements of the temporal structure of repetitive AP firing. Interspike interval (ISI) was defined as the time between the peak overshoots of two consecutive APs, while the instantaneous frequency was the inverse of the ISI. AP frequency adaptation, which is a time-dependent reduction in AP frequency during the stimulus, was quantified as the difference between the instantaneous frequency of the first two APs and the mean instantaneous frequency of the last three APs, while relative AP adaptation (ISI ratio) was defined as the ratio of these two parameters. To characterize the tendency of some neurons to fire an initial burst of APs, we measured the initial AP instantaneous frequency, the mean and SD of the first two ISIs, as well as the change of the ISI interval (Toledo-Rodriguez et al., 2004). The "maximum initial adaptation change" describes the maximum difference of the initial instantaneous frequency for the first AP pair measured for two consecutive depolarization steps divided by the current difference between the two depolarizations. The "maximum initial adaptation change (p2)" describes the maximum initial adaptation change but ignores initial ct values and values following traces with only a single AP. "Peak adaptation level" is the current level at which the maximum initial adaptation change took place and is calculated relative to the ct value. AP firing variability (Cv2) within a spike train was calculated as the mean Cv2 values for all consecutive pairs of ISIs (Holt et al., 1996). To characterize firing variability after the first (Cv2-first AP) or the first two ISI pairs (Cv2-first/second AP), the initial one or two ISI values were excluded for the averaged Cv2 calculation. AP amplitude ratios for the "first/second," "second/third," and "first/last three AP" were derived from the absolute amplitude values of the corresponding APs.

Measurement of temperature sensitivity
To determine the effect of temperature on the intrinsic electrical properties of claustrum neurons, these properties were recorded at both 24°C and 30°C for 13 neurons. The temperature coefficient, Q 10 (see Eq. 1 below), was calculated for all properties in each individual neuron. From all Q 10 values, the interquartile range (IQR) between the first (Q1) and the third (Q3) quartiles was calculated. Q 10 values that were either smaller than Q1-3pIQR or larger than Q313pIQR were considered to be outliers and were discarded; means of the remaining Q 10 values were used to correct parameters measured at 30°C to a temperature of 24°C.

Cell clustering
To identify groups of cells that shared similar features, an unsupervised hierarchical clustering of intrinsic electrical property measurements was done with ClustVis software (Metsalu and Vilo, 2015). Raw data were scaled by the SD of population means for each parameter (Z-score). For the entire population, a list of 38 properties was used (Table 2), and neurons were clustered based on similarity of correlations between their features, with clustering distances between neurons calculated by the Pearson's correlation. To separate neurons into distinct clusters with increasing dissimilarity, an average linkage criterion was used. To separate INs and PNs into distinct subclusters, Euclidean distances with Ward linkage were used. For IN subclustering, an extended list of 63 features was used. Dendrograms and Z-score maps were then generated in ClustVis according to similarities in their features, with the most similar cell pairs at the base of a branch. To identify an optimal number of cell clusters and validate the quality of the classification scheme, a silhouette analysis was performed and the average silhouette index (SI) was calculated using the package "cluster" in the R programming language (Maechler, 2018;R-Core-Team, 2018). The silhouette plot illustrates a measure, ranging from À1 to 1, of proximity between neighboring clusters; a width of 1 indicates that a cluster is highly separated from neighboring clusters, while a value of À1 indicates poor separation. The average silhouette width for all possible clusters was calculated and the optimal number of clusters was selected based on the maximum average width, which represents the optimal separation of clusters. As alternative clustering methods, t-distributed stochastic neighbor embedding (t-SNE; van der Maaten and Hinton, 2008) and/or locally linear embedding (LLE; Roweis and Saul, 2000) were applied using Orange3 (Demsar et al., 2013). All datasets were normalized, Euclidean distances were used as distance measure, data points were initialized with PCA, the learning rate was 200, and the number of maximum iterations was kept at 1000. Dataset-specific t-SNE settings were: "all neurons": perplexity: 25, exaggeration: 1; "PN" set: perplexity: 8, exaggeration: 7; "IN" set: perplexity: 7, exaggeration: 7. The "PN" and "IN" set was additionally analyzed using LLE with the following settings: modified LLE, number of neighbors: IN: 13; PN: 14, max number of iterations 100.

Image acquisition and processing
Images were acquired on an Olympus FV-1000 twophoton microscope. Image analysis and processing were done with Fiji/ImageJ software (Schindelin et al., 2012). As necessary, background subtraction (rolling ball radius 50 pixels) and/or a 5 Â 5 convolved kernel filter (center: 24; remaining: À1) were used to optimize image signal/ noise ratio.

