Adult-born granule cells improve stimulus encoding and discrimination in the dentate gyrus

  1. Diego M Arribas
  2. Antonia Marin-Burgin
  3. Luis G Morelli

  • Curated by eLife

    eLife logo

    eLife assessment

    This paper is of potential interest to both the hippocampal and computational neuroscience fields because it provides a framework for understanding how adult-born granule cells in the hippocampus contribute to network processing. It contains novel interesting ideas, such as the analysis of input-output transformation by SRM models and the establishment of "greedy networks". However, not all major conclusions are sufficiently supported by the data. The paper demonstrates that mixed networks show better encoding performance than pure networks, but the differences are small and only visible with specific performance metrics. Intuitive explanations are not provided.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Heterogeneity plays an important role in diversifying neural responses to support brain function. Adult neurogenesis provides the dentate gyrus with a heterogeneous population of granule cells (GCs) that were born and developed their properties at different times. Immature GCs have distinct intrinsic and synaptic properties than mature GCs and are needed for correct encoding and discrimination in spatial tasks. How immature GCs enhance the encoding of information to support these functions is not well understood. Here, we record the responses to fluctuating current injections of GCs of different ages in mouse hippocampal slices to study how they encode stimuli. Immature GCs produce unreliable responses compared to mature GCs, exhibiting imprecise spike timings across repeated stimulation. We use a statistical model to describe the stimulus-response transformation performed by GCs of different ages. We fit this model to the data and obtain parameters that capture GCs’ encoding properties. Parameter values from this fit reflect the maturational differences of the population and indicate that immature GCs perform a differential encoding of stimuli. To study how this age heterogeneity influences encoding by a population, we perform stimulus decoding using populations that contain GCs of different ages. We find that, despite their individual unreliability, immature GCs enhance the fidelity of the signal encoded by the population and improve the discrimination of similar time-dependent stimuli. Thus, the observed heterogeneity confers the population with enhanced encoding capabilities.

Article activity feed

  1. Version published to 10.7554/elife.80250 on eLife
  2. eLife

    Author Response

    Reviewer #2 (Public Review):

    The paper by Arribas et al. examines the coding properties of adult-born granule cells in the hippocampus at both single cell and network level. To address this question, the authors combine electrophysiology and modeling. The main findings are:

    Noisy stimulus patterns produce unreliable spiking in adult-born granule cells, but more reliable responses in mature granule cells.

    Analysis of spike patterns with a spike response model (SRM) demonstrates that adult-born and mature GCs show different coding properties.

    Whereas mature GCs are better decoders on the single cell level, heterogeneous networks comprised of both mature and adult-born cells are better encoders at the network level.

    Based on these results, the authors conclude that granule cell heterogeneity confers enhanced encoding capabilities to the dentate gyrus network.

    Although the manuscript contains interesting ideas and initial data, several major points need to be addressed.

    Major points:

    1. The authors use and noisy stimulation paradigm to activate granule cells at a relatively high frequency. However, in the intact network in vivo, granule cells fire much more sparsely. Furthermore, granule cells often fire in bursts. How these properties affect the coding properties of granule cells proposed in the present paper remains unclear. At the very least, this point needs to be better discussed.

    In vivo whole cell recordings of granule cells are very scarce. In our study, we based the design of our stimulus on recordings from the intact network in vivo (PerniaAndrade and Jonas 2014), which show that granule cells receive a wide range of frequencies, with a power spectrum that exhibits a power law decay. These properties are built in our noisy stimuli. These in vivo recordings have also reported the presence of theta oscillations, showing a peak in the spectrum. However, in our approach we deliberately removed these oscillations from our stimuli because it is best to fit GLMs using white noise or noise with an exponentially decaying autocorrelation (Paninski et al. 2004).

