Computational biology · Fred Hutchinson Cancer Center
I develop machine learning that treats a tissue as a whole: a continuous distribution of cell states, learned from millions of single cells, with the uncertainty of every estimate stated. With it I want to find the regulation that keeps each state in place, see where that regulation erodes in aging and cancer, and design regulatory DNA that reinforces it.
Dominik J. Otto, PhD. Postdoctoral Fellow in the Setty Lab, Basic Sciences Division, Fred Hutchinson Cancer Center, Seattle. NIH K99/R00 Pathway to Independence Award (NIGMS).
Which fate a single blood stem cell will take cannot be predicted, and genetically identical cells in the same environment differ several-fold in the expression of many genes. Yet populations are predictable: the distribution of states a tissue holds, and the rates at which cells change between them, are reproducible. Part of this comes from numbers, as noise averages out across many cells. The rest is regulation encoded in the genome. That is why I model the tissue, not the single cell.
Single-cell genomics measures cells one at a time. Each cell is a point in a space with as many dimensions as there are genes, and cells that differentiate trace out a branching path through it.
Many stable states behave as attractors. These are what we recognize as cell types, and clusters summarize them well. Around and between them, cells vary continuously as they differentiate, divide, respond to stress and cycle. Rates and gradients live between clusters, so they cannot be expressed on them.
Cells accumulate where change slows or where they divide, and thin out where change accelerates or cells die. How densely cells populate each state therefore carries the dynamics that produced it. Stable states are dense; transitions are sparse, and new lineages often emerge through them.
Mellon infers this density from millions of real cells as a smooth, differentiable function, with uncertainty. It can be compared across conditions, followed over time and built into mechanistic models, which turns a snapshot of a tissue into a quantity that can be modeled. Nature Methods, 2024.
A cell's expression could change in as many directions as it has genes, yet healthy cells use only a few of them. Transcription factors, enhancers, chromatin and signals from neighbors pull cells back when they stray. These are the guardrails of cell identity, and when one fails, identity gives way: deleting Pax5 lets committed B cells become T cells, and leukemic cells co-express lineage programs that healthy cells keep apart. Mutations, inflammation and drugs push cells off their usual states, and the guardrails decide whether they return or drift toward disease.
Clusters, trajectories and most models of cells learn from where cells are, and so see the few directions in which cells change. A model of the whole distribution, including how far it spreads, can also learn the many directions in which identity is defended. The goal is the small set of regulators that hold a tissue's states and fail first in disease: the targets for monitoring and intervention.
All are open-source, documented and maintained. Details and links are on the software page.
Ask which cell states are stable and which are transitional, and where along differentiation cells commit, from the density of cells in each state. In human hematopoiesis, B-lineage commitment passes a low-density window in which enhancers are primed and EBF1 activates.
Compare young and old, treated and untreated, or mutant and wild type at every cell state, without first assigning cells to clusters. Changes that are graded along a transition or confined to part of a cell type stay visible, each with its uncertainty. Because nearby states share information, their estimates are correlated, and Kompot accounts for this when it adds up a gene's evidence, so small changes spread over many states count without being counted twice.
Read the accessible and the silenced regulome from a single CUT&Tag2for1 experiment in single cells, separated by statistical inference that treats each signal as continuous along the genome. Developed with the Henikoff laboratory.
Explore which cells and which genes resemble one another in very large single-cell datasets, one cell or gene at a time and linked to expression, in a viewer that reads only what is on screen and never changes the data.
My future laboratory will be computational and will work closely with experimental laboratories. Its program has three directions, each standing on its own and gaining from the others. All three model the tissue: the first follows how the distribution changes in time, the second places each state among its neighbors, and the third turns the model around to produce regulatory sequence from a described function.
I came to biology from mathematics. After a bachelor's thesis in topology, my master's thesis at Freie Universität Berlin corrected a theorem on the determining nodes of gene regulatory networks, showing that not every feedback loop must be broken to control a network. I earned my PhD in informatics at Universität Leipzig, with research at the Fraunhofer Institute for Cell Therapy and Immunology on gene expression deconvolution and prostate cancer markers. Since 2021 I have been a postdoctoral fellow at Fred Hutch, the first member of Manu Setty's laboratory, where Mellon and Kompot were conceived.