Machine-learning assisted automated classification of cell types
Because our classification scheme relies on combinations of multiple intrinsic electrical properties, we have produced an automated, user-friendly method to implement our classification scheme. We created a GUIbased web tool written in the R programming language and Shiny GUI framework (R-Core-Team, 2018) that classifies cell types based on the electrical properties described here.
The software consists of trained classifiers capable of distinguishing claustral cell types using feedforward neural networks with a single hidden layer. PCA was used to reduce the dimensionality of input data, and 27 components that captured 99% of variance in the data were used to train the network. Ten-fold cross validation was used for all three classification steps to choose the Research Article: New Research appropriate L2 regularization parameter and number of hidden neurons over a grid search of hyperparameters. The final model consisted of 10 hidden neurons and a weight decay parameter of 0.001 and achieved out-of-sample accuracy of 96%. The above was implemented using the caret package in the R language (Kuhn, 2008). Fourteen electrophysiological properties described in this paper are extracted from traces automatically and classifiers trained according to the scheme described in this paper are used to distinguish five different subtypes of claustral PNs or three different subtypes of claustrum INs. While classification of PN subtypes is based on the unsupervised scheme described here, classification of INs is based on a database consisting of confirmed IN-marker-expressing neurons (19 confirmed PV-INs, 28 SST-INs, and 30 VIP-INs).

Software and code accessibility
This software is available as a web-based application via https://claustrum.shinyapps.io/online/ or can be downloaded from GitHub via https://github.com/adityanairneuro/ claustrum.

Statistical analysis
Mean values of intrinsic electrical properties measured in different neuronal types were compared using an unpaired t-test with Welch's correction, if these properties were normally distributed. If the properties were not normally distributed, the medians of cell properties of different neuronal types were compared using a Mann-Whitney U test for two separate neuronal groups or a Kruskal-Wallis test and Dunn's multiple comparison test when more than two groups were compared. To estimate the effect size of non-normally distributed parameters between groups, h 2 was calculated (Tomczak and Tomczak, 2014). To compare more than two groups with normally distributed parameters, an ANOVA test with Tukey's post hoc t-test was used. To estimate the effect sizes for between-groups comparisons of normally distributed parameters, R 2 was used.

Claustral neurons differ widely in their intrinsic electrical properties
To distinguish neuron types, we measured the intrinsic electrical properties of a total of 326 claustral neurons in brain slices prepared from 106 mice. All patched neurons were unambiguously located within the claustrum, identified with the procedures described in Materials and Methods. As shown in Figure 2, claustral neurons differed continued Figure 2. Claustral neurons are heterogenous in their responses to depolarizing currents. A, top, Representative membrane potential responses to depolarizing current pulses show that AP firing patterns differed in their frequency adaptation properties. Bottom, Time course of AP frequency adaptation in the traces shown above. Further differences between claustral neurons were found for their AP amplitude adaptation (B), the ISI, (C) and the magnitude of their post-AP AHP (D). widely in their responses to prolonged (1 s) depolarizing current pulses. Noticeable differences were found in the amount of AP frequency adaptation ( Fig. 2A), the adaptation of AP amplitude (Fig. 2B), the ISI (Fig. 2C), as well as the characteristics of AP repolarization (Fig. 2D). This diversity of intrinsic electrical properties indicates multiple types of claustrum neurons.
To identify subpopulations of claustrum neurons, 38 different intrinsic electrical properties were extracted from neuronal responses to depolarizing and hyperpolarizing current pulses (see Materials and Methods). These parameters were then normalized by their Z-scores and grouped using unsupervised agglomerative hierarchical clustering (Metsalu and Vilo, 2015). Claustral neurons (Fig. 3, columns) differed in their intrinsic properties (Fig. 3, rows) over a wide range, as indicated by the color-coded Z-scores. Cells with similar properties (similar colors) were clustered into groups, with more similar groups clustered into branches, and branches then clustered into subgroups according to their similarity, yielding the dendrogram shown in Figure 3, top. At the top hierarchical level, two main clusters of neurons could be distinguished: group 1 (G1; n = 152, 46.6%) and group 2 (G2; n = 174, 53.4%). These two clusters exhibited significant differences in their population medians for 30 out of 38 intrinsic properties (Table 2). To identify properties that differed the most between the groups, the absolute difference in the subpopulation Z-score means were calculated and ranked. The largest group differences in Z-score means for G1 and G2 were found for AHP magnitude at ctp2 and the RMP ( Fig. 4A; Table 2), while the smallest differences were found for the max AP rise rate at ctp2 and AP threshold at ctp2 ( Fig. 4A; Table 2). Plots of the three most divergent features revealed two distinct clusters that correlated well with G1 or G2 (Fig. 4B). As an alternative approach, neurons were also clustered using t-SNE (see Materials and Methods). Similar to the results of hierarchical clustering, G1 neurons clustered separately from G2 neurons (Fig. 4C). Taken together, the significant differences observed for most of the intrinsic electrical properties indicate that the two groups represent two different types of claustrum neurons.

