Selected research & software
Methods, workflows, and applications developed to make complex biomedical data more interpretable, reproducible, and useful to research teams.
01
Active research at LMU Klinikum across medical imaging, proteomics, and spatial analysis.
Developed a reproducible 3D nnU-Net pipeline for segmenting cortical superficial siderosis from multimodal MRI. Designed an uncertainty-driven active-learning workflow to prioritize informative cases and improve annotation efficiency.
Developed HASA, a replicate-aware method and installable application that attributes bulk proteomic signals to cell types using single-cell RNA-seq reference atlases. Applied the method to a mouse cerebral amyloid angiopathy dataset, recovering smooth-muscle-cell-specific deregulation and corresponding biological pathway signatures.
Collaborating with a master’s student and colleagues to develop a Point Transformer–based workflow that predicts locations in the 3D mouse brain Common Coordinate Framework directly from spatial-transcriptomics patches. The current prototype has been evaluated on held-out tissue sections, with ongoing work testing generalization across MERFISH, Xenium and Visium HD datasets.
Developing a local research workbench that connects neuroimaging cohorts with molecular data and published reference maps while making atlas transformations, spatial statistics and provenance explicit. The system is designed to turn region-level interpretation into a traceable and reproducible analytical record.
02
Research methods supported by software, publications, or openly available code.
Developed ImShot, open-source software that integrates MALDI imaging mass spectrometry with shotgun proteomics for probabilistic in-situ protein identification and interactive data visualization. Released the desktop application, R package, test datasets, documentation and video tutorials; the method was published in Molecular & Cellular Proteomics.
Developed WeiseEule, an open-source biomedical question-answering system that uses explicit signals in user queries to retrieve relevant scientific evidence for retrieval-augmented generation. Evaluated on 50 challenging biomedical questions, the published method achieved a median Precision@10 of 0.95 and a median answer-quality score of 2.5/3, outperforming BM25 and embedding-based retrieval.
Developed an R-based workflow for differential enrichment analysis of BioID mass-spectrometry data, including data cleaning, missing-value handling, normalization, limma moderated inference, interactive volcano plots and automated bait–prey network generation. Applied the workflow to data from 40 BioID baits and built the R/JavaScript MiGENet application, supporting the discovery of a mitochondrial gene-expression network and a feedback mechanism regulating COB mRNA translation.
Developed ComplexMiner, a prototype desktop application combining one-shot learning and Siamese neural networks to classify pairs of SEC-SWATH-MS elution profiles. Built the complementary CoreClust algorithm and applied it to approximately 1,400 protein profiles across 42 fractions, identifying 177 candidate complexes and recovering established assemblies including ribosomal subunits and the chaperonin-containing T-complex.
03
Earlier open-source and published work in statistical proteomics, kernel methods, and protein-sequence classification.
Developed an open-source R workflow for two-group differential protein-abundance analysis using limma moderated inference. The workflow supports normalization, configurable handling of zero intensities, interactive volcano plots and tabular outputs for downstream pathway analysis.
Designed a generalized radial-basis-function kernel combining a Euclidean-distance Gaussian component with cosine similarity, allowing the model to use both distance and angular separation. Evaluated the method on noisy MPSK signal recovery, plant-leaf classification and nonlinear system identification; in the classification experiment, the method achieved 100% accuracy after 150 training epochs compared with 3,500 epochs for the conventional Euclidean RBF kernel.
Developed WAD-1 and WAD-2 for distinguishing Class I and Class II PDZ domains directly from amino-acid sequences using wavelet-based feature extraction. WAD-2 combined amino-acid trigram frequencies with the maximal-overlap discrete wavelet transform and achieved better recognition accuracy than the standard wavelet approach.