    Thus, our choice of the stimuli is far from arbitrary, but rooted on experimental evidence from intact network in vivo recordings, together with previous knowledge about GLM/SRM fitting. This comment reveals to us that we did not clarify this enough in the manuscript. We are grateful to the reviewer for revealing this omission, since this is in fact an important aspect of the study strategy. In the revised manuscript, we brought these points up front in the results section when we introduce the stimulus for the first time, and more thoroughly discussed it in the Methods section that describes the stimulus.

    Still, the bursts observed in granule cells are an important feature and they have been observed to be phase locked to the theta-gamma oscillations in vivo (Pernia-Andrade and Jonas 2014). In the revised version of the manuscript we included new experiments and simulations with stimuli that include a peak in theta frequency. We found that immature neurons also improve decoding performance with these theta modulated stimuli.

    1. The authors induce spiking in granule cells by injection of current waveforms. However, in the intact network, neurons are activated by synaptic conductances. As current and conductance have been shown to affect spike output differently, controls with conductance stimuli need to be provided. Dynamic clamp is not a miracle anymore these days.

    The use of dynamic clamp sounds in principle like a good suggestion. However, in the manuscript we have taken a different approach to enable the use of a single neuron GLM that uses currents as inputs. To control for the differences between mature and immature neurons we used currents with amplitude normalized by the input resistance, and both types of neurons were measured with the same technique to allow for the comparison.

    Importantly, the GLM type model that we use assumes that the membrane potential is a linear convolution of the input, which permits a straightforward and robust fitting approach. We argue that this is not a minor issue, since using dynamic clamp would require a drastic modification of the model. Furthermore, the use of conductance stimuli would not allow for the straightforward model fitting we perform with our approach. The key point here is that the membrane potential would not be correctly approximated as a linear function of the conductance stimulus, precluding the fitting strategy.

    Finally, at the moment we do not have the equipment to perform the suggested experiment, so this suggestion would require a big amount of time to acquire the equipment and set up the experiments in mature and immature neurons. In addition, we would have to change the model and develop a different fitting strategy. With the controls that we already have in the manuscript, we do not think dynamic clamp experiments would fundamentally change the conclusions of the manuscript. Thus, we argue that this is beyond a reasonable timeframe for this revision, but could be something to further explore in future. We now mention this possibility in the discussion.

    1. The greedy procedure is a good idea, but there are several issues with its implementation. First, it is unclear how the results depend on the starting value. What we end up with the same mixed network if we would start with adult-born cells? Second, the size of the greedy network is very small. It is unclear whether the main conclusion holds in larger networks, up to the level of biological network size (1 million). Finally, the fraction of adult-born granule cells in the optimal network comes out very large. This is different from the biological network, where clearly four or five-week-old granule cells cannot represent the majority. Much more work is needed to address these issues.

    The reviewer approves the greedy procedure that we apply in our manuscript and poses three issues for consideration.

    First, the reviewer queries what would be the result of starting the procedure with a different pool of simulated neurons, and whether we would obtain “the same mixed network if we would start with adult-born cells”. Let us remark that the outcome of the greedy procedure is not always the same mixed population of neurons. For each different mature neuron that we use to start the procedure, the trajectory (see Fig. 4A) of selected neurons will be different. Thus, the final population (network) will be different, and this is reflected in the error bars that we obtain in Fig. 4. Presumably, starting with adult-born cells will change the outcome of the greedy procedure. However, note that this is not the point of the approach. The motivation to start with mature neurons is to ask whether adult-born cells can contribute something to decoding, given that mature cells on their own perform better.

    Second, the reviewer questions the size of the population that we reach with the greedy procedure. Note that for the population sizes that we show in the manuscript the decoding performance already begins to saturate, Fig. 4F-H. Furthermore, it is unfeasible to construct a 1M neurons population due to the computational cost –the time it takes to run the algorithm. These two facts motivated us to stop at 12 neurons as it strikes a good balance between computational time and saturation. Importantly, as we expand below, the aim of the greedy procedure simulation is not reconstructing the actual network of the dentate gyrus. Rather, we seek to understand whether immature neurons could improve coding in a population.

