What to Expect from Pyr Gameplay

nurro, science, pyr

Two new preprints titled Functional connectomics spanning multiple areas of mouse visual cortex  and Petascale neural circuit reconstruction: automated methods reveal the mindboggling five year endeavor that produced the Pyr dataset and shine light on what to expect from the forthcoming citizen science game.


Pyr volume with two cells connected by a synapse, Amy Sterling, Pyr, citizen science, microns explorer, connectome
Pyr volume showing two proofread cells connected by a synapse. Image: Amy Sterling

The dataset is unique in that it is the most comprehensive functional connectomics resource released to date. It is the largest connectomics dataset by minimum dimension, number of cells, and number of synapses. It contains automated reconstructions of over 200,000 cells and over 524 million synapses coupled with in vivo two-photon calcium imaging of the activity of around 75,000 pyramidal neurons. This technique sends photons into the brain of a mouse whose neurons are genetically modified to express a modified version of a fluorescent protein found in a jellyfish, causing the neurons to literally glow (fluoresce) when calcium (an indicator of synaptic activity) is nearby.  Once proofread, the combined physiological and anatomical circuitry will be “unprecedented in its completeness.”

dendrite, Pyr, electron microscope, microns explorer, microns
Dendrite reconstructed from a tissue volume composed of a stack of electron microscope images.

In the decade since Eyewire launched, the reconstruction AIs have improved considerably, resulting in cells that are automatically mapped out well enough to identify their type. But there are still many errors. For example, as of August 2021 roughly 600 neurons have been extensively proofread by proofreading experts at Princeton’s Seung Lab. Those neurons required 46,241 splits and 38,694 merges to correct errors. We estimate that the entire volume will require something around 80 million edits. The challenge of making these edits will largely be given to citizen scientists playing Pyr. Here’s an idea of what that might look like.

The Pyr dataset was generated by embedding a volume of brain in resin, sectioning the tissue into ~40 nanometer thin slices and imaging each one using electron microscopy (EM). Similarly to Eyewire, those images are then aligned into a 3D volume where machine learning is used to segment out neurons from the tissue. There are 20,000 sections in the 2 TB dataset, which was imaged at ~4 nm per pixel resolution. Folds in these slices cause distortions in the volume and are the primary source of reconstruction errors. Folds look like long black chasms in EM, not entirely dissimilar to the black spills of Eyewire. Tracing thin axons across folds can be quite tricky.

Pyr dataset acquisition and analysis pipeline via MICrONS Consortium

More imaging defects occur both shallower and deeper in the volume, resulting in more errors in the cellular processes found here. Specifically, the surface of the brain, known as pia and the deeper threshold between cerebral cortex and white matter where axons are largely myelinated, will be areas that require more effort and contain challenging reconstructions.

Fun facts (as of Aug 2021):

  • The Pyr volume is roughly 1 cubic mm (1.3 x 0.87 x 0.82 mm^3).
  • The largest proofread excitatory axon measured 22.2 mm (in just 1 mm of tissue!) and formed 1,893 synapses.
  • The largest inhibitory axon formed 10,081 synapses.
  • Pyr contains over 200,000 cells and 75,909 unique functionally-imaged neurons.
  • Pyr contains over 524,000,000 synapses, automatically labeled at the cleft. Each is labeled presynaptic (sending outgoing action potential) and postsynaptic (receiving end of action potential).
  • The volume spans several regions of mouse cortex: primary visual cortex (VISp), lateromedial area (VISIm), anterolateral area (VISal), and rostrolateral area (VISrl).
  • The dataset was generated over 6 months of continuous imaging by The Allen Institute.

pyr, microns, citizen science, pyramidal neurons, amy sterling,


Here are a few perspectives on the MICrONS project from Seung Lab members. 

Sebastian Seung, Princeton Neuroscience Professor, founder of Eyewire, and connectomics pioneer:

We’re back from our five-year mission, and guess what…we’ve brought
back a tiny piece of inner space! Just fire up your web browser, and you
can take a cosmic journey through half a billion synapses in a cubic
millimeter of cerebral cortex.

It’s hard to believe that the MICrONS program has finally drawn to a
close. Our five-year mission has been truly arduous, with an ambitious
goal that many regarded as unattainable. In the first year, a lab member
argued to me that even the Phase 1 pilot would be impossible. Fast
forward to today, and we are releasing a cubic millimeter of
reconstructed cortex, which is 1000 times larger than the Phase 1

I feel privileged to have worked with such an amazing team at Princeton,
and our outstanding partners at the Baylor College of Medicine and the
Allen Institute. We have been rewarded by breathtaking new vistas of the
cortex. As we transition to a new phase of discovery, we are putting our
energies into building a community of researchers who will use the data
in many ways, most of which we cannot predict. Only the collective
efforts of a community can realize the resource’s potential for enabling
new discoveries about the brain.

Will Silversmith, who you may recall as former Eyewire developer sapphiresun: 

The MICrONS dataset is terrifically large. A cubic millimeter of mouse visual cortex that will provide structural insight for future studies for a long time to come. Speaking from my perspective as a software engineer, the technical development that enabled its production has been sharp and swift and has had many fathers and mothers. Engineers and scientists inside the global academy figure prominently, but also the state support, the rapid development and cooperation of cloud computing providers, the contributions of the larger software community that freely provided essential libraries and techniques and support, and the citizen scientists that laboriously and joyously provided the reconstructions for e2198 and Zebrafish who showed that this enterprise was viable.

Initially when MICrONS was announced in 2016, my impression was that it was a literally impossible project to complete. The size of our version of the Eyewire dataset is only about 958 Gigavoxels and had taken painstaking work to assemble and proofread is ongoing. The project was divided into Phases I, II, & III, of which Phase I required assembly of 10% of the final volume or 100 Teravoxels. It was only by radically changing how we approached the problem, using cloud storage, Neuroglancer, neural networks to perform alignment, formalized pipelines, and constant interplay with tracers that this was accomplished. Further work and partnerships were required to scale up another order of magnitude when phase two came along… with quite a bit of learning and multiple attempts to figure out how to handle such huge data. In my own practice, I found the widely available implementations of common algorithms were extremely slow or memory intensive for our workloads and had to write my own from scratch many times.

We still have many problems to solve, but the work MICrONS accomplished puts the field of connectomics in a much stronger technical position to process and reprocess new datasets which seem to be appearing at a surprising clip. A field where everyone knew each others’ name is starting to grow into one where you can be surprised by the innovative work of others you haven’t met. The proofreading tools are still being developed, but have progressed by leaps and bounds, meaning human efforts can be increasingly focused on the hardest problems and whose resolution reconstructs more neuron per click.

It is my hope and expectation that there will come a point when the scientific questions start to be answered with great rapidity as we finish the engineering work and learn how to rapidly extract answers to our questions. I’m confident one day, the work we’re doing will provide the foundation for answering currently unresolved questions about how the brain functions and for creating detailed clinical disease models.

Kai Kuehner, Eyewire developer @kaikue

We are working hard to build a game experience for Pyr that is fun and rewarding. Building on the Neuroglancer project from Google, Pyr will allow for faster, more accurate proofreading by letting players edit entire neurons at once. We’ve designed over 350 virtual items that players can collect via gameplay, with many more to come. Players can use these items to customize their avatar and express their creativity.