Jens Lagergren: Advancements in Variational Inference-Based Phylogenetic Analysis: From VaiPhy to VICTree

Abstract

In this seminar, we will consider the evolving landscape of variational inference (VI) based methods for phylogenetic analysis, a domain gaining prominence in machine learning and computational biology, e.g., in studying cancer development from single-cell data. Probabilistic phylogenetic inference is a machine-learning problem, and the key to unlocking this crucial application can be expected to be found among the machine-learning methodologies. In particular, VI is a strong candidate, considering the general capacity of the VI methodology to deliver impressive performance gains for Bayesian inference, e.g., compared to MCMC.

First, we introduce VaiPhy, a novel VI-based algorithm designed for rapid and efficient approximate posterior inference in augmented tree space. Moreover, we briefly discuss the emerging field of VI-based phylogenetic inference. Next, we focus on VICTree, another novel variational inference-based algorithm. VICTree addresses the complexities in analyzing tumor heterogeneity and cancer evolution in the context of single-cell sequencing. It incorporates a Tree-structured Mixture Hidden Markov Model (TSMHMM) for modeling copy number evolution, effectively handling the inherent noise and dependencies in copy number profiles. By examining VICTree's application to multiple myeloma and breast cancer samples, we will highlight its capabilities in reliable clustering, clonal tree reconstruction, and copy number evolution analysis, along with its advantages in quality and speed of inference compared to other methods.