    Third, the reviewer observes that the fraction of adult born cells in the reconstructed populations using the greedy procedure are large as compared to the biological network. Again, here note that the aim of the whole in-silico experiment is not to recover the biological network, where other aspects are at play. More simply, we query the possible contribution of adult born cells to coding. In fact, if we obtained the same proportion it would be by chance, since we do not think that adult-born cells in the dentate gyrus are chosen according to the greedy algorithm.

    Still, this comment from the reviewer motivated us to include further simulations of the greedy procedure with constraints. In the revised manuscript we show new results using the greedy procedure, but constraining the fraction of immature neurons in the resulting populations, see Figure 4-figure supplement 2.

    More generally, we think that these comments reveal a possible misunderstanding about the approach, its purpose and the interpretation of the results. The point of the greedy procedure is to show that immature neurons do in fact contribute to improve the decoding, despite being generally worse individually. We do not claim that the population obtained with the greedy procedure faithfully reflects the actual shape of the in vivo network. We are aware that it does not. We see that this may have not been clear in the original version. In the revised version, we now explain the purpose of the greedy procedure when we introduced it. Additionally, we comment on the proportion of immature neurons in the same paragraph.

    1. Likewise, the idea of dynamic pattern separation seems quite nice. However, the authors focus on the differences between mixed and pure networks, which are extremely small. Furthermore, the correlation coefficients of "low", "medium", and "high" correlation groups are chosen completely arbitrarily. A correlation coefficient of 0.99, considered low here, would seem extremely high in other contexts. Whether dynamic pattern separation is possible over a wider range of input correlation coefficients is unclear (see O'Reilly and McClelland, 1995, Hippocampus, for a possible relationship). Finally, aren't code expansion and lateral inhibition the key mechanisms underlying pattern separation? None of these potential mechanisms are incorporated here.

    The reviewer positively appreciates the idea of the pattern separation task that we propose in the manuscript, and poses some questions concerning the extent of the contribution of adult-born neurons.

    We agree that code expansion and lateral inhibition are key mechanisms for pattern separation in the DG, and we do not claim that adult-born neurogenesis is the key mechanism behind pattern separation. Rather, in our work we explore the role of adultborn immature neurons in coding in general, and in pattern separation in particular, given that it’s a commonly attributed function to the DG.

    We note that the correlation in O'Reilly and McClelland 1994 (actually, what they call pattern overlap) is of a very different nature than the one we compute in our work. They compute the overlap between different patterns of activation in a population of neurons, that is the probability that a single neuron is active in two different patterns of activation. In our manuscript we compute the correlation between different continuous time-varying stimuli that stimulate single neurons.

    Importantly, previous work has shown that ablating neurogenesis particularly affects fine spatial discrimination, that is when the separation between patterns is small, but not when it is large (Clelland 2009, Science). Hence, we were actually expecting the impact of adult-born neurons to be important only for relatively large correlation coefficient values.

    In the revised manuscript, we now explain the rationale for the choice of correlation values, both in the main text when we introduce the task, and in the Methods when we set the values for the low, medium and high correlation classes. We also added a sentence to the discussion on pattern separation, bringing in the importance of the ideas of lateral inhibition, code expansion, and the work of O’Reilly 1994.

    1. A main conclusion of the paper is that while mature GCs are better decoders on the single cell level, heterogeneity in mixtures improves coding in neuronal networks. However, this seems to be true only for r^2 as a readout criterion (Fig. 4F). For information, the result is less clear (Fig. 4G). The results must be discussed in a more objective way. Furthermore, intuitive explanations for this paradoxical observation are not provided. Saying that "this is an interesting open question for future work" is not enough.

