Using rabies virus for tracing neural connections: caveats and limitations exposed by studies of barrel cortex circuits

Viral methods, in particular attenuated rabies virus (RV), have become a widely used part of the toolbox for mapping mammalian neural circuits. It is often assumed, despite surprisingly little evidence, that RV spreads strictly across synapses. Here I show that, for barrel cortex circuits, circuits mapped using neurophysiology and one-step (‘monosynaptically restricted’) RV mapping show little quantitative correspondence. For example, certain types of GABAergic interneurons that are known to receive little functional input from thalamus show relatively strong ‘input’ from thalamus by RV mapping, suggesting non-synaptic spread, or efficient spread through weak synapses. To gain an understanding of the functional organization of neural circuits, RV mapping studies need to be combined with quantitative methods that actually map synaptic connections. Moreover, to advance viral methods for mapping circuits, a better understanding of the mechanisms of trans-cellular viral spread and the resultant biases is needed.

Below I outline issues with interpreting one-step RV experiments that are well-known to some (‘you are preaching to the choir!’), but not known to many others in the field (‘holy guacamole, are you serious?’). Many investigators are not aware how little is known about the mechanisms underlying viral tracing for mapping synaptic circuits (‘I assumed enough was known about the virology that the issue of transsynaptic spread was settled’). I have two main motivations for writing this piece: i) Dampen enthusiasm for one-step RV tracing as a silver bullet for the classic hard problem in brain research, namely to map synaptically coupled neurons over all spatial scales. ii) Stimulate work to characterize the mechanisms of trans-cellular viral spread and to improve the efficiency and specificity of these promising tools. Note that this is an opinion piece, not a research paper, nor a review of the field, so coverage is  idiosyncratic and referencing is sparse.


The flow of information in neural networks is determined in large part by synaptic connections. Synapses require overlap of axonal and dendritic arbors (or dendrites, for dendrodendritic synapses). The shapes of axonal and dendritic arbors therefore provide an important source of specificity in neural networks. A classic view is that the overlap of axons and dendrites defines synaptic circuits (i.e. ‘Peters rule’): where axons and dendrites are sufficiently close, synaptic connections occur. Quantitative versions of this principle have long been used to produce approximate circuit diagrams in the cortex (e.g. (Binzegger et al., 2004)). The data underlying this structural view of connectivity is based on a rich toolbox of anterograde and retrograde tracers, as well as intracellular labeling. Over the last decade various engineered viruses have become widely used for anterograde and retrograde tracing.   

The structural view of connectivity is an over-simplification. First, connectivity is cell type-specific: one cell type preferentially connects to a subset of specific cell types and less with other cell types, despite similar overlap of axonal and dendritic arbors. For example, thalamocortical axons from VPM target L4 of the barrel cortex. VPM axons make strong and numerous synapses with parvalbumin (PV) expressing fast-spiking (FS) interneurons and excitatory stellate cells, but make only weak functional connections with somatostatin (Sst) interneurons (Cruikshank et al., 2010). Similarly, in cortical local circuits the major classes of interneurons interact in a cell type-specific manner (Pfeffer et al., 2013).

Second, beyond specificity at the level of  cell types, cortical circuits show so-called fine-scale specificity. This specificity manifests in various ways. Nearby pyramidal neurons have low connection probability (on the order of 0.1). However, pairs of pyramidal neurons that are connected in one direction are also highly likely to be connected in the reciprocal directions (Song et al., 2005). The receptive fields of neighboring pyramidal neurons can have very different properties, suggesting specific inputs to individual neurons. For example, in the mouse visual cortex, nearby neurons have nearly uncorrelated tuning to the orientation and direction of moving stimuli. Fine-scale specificity in synaptic connectivity underlies selectivity in these neuronal responses and likely is critical to explain the computational power of the brain (Ko et al., 2011).

There is great interest in methods that map connections between individual neurons, or between groups of genetically defined neurons. In particular, methods that produce transgene expression in synaptically-coupled circuits would be very powerful. Expression of fluorescent proteins then enables brain-wide anatomy. Imaging activity using protein-based probes for calcium and other sensors would help delineate how activity in input neurons shape activity in their targets, a major goal of neuroscience. Expression of optogenetic reagents would allow testing of hypotheses about how neurons influence each other and how activity in defined neurons influences behavior.

