Graph neural networks (GNNs) are a potentially powerful method for discovering connectivity in geometrically complex datasets. The DevoWorm group has developed an open-source GNN framework for embryogenetic data called DevoGraph. Developmental GNNs (D-GNNs) allow us to characterize a growing network that undergoes shape transformations along with increases in size. During GSoC 2022, we developed a roadmap for progress in this area, but were not able to develop full integration with our Deep Learning-based pre-trained model (DevoLearn). Ultimately, we aim to tie our D-GNN work into the group’s work on embryo networks, developmental connectomes, and embryo differentiation.
During the project period, you will be involved in three activities: 1) refining a means to segment raw data and incorporate it into the DevoGraph pipeline, 2) refining our method for deriving graph embeddings, using techniques from topological data analysis and complex network theory, and 3) more tightly integrating DevoGraph as a network structure discovery module of DevoLearn. Achieving 1) will require refactoring CNN models and understanding biological training datasets. Activities 2) and 3) require the ability to work with mathematical models and associated algorithms. Knowledge of graph and/or network theory is helpful, but not required.
What can I do before GSoC?
You can ask one of the mentors to direct you to the data source and you can start working on it. Please feel free to join the OpenWorm Slack or attend our meetings to raise questions/discussions regarding your approach to the problem.
OpenWorm Foundation: https://openworm.org/
DevoWorm website: https://devoworm.weebly.com/
DevoGraph (Github): GitHub - DevoLearn/DevoGraph
DevoWorm AI: DevoWorm.AI
Learning on Graphs (LoG) Conference Recap:
http://syntheticdaisies.blogspot.com/2022/12/ learning-on-graphs-log-conference-
Skill level: Advanced
Required skills: All of our existing models are built for PyTorch, so experience with Python and
PyTorch/Tensorflow workflows is preferred. The ability to work with datasets, such as
segmenting video and generating graph visualizations is essential. An ability to build
web interfaces, UI design, basic knowledge of biology, open-source practices, and
applied mathematical tools will also be useful.
Time commitment: Full-time (350 hours)
Lead mentor: Bradly Alicea
Project website: https://devoworm.weebly.com/
Backup mentors: Jiahang Li
Tech keywords: GNNs, Computational Biology, Graph Theory, PyTorch