    This is an interesting point raised by the Reviewer. While r^2 is quantified by comparing the decoded stimuli with the true stimuli, mutual information is related to the uncertainty about the decoding. That is, it quantifies the correspondence between decoded and true stimuli, but does not tell us whether it is a good approximation to it. For example, a decoder could achieve perfect mutual information but result in a poor reconstruction by performing a perfectly scrambled one-to-one mapping of the true stimulus [Schneidman et al. 2003], see also our reply to point [5] by Reviewer #1 above.

    We agree that this is an important point and we realize that it was not clear in the original version of the manuscript. In the revised manuscript we added some sentences to clarify this point.

    1. The authors ignore possible differences in the output of mature and adult-born granule cells in their thinking. If mature and adult-born granule cells had different outputs, this could affect their contributions to the code (either positively or negatively). At the very least, this possibility should be discussed.

    Newborn neurons contact the same targets as mature neurons, born during development: pyramidal cells in CA3, and interneurons in CA3 and the DG. During the maturation, there is a sequence of connectivity with CA3 and within the DG (Toni et a. 2008). At 4 weeks, newborn cells are already contacting their postsynaptic targets. Still, there may be subtle differences in the strength of these connections compared to mature neurons.

    So, although the targets are the same, there may be quantitative differences in the way they contribute to the code. Thus the point raised by the reviewer is interesting, so we decided to discuss it further in the revision.

  3. eLife

    eLife assessment

    This paper is of potential interest to both the hippocampal and computational neuroscience fields because it provides a framework for understanding how adult-born granule cells in the hippocampus contribute to network processing. It contains novel interesting ideas, such as the analysis of input-output transformation by SRM models and the establishment of "greedy networks". However, not all major conclusions are sufficiently supported by the data. The paper demonstrates that mixed networks show better encoding performance than pure networks, but the differences are small and only visible with specific performance metrics. Intuitive explanations are not provided.

  4. eLife

    Reviewer #1 (Public Review):

    In this paper, the authors use patch-clamp recordings of immature (4w and 5w) and mature granule cells (GC) in hippocampal slices to study stimulus-response properties at different cell ages. First analyzing spike trains generated by a fluctuating stimulus, they show that the reliability of spiking responses increases with cell age. They then fit a Spike Response Model (SRM), a type of GLM that translates inputs to membrane potential and then membrane potential to spikes. Using this model they compare the model parameters from different cells. Time constants for the input-voltage filter are faster for the mDGCs than the 4w, with 5w intermediate, and time constants across all cells appear to be faster when reliability is higher. They analyze stimulus reconstruction and stimulus-response information using the recordings and then extend this to pseudo-populations to test how heterogeneous properties contribute to coding efficacy. They find that mixed pools of neurons, including cells of multiple ages, decode stimuli more accurately.

    Overall, this is a cleverly designed study with sound methodology. A major contribution of the paper is demonstrating with precise, quantitative methods how a degree of heterogeneity that naturally arises in neural populations may be beneficial to decoding the stimulus, despite the fact that some of the heterogeneity arises from variability in single cells. This is an intriguing result showing how neural coding and decoding may actually benefit from heterogeneous response properties rather than only be hindered by variation.

    The paper has a couple of weaknesses. First, it is difficult to assess how meaningful the effects that the authors measure are. For example, is a 3% improvement in decoding (Fig. 4H) with mixed populations of GCs substantial? A second issue not currently addressed in the paper is the relative roles of age-dependent variability and within-group variability: how much of the improvement in stimulus decoding/information encoding is achieved by heterogeneity across model parameters that appears in each age group? Further analyses and clarifications in this vein are suggested.

  5. eLife

    Reviewer #2 (Public Review):

    The paper by Arribas et al. examines the coding properties of adult-born granule cells in the hippocampus at both single cell and network level. To address this question, the authors combine electrophysiology and modeling. The main findings are:
    - Noisy stimulus patterns produce unreliable spiking in adult-born granule cells, but more reliable responses in mature granule cells.
    - Analysis of spike patterns with a spike response model (SRM) demonstrates that adult-born and mature GCs show different coding properties.
    - Whereas mature GCs are better decoders on the single cell level, heterogeneous networks comprised of both mature and adult-born cells are better encoders at the network level.

