Research

Machine Learning on Graphs

We develop machine learning methods for graph-structured data, spanning theory, representation learning and generative modelling. Building on the Weisfeiler-Lehman graph kernel, a precursor of message passing, we study what graph neural networks (GNNs) can and cannot express. We design GNNs and graph transformers that predict graph properties, and generative models that create realistic graphs at scale. Beyond these methodological contributions, we apply our methods to real-world data such as molecules, proteins and biological networks.

Machine Learning on Graphs

  • Graph Kernels Based on the Weisfeiler-Lehman Test
    Shervashidze, N., Schweitzer, P., van Leeuwen, E. J., Mehlhorn, K., & Borgwardt, K. M. (2011). Weisfeiler-Lehman Graph Kernels. Journal of Machine Learning Research, 12, 2539-2561.

  • Structure-Aware Graph Transformers
    Chen, D., O’Bray, L., & Borgwardt, K. (2022). Structure-Aware Transformer for Graph Representation Learning. Proceedings of the 39th International Conference on Machine Learning (ICML 2022).

  • Capturing Long-Range Dependencies via Random Walks
    Chen, D., Schulz, T., & Borgwardt, K. (2025). Learning Long Range Dependencies on Graphs via Random Walks. 13th International Conference on Learning Representations (ICLR 2025).

  • Substructure Counting with Graph Neural Networks
    Pellizzoni, P., Schulz, T., Chen, D., & Borgwardt, K. (2025). Graph Neural Networks Can (Often) Count Substructures. 13th International Conference on Learning Representations (ICLR 2025).

  • Scalable Graph Generation with Sequence Transformers
    Chen, D., Krimmel, M., & Borgwardt, K. (2025). Flatten Graphs as Sequences: Transformers are Scalable Graph Generators. 39th Annual Conference on Neural Information Processing Systems (NeurIPS 2025).

  • Classifier-Based Evaluation of Graph Generative Models
    Krimmel, M., Hartout, P., Borgwardt, K., & Chen, D. (2026). PolyGraph Discrepancy: A Classifier-Based Metric for Graph Generation. 14th International Conference on Learning Representations (ICLR 2026).

 

Machine Learning on Proteins

We develop machine learning methods for proteins, spanning approaches from representation learning to generative modeling. Our work covers multiple modalities, from structured information represented as graphs and sequences to spatial organization information captured as point clouds and images. In this line of work, we often start with the biological problem and draw from a diverse set of modeling approaches to develop the methods needed to solve it.

Machine Learning on Proteins
  • Protein Engineering
    Data-Driven Protease Engineering by DNA-Recording and Epistasis-Aware Machine Learning (Huber, L., Kucera, T., Höllerer, S., Borgwardt, K., Panke, S., & Jeschek, M. (2025). Data-driven protease engineering by DNA-recording and epistasisaware machine learning. Nature Communications, 16, 5466. doi:10.1038/s41467-025-60622-7.)
  • Representation Learning on Proteins Folds 
    Protein Fold Classification at Scale: Benchmarking and Pretraining (Chen, D., Manolache, A., Niepert, M., & Borgwardt, K. (2026). Protein Fold Classification at Scale: Benchmarking and Pretraining. Proceedings of the 43rd International Conference on Machine Learning (ICML 2026), PMLR 306. doi:10.48550/arXiv.2605.18552.)
  • Protein Language Models
    Endowing Protein Language Models with Structural Knowledge (Hartout, P., Chen, D., Pellizzoni, P., Oliver, C., & Borgwardt, K. (2025). Endowing protein language models with structural knowledge. Bioinformatics, 41(11), btaf582. doi:10.1093/bioinformatics/btaf582.)
  • Spatial Proteomics
    Single-Cell Spatial Proteomics Maps Human Liver Zonation Patterns and Their Vulnerability to Disruption in Tissue Architecture (Weiss, C. A. M., Brown, L. A., Miranda, L., Pellizzoni, P., Steigerwald, S., Remmert, K., Hernandez, J. M., Borgwardt, K., Kleiner, D. E., Ben-Moshe, S., Rosenberger, F. A., Porat-Shliom, N., & Mann, M. (2026). Single-cell spatial proteomics maps human liver zonation patterns and their vulnerability to disruption in tissue architecture. Nature Metabolism, 8, 741–756. doi:10.1038/s42255-026-01459-2.)
  • Substitution Matrix Learning
    Structure- and Function-Aware Substitution Matrices via Learnable Graph Matching (Pellizzoni, P., Oliver, C. G., & Borgwardt, K. M. (2024). Structure- and Function-Aware Substitution Matrices via Learnable Graph Matching. Research in Computational Molecular Biology (RECOMB 2024), 288–307. doi:10.1007/978- 1-0716-3989-4_18.)
  • Function Prediction
    Bacterial Protein Function Prediction via Multimodal Deep Learning (Muzio, G., Adamer, M., Fernandez, L., Miklautz, L., Borgwardt, K., & Avican, K. (2026). Bacterial protein function prediction via multimodal deep learning. Research in Computational Molecular Biology (RECOMB 2026). doi:10.1101/2024.10.30.621035.)

