I'm a physicist by training. My Bachelor's and Master's in physics took me from mathematical modeling of complex systems to machine learning. Now, in my PhD, I combine machine learning with mechanistic modeling to study complex systems in life sciences.

Research

Teaching machine learning models what science already knows

Classical differential equation models are interpretable but only know what we already understand, while deep learning can learn the rest but is data-hungry and opaque. Neural differential equations combine the strengths of both, and my research aims to show that they are a practical modeling framework for complex systems in life sciences. With a focus on data efficiency, interpretability and forecasting in areas such as cell signaling and epidemiology, I develop and benchmark informed models that keep the known mechanisms as equations and let neural networks fill in the rest. In other words, I teach machine learning models what science already knows, so they only have to learn what it doesn't.

Keywords: Neural Differential Equations, Informed Machine Learning, Mathematical Modeling, Neural Ordinary Differential Equations, Neural Controlled Differential Equations, Uncertainty Quantification, Explainable AI, Computational Statistics, Monte Carlo Methods, Active Learning, Systems Biology, Epidemiological Modeling

Projects

• UDE-epi-forecast (ongoing) – Building a theoretical foundation for universal differential equations in epidemiology and benchmarking how well they forecast disease dynamics.
• Informed neural CDEs (ongoing) – Embedding known mechanistic ODEs into neural controlled differential equations, so models can continuously learn from irregular, multi-channel time series while retaining interpretability and explainability.
• Active informed ML (ongoing) – Combining informed machine learning with active learning, so that models built on prior scientific knowledge can also choose which data or experiments would be most informative to collect next.
• NF-κB – p53 signaling crosstalk (completed) – Universal differential equations for quantifying crosstalk between the NF-κB and p53 pathways (Bioinformatics 2026).

Publications

• Umur Can Kaya, Xhemal Kodragjini, Samuel Zambrano, Katharina Baum. "Universal differential equations for quantifying NF-κB – p53 signaling crosstalk", Bioinformatics 42 (Supplement_2), btag449 (ECCB 2026 Proceedings), 2026. paper

Talks & Outreach

• Talk (upcoming) at the Zuse Institute Berlin (ZIB), Berlin, Germany, 2026: "Applications of Universal Differential Equations for Systems Biology"
• Talk at the 25th European Conference on Computational Biology (ECCB), Geneva, Switzerland, 2026: "Universal differential equations for quantifying NF-κB – p53 signaling crosstalk"
• Workshop talk (internal) at the Hasso Plattner Institute, Potsdam, Germany, 2026: "Vibe coding 101", an introduction to using AI coding agents such as Claude Code for scientific programming
• Poster at the German Conference on Bioinformatics (GCB), Düsseldorf, Germany, 2025: "One model is worth a thousand datasets: Benchmarking informed Neural ODEs for data-efficient prediction of biological dynamics"

Teaching

• Softwareprojekt: Machine Learning for data from the life sciences (Neural differential equations), Winter Term 2025/26