Representations of structured biological data and prototype-based machine learning
Biological data, such as protein structures and DNA sequences, is inherently complex—varying widely in shape and size. This complexity challenges standard machine learning (ML) tools, which typically require simple, fixed-size inputs. In sensitive fields like biomedicine, it is vital that our models are not only accurate but also interpretable; we must understand why a classification decision was made.
In her research, Katrin Bohnsack focuses on two key strategies to bridge that gap: smart data representation and interpretable classification learning in the context of bioinformatics.
For data representation, Bohnsack explored methods to convert complex structures into efficient numerical descriptors. This includes new sequence signatures based on information theory, and a novel reference-based embedding approach for molecular data. This technique characterizes a complex object by its proximity to a small, select group of reference points, allowing for highly efficient and robust comparisons.
For classification, Bohnsack uses a variant of prototype-based classifiers, namely Matrix Learning Vector Quantization. This model provides transparency in the learning and prediction processes and helps recognize which data features actually contributed to solving the classification problem at hand.
Ultimately, this work supports a more trustworthy and informed use of ML with structured biological data.