Identification of PNs and INs
In many brain regions, PNs and INs differ in their intrinsic electrical properties (McCormick et al., 1985). We next asked whether such differences underlie the two different groups of claustrum neurons. To identify PNs, fluorescent beads were injected into cortical areas (prefrontal, orbitofrontal, auditory, visual, motor cortex) or the subcortex (thalamus, habenula, hippocampus; Fig. 5A) of mice where the claustrum core could be identified by eGFP labeling (PV-cre Â ChR2-eGFP; Fig. 5B). The cell bodies of PNs that retrogradely transported these beads could be visualized in live brain slices (Kim et al., 2016;Chia et al., 2017) and were located within the PV-enriched claustral core ( Fig. 5B-D). Whole-cell patch clamp recordings were then used to determine the electrical properties of 44 of these labeled PNs. Remarkably, all PNs had G2 characteristics ( Fig. 5J, orange), while G1 neurons were not labeled.
Because G1 cells did not appear to be PN, we determined whether these cells were INs by examining their expression of IN marker genes. The electrophysiological criteria described above were used to identify G1 cells, while the presence of IN marker genes was detected via two strategies. First, the fluorescence of eGFP expressed behind various IN promoters could be imaged in live slices to target IN for patch clamp recordings  . Differences between group 1 and group 2 claustrum neurons. A, Ranking of absolute differences between the subpopulation Z-score means of G1 and G2 neurons. Largest differences were observed for AHP amplitude at ctx2, RMP and ratio of amplitudes of first and second APs at ctx2. Dark gray shading indicates medians that differed significantly between both groups, while light gray indicates non-significant differences; 30 out of 38 features showed a significant difference. B, Comparison of the RMP, ratio of first/second AP amplitudes, and AHP amplitude at ctx2 for G1 and G2 neurons. Both groups were clearly separate, with the axis of the G1 cluster orthogonal to the G2 cluster axis. C, Comparison of the intrinsic properties of G1 and G2 neurons, using t-SNE-based clustering. These groups did not overlap in the t-SNE feature space.
Research Article: New Research retrobead-labeled neuron was in G2 (44/44). Because no G1 cells were labeled by retrobeads and a large fraction of these neurons were labeled by IN markers, it is likely that all G1 neurons are INs. Further, all bead-labeled PNs had G2 properties and were almost never labeled by IN markers. We therefore conclude that G1 neurons are very likely to be INs and G2 neurons are very likely to be PNs, so we will refer to G1 cells as INs and G2 cells as PNs.
In summary, significant differences in most intrinsic electrical properties allow reliable identification of claustral INs and PNs. IN marker gene expression and the presence of long-range anatomical connections in retrobead-labeled neurons support this conclusion.

