Connectomics of the Octopus vulgaris vertical lobe provides insight into conserved and novel principles of a memory acquisition network

  1. Flavie Bidel
  2. Yaron Meirovitch
  3. Richard Lee Schalek
  4. Xiaotang Lu
  5. Elisa Catherine Pavarino
  6. Fuming Yang
  7. Adi Peleg
  8. Yuelong Wu
  9. Tal Shomrat
  10. Daniel Raimund Berger
  11. Adi Shaked
  12. Jeff William Lichtman
  13. Binyamin Hochner

  • Curated by eLife

    eLife logo

    eLife assessment

    This extraordinary study mapped the circuit architecture of a brain module for learning and memory in the octopus brain. In particular, one gyrus of the Octopus vulgaris brain vertical lobe was imaged with volume electron microscopy, its neurons reconstructed and their synapses mapped. The acquisition of this pioneering data set was followed by a very convincing analysis of the circuits supporting learning and memory, and therefore behavioral plasticity, in this animal. The data and findings establish an important point of comparison with analogous brain structures in other organisms, such as the vertebrate cerebellum and the arthropod mushroom body, offering a new neural circuit architecture to support the study of behavior and inspire the design of artificial neural networks.

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

Here, we present the first analysis of the connectome of a small volume of the Octopus vulgaris vertical lobe (VL) , a brain structure mediating the acquisition of long-term memory in this behaviorally advanced mollusk. Serial section electron microscopy revealed new types of interneurons, cellular components of extensive modulatory systems, and multiple synaptic motifs. The sensory input to the VL is conveyed via~1.8 × 10 6 axons that sparsely innervate two parallel and interconnected feedforward networks formed by the two types of amacrine interneurons (AM), simple AMs (SAMs) and complex AMs (CAMs). SAMs make up 89.3% of the~25 × 10 6 VL cells, each receiving a synaptic input from only a single input neuron on its non-bifurcating primary neurite, suggesting that each input neuron is represented in only~12 ± 3.4SAMs. This synaptic site is likely a ‘memory site’ as it is endowed with LTP. The CAMs, a newly described AM type, comprise 1.6% of the VL cells. Their bifurcating neurites integrate multiple inputs from the input axons and SAMs. While the SAM network appears to feedforward sparse ‘memorizable’ sensory representations to the VL output layer, the CAMs appear to monitor global activity and feedforward a balancing inhibition for ‘sharpening’ the stimulus-specific VL output. While sharing morphological and wiring features with circuits supporting associative learning in other animals, the VL has evolved a unique circuit that enables associative learning based on feedforward information flow.

Article activity feed

  1. Version published to 10.7554/elife.84257 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    This paper raises an interesting question about learning signals. The most intriguing property of this system is the one-to-one convergence, plasticity, and apparently linear input/output function of the SFL-to-SAM relay. These properties suggest that, unlike structures like the insect mushroom body or mammalian cerebellum, in which the intermediate layer is thought to increase the dimensionality of the representation, the SAMs should be thought of more like the weights of a linear readout of the SFL inputs by the LNs.

    What learning signal guarantees appropriate weight changes? In a few places (the section on "associativity" and the section on AFs), it is suggested that SAMs can themselves, through coordinated local activity, cause LTP, which the authors call "self LTP-induction." But what is the purpose of such plasticity? It doesn't seem like it would permit, for example, LTP which associates a pattern of SFL activity with the appropriate LNs for the correct vs. the incorrect action. Presumably, appropriately routed information from the NMs and AFs sends the appropriate learning signals to the right places. Does the pattern of innervation of NMs and AFs reveal how these signals are distributed across association modules? Does this lead to a prediction for the logic of the organization of the association modules?

    We extended the discussion section to clarify some of these points. One paragraph describes our idea of “self-LTP induction” (L712-744). In addition, we address the potential role of the neuromodulatory fibers (NMs) and ascending fibers (AFs) in a paragraph titled "Perspective on the involvement of the ascending fibers (AF) and the neuromodulatory fibers (NM) in the supervision of learning" (L786). Answering how these signals manifest across different association modules requires a larger reconstruction.

    One challenge for a reader who is not an expert on the VL is that the manuscript in its present form lacks discussion about the impact (or hypothesized impact) of the VL on behavior. There is a reference to a role for LNs suppressing attack behavior, but a more comprehensive picture of what the readout layer of this system is likely controlling would be helpful.

