Dr. Zweig is modeling how gene regulation changes across time and space using advanced geometry, deep learning models, and CRISPR gene-editing technology. With these tools, it is possible to learn how RNA expression evolves in a cell over time and, ideally, predict evolution in regions susceptible to cancerous mutation(s). This work has the potential to apply to many types of cancer but is especially applicable for acute myeloid leukemia (AML). This is in part because there are known “precursor” states—when cells are at higher risk of becoming cancerous—where understanding gene regulation is most impactful, and in part because a common element of AML treatment is a stem cell transplant, which may attack host cells inconsistently in different regions of the same organ.
The project models temporal gene dynamics via stochastic differential equations, parameterizing the drift vector field with linear networks or shallow neural networks to guarantee provable identifiability, trained via differentiation through the adjoint method. Spatial interactions between spot clusters are characterized with graph neural networks, symmetric functions parameterizing summary statistics, and self-attention modules on latent gene embeddings. The latent embeddings are also defined through a variational autoencoder integrating RNA with other modalities.