Research

We develop the next generation of machine learning (ML) methods for addressing fundamental problems in biology.

Our major motivation is the rapid development of new biological technologies – including single-cell/spatial sequencing and CRISPR – which measure diverse molecular modalities at unprecedented throughput and resolution. ML methods are essential for analyzing and interpreting these large, high-dimensional, and multi-modal biological datasets. However, standard “off-the-shelf” ML methods are severely challenged by the high noise, sparsity, heterogeneity, and other limitations of modern technologies.

We address these challenges by developing ML methods that can extract meaningful biological insights from noisy biomedical data. Our research involes developing both new mathematical theory and practical algorithms, depending on the problem domain. We draw on techniques from many different disciplines including deep learning, graph theory, statistical inference, and complex analysis.

ML for spatial biology

Next-generation spatial sequencing technologies measures both high-throughput cellular measurements (e.g. gene expression) and the spatial location of measured cells. We develop spatial ML methodologies that model the latent geometric structure of biological tissues from sparse spatial sequencing data. Our goal is to quantify cellular dynamics over space and time.

Key words: spatial transcriptomics, cell-cell interactions, spatial gradients, deep learning, copula, time-series analysis

Gene expression topography: a new mathematical and deep learning framework for learning topographic maps of 2-D tissue slices which reveal spatial gradients and tissue geometry.
- GASTON (Nat Meth 2025) (neural field model) and GASTON-Mix (Bioinformatics 2025) (spatial mixture-of-experts, MoE)
- Belayer (Cell Syst 2022): introduced complex analysis to spatial biology.
Cellular interactions: statistical/causal inference methods for learning cell-cell interactions from sparse spatial transcriptomics data.
- Copulacci (Bioinformatics 2024): using bivariate copula distributions to learn local cellular interactions.
- GLACIER (RECOMB-seq 2025) (w/ Rohit Singh lab): DAG-based Granger causality and gene expression topography to identify global gene-gene correlation patterns.

Biological networks

Biological networks underlie many aspects of human health and disease, but existing networks are highly incomplete (missing ≈90% of edges). We develop statistical methodologies for inferring biological interaction networks from noisy data (e.g. genetic perturbations or mutation). We have also developed theoretical approaches for analyzing networks in biology and other disciplines.

Key words: graphs/networks, hypergraphs, epistasis, multivariate Bernoulli, genetic perturbations

DIALECT (RECOMB-CCB 2024): a latent variable model for learning interactions between cancer driver genes using somatic mutation data.
A multivariate Bernoulli-based framework (Nat Comms 2025) that unifies the different approaches for quantifying higher-order (>3-way) genetic interactions (epistasis).
Mathematical frameworks for stochastic processes on hypergraphs (ICML 2019) and social networks (WSDM 2020).

Anomaly detection

Anomalous interactions between genes/proteins underlie many complex diseases. We develop methods for anomaly detection in networks and other structured data (e.g. time series, matrices).

Key words: protein-protein interaction (PPI) network, network anomaly detection, scan statistics, statistical bias

NetMix (RECOMB/JCB 2020) and NetMix2 (RECOMB/JCB 2022): asymptotically unbiased combinatorial optimization algorithms for anomaly detection in protein-protein interaction (PPI) networks.
- We prove that many existing network anomaly detection methods are statistically biased. By correcting this bias, we achieve state-of-the-art performance in disease gene identification.
A statistical framework for structured anomaly detection (ICML 2021) that unifies existing anomaly detection approaches in biology, statistics, and epidemiology.