 

Machine Learning for Biomarker Discovery

We develop machine learning methods that identify the molecular features associated with a disease or trait, from proteins and patterns in clinical time series to genetic variants and their interactions. Together with clinical partners, we apply interpretable methods to omics data from patient cohorts. In genetics, where individual signals are often weak, we draw on biological networks as prior knowledge to aggregate evidence across neighbouring genes. Across all of these settings the search space is enormous and cohorts are often small, so our methods are built to stay computationally tractable while providing statistical guarantees.

Machine Learning for Biomarker Discovery
 
Biomarkers from Clinical Omics Data
 
  • Long Covid Biomarkers
    (Cervia-Hasler, C., Brüningk, S. C., Hoch, T., Fan, B., Muzio, G., Thompson, R. C.,... & Boyman, O. (2024). Persistent complement dysregulation with signs of thromboinflammation in active Long Covid. Science, 383(6680), eadg7942.)
  • Biomarker Identification by Interpretable Maximum Mean Discrepancy
    (Adamer, M. F., Brüningk, S. C., Chen, D., & Borgwardt, K. (2024). Biomarker identification by interpretable maximum mean discrepancy. Bioinformatics, 40 (Suppl. 1), i501-i510.)
  • Temporal Biomarkers
    (Gumbsch, T., Bock, C., Moor, M., Rieck, B., & Borgwardt, K. (2020). Enhancing statistical power in temporal biomarker discovery through representative shapelet mining. Bioinformatics, 36 (Suppl. 2), i840-i848.)

Network-Based Association Mapping

  • Network-Based GWAS
    (Muzio, G., O’Bray, L., Meng-Papaxanthos, L., Klatt, J., Fischer, K., & Borgwardt, K. (2023). networkGWAS: A network-based approach to discover genetic associations. Bioinformatics, 39(6), btad370.)
  • Higher-Order Genetic Interactions
    (Pellizzoni, P., Muzio, G., & Borgwardt, K. (2023). Higher-order genetic interaction discovery with network-based biological priors. Bioinformatics, 39 (Suppl. 1), i523-i533.)
  • Genetic Heterogeneity Between Gene Pairs
    (Gumpinger, A. C., Rieck, B., Grimm, D. G., International Headache Genetics Consortium, & Borgwardt, K. (2021). Network-guided search for genetic heterogeneity between gene pairs. Bioinformatics, 37(1), 57-65.)

Significant Pattern Mining

  • Significant Pattern Mining (Westfall-Young Light) 
    (Llinares-López, F., Sugiyama, M., Papaxanthos, L., & Borgwardt, K. (2015). Fast and Memory-Efficient Significant Pattern Mining via Permutation Testing. KDD '15: The 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 725-734. doi:10.1145/2783258.2783363.)
  • Finding Genomic Intervals of Genetic Heterogeneity (FAIS) 
    (Llinares-López, F., Grimm, D. G., Bodenham, D. A., Gieraths, U., Sugiyama, M., Rowan, B., et al. (2015). Genome-wide detection of intervals of genetic heterogeneity associated with complex traits. Bioinformatics, 31(12), i240- i249. doi:10.1093/bioinformatics/btv263.)
  • Significant Pattern Mining with Covariates (FACS) 
    (Papaxanthos, L., Llinares-López, F., Bodenham, D., & Borgwardt, K. (2016). Finding significant combinations of features in the presence of categorical covariates. Advances in Neural Information Processing Systems 29 (NIPS 2016), 2271-2279.)

