About my work

Research focus and identity

My work starts from a common problem in biomedical research: data are often effectively small relative to their complexity. Measurements are high-dimensional, partial or indirect, and limited by ethical, clinical, or financial constraints. I develop methods that combine statistical modelling and machine learning to use stronger biological or clinical assumptions while staying flexible enough to accommodate the high dimensionality and heterogeneity of complex biomedical data.

I enjoy collaborating across disciplines and have experience spanning single-cell omics, deep generative models, dynamic modelling, and computational software.

I am now bringing some of what I learned about knowledge- and data-driven modelling and structured representation learning into spatial omics, where a lot of biological structure is visible in the tissue, but so far only partially used in downstream spatial modelling.

PhD research

During my PhD, I worked mainly in two application areas: single-cell omics and dynamic modelling in clinical registries. These may look like different domains, but I became interested in the structural challenges they share. Recognizing such common structures helped me transfer concepts between fields. My PhD thesis, Capturing dynamics under misspecification and potential transformation, compiles a few of these challenges:
How to capture dynamics when

• models are necessarily mis-specified because they have to be simplified or specified at lower resolution than the task would demand due to data limitations,
• observations live in a much higher-dimensional space than the underlying process of interest,
• and noise, heterogeneity, missing data, or lack of one-to-one correspondence of observations across time points make robust estimation very challenging.

Encoding biological knowledge into single-cell omics representations

In single-cell omics, I have worked on representations that make allow for explicitly encoding biological assumptions or incorporating additional biological knowledge. This includes sparse dimension reduction for scRNA-seq, described here (code), and work on discrepancies in time-series single-cell representations, described here (code). In related student projects that I (co-)supervised, we extended the idea of incorporating knowledge to benchmarking multimodal omics-plus-text representations with mmContext and incorporating biomedical literature with alias.

Combining mathematical modelling and machine learning for modelling disease trajectories

In clinical and registry data, I have focused on disease trajectories that are only indirectly observed through sparse, heterogeneous measurements. I developed latent dynamic models that combine neural-network representations with ordinary differential equations and statistical modelling (main publication, package).
My interest in these questions started back in the days when I was writing my Master’s thesis on Temporal dynamics in generative models. There, I have investigated what can still be learned about latent trajectories from just two time points. Later, we have extended this to related questions such as how to connect different clinical measurement instruments, and how to investigate the effect of treatment switches using latent mixed-effects models.

Statistical computing and scientific software

Across these projects, I have translated methods into computational tools, bringing together statistical models, neural networks, differential equations, and automatic differentiation. In particular, I have built computational tools for latent-variable modelling in Julia (scVI.jl, LatentDynamics.jl), worked on differentiable programming for flexible statistical modelling and interoperability across programming languages with the JuliaConnectoR.

Current direction: spatial omics

My current work brings these threads to spatial omics as an EIPOD-LinC Postdoctoral Fellow in the Huber group at EMBL, in collaboration with the Risso lab at the University of Padova. Spatial transcriptomics, spatial proteomics, histology, and pathology images make tissue structure explicit, but downstream models still often use generic neighbourhoods, grids, or absolute coordinates. I want to incorporate tissue structure into spatial modelling, using mathematical modelling, machine learning, and scientific computing to identify interpretable and biologically meaningful spatial phenotypes around landmarks such as tissue interfaces, vessels, membranes, and cellular niches.