Viral methods have been used for more than 20 years for transneuronal labeling. In the absence of evidence, I use the term ‘transneuronal’ rather than ‘transsynaptic’ (more below). Retrogradely transported rabies virus (RV) and pseudo-rabies virus (PRV) spreads from starter cells to nearby neurons. These viruses have been workhorses in neuroanatomy (Card and Enquist, 2001; Ugolini, 1995). They have been used to discover new macro-scale circuits (Dum and Strick, 2013). However, viral spread across chains of multiple neurons can complicate interpretation.  

An exciting and conceptually clever development was the invention of schemes that limit the spread of RV to one-step of transmission. To restrict viral spread, the gene encoding the glycoprotein (G), which forms the viral capsid, was deleted from the RV genome (RVdG) (Callaway and Luo, 2015; Wickersham et al., 2007). Expression of G and RVdG infection defines “starter cells” which can produce infectious virus. The virus then exits cells (presumably dendrites) and spreads to putative “input cells”. However, these input cells do not produce G and spread from the input cells does not occur. The number and spatial distribution of different types of infected ‘input cells’ is the primary measurement in the experiment. The somata of input cells can be anywhere in the brain, facilitating brain-wide analysis of putative connectivity. I refer to this scheme ‘one-step RV’ (often referred to as ‘monosynaptically restricted RV’). Hundred of studies have been published using one-step RV to trace neural circuits.

Despite this profusion of activity, it is unknown how RV spreads from cell to cell. Direct evidence for synapse-specific spread is sparse and circumstantial. This was the case in 2008, when I had the privilege to coauthor a review on circuit cracking methods, including viral methods, with Ed Callaway and Liqun Luo (Luo et al., 2008). A decade later Ed, Liqun and I wrote an update on the same topic (Luo et al., 2018). In the intervening decade one-step RV has gained great popularity, but little had changed in terms of our understanding of what the collection of ‘input cells’ actually represents in terms of neural circuits.

The barrel cortex is one of the best understood central neural circuit in terms of cell types and their functional connections. Cell type-specific connections have been mapped by multiple laboratories using quantitative neurophysiological methods, producing a consensus view of major features of the circuit. One-step RV mapping experiments have produced results that are in dissonance with the field’s understanding of the organization of barrel cortex circuits. Below I briefly compare barrel cortex circuits mapped with  one-step RV. I focus on the state-of-the-art work from Ed’s and Liqun’s labs, because these studies have been done with sufficient care that quantitative comparisons with neurophysiological studies are possible (in particular, the number and distributions of starter cells were quantified). These comparisons reveal that one-step RV-based tracing fails to capture key information about functional connectivity at the level of defined cell types. I focus on barrel cortex simply because a lot of detailed circuit mapping has been done using different experimental modes, but the issues that emerge are general.

Specificity of viral transfer

The basic idea of RV-based circuit mapping is that RV spreads strictly across synapses, and more so through strongly connected pathways, so that the patterns of labeled input cells would provide quantitative information about the synaptic input to the starter cells. In fact, the moniker ‘monosynaptic RV’ has taken hold for one-step RV. It’s useful to ask what synapse-specificity might mean for the analysis of neural circuits? Let us qualitatively distinguish four levels of synapse-specificity.

  1. Strong synapse-specificity (SSS). The number of synapses and/or synaptic strength are the primary determinants of viral spread. Cell type-specific viral tropism and cell biology play no role in viral spread. In other words, given a starter cell, the number of ‘input cells’ is proportional to the number of synapses made with the input cells and the strength of the synapses. Note that this definition of ‘strong synapse-specificity’ does not require efficient transfer to input cells. Rather the spread to different types of input cells should be proportional to overall synaptic strength, and independent of the identity of the input cell.
  2. Medium synapse-specificity (MSS). Same as above, but allowing for biases, such as moderate tropism for specific cell types.  The proportion of different types of input cells would still reflect the rank order of overall synaptic connectivity with different types of input cells.
  3. Weak synapse-specificity (WSS). Same as above, but allowing for strong biases, such as pronounced tropism for one input cell type over another. The rank order of synaptic strengths is not reflected in the number of input cells. The key requirement is that RV spread is limited to cells that are synaptically coupled with the input cells.
  4. No synapse-specificity (NSS). The virus spreads across non-synaptic junctions by ‘proximity’, perhaps in addition to synaptic junctions. It is unlikely that RV can spread far in the extracellular space, given its large size. Therefore proximity here refers to distances on the order of 1 micrometers. However, on length scales of 1 micrometer the substrates for spread by proximity are abundant. For example,  in cortical tissue appositions (i.e. apposed membranes) of axon and dendrite are mostly non-synaptic (Figure 1). In addition, exocytosis happens along the dendrite in non-synaptic membranes (Patterson et al., 2010). For these reasons a viral particle leaving a starter cell might well first encounter an axon or presynaptic terminal that is not actually connected to the starter cell.    