Subtypes of claustral PNs
To identify subtypes of claustrum PNs and INs, hierarchical clustering was extended to subpopulations of neurons within the IN and PN groups. PNs were clustered using the same intrinsic electrical parameters as above, yielding the dendrogram shown in Figure 6. PNs were separated into two large subclusters that differed mainly in their probability of evincing an ADP at the current threshold, ct (first branch: 22%; second branch: 90%), the ratio of the amplitudes of first and last AP at ctp2 (first branch median: 1.43, second branch median: 1.15), and the latency to fire the first AP at ctp2 (first branch median: 16.1 ms; second branch median: 23.9 ms). After this initial bifurcation, the dendrogram split into multiple finer clusters. To determine the optimal number of separate clusters, a silhouette analysis was performed for two clusters (hierarchical level 1) to nine clusters (hierarchical level 8). An average SI width below 0.25 indicates random structure, while higher values indicate clustering (Kaufman and Rousseeuw, 1990). Higher SI values also indicate a higher degree of similarity of neurons within a cluster and better separation from other neurons. The mean SI width peaked at 0.27 for six clusters (Fig.  7A), indicating that optimal cell clustering occurred when the left branch was separated into two PN subtypes (PN1/ PN2), and the right branch was separated into three PN subtypes (Fig. 7B,C). The sixth cluster was a single outlier neuron, the VIP-expressing IN, adjacent to PN subtype 5 (Fig.  7B, circle, Fig. 7C, arrow).
To validate the hierarchical clustering results, neurons were also clustered with three other approaches using their Z-scored intrinsic properties (see Materials and Methods): PC clustering, t-SNE, and LLE. Considering the first three principle components (PCs) of all Z-scored intrinsic properties, PNs segregated in line with the hierarchical clustering results (Fig. 7D) and the outlier VIP neuron was separated from the remaining PN clusters (circle). Because the three first PCs account for only ;60% of the observed variance, there was some overlap between neighboring clusters. In both t-SNE (Fig. 7E) and LLE (Fig. 7F), PN1-3 neuron subtypes were clearly separated from other groups, while PN4 and PN5 subtypes partially overlapped. This recapitulated the hierarchical clustering, which could only distinguish these subtypes at hierarchical level 4 (Fig. 7C). The single VIP-IN (circle) could not be resolved from PN5 neurons using t-SNE but was distinguished in LLE. Taken together, all clustering approaches resulted in a comparable separation of PN subtypes, with neurons being clustered into identical subtypes in all cases.
A comparison of the properties of these PN neuron subtypes is shown in Figure 8. All subtypes, except for PN5 and the outlier VIP-IN, included neurons labeled with retrobeads, indicating that these subtypes are PNs. Remarkably, 92.3% of all subcortical-projecting neurons (12/13) were clustered into PN1, while 100% of the cortical-projecting neurons (28/28) were either PN2, PN3, or PN4. This indicates that the intrinsic electrical properties of claustral PNs allows these cells to be classified into at least two network-specific PN types. PN5 was unlabeled by either retrobeads or IN markers (Fig. 8A). Because .98% of the confirmed INs were correctly separated from PNs (Fig. 5J), and the only confirmed VIP-IN was flagged as an outlier, we presume that PN5 are PNs that project to targets that were not injected with retrobeads. The single outlier VIP-IN, the only confirmed IN within the PN group, was assigned to group 2 because its AP waveform was similar to that of genuine PNs, but was assigned its own PN subtype because of its unusually irregular AP firing.
A detailed comparison of the properties of all PN subtypes is presented in Tables 3, 4. The medians of the intrinsic electrical properties of all six groups differed significantly from each other (Kruskal-Wallis test), and an overview of their significant differences is shown in Table  5. PN1 neurons had minimal adaptation of AP frequency ( Fig. 8B) and relatively strong adaptation of AP amplitude over time; they also had no ADP (Fig. 8C). A radar plot profile of the mean Z-scores of PN1 intrinsic electrical properties indicated that PN1 differed substantially from the population mean of all PN (Fig. 8D, black circle) in many properties, indicating that PN1 neurons were relatively distinct from other PNs. The most distinctive difference between PN1 and other PN subtypes was a relative high ISI ratio ( Fig. 8D; Table 3, property #31). PN2 neurons rarely generated an ADP when stimulated at the ct (Fig.  8F), although ADPs were evoked at higher levels of depolarization. This ADP caused modest initial doublet-spiking at higher depolarization levels (Fig. 8E) and reduced the amplitude of the second AP in a train. The most distinctive feature of PN2 cells was their high current threshold ( Fig.  8G; Table 3, property #3), presumably caused by their low membrane resistance (see Table 3). PN3-PN5, as well as the errant VIP-IN, were characterized by a higher probability of generating an ADP at the current threshold (   Table 3, property #29), indicating pronounced AP doublet firing, and their initial AP burst was followed by relatively constant-frequency AP firing. PN4 and PN5 neurons showed an AP doublet at the onset of the AP train but subsequent AP firing was less consistent than the other PN groups (Fig. 8K,N). PN4 neurons were readily distinguished by the low amplitude of their second AP relative to the third AP (Fig. 8K,M; Table 4, property #27). PN5 were  Tables 3-5. distinguished by their delayed AP firing at ctp2 (Fig. 8P; Table 4, property #18) and also had relatively high membrane resistance and low current threshold (properties #2 and #3 in Table 4). The VIP-IN also had an ADP at the ct and doublet-spiking behavior at higher depolarization levels. However, this cell was distinguished by its high degree of AP firing variability; the Cv2 and the SD of the first two ISIs were larger than for any of the genuine PN ( Fig. 8S; Table 4, properties #33 and #36).
In summary, our analysis of intrinsic electrical properties revealed multiple subtypes of claustral PNs that differed in their projection targets: cortical (PN2, PN3, PN4) or subcortical (PN1) areas. The target of PN5 cells remains to be determined.

