Drug Bioinformatics

By analyzing the spatial distribution of genetic variants in three-dimensional structures of proteins harboring them and their homologs, we can produce hypotheses about the functional consequences of these variants. For example, if a mutation caused by a single-nucleotide polymorphism lies on an interaction interface with another protein or in a ligand-binding pocket, it may affect the corresponding binding affinity, and mutations lying in the protein core can be detrimental for its stability. We develop methods that can annotate very large datasets in this way, providing insight into the relation between mutations’ annotated pathogenic or functional effect and their location in the three-dimensional structures of proteins and their complexes.

We build machine-learning methods for predicting the impact of mutations using a variety of features related to protein three-dimensional structures, interactions, and evolution. The methods can be trained to predict the impact on protein function, as well as their pathogenicity, which correlates with protein function. Additionally, we explore the possibility of training such methods to predict the impact on more specific phenotypes, such as resistance towards antibacterial compounds.

In this BMBF-funded project in cooperation with the Hamburg University and University Hospital Greifswald, we apply our expertise in structural modelling and annotation to investigate novel mechanisms of pathogenesis in cardiac and renal diseases, focussing on the alterations of protein sequences caused by disease-specific alternative splicing events.

Drug side effects are widespread, yet their mechanisms are poorly understood. In this BMBF-funded project, we aim to address the molecular basis and population importance of such mechanisms in cooperation with the Hamburg University and DESY. We combine systems biology, structural bioinformatics, cheminformatics and machine learning to investigate the effects of off-target drug binding and the influence of genetic variation on it.

Antimicrobial Resistance (AMR) is perhaps the most urgent threat to human health. While individual resistance mutations are well-researched, knowing which new mutations can cause antimicrobial resistance is key to developing drugs that reliably sidestep microbial defenses. In a HelmholtzAI-funded project in cooperation with CISPA, we investigate AMR via explainable artificial intelligence, by developing and applying novel methods for discovering easily interpretable local patterns that are significant with regard to one or multiple classes of resistance. We learn a small set of easily interpretable models that together explain the resistance mechanisms in the data, using statistically robust methods for discovering significant subgroups, as well as information theoretic approaches to discovering succinct sets of noise-robust rules.

Recent developments in the protein structure prediction field led to a drastic increase in the number of available protein three-dimensional structures. This creates a challenge and presents an opportunity for discovering fitting approaches to utilise such new datasets in various machine learning settings. On the other hand, large language models (LLMs) trained on protein sequences, called protein large language models (PLLMs), have proven to be useful in various bioinformatics problem settings. In our work, we are interested in applying PLLMs and extending them by considering protein three-dimensional structure. We use sequence-based pre-trained PLLMs and our own structure-based representations, separately and combined, and try to interpret their behaviour on our data. We have developed a self-supervised learning approach for protein three-dimensional structures based on convolutional graph neural networks and graph transformers that creates meaningful embeddings of protein structures. We demonstrated its utility in a variety of downstream tasks, including the prediction of drug-target interactions and predicting products of biosynthetic gene clusters.

We employ the latest developments in protein-based pre-trained models to create efficient deep learning models for predicting protein-drug interactions. In these models, we use representations of not only drugs, but also protein targets, in particular, of their three-dimensional structures. This makes our models unique in the field. To make these models even more useful for biomedical research, in cooperation with the NextAID project at the Saarland University, we explore explainability approaches to identify amino acids in the target proteins that most contribute to interactions. A similar approach is taken in a HelmholtzAI-funded project XAI-Graph with UFZ, where we explore explainable graph-based drug-target interaction models applied to toxicology research.