Talk Description

Bayesian nonparametric models provide a formal mechanism for encoding probabilistic assumptions about the data generation process where the dimension of the latent space is unknown a priori or may grow with additional samples. A common limitation of these models is that posterior inference is computationally intensive, particularly for nonconjugate models or when integrating over combinatorial structures. In this talk, I will introduce two hierarchical Bayesian nonparametric models and inference algorithms that scale to large genomics and transcriptomics data.

First, I will present our genetic sequence clustering model based on fragmentation coagulation processes and how we scale nonconjugate model inference using maximization-expectation. Then, I will describe our mixed-membership model for alternatively spliced transcript discovery with explicit sparsity and inference algorithms based on stochastic variational inference. I will demonstrate the advantages of our Bayesian nonparametric approach when compared to state-of-the-art methods on simulated and experimental data. Time permitting, I will describe recent extensions to our alternative splicing methods that integrate graph theoretic algorithms with probabilistic modelling.