So, why does the community refer to  ‘one-step RV’ as ‘monosynaptic RV’? A classic RV study mapping the circuits underlying facial motor control is often cited in support for synapse-specificity for RV virus (Ugolini, 1995). RV was injected into the hypoglossal nerve. Early after infection labeling was limited to the hypoglossal nucleus. Labeling then spread to premotor neurons in the medulla and later still to the mesencephalon and the cortex. The paper shows that RV infection is ‘specific’ in the sense that i) the virus does not spread promiscuously to all neighboring neurons and glia from infected neurons and ii) the locations of infected input cells are roughly in line with expectations (e.g. motor cortex but not somatosensory cortex). These results are entirely consistent with spread by proximity.

The original one-step RV paper contained some paired electrophysiological recordings of starter cells and input cells in slice cultures (Wickersham et al., 2007). These experiments were suggestive of preference for synaptic spread. But this study was not sufficiently comprehensive to be conclusive and the slice culture preparation is not representative for the situation in the intact brain.

Additional evidence comes from a couple of heroic studies, in which the receptive field of a mouse visual cortex starter cell was characterized (Wertz et al., 2015) (Rossi et al, Input cells were transduced with RV expressing GCaMP, a calcium indicator, so that the receptive fields of the input cells could also be assessed. In some cases, the input cells shared orientation selectivity with the starter cell (especially in the Rossi et al study). This shows that input cells with specific receptive fields can be preferentially labeled, providing some evidence for preference for connected neurons within a class. However, it is still possible, even likely, that the axons of the labeled input cells are locally more intertwined with the dendrites of the starter cells, which would make synapses as well as viral spread more probable, independent of the RV route. In other words, this experiment still tells us relatively little about specificity in RV spread.   

I note again that substrates for spread by proximity (rather than synapses) are abundant. Dendritic exocytosis occurs at non-synaptic sites (Patterson et al., 2010). Axons often touch dendrites without making synapses. In fact, in cortical circuits non-synaptic touches outnumber synaptic touches by several-fold (Mishchenko et al., 2010). This point is sometimes hard to convey. For a clear explanation and demonstration see the gorgeous EM study on mouse cortex from the Lichtman lab (Kasthuri et al., 2015). A rabies particle released by a starter cell therefore likely first encounters an axon from a cell that is not synaptically connected.

Figure 1. Schematic showing possible mechanisms for non-synaptic spread and cell type-specific biases in the efficiency of spread. Left, the process of transcellular spread from starter cell (magenta) to input cell (green). The viral production (1), retrograde transport (2), exocytosis (3), endocytosis (4), and axonal transport (5) etc all could have rates that differ across different cell types. Right, blow-up shows how an exocytosed viral particle likely encounters axons that are not connected to the starter cell (the green axon), at least in neocortex and hippocampus where the relevant ultrastructural analysis has been done. This provides a pathway for non-synaptic spread by proximity. Proximity refers here to distances on the order of 1 micrometers. It is unlikely that RV can spread far in the extracellular space, given its large size.

In addition, potential mechanisms for strong cell type-specific biases, independent of synaptic connectivity, are also abundant (Figure 1). Viral production likely proceeds at different rates in different types of starter cells. Starter cells might also differ in mechanisms and efficiency of dendritic transport and exocytosis. Viruses naturally tend to display tropism for one cell type over another, as a rule rather than the exception. For example, investigators report different results with different strains of RV (e.g. N2c vs SADB19). Tropism is especially acute at very low titers, such as those produced by sparse starter cells. These considerations alone make SSS and MSS unlikely, and suggest WSS or NSS for RV tracing.

Barrel cortex circuits

Indeed, comparisons of barrel cortex circuits mapped with one-step RV and neurophysiological methods indicate that one-step RV-based tracing does not show SSS or MSS, instead demonstrating WSS or NSS.

