(14/14) Thanks to Allyson Q Ryan whose ToSkA method inspired SpinePy. Thanks to @jesseveenvliet.bsky.social for supervision, @campaslab.bsky.social & @carldmodes.bsky.social for input and guidance, the @mpi-cbg.de & @poldresden.bsky.social and the #EIC for funding @sumoconsortium.bsky.social

(13/14) Thank you to everyone involved, especially my fellow PhD students Alba Villaronga Luque and @mtrani.bsky.social for sharing their data and contributing to the development but also @yonitmms.bsky.social the rest of our lab for their input and help with annotation.

(12/14) If you want to find out more about the synthetic data generation, see comparisons to 2D patterning approaches, or just want to dive into the details give the preprint a read and contact me or @jesseveenvliet.bsky.social if you’re interested in using the method!

(11/14)
Building on this we can identify features that correlate with the observed patterning classes. The 3D patterning maps thus provide framework to systematically investigate gastruloid patterning spaces!

(10/14) Alba Villaronga Luque generously shared her data investigating how initial cell number (N0) influences patterning. We extracted patterning maps of 54 gastruloids and were able to detect distinct patterning phenotypes using dimensionality reduction and clustering.

(9/14) With this normalized reference frame in hand we can use data with multiple developmental markers and quantify patterning maps using SLIC segmentation. Maps are a 2D representation of 3D patterning data, allowing comparison of patterning types in a unified reference frame.

(8/14) A core feature of SpinePy is the generation of a common reference frame, no matter the gastruloids size, shape, or orientation under the microscope. We define the AP and core-to-surface (CS) position and can use it to map the signals into a common space!

(7/14) Another neat feature is the quantification of scalar fields along the axis. We use density as an example but any signal of interest could be used to generate profiles along the AP axis. Again, the synthetic data facilitates verification of the pipeline!

(6/14) One core biological feature to quantify is gross morphology. Using optimized, non-intersecting planes, we measure radial profiles over the AP axis for fixed and live gastruloids giving insights into 3D morphodynamics. Big thanks to @mtrani.bsky.social who provided data and insight!

(5/14) This allowed me to create hundreds of gastruloids in silico, with defined spines and thickness profiles to benchmark SpinePy. We show that the spine detection performs well with low relative errors compared to the ground truth!

(4/14) Using manual annotations to benchmark the spine is a crucial but in 3D annotation can be challenging. To have a second method of verification I generated synthetic gastruloids with some Perlin noise to generate realistic structures (like this wobbly gastruloid)

(3/14) The first step is detecting the anteroposterior (AP) axis in gastruloids, based on gross morphology. We can either use a skeletonization approach using segmentation or a surface mesh approach using Non-linear PCA, which was important for timelapse data.

(2/14) Its modular architecture allows it to plop into pre-existing workflows. This means you can map any signal of interest into a common reference frame across gastruloids!

🙌🏻🙌🏻🙌🏻