I want to thank all my co-authors from UCSF, the Laboratory for Genomics Research, and GSK for making this project a reality. I especially want to express my deepest appreciation for @kampmann.bsky.social , Adam Yala, and Michael Keiser for being such incredible mentors!

Please dive into the preprint to learn more! We hope this work inspires the further development of representation learning methods for dynamic biological phenotypes and that Plexus can be a useful tool for gaining deeper insights from high-content screens. 8/

Projecting the gene knockdown-induced phenotypes onto this axis pointed toward the role of dysregulated excitatory synaptic activity in driving the aberrant activity phenotype. 7/

Finally, we incorporated an iPSC-derived isogenic cell-line model of mutant MAPT to study a disease-relevant aberrant activity phenotype. Leveraging the learned embeddings, we generated a “disease axis” that best linearly explains the differences between wild-type and mutant MAPT phenotypes. 6/

We then showed that we can map relevant signal processing features onto the learned features to enable physiological interpretability, using the knockdown of the gene KCNQ2 as an example. 5/

We then developed a CRISPRi-enabled iPSC-derived astrocyte and neuron co-culture model to study the effect of 52 genetic perturbations on neuronal activity dynamics. Using the learned embeddings, we recovered ~1.5x as many generalizable phenotypes compared to manual feature engineering. 4/

We validate that these learned embeddings outperform signal processing-based manual features and traditional masked autoencoders when used for the linear classification of distinct simulated and experimental neuronal activity phenotypes. 3/

In this project, we developed a self-supervised model we call Plexus, which uses a novel single-cell encoding strategy to efficiently learn patterns of both intrinsic excitibility and network-level neuronal activity measured via calcium imaging. 2/