Barrel cortex circuits have been studied for decades with quantitative neurophysiological methods that have synaptic resolution. Paired recordings in brain slices have measured connection probabilities and unitary synaptic potentials (uEPSP, a measure of synaptic strength) between nearby neurons. Many synaptic connections have been probed in multiple studies in mouse and rat, with largely consistent results (Lefort et al., 2009; Ma et al., 2012). In local circuits the results from paired recordings generally agree with ultrastructural analysis at the level of synapses (e.g. stronger uEPSP between cells implies larger numbers of synapses; e.g.  (Markram et al., 1997)). Brain slice neurophysiological studies have to be interpreted with some caution, because longer pathways are more likely to be reduced by the brain slicing procedure. I therefore restrict the analysis of brain slice studies to circuits that span short distances and comparisons over similar spatial scales  (i.e. L2 → L3 & L2 → L2).

A widely used method for measuring long-range synaptic connectivity is ChR-assisted circuit mapping (CRACM). CRACM measures connections between presynaptic ChR-positive neurons and postsynaptic neurons that are targeted by whole-cell recordings (Petreanu et al., 2007). Because ChR-positive axons can be photostimulated even if they are severed from their parent somata, CRACM can be used to quantify connections that are not preserved in brain slices. CRACM provides quantitative comparisons of inputs from the same stimulated presynaptic input across different postsynaptic neuronal populations, which can be compared to one-step RV tracing.

In addition, a great deal is known about the propagation of sensory stimuli through barrel cortex circuits (Figure 2). For example, in response to touch, neurons in the ventral posterior medial nucleus of the thalamus (VPM) produce a single spike with high probability, short latency (3 ms), and low latency jitter (< 1 ms) (Figure 2D). Similarly, short-latency spikes and EPSPs are detected in L4 & 5 excitatory and PV neurons. In contrast, in Sst neurons touch evokes spikes and EPSPS with a substantial delay, or even shows hyperpolarization after touch (Biorxiv 554949) (Gentet et al., 2012). Similarly, VIP interneurons are not excited by touch (Biorxiv 554949) . This suggests that L4 & 5 excitatory and PV neurons receive strong driving input from VPM, but Sst and VIP neurons do not. Below I compare RV tracing and ephys for three cases.

1. RV mapping of inputs to GABAergic neurons

Multiple studies have reported one-step RV tracing in the barrel cortex. Most use EnVA pseudotyped RVdG, which can only infect mammalian cells expressing the avian receptor TVA. Wall et al (2016) analyze the long-range inputs to GABAergic interneurons (Wall et al., 2016). They use PV, Sst or VIP neurons as starter cells. The paper represents an impressive amount of work. The analysis involves careful quantification of starter cells (expressing protein G and also mCherry encoded in the RVdG genome) and input cells (expressing mCherry).

Figure 2. Comparison of connections measured by one-step RV and electrophysiology. A, Inputs per starter cell for Sst, FS interneurons (PV) and VIP interneurons (Wall et al., 2016). NOTE: Color scheme is different in panels A vs B-D. B, Circuit based on one-step RV, which reports powerful input from VPM to fast-spiking (FS) parvalbumin-positive interneurons, as well as Sst- and VIP-expressing interneurons. C, Circuit based on in vivo and in vitro electrophysiology, which shows powerful VPM input to FS interneurons, but not to Sst or VIP interneurons. D, Touch-evoked responses in VPM, barrel cortex (L4 FS and Sst, excitatory (E) neurons are shown, but other layers are similar) and VIP interneurons across layers (cyan) (Biorxiv 554949). Note that Sst neurons are excited long after FS and E neurons, reflecting input from E neurons but not VPM neurons (for consistent data from brain slices see for example (Cruikshank et al., 2010)).

Wall et al find that VPM input to PV and Sst neurons is similar (0.8 vs 0.6 inputs per starter cell; statistically indistinguishable). This finding is difficult to square with neurophysiological experiments. First, CRACM experiments have shown that input from VPM is 20-fold stronger to PV neurons than for Sst neurons (Cruikshank et al., 2010). Second, PV neurons and stellate cells show short-latency PSPs and bursts of action potentials after touch (Biorxiv 554949). Some Sst neurons show PSPs and bursts of action potentials after touch; but these responses are delayed, caused by excitatory drive from L4 stellate cells, but not VPM cells (Figure 2) (see also whole cell recordings in Biorxiv 554949).

Furthermore, Wall et al report even stronger input from VPM to VIP neurons (1.7 inputs per starter cell). However, VIP do not show short-latency touch responses in vivo. Many are inhibited by touch (Biorxiv 554949) (Figure 2D). The electrophysiological and one-step RV data are therefore difficult to reconcile (cf Figure 2B, C).