 

Machine Learning in Medical Informatics

We develop machine learning methods to support high-stakes clinical decisionmaking, with a particular emphasis on reliability under real-world conditions. Our
work spans antimicrobial resistance (AMR) prediction from MALDI-TOF mass spectra, ICU and health-outcome prediction such as sepsis and cardiovascular disease diagnosis. Methodologically, we draw on neural networks, reinforcement learning, and uncertainty quantification to ensure that predictions are trustworthy enough to inform clinical practice.

Machine Learning in Medical Informatics
  • AMR prediction from MALDI-TOF mass spectra 
    Weis, C., Cuénod, A., Rieck, B., Dubuis, O., Graf, S., Lang, C., ... & Egli, A. (2022). Direct antimicrobial resistance prediction from clinical MALDI-TOF mass spectra using machine learning. Nature medicine, 28(1), 164-174.
  • Uncertainty quantification for AMR prediction 
    Corvelo Benz, N., Miranda, L., Chen, D., Sattler, J., & Borgwardt, K. (2026). Conformal Prediction with Knowledge Graphs for Reliable Antimicrobial Resistance Detection with MALDI-TOF Mass Spectra. Journal of Computational Biology, 33(1), 19-35.
  • Early antibiotic susceptibility detection 
    Pichl, T., Miranda, L., Warkotsch, M. T., Wantia, N., Borgwardt, K., & Sattler, J. (2026). Early detection of ampicillin susceptibility in Enterococcus faecium with MALDI-TOF MS and machine learning. Journal of Global Antimicrobial Resistance.
  • Sepsis prediction 
    Moor, M., Bennett, N., Plečko, D., Horn, M., Rieck, B., Meinshausen, N., Bühlmann, P., & Borgwardt, K. (2023). Predicting sepsis using deep learning across international sites: a retrospective development and validation study. EClinicalMedicine, 62.
  • Coronary artery disease diagnosis 
    Bock, C., Walter, J. E., Rieck, B., Strebel, I., Rumora, K., Schaefer, I., Zellweger, M.J., Borgwardt, K., & Mueller, C. (2024). Enhancing the diagnosis of functionally relevant coronary artery disease with machine learning. Nature Communications, 15(1), 5034.

 

 

 

Earlier Topics

 

Earlier Topics
 



Big data analysis and biomedical research meet in our lab: We develop novel data mining algorithms for detecting patterns and statistical dependencies in large datasets from the life sciences. 

The ultimate goal in our work is to contribute to two big goals of science in the 21st century: To enable the automatic generation of new knowledge from big data through machine learning, and to help to gain an understanding of the relationship between diseases and molecular properties of patients, thereby enabling precision medicine.

Below you can find further information for some of our projects:

 

Machine Learning: Comparing Structured Data

We develop methods for comparing and classifying high-​dimensional objects. One prominent example are graph kernels, i.e. efficient distance functions between graphs.  

   

Machine Learning: High-​Dimensional Correlations

We develop methods for measuring statistical dependence between high dimensional variables, two-​sample tests to tell whether two samples were drawn from the same distribution, outlier detection algorithms to tell find "unusual" observations in a given dataset, and approaches that detect non-​linear dependence between variables.

 

 

   

Machine Learning: Significant Pattern Mining

We develop methods that discover significant patterns in high dimensional datasets while being runtime efficient and statistically sound. Our algorithms can be applied to graphs or collections of sequences and allow to account for dependencies between objects, to control the Family-​Wise Error Rate and to correct for categorical covariates.

 

   

Computational Biology: Genome-​Wide Association Studies

We develop efficient multivariate approaches for the genome-​wide discovery of genetic loci that are associated with a phenotype, thereby trying to elucidate the multicausal basis of complex traits.

 

 

Computational Biology: Genome Annotation

We have developed methods for detecting genomic insertions and deletions using next-​generation sequencing, and thoroughly assessed the difficulty of comparing the performance of variant pathogenicity prediction tools.

 

   

Computational Biology: Molecular Graph Classification via Graph Kernels

We developed new, fast and scalable similarity measures on graphs, so-​called graph kernels. Their prime purpose is to compare molecular graphs or protein structures and to classify them into functional categories.

 

 

Personalized Medicine

We have coordinated several national and international networks on personalized medicine:

 

Web-View
Print Page
Open in new window
Estimated DIN-A4 page-width
Go to Editor View