    To contextualize how the VL circuitry can allow for the coincident detection of visual stimuli and environmental cues (punishing or rewarding) to control the stereotypic attack behavior of the octopus, we added two discussion sections: "Perspective on the VL involvement in octopus associative learning" (L774) and "Perspective on the involvement of ascending fibers (AF) and neuromodulatory fibers (NM) in the supervision of learning" (L786).

    The authors do a thorough job of characterizing the "fan-out" architecture from SFL axons to SAMs and CAMs. A few key numbers remain to characterize the "fan-in" architecture of LNs. There appears to be a 400:1 convergence from AMs to LNs. Is it possible to estimate the approximate number of presynaptic inputs per LN? The text around Figure 7 states a median of 162 sites per 100μm dendrite length. One could combine this with an estimate of the total dendritic length for one of these cells from previously available data to estimate the number of inputs per LN. This would help determine the degree of overlap of different association modules in Figure 11, which would be interesting from a computational perspective.

    Due to the limitations associated with a small EM volume, our study focused on the fanout of the VL network. We agree that a better understanding of the fan-in part of the VL network is crucial. To the best of our knowledge, previous data have not provided estimates on the dendritic length of the LNs, due to low-resolution images or lack of 3D imaging (Hochner/Shomrat/Young experiment). We intentionally avoided making a largely inaccurate estimate of the fan-in part of the network based on our data. We believe that future research can aim to combine neuron labeling with EM, or other super-resolution techniques, to allow for detailed assessment of the large neuron arborization.

    Reviewer #2 (Public Review):

    Octopuses are known for their abilities in solving complex tasks and numerous apparently complex cognitive behaviours such as astonishment at octopuses learning how to open jars by watching others and the mind-boggling camouflage. They are very clever molluscs. The octopus shows the famously advanced brain plan but it is one that has little research progress due to its large size and structural complexity. This was originally recognised by the work of BB Boycott, JZ Young, EG Gray, and others in mid last century. Since then, however, little progress has been achieved towards a modernday description of the octopus neural network particularly in the higher-order brain lobe, despite intense interest and indeed research progress concerning their complex behavioural and cognitive abilities.

    This study applied a combination of EM-based imaging, neural tracing, and analyses to start revealing a further detailed view of a part of the lateral gyrus of the vertical lobe (learning and memory centre) of the common European octopus. It is a long overdue contribution and starts to bring octopus neuroscience a step close to the details of some vertebrates achieved. The new findings of neurons and the associated network provide new insights into this very complex but unfamiliar brain, allowing to propose a functional network that may link to the octopus memory formation. Also, this work could be of potential interest to a broad audience of neuroscientists and marine biologists as well as those in bio-imaging and deep-learning fields.

    Strengths:

    Current knowledge of the neuroanatomy and the associating network of the octopus vertical lobe (learning and memory centre) remains largely based on the pioneering neuroanatomical studies in the '70s, this work indeed provides a rich and new dataset using modern-day imaging technology and reveals numerous previously-unknown neuron types and the resulting further complex network than we thought before. This new dataset reveals hundreds of cell processes from seven types of neurons located in one gyrus of the vertical lobe and can be useful for planning further approaches for advanced microscopy and other approaches including electrophysiological and molecular studies.

    Another strength of this study is to apply the current fashion of the deep learning technique to accelerate the imaging process on this octopus complex neural network. This could trigger some inventions to develop new algorithms for further applications on those non-model animals.

    Weakness/limitations:

    In an effort to match the key claims of the first connectome of the octopus vertical lobe, mapping up an entire vertical lobe is essential. However, also understandably, given challenges in imaging a large-sized brain region, this study managed to image a very small proportion of the anterior part of the lateral gyrus. Along with the current limited dataset, a partially reconstructed neural network of one gyrus, it is unclear whether the wiring pattern found in this study would appear as a similar arrangement throughout an entire lateral gyrus. Furthermore, it is also unknown if another 4 gyri might keep a similar pattern of neural network as it found in the lateral gyrus. Considering some recent immunochemistry evidence that showed distinct different signals in different gyri in terms of heterogeneity of neuron types amongst gryi, to assume this newly discovered network can represent the wiring pattern across an entire 5-gyrus vertical lobe is inadequate.

    We revised the introduction (L106-113) to address this important point, and added discussion section titled "How well does a partial connectome of a small portion of one VL lateral lobule represent the connectivity patterns across all five VL lobuli?” (L894). We clarify what we believe is likely conserved across VL gyri and what is more likely to differ.