A few other features of the one-step RV tracing of long-range inputs to barrel cortex interneurons are cause for concern. First, efficiency of spread is low on average (on the order of 1 neuron labeled per input cell). In contrast, individual L4 PV neurons likely receive input from nearly 200 VPM neurons (Bruno and Sakmann, 2006). Second, the efficiency of spread (input cells per starter cell) varies greatly across individual experiments (by a factor of nearly 10). The mechanisms underlying this variability are not understood and these factors could skew RV spread in the context of circuit mapping. Third, across all inputs and interneuron classes, VIP neurons receive the strongest input. This is indicative of cell-autonomous mechanisms related to the starter cell determining the efficiency of viral spread, rather than synaptic connectivity (e.g. more efficient overall production and release of RV particles from VIP neurons compared to other GABAergic interneurons).

2. RV mapping of interlaminar connections

Let’s consider local inputs (i.e. within < 200 micrometers or so) to excitatory starter cells in defined barrel cortex layers (DeNardo et al., 2015). This paper contains many interesting experiments that are not relevant to the discussion here. I focus on local excitatory circuits within a layer, or across neighboring layers. For these local circuits, RV mapping can be directly compared with results from neurophysiological circuit mapping in brain slices.  In short, in these studies RV does not recapitulate the rank order of strength of input based on synaptic connectivity. Below is an illustration of the kind of calculation one needs to do to compare RV tracing and ephys. Skip to the next section (Inputs from GABAergic neurons), unless you are really interested in the details.         

In one experiment, L2 & 3 pyramidal neurons were used as starter cells and produced labeling of L2 & 3 pyramidal neurons (neurons in other layers were also labeled, but I will not treat these because of limitations of the neurophysiological methods in brain slices over longer distances). Based on classical anatomy, axons of L2 and L3 pyramidal neurons both overlap with the dendrites of L2 and L3 pyramidal neurons and connections are expected (I am not aware of exceptions for excitatory neurons). The question is if RV can predict the rank order of connectivity strength between these cell groups.   

In all cases the overall proportion of neurons that were either starter cells or input cells was low. Confusion of real starter cells with neurons that express G but were later labeled via other starter cells is therefore not likely to be a concern.

Based on the proportions of starter cells across L2 & L3, neurophysiology measurements allow us to predict the strength of input to L2 vs L3. This is a prediction of the distribution of input across L2 & L3 produced by this configuration of starter cells. The calculation can be done using measured connection probability, or can in addition take the strength of unitary input into account (since the unitary EPSPS across the neuronal populations is similar, these estimates are nearly identical) (data from (Lefort et al., 2009)) (Figure 3). The predicted ratios of input cells are L2: 0.56 & L3: 0.44 . The measured numbers based on one-step RV are L2: 0.19 & L3: 0.81. RV does not recapitulate the rank order of strength of input based on synaptic connectivity.        

Input strength from one-step RVProportions of starter cellsProportion of input cells of all layersProportion of input to Layer (from L2 & 3)
Layer 20.750.070.19

Connection data from LeFort et al 2009:

L2–>L2: Connection probability, p_22 = 0.09; uEPSP, u_22 = 0.5  

L3–>L2: Connection probability, p_32 = 0.12; uEPSP, u_32 = 0.6

L2–>L3: Connection probability, p_23 = 0.05; uEPSP, u_23 = 0.4

L3–>L3: Connection probability, p_33 = 0.19; uEPSP, u_33 = 0.5

Projection strength from ephysProportions of starter cellsInput predicted by paired recordings (LeFort et al 2009) Proportion of input to Layer (from L2 & 3)
Layer 2Prop. of L2 starter cells [L2]=0.75[L2]*p_22*u_22+[L3]*p_32*u_32=0.75*0.09*0.5+0.23*0.12*0.6=0.050.56
Layer 3Prop. of L3 starter cells [L3]=0.23[L2]*p_23*u_23+[L3]*p_33*u_33=0.75*0.05*0.4+0.23*0.19*0.5=0.040.44

Figure 3. Comparison of intracortical connections measured by one-step RV and electrophysiology in brain slices. Top, figure panel from DeNardo et al 2015, showing starter cells across layers (left) and input cells across layers (right). Orange bars show predicted inputs based on electrophysiology overlaid on the corresponding measurements. Bottom, tables show input strengths from one-step RV (blue) and predicted based on electrophysiology (red) (data based on paired recordings from LeFort et al 2009). If RV measures strength of inputs then the relative strengths of inputs calculated from ephys (red) and RV (blue) should correspond.