Subtypes of claustral IN
INs in the CNS are diverse and consist of numerous subtypes (Ascoli et al., 2008). To identify claustrum IN subtypes, we analyzed 63 intrinsic electrical properties and measured responses to three different depolarizing currents: at threshold, twice the threshold current, and at levels that evoked maximal AP firing. Hierarchical clustering of these IN intrinsic properties, as done for PNs, revealed one smaller cluster of INs on the left side of the dendrogram (IN1) that was distinct and differed in most properties from a larger cluster on the right (IN2; Fig. 9). Neurons in the smaller cluster had higher AP rise rates, more negative AP decay rates, and shorter AP halfwidths. To determine whether additional subgroups of IN could be distinguished, mean SI width was calculated for two (hierarchical level 1) to nine (hierarchical level 8) IN subclusters (Fig. 10A). The highest mean SI width was 0.33, at hierarchical level 1, with an average SI value of 0.26 at hierarchical level 2 and values below the threshold of 0.25 for larger numbers of clusters. The maximum SI value at hierarchical level 1 reflects the two main subtypes readily apparent in Figure 9, while a further subclassification into three groups is plausible. At hierarchical level 1, the first cell cluster accounted for 28.8% of all neurons (34/152), and their single-neuron SI values were generally high, indicating relatively homogenous intrinsic properties that allowed for clear separation of IN1 from IN2 (Fig. 10B). Neurons within the second cluster accounted for 71.2% of all IN and their single-neuron SI values varied very broadly, indicating a more heterogenous population. A plot of the three most distinctive properties, determined from their absolute differences within subpopulation Z-score means, again demonstrated that the two IN subtypes were clearly separable (Fig. 10C,D).
Cortical INs consist of at least three distinct types: those expressing PV, SST, or the type 3a serotonin receptor (5HT3aR), with the latter including a large subpopulation of VIP-IN (Lee et al., 2010;Rudy et al., 2011). We next asked whether expression of PV, SST, or VIP correlated with claustrum IN intrinsic electrical properties by mapping IN marker expression onto the hierarchical clustering dendrogram (Fig. 11A) (Fig. 11B,  left), a VIP-IN (Fig. 11B, middle), and a SST-IN (Fig. 11B, right) to depolarizing current pulses illustrate that PV-INs  were clearly distinguished by their rapid and regular AP firing, as previously described (Kim et al., 2016;. A detailed analysis of the first three PCs of all Z-scored electrical parameters showed that the PV-enriched IN1 cluster separated fairly well from the IN2 cluster enriched in VIP-INs and SST-INs (Fig. 11C). The alternative clustering approaches t-SNE (Fig. 11D) and LLE  the mean SI width at two clusters to be higher than that for three clusters (Fig. 10A). Despite this heterogeneity, SST-INs and VIP-INs still appeared in distinct clusters at hierarchical level 2 of the IN2 subpopulation: 90% of all confirmed VIP-INs were clustered into the left branch, while all confirmed SST-INs were in the right branch (Fig. 11A). Similarly, VIP-IN and SST-IN subtypes were largely distinct based on their first three PCs (Fig. 11F), indicating some separation of all three IN subtypes. Indeed, both t-SNE (Fig. 11G) and LLE (Fig. 11H) (Fig. 12, black circle). VIP-INs were the most excitable, with lowest ct, most depolarized RMP and highest Rm. PV-INs were the least excitable, with the most hyperpolarized RMP, lowest Rm and highest ct, as well as having the shortest AP half-width and highest AP frequency. Characteristic features of SST-INs were relatively low amounts of adaptation and regular AP firing in response to all three stimulus protocols. In summary, the intrinsic electrical properties of three IHC-defined IN subtypes showed significant differences that could be used to identify these IN subtypes.
To enhance our ability to distinguish IN subtypes, an artificial neural network was trained with the 63 intrinsic electrical properties of 77 confirmed . Overall, the cross-validation prediction accuracy for distinguishing IN subtypes via this approach was 93%, indicating reliable resolution of the three subtypes of claustrum INs.
In conclusion, measurements of intrinsic electric properties allowed us to distinguish claustral PNs from INs and to resolve subtypes of both PNs and INs. The anatomical projections and expression of marker genes in these neurons largely respected the electrophysiological distinctions between cell types.

Effects of temperature
Because the recordings described above were done at room temperature (24°C), we also asked whether our classification scheme could be extrapolated to other conditions by determining how warmer temperatures affected cell electrical properties. A representative example of the activity of a PN3 neuron recorded at two different temperatures (24°C and 30°C) is shown in Figure 13A. It is  evident that the warming the temperature affected the electrical properties of this neuron, for example, altering AP frequency and the amount of current required to evoke APs. For all 13 neurons examined at these two temperatures, we quantified the temperature sensitivity of their intrinsic electrical properties by calculating the temperature coefficient (Q 10 ), a measure of how much each property changes for a ten degree rise in temperature: where P 2 is the value of each parameter at temperature T 2 , and P 1 is the same parameter at a lower temperature, T 1 . A Q 10 . 1 indicates a parameter that increases with temperature, while values below one indicate parameters that decrease with temperature (Janssen, 1992;Tang et al., 2010). Most of the intrinsic electrical properties of claustral neurons had Q 10 values ranging from 0.25 to 4 ( Fig. 13B; Table 8). Exceptions were the Q 10 values of some ratiometric parameters, which were extreme due to the effects of ratioing the Q 10 values of single parameters. To determine whether these Q 10 values could be used to adapt our cell classification scheme to different temperatures, we examined the ability of the classification scheme to correctly predict the subtypes of the 13 neurons that were examined at the two temperatures (Fig. 13C). At 30°C, the scheme often misclassified neurons as strongly-adapting (SA) PN3 or IN subtypes because of the higher AP rise and decay times and faster frequency adaptation at the higher temperature (Fig. 13C, center). However, after using the mean Q 10 values shown in Figure 13B to adjust the cellular properties measured at 30°C, cell type identity was predicted with high fidelity: 100% of the neurons recorded at 30°C were correctly predicted to be either PNs or INs (Fig. 13C, right). This temperature correction was also 100% accurate at identifying the three IN types and in separating subcortical PNs from cortical PNs. When separating all neurons into their previously defined cell types, the accuracy was reduced to a still acceptable 85%, due to some errors in distinguishing subtypes of cortical PNs.
In summary, because the intrinsic electrical properties of claustrum neurons scaled predictably with temperature, our classification scheme could accurately recognize different subtypes of neurons at temperatures other than room temperature.

