Research
Low-dimensional representation learning and data visualization
We develop self-supervised learning methods that automatically learn structure in large datasets and represent them in low-dimensional spaces. A key goal is to produce low-dimensional representations and visual maps that help researchers explore complex genomic, image, or text data. As explorers of the 15th century required compasses and astrolabes to chart the surface of the Earth, researchers today require tools to construct maps, make atlases, and create charts to navigate their data and their models. Our ambition is to provide them with appropriate tools.
Our work includes developing contrastive-learning approaches for visualizing image datasets, neighbor embeddings for graph representation learning, and multimodal embeddings that integrate various single-cell data modalities in a shared space. In parallel, we study theoretical foundations of low-dimensional representations, aiming to understand when such visualizations can faithfully reflect the underlying data structure.
Computational biology and single-cell data
A major focus here is developing statistically grounded analysis approaches for single-cell RNA-seq data, improving common tasks such as normalization, dimensionality reduction, clustering, and visualization.
We also study topological methods to detect complex structures in single-cell datasets in the presence of high-dimensional noise. Furthermore, we develop machine-learning approaches that integrate multimodal biological measurements, such as gene expression, morphology, and neurophysiology, to build unified representations of cellular states.
Selected publications
Unsupervised visualization of image datasets using contrastive learning
In this paper, we combined ideas from the field of self-supervised learning with ideas from the field of manifold learning to develop t-SimCNE: the first ever algorithm for 2D visualization of image datasets.
Analytic Pearson residuals for normalization of single-cell RNA-seq UMI data
In this paper, we suggested a procedure for normalization and transformation of scRNA-seq UMI counts based on the statistical model of count noise. Our algorithm has been implemented in popular ScanPy library.
Tracking excess mortality across countries during the COVID-19 pandemic with the World Mortality Dataset
Since the start of the Covid-19 pandemic, we have been maintaining the World Mortality Dataset collecting weekly and monthly all-cause mortality numbers from countries across the world. Our data and our estimates of excess mortality have been incorporated into aggregators like Our World in Data, and have been widely used.
Phenotypic variation of transcriptomic cell types in mouse motor cortex
We related phenotypic and transcriptomic variation of single neurons in mouse motor cortex using experimental technique called Patch-seq.
The art of using t-SNE for single-cell transcriptomics
This paper describes best practices for using 2D embeddings in single-cell biology. Some of our recommendations have since then been made default in popular Python implementations.