In a separate set of experiments, L6 neurons were used as starter cells. RV mapping revealed numerous input cells in L3, suggesting strong input from L3 (Fig 1d, (DeNardo et al., 2015)). However, this connection is known to be  almost undetectable weak (Fig 1g, (DeNardo et al., 2015)) (Hooks et al., 2011; Lefort et al., 2009).

3. Inputs from GABAergic neurons to pyramidal neurons

Finally, studies in the barrel cortex have revealed a qualitative mismatch between one-step RV and electrophysiology at the level of entire cell populations. L1 neurons consist of up to four distinct GABAergic types. These neurons powerfully inhibit L23 pyramidal neurons. This is one of the strongest measured connections in the barrel cortex (Schuman et al., 2019) (high connection probability and IPSP amplitude from two distinct cell classes). Yet RV tracing with L23 pyramidal neuron starter cells labels no input cells in L1 (Fig. 2; Supp Fig. 1 in (Yetman et al., 2019); see also (DeNardo et al., 2015)). This implies a false negative problem at the level of entire cell classes.

Conclusions from RV tracing in BC

RV by itself does not provide a rank order of the strengths of inputs from different neuronal populations and fails to spread to entire cell classes. These comparisons imply that RV has at best WSS.  

The barrel cortex results are consistent with biased non-synaptic spread (i.e. proximity). But it is difficult to be sure. Specificity between cell types is not absolute, but quantitative. For example, Sst neurons in L4 still receive some input from VPM, but it’s functionally 20-fold weaker than for PV neurons (Cruikshank et al., 2010). It is possible that RV spreads across Sst ← VPM synapses much more efficiently than across  PV ← VPM synapses. In my view the assumption should be that, unless proven otherwise, RV spreads by proximity and possibly also via synapses (‘assume the worst, hope for the best’).

It is not straightforward to relate electrophysiological measurements to one-step RV. For example, electrophysiological measurements are less sensitive to electrotonically distant connections; or synapses might be functionally silent due to presynaptic modulation. For example, It is theoretically possible that Sst neurons make many ‘effectively silent’ synapses with VPM axons.  An alternative analysis would compare one-step RV to anatomy, to analyze the neuropil in the vicinity of a single starter cell with high (synapse-level) resolution (i.e. a detailed and realistic view of the blow-up in Figure 1). This would require electron microscopy or super-resolution optical microscopy. Expansion microscopy is particularly interesting, since it can be performed in thick tissues and with standard, high-throughput fluorescence microscopy (Chen et al., 2015). Currently the anatomical data simply do not exist.

What is clear is that too little is known about what the population of ‘input cells’ represents in one-step RV experiments.

Clearly, one-step RV has been used to make interesting discoveries, especially in studies that combine viral tracing with other methods. For example (and there are many others), Beier et al have screened for cocaine-induced changes in transcellular spread from dopaminergic VTA neurons (Beier et al., 2017). They found enhanced labeling of GPe → VTA projections and tracked this down to a cocaine-dependent increase in GPe activity. Here the authors exploited modulation of transneuronal spread by neural activity. Of course, this same modulation is expected to confound studies of connectivity.

Why is one-step RV so widely used? The potential for mapping brain-wide connectivity is alluring. It’s a very clever idea and has a ‘modern’ feel. The patterns of input cells kinda-sorta look right. As noted above, it is not easy to refute a one-step RV pattern because in the vast majority of cases quantitative data on synaptic connectivity (functional or structural) are simply not available. But the truthiness of the input patterns does not imply that they represent a map of synaptic connectivity and they probably should not be reported as such.

What to do?

As we stated in our recent review: ‘Results of rabies tracing should therefore be treated as a roadmap for further studies’ (Luo et al., 2018). One-step RV tracing studies have to be followed up by methods with synaptic sensitivity. This could be electrophysiology or microscopy with synaptic resolution.

But why start with one-step RV tracing in the first place? Some studies would better start with standard retrograde tracing experiments using attenuated RV (Chatterjee et al., 2018), AAV retro (Tervo et al., 2016), CAV2 (Junyent and Kremer, 2015), or some of the excellent synthetic retrograde labels (Tsuriel et al., 2015) (ideally combinations thereof, to deal with the aforementioned tropism issues). These tools are cheaper, easier to use and likely provide less biased data and starting hypotheses about connectivity (i.e. in Figure 1, only processes 4 & 5 would produce cell type-specific biases; processes 1 – 3 are not involved). In some cases it may make sense to use one-step RV as a tool to simply restrict expression of a retrograde tracer to a spatially restricted group of starter cells.  

