We combine computational/engineering techniques with high-throughput experimental data to model the behaviour of cellular systems, the progression of diseases, and the effect of drugs.
We design and build custom devices. The main areas of research are: (i) Proteomic platforms and Point of Care Systems and (ii) Device for real-time monitoring the mechanical properties of cartilage and tissue-engineering constructs.
We employ biomechanics for monitoring tissue mechanical properties in-vitro and real time. Our focus is particulartly on cartilage degeneration using a combination of systems biology and biomechanical approaches.
Modeling Cartilage Degeneration by Combining Systems Biology and Biomechanical Approaches
Osteoarthritis, the leading cause for disability affecting one of every six people in Europe, is characterized by the painful deterioration and destruction of articular cartilage. The causes for this disease are multifactorial involving mechanical factors like cartilage wear and destruction as well as a complex interaction of pathological cellular mechanisms. Except for pain-relief treatments there are no medication strategies for osteoarthritis and disease modifying drugs are still missing. The research objective is to link the mechanical degeneration of articular cartilage to the underlying biological processes and build biomechanical and systems biology models to evaluate pharmacological interventions on the cellular and on the tissue level.
Funded by: German Research Foundation
Compound prioritization based on transcription factor activity
Intro: Accurate and effective in-silico compound prioritization plays a pivotal role in identifying promising drug candidates in early stages of drug discovery. Since the release of the first Connectivity Map (CMAP), the use of transcriptomic signatures for characterizing the systematic effects of compounds has gained considerable attraction, resulting in the release of a plethora of large scale data sets and computational methods for their analysis.
Objective: In this project, recently available transcriptomic data, arising from genome wide gene knock downs (KDs) and compound perturbations, along with newly developed methods for transcription factor (TF) activity inference, are utilized to prioritize compounds that modulate the activity of a specific TF. As an application, this newly developed platform was tasked to identify compounds that indirectly inhibit the activity of Myc, a proto-oncogene that controls key cell cycle mechanisms.
Drug-Drug biological effect similarity prediction
Intro: Association of drug pairs based on their biological effect similarity plays an important role in the process of drug repurposing and mode of action identification. Since the release of the first Connectivity Map, several computational methods have been proposed to identify similar compound perturbations based on their in-vitro transcriptomic signatures. However, the proposed similarities at the gene expression level exhibit no correlation with structure based similarity approaches, which are widely used in the field of medicinal chemistry. A possible explanation for this lack of correlation is the featurization and similarity functions used in structure-based approaches, which have been finely tuned for QSAR and physico-chemical properties prediction.
Objective: The goal of this project is first to investigate the similarity of transcriptomic signatures, arising from compound perturbations, at different biological levels, mainly at the gene, pathway, transcription factor activity, GO term and signaling network level. Finally, a deep learning model will learn the appropriate molecular representation and similarity function to predict the bio-effect similarity of drug pairs given their structure.
Drug-Kinase binding prediction
Intro: Prediction of novel drug-protein interaction plays a crucial role in the drug discovery process. The majority of computational approaches so far have focused on fixed engineered features for the representation of drugs and proteins and have treated the binding prediction problem as a binary classification task. However, protein-ligand interactions assume a continuum of binding strength values, also called binding affinity and predicting this value still remains a challenge. Furthermore, the recent increase of publicly available bioactivity data have enabled the use of end-to-end deep learning architectures that are able to learn task specific representations of the input molecules for the binding affinity prediction.
Objective: The goal of this project is to build an end-to-end deep regression model that takes as input the 2D representation of a compound, encoded as a molecular graph, and the amino acid sequence of a protein and outputs the KD (equilibrium dissociation constant) of the compound-protein pair. A considerable amount of effort is focused on utilizing models and techniques that are able to quantify predictive uncertainty, e.g. MC-dropout, deep ensembles and Bayesian deep learning.