Neural network assisted claustral cell classification
While intrinsic electrical properties can unambiguously identify subtypes of claustral neurons, to achieve high accuracy it is necessary to consider multiple cellular properties, particularly for distinguishing IN subtypes. To simplify the process of identifying claustrum neurons according to their intrinsic electrical properties, we trained a neural network to classify claustral cells (Fig. 14) using the open-source programming language R (R-Core-Team, 2018). The tool analyzes experimental recordings saved in the .abf format (Fig.  14, step 1) and extracts 14 key intrinsic electrical properties from raw data (Fig. 14, step 2). The extracted properties are then automatically compared with our database of five PN and three IN subtypes, and the network-assisted cell classification procedure (Fig. 14, step 3) defines the probability that a neuron is one of the eight cell subtypes (Fig. 14, step  4). Tests of the classifier revealed that its prediction accuracy for initially identifying a neuron as IN versus PN was 97%, for the identification of PN subtypes was 89%, and for IN subtypes the classifier had a prediction accuracy of 93%. In summary, the classifier provides a user-friendly approach that automatically extracts relevant raw data that are then used to classify claustral neurons. This tool is freely available at https://claustrum.shinyapps.io/classifier/ or https://github.com/adityanairneuro/claustrum.

Discussion
We have described an electrophysiology-based scheme for classifying claustral neurons that was augmented by anatomical tracer studies and analyses of marker gene expression. Five subtypes of claustral PNs could be distinguished, with PN subtypes that predominantly projected to either cortical or subcortical targets, as well as three subtypes of INs that expressed PV, SST, or VIP. In addition, we developed a cell classifier tool that facilitates data extraction and identification of claustral neurons according to their intrinsic electrical properties.

Intrinsic electrical properties of claustral neurons
Our work substantially extends previous efforts to characterize the intrinsic electrical properties of claustrum neurons.

PNs
Kim et al. (2016) described a few of the intrinsic electrical properties of claustro-cortical PNs, identified via retrograde labeling, that exhibited variable degrees of AP frequency adaptation. Although they did not attempt to Cell property values are given as means and their SEM and medians and their median absolute deviation (MAD). Cell features are sorted according to the radar plots in Figure 12. distinguish PN subtypes, their results likely correspond to our PN subtypes 2-4. Shibuya and Yamamoto (1998) identified two distinct neuron types in rat claustrum that differed in their frequency adaptation and called these "slow-adapting neurons" and "fast-adapting neurons." Both of these resemble our subtypes of claustro-cortical PN: their slow-adapting neurons likely correspond to our cortical-projecting type PN2 or PN3, while the fast-adapting neurons most closely resemble PN4 and PN5. Our results also improve on the classification scheme of Chia et al. (2017) that distinguished strongly-adapting (SA) and mildly-adapting (MA) populations of anterior cingulate cortex-projecting claustral neurons. Both of our approaches agree that the claustrum contains multiple subtypes of PNs that can be distinguished based on their AP frequency adaptation: the SA neurons of Chia et al. (2017) likely correspond to PN3-5, while the MA neurons likely correspond to our cortical-projecting PN2 subtype, which only show adaptation at relatively high levels of depolarization. Chia et al. (2017) also reported a sex-specific enrichment of distinct PN subtypes, depending on the rostrocaudal location of the PN subtype. Although we observed a trend for certain PN types to be more enriched in males or females, none of those trends were statistically significant, presumably because our analysis did not sort neurons according to their locations. Our classification scheme also extends the results of , who described two types of claustral PN that differed in their intrinsic properties and AP frequency adaptation. Their type 2 neurons are comparable to the vigorous doublet-AP firing PN3 neurons we have described, while   Table 6. The identity of every tenth property is also indicated.
their type 1 neurons are likely to be PN2 neurons that show less frequency adaptation and reduced initial doublet spiking. No PN1 subtype neurons have been described previously; this is due to the fact that no previous analyses considered claustral PN projecting to subcortical regions. It was somewhat surprising to find claustrum neurons that projected to subcortical structures. Previous work has suggested that the primary subcortical target of the claustrum is the amygdala (Zingg et al., 2018), which we did not examine in our experiments. Although injecting retrobeads into the subcortex mainly labeled insular neurons (Mathur et al., 2009), there was also sparse labeling of neurons within the PV-enriched claustral core/shell region and we only recorded from neurons that were clearly within this region. Injection of retrobeads into subcortical regions also yielded some bead leakage into cortical structures along the injection track. However, because our subcortical bead injections almost exclusively labeled PN1 cells, which were not labeled by cortical bead injections, we can deduce that leakage of tracer into the cortex had minimal impact on our experimental results. Thus, we conclude that PN1 cells genuinely project to subcortical structures, such as the mediodorsal thalamus, central medial thalamic nucleus, ventromedial thalamic nucleus, or the habenula (Table 1), and that the function of claustral PN1 neurons is to provide input to these structures.
In summary, our classification scheme largely unifies previous reports and extends our understanding by distinguishing multiple subtypes of claustral PNs that project to different brain areas.