Importantly, more effort has to go into understanding circuit mapping tools such as one-step RV. Combining one-step RV with electrophysiological methods or super-resolution imaging is key to learn about biased in a more comprehensive manner. The cell biology of the RV life cycle needs more attention. How is RV transported into dendrites and exocytosed, what is its residence time in the extracellular space, and what are the mechanisms of endocytosis into axons and retrograde axonal transport? With better cell biological understanding, synapse specificity could possibly be engineered into viral reagents. Finally, it may be possible to develop high-throughput screens for viruses showing synapse-specificity in dendritic exocytosis and axonal endocytosis.


Beier, K.T., Kim, C.K., Hoerbelt, P., Hung, L.W., Heifets, B.D., DeLoach, K.E., Mosca, T.J., Neuner, S., Deisseroth, K., Luo, L., et al. (2017). Rabies screen reveals GPe control of cocaine-triggered plasticity. Nature 549, 345–350.

Binzegger, T., Douglas, R.J., and Martin, K.A. (2004). A quantitative map of the circuit of cat primary visual cortex. J Neurosci 24, 8441–8453.

Bruno, R.M., and Sakmann, B. (2006). Cortex is driven by weak but synchronously active thalamocortical synapses. Science 312, 1622–1627.

Callaway, E.M., and Luo, L. (2015). Monosynaptic Circuit Tracing with Glycoprotein-Deleted Rabies Viruses. J. Neurosci. Off. J. Soc. Neurosci. 35, 8979–8985.

Card, J.P., and Enquist, L.W. (2001). Transneuronal circuit analysis with pseudorabies viruses. Curr Protoc Neurosci Chapter 1, Unit1 5.

Chatterjee, S., Sullivan, H.A., MacLennan, B.J., Xu, R., Hou, Y., Lavin, T.K., Lea, N.E., Michalski, J.E., Babcock, K.R., Dietrich, S., et al. (2018). Nontoxic, double-deletion-mutant rabies viral vectors for retrograde targeting of projection neurons. Nat. Neurosci. 21, 638–646.

Chen, F., Tillberg, P.W., and Boyden, E.S. (2015). Optical imaging. Expansion microscopy. Science 347, 543–548.

Cruikshank, S.J., Urabe, H., Nurmikko, A.V., and Connors, B.W. (2010). Pathway-specific feedforward circuits between thalamus and neocortex revealed by selective optical stimulation of axons. Neuron 65, 230–245.

DeNardo, L.A., Berns, D.S., DeLoach, K., and Luo, L. (2015). Connectivity of mouse somatosensory and prefrontal cortex examined with trans-synaptic tracing. Nat. Neurosci. 18, 1687–1697.

Dum, R.P., and Strick, P.L. (2013). Transneuronal tracing with neurotropic viruses reveals network macroarchitecture. Curr. Opin. Neurobiol. 23, 245–249.

Gentet, L.J., Kremer, Y., Taniguchi, H., Huang, Z.J., Staiger, J.F., and Petersen, C.C.H. (2012). Unique functional properties of somatostatin-expressing GABAergic neurons in mouse barrel cortex. Nat. Neurosci. 15, 607–612.

Hooks, B.M., Hires, S.A., Zhang, Y.X., Huber, D., Petreanu, L., Svoboda, K., and Shepherd, G.M. (2011). Laminar analysis of excitatory local circuits in vibrissal motor and sensory cortical areas. PLoS Biol 9, e1000572.

Junyent, F., and Kremer, E.J. (2015). CAV-2–why a canine virus is a neurobiologist’s best friend. Curr. Opin. Pharmacol. 24, 86–93.

Kasthuri, N., Hayworth, K.J., Berger, D.R., Schalek, R.L., Conchello, J.A., Knowles-Barley, S., Lee, D., Vázquez-Reina, A., Kaynig, V., Jones, T.R., et al. (2015). Saturated Reconstruction of a Volume of Neocortex. Cell 162, 648–661.

Kerlin, A.M., Andermann, M.L., Berezovskii, V.K., and Reid, R.C. (2010). Broadly tuned response properties of diverse inhibitory neuron subtypes in mouse visual cortex. Neuron 67, 858–871.

Lefort, S., Tomm, C., Floyd Sarria, J.C., and Petersen, C.C. (2009). The excitatory neuronal network of the C2 barrel column in mouse primary somatosensory cortex. Neuron 61, 301–316.