    As this study is the first big step to reveal the complex network in the octopus vertical lobe system, the title may be changed to "Toward the connectome of the Octopus vulgaris vertical lobe - new insights into a memory acquisition network".

    We appreciate the reviewer's suggestion for a new title for our manuscript. We feel, however, that the current title reflects the scope of our work and the significant step it makes toward understanding the neuronal network of the octopus vertical lobe. We do not claim to provide the octopus' VL connectome but how its connectomics unravels its workings and underlying principles. After deliberations, we decided to leave the title unchanged.

  3. eLife

    eLife assessment

    This extraordinary study mapped the circuit architecture of a brain module for learning and memory in the octopus brain. In particular, one gyrus of the Octopus vulgaris brain vertical lobe was imaged with volume electron microscopy, its neurons reconstructed and their synapses mapped. The acquisition of this pioneering data set was followed by a very convincing analysis of the circuits supporting learning and memory, and therefore behavioral plasticity, in this animal. The data and findings establish an important point of comparison with analogous brain structures in other organisms, such as the vertebrate cerebellum and the arthropod mushroom body, offering a new neural circuit architecture to support the study of behavior and inspire the design of artificial neural networks.

  4. eLife

    Reviewer #1 (Public Review):

    This manuscript presents a fascinating "connectome" dataset of the Octopus vulgaris vertical lobe (VL), a brain region involved in learning and memory with a unique structure. It presents the cell types and connectivity of several major classes of cells in this region. One of the most notable findings is that the most numerous neurons, the SAMs, receive only one synaptic input, while another much less numerous class, the CAMs, receive many. Both of these feed onto an output layer of neurons named LNs. This organization is strikingly different from many other associative learning areas in other species.

    Overall, the paper presents an interesting and important collection of anatomical results that will be of interest to those working on this system, as well as (at least at first glance) related systems like the insect mushroom body or mammalian cerebellum. The authors do a good job of highlighting the key properties of this system and contrasting them to other systems. My detailed suggestions are largely about the presentation, but I do have some conceptual comments.

    This paper raises an interesting question about learning signals. The most intriguing property of this system is the one-to-one convergence, plasticity, and apparently linear input/output function of the SFL-to-SAM relay. These properties suggest that, unlike structures like the insect mushroom body or mammalian cerebellum, in which the intermediate layer is thought to increase the dimensionality of the representation, the SAMs should be thought of more like the weights of a linear readout of the SFL inputs by the LNs. What learning signal guarantees appropriate weight changes? In a few places (the section on "associativity" and the section on AFs), it is suggested that SAMs can themselves, through coordinated local activity, cause LTP, which the authors call "self LTP-induction." But what is the purpose of such plasticity? It doesn't seem like it would permit, for example, LTP which associates a pattern of SFL activity with the appropriate LNs for the correct vs. the incorrect action. Presumably, appropriately routed information from the NMs and AFs sends the appropriate learning signals to the right places. Does the pattern of innervation of NMs and AFs reveal how these signals are distributed across association modules? Does this lead to a prediction for the logic of the organization of the association modules?

    One challenge for a reader who is not an expert on the VL is that the manuscript in its present form lacks discussion about the impact (or hypothesized impact) of the VL on behavior. There is a reference to a role for LNs suppressing attack behavior, but a more comprehensive picture of what the readout layer of this system is likely controlling would be helpful.

    The authors do a thorough job of characterizing the "fan-out" architecture from SFL axons to SAMs and CAMs. A few key numbers remain to characterize the "fan-in" architecture of LNs. There appears to be a 400:1 convergence from AMs to LNs. Is it possible to estimate the approximate number of presynaptic inputs per LN? The text around Figure 7 states a median of 162 sites per 100μm dendrite length. One could combine this with an estimate of the total dendritic length for one of these cells from previously available data to estimate the number of inputs per LN. This would help determine the degree of overlap of different association modules in Figure 11, which would be interesting from a computational perspective.

    This is an exciting and intriguing set of results that contributes significantly to our knowledge about the brain regions that control learning and memory.

