Our Research
We seek to identify conserved, mechanistic principles of signalling heterogeneity by integrating quantitative perturbation proteomics with scientific machine learning (SciML). We turn biological knowledge into equations and make discoveries where these models break: when a model fails, the discrepancy points to regulation from outside the modelled pathway, or to biology that was lost in translating knowledge into equations. Our current projects apply this approach to cancer cell plasticity, with a focus on ERK and AKT signalling in breast and colorectal cancer cell lines and organoids. Two of our four themes address our core biological and modelling questions; the other two build the software and experimental platforms they rely on.
Protein Dynamics Across Regulatory Layers
Many cellular phenotypes arise from coupled regulatory layers that operate across spatial and temporal scales, yet few mechanistic modelling frameworks work coherently across these layers. We study the interplay between three layers central to cellular plasticity and therapy resistance: the cell cycle, intracellular signalling and differentiation. We combine mathematical descriptions of individual layers with machine learning that contextualises them using information from the other layers [Fabrini & Fröhlich 2026], and use pseudotime approaches to extract temporal dynamics from static snapshots [Cortés-Ríos et al. 2025].
Rethinking Rules of Biological Modelling
Mechanistic models are usually built from literature curation, modelling dogma such as mass-action kinetics, and many subjective, rarely documented decisions. We aim to augment this process with data-driven methods, such as symbolic regression [de Pomereu & Fröhlich 2026] and universal differential equations [Persson et al. 2026a, 2026b], and with agentic approaches that combine them with existing mechanistic knowledge [de Pomereu et al. 2026].
SciML Software
SciML methods are often developed within individual projects and demand substantial numerical expertise, which limits their reuse and adoption. We turn them into reusable community software, from data processing and problem specification to model training. PEtab SciML extends the PEtab standard to hybrid mechanistic and machine learning models, with software support in Python (JAX) and Julia [Persson et al. 2026a], and builds on the second version of PEtab [Pathirana et al. 2026]. Curriculum multiple shooting makes training neural and universal differential equations robust [Persson et al. 2026b]. We also maintain established tools for simulation and calibration, including AMICI [Fröhlich et al. 2021] and pyPESTO [Schälte et al. 2023]. More recently, we have started to integrate these tools into agentic workflows for mechanistic world modelling [de Pomereu et al. 2026].
High-Throughput Single-Cell Perturbation Proteomics
Our models need datasets at a scale and level of multiplexing beyond what conventional wet-lab workflows produce. We are building high-throughput, reproducible pipelines for single-cell perturbation proteomics, centred on mass cytometry: it resolves single cells, scales to thousands of conditions through barcoding, and offers established markers of cell cycle state, differentiation and signalling.