Luo, L., Callaway, E.M., and Svoboda, K. (2008). Genetic dissection of neural circuits. Neuron 57, 634–660.

Luo, L., Callaway, E.M., and Svoboda, K. (2018). Genetic Dissection of Neural Circuits: A Decade of Progress. Neuron 98, 865.

Ma, Y., Hu, H., and Agmon, A. (2012). Short-term plasticity of unitary inhibitory-to-inhibitory synapses depends on the presynaptic interneuron subtype. J. Neurosci. Off. J. Soc. Neurosci. 32, 983–988.

Markram, H., Lubke, J., Frotscher, M., Roth, A., and Sakmann, B. (1997). Physiology and anatomy of synaptic connections between thick tufted pyramidal neurones in the developing rat neocortex. J Physiol 500 ( Pt 2), 409–440.

Mishchenko, Y., Hu, T., Spacek, J., Mendenhall, J., Harris, K.M., and Chklovskii, D.B. (2010). Ultrastructural analysis of hippocampal neuropil from the connectomics perspective. Neuron 67, 1009–1020.

Patterson, M.A., Szatmari, E.M., and Yasuda, R. (2010). AMPA receptors are exocytosed in stimulated spines and adjacent dendrites in a Ras-ERK-dependent manner during long-term potentiation. Proc. Natl. Acad. Sci. U. S. A. 107, 15951–15956.

Petreanu, L., Huber, D., Sobczyk, A., and Svoboda, K. (2007). Channelrhodopsin-2-assisted circuit mapping of long-range callosal projections. Nat Neurosci 10, 663–668.

Pfeffer, C.K., Xue, M., He, M., Huang, Z.J., and Scanziani, M. (2013). Inhibition of inhibition in visual cortex: the logic of connections between molecularly distinct interneurons. Nat. Neurosci. 16, 1068–1076.

Schuman, B., Machold, R.P., Hashikawa, Y., Fuzik, J., Fishell, G.J., and Rudy, B. (2019). Four Unique Interneuron Populations Reside in Neocortical Layer 1. J. Neurosci. Off. J. Soc. Neurosci. 39, 125–139.

Song, S., Sjostrom, P.J., Reigl, M., Nelson, S., and Chklovskii, D.B. (2005). Highly nonrandom features of synaptic connectivity in local cortical circuits. PLoS Biol 3, 1–13.

Tervo, D.G., Hwang, B.Y., Viswanathan, S., Gaj, T., Lavzin, M., Ritola, K.D., Lindo, S., Michael, S., Kuleshova, E., Ojala, D., et al. (2016). A Designer AAV Variant Permits Efficient Retrograde Access to Projection Neurons. Neuron 92, 372–382.

Tsuriel, S., Gudes, S., Draft, R.W., Binshtok, A.M., and Lichtman, J.W. (2015). Multispectral labeling technique to map many neighboring axonal projections in the same tissue. Nat Methods 12, 547–552.

Ugolini, G. (1995). Specificity of rabies virus as a transneuronal tracer of motor networks: transfer from hypoglossal motoneurons to connected second-order and higher order central nervous system cell groups. J. Comp. Neurol. 356, 457–480.

Wall, N.R., De La Parra, M., Sorokin, J.M., Taniguchi, H., Huang, Z.J., and Callaway, E.M. (2016). Brain-Wide Maps of Synaptic Input to Cortical Interneurons. J. Neurosci. Off. J. Soc. Neurosci. 36, 4000–4009.

Wertz, A., Trenholm, S., Yonehara, K., Hillier, D., Raics, Z., Leinweber, M., Szalay, G., Ghanem, A., Keller, G., Rozsa, B., et al. (2015). PRESYNAPTIC NETWORKS. Single-cell-initiated monosynaptic tracing reveals layer-specific cortical network modules. Science 349, 70–74.

Wickersham, I.R., Lyon, D.C., Barnard, R.J., Mori, T., Finke, S., Conzelmann, K.K., Young, J.A., and Callaway, E.M. (2007). Monosynaptic restriction of transsynaptic tracing from single, genetically targeted neurons. Neuron 53, 639–647.

Yetman, M.J., Washburn, E., Hyun, J.H., Osakada, F., Hayano, Y., Zeng, H., Callaway, E.M., Kwon, H.-B., and Taniguchi, H. (2019). Intersectional monosynaptic tracing for dissecting subtype-specific organization of GABAergic interneuron inputs. Nat. Neurosci. 22, 492–502.