Chem. Engg. Seminar Series: Shaon Chakrabarti

From cancer to single-cell circadian clocks: inferring biology from the statistics of low-dimensional Latent spaces 

Variability between single cells arising from stochastic molecular processes is a defining feature of living systems. This variability often contains information about the underlying biology, hidden in the structure of low-dimensional latent states. A central challenge is to infer these states from noisy, destructive measurements and connect them to cellular function and fate. In today’s talk, I will present work from our lab exploiting the low-dimensional structure of two biological systems — heritable ‘memory’ states in cancer and circular phase variables in circadian oscillations.

In cancer, I will argue that non-genetic variability enables pre-existing latent states to drive survival and therapy tolerance. As inferred from statistical structures in fate correlations, these states are transiently heritable across multiple generations, giving rise to “memory” across cell divisions. I will present ‘Power-Seek’, an algorithm based on Random Matrix Theory that identifies memory genes from a single snapshot scRNA-seq dataset, and briefly discuss a complementary method that uses these genes to accurately reconstruct cellular lineages.

While heritability is of key importance in cancer, inferring the circular geometry (phase) of the latent state is an open challenge in circadian biology. Unlike the cell cycle, circadian phase inference from single-cell RNA levels is significantly more challenging due to low expression of core clock genes, cellular desynchrony, and a very limited set of genes that robustly oscillate across conditions. I will describe our experimental and computational efforts to address these problems, including a new algorithm ‘DECIPHR’, which combines Variational Auto-Encoders with Deep Embedding on circular spaces to infer circadian phase accurately using only 3–4 core clock genes. These examples serve to highlight the importance of learning biologically constrained latent spaces. 

Bio:

Shaon is an Associate Professor at NCBS, Bangalore, with research interests spanning experimental, theoretical and computational biology. His lab is broadly interested in questions on cell proliferation, circadian clocks and drug resistance in cancer. Prior to this, Shaon did a postdoc at Harvard University and Dana Farber Cancer Institute where he worked on cancer evolution. His PhD at the University of Maryland College Park was on the non-equilibrium statistical physics of molecular motors.