INs
Neither Shibuya and Yamamoto (1998) Figure 13. Effects of temperature on claustrum cell classification. A, Representative example of the activity of a single PN3 neuron to a depolarizing current pulse at both 24°C (left) and at 30°C (right). B, Relative changes in cell excitability (gray), first AP temporal properties (brown), AP frequency properties (pale green), AP and AHP amplitude (blue), ADP characteristics (green), and firing regularity (violet), expressed as Q 10 values. Parameters are arranged as in Table 8, where precise Q 10 values can be found. C, Cell prediction results at 24°C (left) and 30°C, without (center) or with adjustment of the intrinsic cell features according to the Q 10 values (right). Cell types are indicated on the right, IN are outlined by a red box, PN by a blue box. Adjustment of the intrinsic features by their mean Q 10 values improved the prediction accuracy. which have more hyperpolarized RMP and deeper AHPs than their type 5 cells. We suspect this discrepancy arises because the Cre-driver mouse line used by  ectopically expresses Cre in non-PV cells, including PNs (Tanahira et al., 2009). Our classifier provides the first reliable way to use intrinsic electrical properties to distinguish claustral PV-INs, SST-INs, and VIP-INs. In summary, our results characterize many different subtypes of claustrum neurons, including previously unknown PN and IN subtypes, and lay the foundation for future studies of claustrum neurons and their functions. From a technical perspective, our work substantially advances previous efforts to identify types of claustrum neurons by systematically analyzing up to 63 intrinsic electrical properties of claustrum neurons and by integrating these properties with information about projection targets and expression of IN marker genes. Our efforts have also yielded an objective algorithm for classifying claustrum neurons that was validated by multiple non-overlapping approaches. Although our classification scheme was based on recordings done under one set of experimental conditions, at 24°C and in adult mice (postnatal age 65), it is largely applicable under other experimental conditions. Specifically, we found that our classification scheme works at higher temperatures, after taking into account the temperature dependence of intrinsic electrical properties. The characteristic firing patterns of claustrum neurons, such as strong frequency adaptation with an initial AP doublet or amplitude adaptation with mild frequency adaptation, are also observed in neurons from younger age mice (mean age: postnatal age 21 6 0.4; unpublished results of Y. Tang and M. Graf), and we anticipate that our classification scheme will apply over a range of ages. Hence, our classification rules can be used under a variety of experimental conditions.

Comparison of claustral and cortical neurons
In cerebral cortex, INs differ fundamentally from PNs (McCormick et al., 1985;Gouwens et al., 2019) and can be grouped into three main subtypes: PV-INs, SST-INs, and 5Ht3aR-INs, with the latter further subdivided into VIP-expressing and non-expressing INs (Lee et al., 2010;Rudy et al., 2011). Although the intrinsic properties of different cortical IN types are variable, certain combinations of properties are enriched in distinct IN types and can be used to cluster them with unsupervised methods (Gouwens et al., 2019). As in the cortex, our trained neural network was able to identify rules that could be used to  Step 1, Input raw data file.
Step 2, Automated extraction of intrinsic electrical properties.
Step 3, Neuronal network assisted cell classification.
Step 4, The classifier returns the likelihood that the classified cell belongs to one of the eight cell subtypes.
distinguish fast-spiking PV-INs from irregular-firing VIP-INs and regular-firing SST-INs. How do the intrinsic electrical properties of neurons in the subcortical claustrum compare to those of cortical neurons? Overall, claustral neurons show trends very similar to those of cortical cells. As in the cortex, the intrinsic features of claustral PNs correlate with differences in their projection targets. For example, there are pronounced differences in intrinsic electrical properties between intratelencephalic PNs that connect to cortical areas (and the striatum) and extratelencephalic PNs that additionally project to areas outside the telencephalon such as brain stem or thalamus (Baker et al., 2018). Similarly, subcortical projecting claustral neurons (PN1) showed less frequency adaptation than neurons that connect to the cortex (PN2-PN4) and PN5. Cortical PN that target different areas also differ in their AP threshold (for review, see Baker et al., 2018). Claustro-cortical PN subtypes differed in some of their electrical properties, suggesting that these subtypes may differ in their cortical projections. Future studies using refined retrograde labeling procedures will be required to consider this possibility. Claustral PV-INs, SST-INs, and VIP-INs generally are comparable to their cortical counterparts and differences between these IN subtypes are consistent between claustral and cortical INs, with PV-INs and VIP-INs differing the most and SST-INs intermediate.
The high similarity of claustral and cortical neurons could arise from their similar developmental origins (Watson and Puelles, 2017;Binks et al., 2019): both cortical and claustral PNs are of pallial origin, while INs originate in subpallial areas and then migrate into cortex or claustrum. As a result of their shared developmental origin, claustral and cortical neurons perhaps could also show similarities in their connectivity patterns.