  5. eLife

    Reviewer #2 (Public Review):

    Octopuses are known for their abilities in solving complex tasks and numerous apparently complex cognitive behaviours such as astonishment at octopuses learning how to open jars by watching others and the mind-boggling camouflage. They are very clever molluscs. The octopus shows the famously advanced brain plan but it is one that has little research progress due to its large size and structural complexity. This was originally recognised by the work of BB Boycott, JZ Young, EG Gray, and others in mid last century. Since then, however, little progress has been achieved towards a modern-day description of the octopus neural network particularly in the higher-order brain lobe, despite intense interest and indeed research progress concerning their complex behavioural and cognitive abilities.

    This study applied a combination of EM-based imaging, neural tracing, and analyses to start revealing a further detailed view of a part of the lateral gyrus of the vertical lobe (learning and memory centre) of the common European octopus. It is a long overdue contribution and starts to bring octopus neuroscience a step close to the details of some vertebrates achieved. The new findings of neurons and the associated network provide new insights into this very complex but unfamiliar brain, allowing to propose a functional network that may link to the octopus memory formation. Also, this work could be of potential interest to a broad audience of neuroscientists and marine biologists as well as those in bio-imaging and deep-learning fields.

    Strengths:
    Current knowledge of the neuroanatomy and the associating network of the octopus vertical lobe (learning and memory centre) remains largely based on the pioneering neuroanatomical studies in the '70s, this work indeed provides a rich and new dataset using modern-day imaging technology and reveals numerous previously-unknown neuron types and the resulting further complex network than we thought before. This new dataset reveals hundreds of cell processes from seven types of neurons located in one gyrus of the vertical lobe and can be useful for planning further approaches for advanced microscopy and other approaches including electrophysiological and molecular studies.
    Another strength of this study is to apply the current fashion of the deep learning technique to accelerate the imaging process on this octopus complex neural network. This could trigger some inventions to develop new algorithms for further applications on those non-model animals.

    Weakness/limitations:
    In an effort to match the key claims of the first connectome of the octopus vertical lobe, mapping up an entire vertical lobe is essential. However, also understandably, given challenges in imaging a large-sized brain region, this study managed to image a very small proportion of the anterior part of the lateral gyrus. Along with the current limited dataset, a partially reconstructed neural network of one gyrus, it is unclear whether the wiring pattern found in this study would appear as a similar arrangement throughout an entire lateral gyrus. Furthermore, it is also unknown if another 4 gyri might keep a similar pattern of neural network as it found in the lateral gyrus. Considering some recent immunochemistry evidence that showed distinct different signals in different gyri in terms of heterogeneity of neuron types amongst gryi, to assume this newly-discovered network can represent the wiring pattern across an entire 5-gyrus vertical lobe is inadequate. As this study is the first big step to reveal the complex network in the octopus vertical lobe system, the title may be changed to "Toward connectomics of the Octopus vulgaris vertical lobe - new insights of memory acquisition network".

  6. eLife

    Reviewer #3 (Public Review):

    The manuscript 'Connectomics of the Octopus vulgaris vertical lobe provides insight into conserved and novel principles of a memory acquisition network' by Bidel et al. uncovers the connectivity of the vertical lobe (VL) of the octopus' central brain. Using serial section electron microscopy, the authors report several cell types and connectivity patterns consistent with their previous work and the classic work of Young and Gray. They also uncover novel cell types, including a set of complex amacrine cells (CAMs), with far less abundance compared to simple amacrine cells (SAMs). Importantly, CAMs are proposed to be GABAergic and inhibitory and plausibly suggested to be involved in pattern sharpening - while SAMs are cholinergic and excitatory. SAMs receive single inputs from diverging SFL input, while CAMs receive multiple afferent inputs and additionally pool inputs from SAMs. Both SAMs and CAMs converge onto LNs that form the output layer of the VL. Finally, the authors describe putative neuromodulatory connections.

    This study is equally impressive as important - using high-resolution anatomy it uncovers putative computational motifs at high resolution. The described network reveals a novel computational logic and highlights how different biological computational networks can be made up. Indeed, comparison to the Drosophila mushroom bodies - a structure following a fan-out, fan-in logic - will allow more in-depth cross-species comparisons in the future, both regarding commonalities and differences in network architecture. Importantly, this study additionally describes, at high resolution, synaptic motifs (palms) that appear quite different from motifs in other systems, including putative direct feedforward connections via SAMs to CAMs and organelle distributions.

  7. Version published to 10.1101/2022.10.03.510303v2 on bioRxiv
  8. Version published to 10.1101/2022.10.03.510303v1 on bioRxiv