    Based on these results, the authors conclude that granule cell heterogeneity confers enhanced encoding capabilities to the dentate gyrus network.

    Although the manuscript contains interesting ideas and initial data, several major points need to be addressed.

    Major points:
    1. The authors use and noisy stimulation paradigm to activate granule cells at a relatively high frequency. However, in the intact network in vivo, granule cells fire much more sparsely. Furthermore, granule cells often fire in bursts. How these properties affect the coding properties of granule cells proposed in the present paper remains unclear. At the very least, this point needs to be better discussed.

    2. The authors induce spiking in granule cells by injection of current waveforms. However, in the intact network, neurons are activated by synaptic conductances. As current and conductance have been shown to affect spike output differently, controls with conductance stimuli need to be provided. Dynamic clamp is not a miracle anymore these days.

    3. The greedy procedure is a good idea, but there are several issues with its implementation. First, it is unclear how the results depend on the starting value. What we end up with the same mixed network if we would start with adult-born cells? Second, the size of the greedy network is very small. It is unclear whether the main conclusion holds in larger networks, up to the level of biological network size (1 million). Finally, the fraction of adult-born granule cells in the optimal network comes out very large. This is different from the biological network, where clearly four or five-week-old granule cells cannot represent the majority. Much more work is needed to address these issues.

    4. Likewise, the idea of dynamic pattern separation seems quite nice. However, the authors focus on the differences between mixed and pure networks, which are extremely small. Furthermore, the correlation coefficients of "low", "medium", and "high" correlation groups are chosen completely arbitrarily. A correlation coefficient of 0.99, considered low here, would seem extremely high in other contexts. Whether dynamic pattern separation is possible over a wider range of input correlation coefficients is unclear (see O'Reilly and McClelland, 1995, Hippocampus, for a possible relationship). Finally, aren't code expansion and lateral inhibition the key mechanisms underlying pattern separation? None of these potential mechanisms are incorporated here.

    5. A main conclusion of the paper is that while mature GCs are better decoders on the single cell level, heterogeneity in mixtures improves coding in neuronal networks. However, this seems to be true only for r^2 as a readout criterion (Fig. 4F). For information, the result is less clear (Fig. 4G). The results must be discussed in a more objective way. Furthermore, intuitive explanations for this paradoxical observation are not provided. Saying that "this is an interesting open question for future work" is not enough.

    6. The authors ignore possible differences in the output of mature and adult-born granule cells in their thinking. If mature and adult-born granule cells had different outputs, this could affect their contributions to the code (either positively or negatively). At the very least, this possibility should be discussed.

  6. eLife

    Reviewer #3 (Public Review):

    This is a highly interesting paper that comprehensively investigates the electrophysiological properties of granule cells in the dentate gyrus at different developmental stages. Using state-of-the-art in vitro electrophysiological techniques, the authors record granule cell responses to fluctuating current injections to study how they encode stimuli. The authors find that while immature granule cells produce less reliable stimulus responses and worse stimulus representations than mature cells (8wks and older), cell populations containing neurons of mixed ages improve overall stimulus reconstruction. These data suggest that the cellular diversity contributed by immature granule cells could be beneficial for transmitting distinct properties of stimuli with rich temporal structure, potentially improving the cellular process of pattern separation.
    Major strengths of the paper lie in the precise age determination of immature neurons in Ascl1-CreERT2-Tom mice, recordings of immature neurons, which are rare in in vivo and in vitro studies, precise control over cell-intrinsic properties by blocking excitatory and inhibitory inputs in vitro, and characterization of encoding properties using a spike response model (SRM).

    The conclusions drawn are supported by the data, and the results are likely of great interest to a specialist community of hippocampal electrophysiologists.

  7. Version published to 10.1101/2022.05.12.491644v1 on bioRxiv