Inference at the Limits of Microscopy: Turning Photons into Biological Insight with Machine Learning

Modern microscopy is often an inverse problem at the edge of measurement: the instrument records photons, but the scientific question is about the hidden object or process that produced them. In biology, this problem appears in different forms. Sometimes the photons form a movie-like time series, where the goal is to infer how an individual molecule changes shape over time. Sometimes they form an image, where the goal is to recover cellular structures that are blurred by diffraction and corrupted by noise. My work develops Bayesian machine learning methods for both settings by modeling the physics of how microscopy data are generated. For time-series measurements, I will show how probabilistic inference can recover molecular states, transition rates, and heterogeneity from noisy single-molecule signals, with applications to DNA Holliday junctions involved in DNA repair. For imaging measurements, I will discuss image deconvolution and structured illumination microscopy, where computation is used to recover light-emitting structures, including features beyond the usual diffraction limit. Across these examples, the goal is not simply to make cleaner pictures or smoother traces, but to turn sparse photons into quantitative evidence for how living systems move, organize, and function.

Bio
Ayush Saurabh is a scientist in the Center for Single-Molecule Biophysics at the Biodesign Institute. He develops Bayesian machine learning methods that turn sparse microscopy data into quantitative insight about biological systems. He received his Ph.D. in Physics from Arizona State University, where he used numerical simulations of SU(2) and U(1) gauge theories to study topological solitons that may have formed in the early universe. His research background spans fluid mechanics, cosmology, particle physics, and biophysics, with a common thread of using mathematical modeling and computation to understand complex physical systems. His current work brings that perspective to microscopy, developing probabilistic methods to learn biomolecular dynamics from single-molecule time series, recover cellular structure from low-photon images, infer energy landscapes, understand molecular organization, and probe physical environments inside living cells.