New insights into claustrum function
Although our study was intended to lay a cellular foundation for future system-level studies of claustrum function, our work already provides some initial and novel insights into how the claustrum processes information.

PNs
AP doublet firing by cortical-projecting claustrum PN2-4 should have significant consequences for information encoding by these cells. Because the depolarization required for initial ADP generation in PN2-PN4 was 1.5 times higher than AP threshold, multiple spiking in these neurons depends non-linearly on input strength. Further, double spiking can cause synaptic facilitation that transiently increases presynaptic release probability (Lisman, 1997). Thus, such doublet-firing properties could introduce non-linearities in transmission to downstream targets. For example, if two separate populations of claustro-cortical PNs receive signals from two different presynaptic inputs, the neurons with weaker input would encode incoming information more linearly and with a lower likelihood of eliciting large postsynaptic potentials, while the supralinear properties associated with double spiking would cause neurons receiving stronger input to have a disproportionately higher probability of transmitting this information to their postsynaptic targets. This could provide a mechanism to differentiate between relevant and irrelevant signals in the claustro-cortical network and might underlie the reported ability of claustro-cortical neurons to serve as a selective attention filter that depends on the strength of their inputs (Mathur, 2014;Remedios et al., 2014;Goll et al., 2015).
In contrast, subcortical projecting PN1 lack AP doublet firing. The resulting reduction in AP frequency adaptation will confer more linear transmission properties on these neurons during initial AP firing. In addition, PN1 AP amplitude often adapts during repetitive firing. Because synaptic transmission is a sensitive function of AP amplitude (Katz and Miledi, 1967), amplitude adaptation would yield a time-dependent reduction in the ability of claustro-subcortical cells to excite their postsynaptic target cells. In summary, the intrinsic properties of PNs could significantly shape claustrum network activity more globally. Our classification scheme provides a way to identify these subpopulations of PNs and, thereby, to design future experiments that determine how differences in their AP firing contribute to claustrum function.

INs
Our discovery that each subtype of claustral IN has unique intrinsic electrical properties also suggests potential differences in the function of these INs. In particular, VIP-INs are the most excitable neurons (highest Rm, lowest ct, relative depolarized RMP) and therefore could be preferentially engaged in claustral responses to weak excitatory input.
Considering that the activity of VIP neurons is modulated by expectation (Krabbe et al., 2019;Garrett et al., 2020), and that claustral VIP neurons disinhibit PNs  in a fashion similar to what has been described for cortical areas (Lee et al., 2013;Pfeffer et al., 2013;Pi et al., 2013), the function of VIP-INs might be to increase the responsiveness of claustral PNs to weak stimuli during top-down modulated attention. Thus, in the context of the proposed role of the claustrum in defining salient information from noise (Mathur, 2014;Remedios et al., 2014;Goll et al., 2015), the unique electrical properties of VIP-INs could allow claustral PNs to respond to weak signals that are otherwise insufficient to activate claustral PNs and thereby determine whether or not an incoming signal is attention worthy.

Conclusions
Our results provide a comprehensive analysis of the intrinsic electrical properties of claustrum neurons and demonstrate that these properties can be used to distinguish different subtypes of claustral neurons. This classification scheme will serve as the foundation for future studies of the cell-type-specific functions of claustrum neurons. For example, our lab is now using it successfully to discover cell-specific differences in the local inhibitory microcircuitry of the claustrum , cholinergic gain control of PNs (Nair et al., 2019), and serotoninergic modulation of claustral cell activity (Wong and Augustine, 2019). Finally, by analogy to the Blue Brain Project (Markram et al., 2015), this refined understanding of the properties of claustral neuron subtypes will facilitate development of large-scale models of claustrum networks and will enable a better understanding of higher-level signal processing within the claustrum, as well as between the claustrum and its connected structures.