Machine Learning in Life Science

Scalable Bayesian deep learning.

Spearheading of Bayesian approximations of large contemporary deep neural networks.

• Fundamental mathematical theories to build identifiable distributions of parameters in overparametrized models (NeurIPS 2024).
• Efficient sampling algorithms for fully correlated posterior approximations in overparametrized models (AISTATS 2025) and scalable variational training algorithms (NeurIPS 2025).

Rich and reliable representations.

Contribution to fundamental advances in representation learning.

• Demonstration that differential geometric representations provide a unique solution to the underlying identifiability problem, thereby leading to more reliable latent interpretations (ICML 2025), which can recover, otherwise obscured, biological insights (Nature Communications 2022).
• Overview of research communities focusing on the above area in a recent edition of Current Opinion in Structural Biology (2025).

Biological sequence analysis.

Advancing state-of-the-art popular biological prediction tools and genomic language models.

• New versions of DeepLoc, DeepTMHMM (first update of TMHMM since 2001), and NetStart (first update since 1997).
• Impactful contributions to the prediction of protein stability — papers in eLife (2023) and Nature Communications (2024).
• Contribution to the recent development of genomic language models and establishment of the first biologically relevant benchmark for such models (ICLR 2024).
• Pilot project on the Gefion supercomputer on genomic language models.

Protein engineering and fitness landscapes.

Creating an overview of challenges and development of tools to stepwise resolution of these.

• Systematic analysis of methodological challenges in the regression of fitness values from protein sequences (PLoS Comp Bio 2024).
• A novel technique for describing protein fitness functions (Kermut, NeurIPS 2025a), ranked as the current best method according to established community benchmarks (ProteinGym).
• A software library for systematic comparisons of the performance of high-dimensional Bayesian Optimization procedures in the context of molecular and protein optimization (NeurIPS 2025b).

Interpretability in terms of underlying physics.

Starting zero-shot performance improvement of inverse-folding models for protein design.

• Establishment of a theoretical connection between the inverse-folding models used for protein design, and the free-energy considerations underlying protein thermodynamic stability (NeurIPS 2026).
• Understanding of the origin of high correlation between likelihood and thermodynamic stability.
• Start of zero-shot performance of these models, without any retraining of the models.

Uncovering limitations of “trustworthy AI” tools.

Displaying how naïve use of explainable AI (XAI) and algorithmic fairness can lead to failure.

• Evidence of misleading representation visualization in equivariant models (ICML 2024).
• Demonstration that label bias makes algorithmic fairness and XAI unreliable (EWAF 2025, ISBI 2026).

Tools and understanding of shortcut learning

• Slice discovery for image segmentation (MIDL 2026) showcasing systematic errors beyond algorithmic fairness.
• Weight space correlation analysis to assess how different acquisition, clinical and demographic properties are used for predictions (MIDL 2026).
• Image segmentation is also susceptible to shortcut learning, against most people’s intuition (MICCAI 2024).
• Detection of shortcut learning via fast counterfactual explanations (ECCV oral 2024).

Uncertainty quantification in real-world tasks

• Showing that OOD detection is a poor metric for model selection if the downstream goal is to avoid errors (MICCAI 2025).
• Improving epistemic UQ for medical image segmentation through better aleatoric uncertainty modeling (MICCAI 2024).
• Label style affects medical image segmentation uncertainty quantification (ICLR 2023).

Explainable AI

• Inherently explainable and steerable diffusion models via prototypes (ICLR 2026).

Community building.

MLLS is a cornerstone of the Danish and European AI community and hosts many events.

• Continuous seminar series with national and internationally renowned speakers from the machine learning community.
• Local and regular community-building events, such as annual retreats.
• Yearly conference on Generative Models and Uncertainty Quantification (GenU) attracts elite researchers within the field to Copenhagen.
• Conferences on Machine Learning and Molecules, and Generative AI in Life Science for both the national and international community.
• Central role in creating and organizing the first EurIPS conference that brought a European venue to the elite NeurIPS conference.
• Contributing to ELLIS Unit Copenhagen.