Abstract:

The emerging technology of single-cell RNA-seq gives access to gene expression measurements for thousands of cells, allowing discovery & characterisation of cell types via clustering. However, the data is confounded by technical variation emanating from experimental errors & cell type-specific biases. Current clustering approaches perform a global normalisation prior to analysing biological signals, which does not resolve missing data or variation dependent on latent cell types.

In this talk, I will discuss an iterative normalisation & clustering method for single-cell gene expression data called BISCUIT (ICML 2016, CELL 2018). The model is formulated as a hierarchical Bayesian mixture model with cell-specific scalings that aid the iterative normalisation & clustering of cells, teasing apart technical variation from biological signals. The approach is superior to global normalisation followed by clustering. Identifiability & weak convergence guarantees & a scalable Gibbs inference algorithm will be presented. This method improves cluster inference in both synthetic & real single-cell data compared with previous methods, & allows easy interpretation & recovery of the underlying structure & cell types.

With the launch of Human Cell Atlas (HCA) & Human Tumor Atlas (HTA) consortia, vast amounts of public single-cell data will be generated in the next decade, presenting opportunities for developing & applying ML & DL techniques appropriate to the complexity of biological systems & the challenges inherent to single-cell data. With this talk, my goal is also to bring awareness & encourage interdisciplinary efforts between theory & application in the Computational Biology domain.

Speaker bio

Sandhya Prabhakaran (https://sandhya212.github.io) is a Research Fellow at Memorial Sloan Kettering Cancer Centre, NYC. Her research deals with developing & applying statistical models to problems in Computational Biology, particularly in analysing both single-cell sequencing & imaging data. She works on clustering, network inference & sparsity selection models. She obtained her Ph.D degree from the Department of Mathematics & Computer Science, University of Basel, Switzerland & her Masters in Artificial Intelligence from University of Edinburgh, Scotland. Sandhya has received the Best Student Paper Award in ACML 2012, Best Paper Award Runner-Up in ICML 2010 & was one of the 23 global recipients of the Scottish International Scholarship 2008. She is also an avid hiker & runner.