I'm a Flatiron Research Fellow in the Center for Computational Biology at the Flatiron Institute working with Prof. Michael J. Shelley. I obtained my Ph.D. from the Max Planck Institute for Molecular Cell Biology and Genetics /TU Dresden where I was mentored by Prof. Ivo F. Sbalzarini.
My research combines classical first-principle based approaches with modern AI/ML techniques to develop predictive, interpretable methods for solving forward and inverse problems in the life sciences.
During my Ph.D, I developed robust statistical techniques for inferring physically consistent PDEs and ODEs from noisy, limited data. I developed a new sparsity promoting algorithm called Iterative Hard Thresholding with debiasing (IHT-d) which in combination with the theory of stability selection enabled learning differential equations with minimal parameter tuning directly from microscopy videos.
Physics-Informed Neural Networks serve as promising means to integrate physical priors based on differential equations with the universal approximation properties of neural networks to solve forward and inverse problems in physical and life-sciences.
Using rotationally and translationally invariant representation of second-order tensorial functions, we learn closure models that outperform classical closures. To ensure temporal stability of the inferred closures, we use a Model Predictive Control strategy and adjoint optimization to learn closures that are fast, accurate and stable.
In collaboration with experimentalist at Prof Daniel Needleman's lab, we investigated the physical basis of spindle organization in HeLa cells. Using static ultrastructural data from electron microscopy and dynamic data from light microscopy, we condition a minimal coarse-grain model to infer key materials properties of the spindle.
Leveraging recent advances in score-based generative modeling, we propose a novel strategy for bridging cross-sectional data acquired from RNAseq. Our method termed Probability Flow Inference (PFI) faithfully interpolates between the observed marginals while accommodating for general noise forms. Using biological curated models, we illustrate how an incorrect choice of noise model leads to inaccurate inference of the underlying genetic program.
