Overview#

What is FAFB?#

The FAFB (short for “full adult fly brain”) dataset is an serial-section transmission electron microscopy (ssTEM) image data set of an entire female Drosophila melanogaster brain imaged by Zhihao Zheng and Davi Bock (see Zheng et al., 2018). The raw image data can be downloaded from https://temca2data.org/.

From around 2016, a consortium of labs manually reconstructed neurons and their connectivity in this dataset using CATMAID. Published data from this effort is hosted by Virtual Fly Brain. That manual reconstruction effort has now been superseded by the automatic segmentations described below.

FAFB segmentations#

Generally speaking, “segmentation” refers to the automatic labelling of some feature in image data using machine learning algorithms. These features can be neurons but also ultrastructural structures such as synapses.

FlyWire neuron segmentation#

In mid 2020, the Seung and Murthy labs at Princeton made their segmentation of FAFB public through “FlyWire” and many labs have since contributed to proofreading the dataset.

A first version of the proofread dataset was publicly released in July 2023 as version “630”. A second version with additional proofreading will be made available in early 2024 as version “783”. See Dorkenwald et al. and Schlegel et al. for reference.

Note

On a sidenote: the FAFB image data was realigned for the FlyWire segmentation. This means that x/y/z coordinates typically vary by a micron or so between original FAFB (also called “FAFB v14” or just “FAFB14”) and FlyWire (also called “FAFB v14.1” or “FAFB14.1”). Keep that in mind when comparing data.

Deformation fields mapping between FAFB14 and FlyWire have been kindly provided by the Seung lab and can be used via fafbseg (see tutorials).

Google neuron segmentation#

In 2018/19, Peter Li (Google) auto-segmented the FAFB dataset using Google’s flood filling algorithm. See this website for the paper, examples and data. Peter kindly shared skeletons derived from the segmentation with the FAFB tracing community early on. These skeletons were loaded into CATMAID instances by Tom Kazimiers and Eric Perlman.

Synaptic partner predictions#

Building on their own segmentation of synaptic clefts in FAFB (Heinrich et al., 2018), the Funke lab (Janelia Research Campus) produced a synaptic partners prediction for FAFB (Buhmann et al., 2019).

This data is effectively represented by pairs of x/y/z coordinates where one coordinate is pre- and the other one is postsynaptic. In combination with either the Google or the FlyWire segmentation, we can map these connections onto neurons to compile connectivity tables.

The raw data is publicly available: see this repository by the Funke lab. To make it easier for researcher to access, Sven Dorkenwald, Forrest Collman et al. have loaded the data into their CAVE annotation backend.

Bringing it all together#

As the introduction above shows, there are various types of FAFB data (skeletons, meshes, segmentations, synapses, synaptic partners) available on various platforms (CATMAID, FlyWire and other web services) and in two different spaces (FAFB14 and FlyWire/FAFB14.1).

FAFBseg and its R analog (link) provide a single interface to draw from and combine all the FAFB data.

General layout#

Currently, FAFBseg is divided into two main modules:

  • fafbseg.flywire: work with FlyWire segmentation of FAFB

  • fafbseg.google: work with Google segmentation of FAFB

You will find that certain functions have a version for Google and for FlyWire. For example, fafbseg.flywire.locs_to_segments() and fafbseg.google.locs_to_segments() let you map locations to segmentation IDs for the respective data set. That said: at this point the Google segmentation has effectively been superseded by FlyWire and has therefore seen less attention.

Please check out the tutorials for details!