by Billy Morris
One of the key benefits of connectomics is that we establish the layout of the nervous system and its constituent elements. Knowing how the elements that make up the brain are connected simplifies the challenge of understanding how the brain processes information, executes behaviour and is modulated by internal factors. However, the key downside to connectomics is that generating a usable dataset is intensely laborious: it would take decades for us to manually reconstruct the elements of a whole brain, even that of an animal as small as the fruit fly. Due to this, great efforts have been made to develop state of the art computational methods to perform this task automatically. This is what has been achieved in the hemibrain dataset published by Xu et al., in collaboration with our team member Philipp Schlegel.
The FIB-hemibrain dataset (Xu et al. 2020) is the most comprehensive insect EM reconstruction to date. It covers approximately a third of the fly brain, encompassing several key regions. These include areas devoted to navigation and orientation (the central complex), memory formation (mushroom body), and higher odor processing (lateral horn).
This study used AI derived flood-filling networks to automatically segment neurons from the image data. In combination with a synapse detection algorithm, predicting roughly 20 million chemical synapses between neurons, this created a directed network of neurons spanning the hemibrain. These efforts subsequently underwent extensive manual curation for quality control, joining fragments that belonged to the same neuron or splitting those that had been incorrectly agglomerated. The dataset was then annotated by human experts who matched each segmented neuron to its specific type; this leveraged one of the strengths of insect neuroscience, as we can identify the exact same types, and even individual neurons, across animals. This identification was achieved by first grouping the neurons by the fibre running to their cell bodies (which reflects a common developmental origin) and then subdividing these groups based on the exact neuronal morphologies and their connections to other neurons. Through these processes the authors identified 20,607 neurons across 4788 cell types, and established the connectivity between each element.
The degree to which this dataset is complete varies across the different regions of the central brain. There had to be some prioritisation in sifting through this vast amount of data. Key regions of especial interest were heavily curated and are estimated to be highly complete, such as the ellipsoid body (~81%), whilst others of lower priority are relatively incomplete, e.g. the inferior clamp (~22%). Whilst this dataset is the most comprehensive of its kind across insects, in some areas where the completeness is lowest the connections between neurons are only a sample, enough to provide an indication of the strongest synaptic partners. This means that there is even more that we can take from this dataset in the future.
This is an incredible piece of work, providing an invaluable resource for neurobiologists to utilise in the coming years. In addition to its intrinsic value, combining this study with our partial, manual reconstruction in another EM dataset, FAFB, enables us to address a key drawback of connectomics studies: that they rely on a sample size of one. We have to assume that the brain we work on is typical, not deformed by any genetic or developmental abnormalities, and errors in reconstruction might go undetected. However, if we can see the same trend in multiple connectomic datasets, we can be more confident that observations in each one are accurate and reflect typical neuroanatomy.
Cross comparison between the two datasets supports the veracity of our conclusions, and tantalises us with the possibility of experimental connectomics, in which we could manipulate genes and observe their effect on brain organisation, in the not-too-distant future.