David Bednář

Proteins with deleterious mutations may be targeted by known therapeutics, or even novel treatment can be identified using the drug repurposing strategy. Unfortunately, many of the available tools and databases are hard to find, install, and operate, because many of them were not even developed for medical purposes. Interpretation of the results needs to be done carefully with expert knowledge of the bioinformatics principles. Therefore, automation of data collection and their user-friendly presentation is essential for busy and often overwhelmed oncologists. Moreover, due to the growing amount of biological and clinical data and the complexity of the underlying mechanisms, machine learning methods have increasingly been used in oncology. These methods are based on finding generalizable patterns in the available data using flexible, general-purpose algorithms such as artificial neural networks, decision trees, random forests, and kernel methods to name a few. In oncology, these algorithms have already been applied for a wide range of problems, e.g., for oncological risk assessment, automated lesion detection, characterization, grading and staging, prediction of prognosis, and therapy response. Both clinical and imaging data are now used for training such algorithms to improve the accuracy of the established screening strategies in neuro-oncology, prostate, lung, pancreatic, breast, and many other types of cancer. In recent years, high-throughput sequencing technologies augmented the clinical and imaging data with samples at multiple molecular levels, promoting the use of multi-omics in oncology to address such problems as patients’ phenotyping, biomarker discovery, and drug repurposing.

In Loschmidt Laboratories, we already have extensive expertise in predicting the effects of mutations on protein function, stability, and solubility, which are often linked to the molecular mechanisms studied in oncology. Moreover, our web servers are often composed of complex workflows combining dozens of tools and databases, providing data in an understandable user-friendly way. In the course of this project, we will develop web-based software that will combine all our knowledge in the analysis of proteins and their mutations with information obtained in available databases. We will focus on the concentration of only relevant information in one tool and provide them in an easy-to-use way. Moreover, the tool should provide information on known drugs and apply virtual screening to identify new potential candidates using the drug repurposing strategy. In the second version, we want to expand our expertise and further apply state-of-the-art machine learning methods in two directions: i) drug identification and ii) interaction network analysis. The former will allow us to build on our expertise in structural and computational biology to leverage protein structural and molecular dynamics data in predicting protein-drug interactions for known protein targets in oncology. The latter will enable us to further expand the set of targets by the use of recent network-based methods of machine learning to detect functional correlations among different metabolic pathways. Thus, the combination of structure-based bioinformatics with the tremendous potential of machine learning will unlock the understanding of the underlying molecular mechanisms, predict of effect of mutations, and help identify effective drugs in precision oncology.

Research objectives

1. Development of software for clinical oncology.
2. Application of machine learning and artificial intelligence in oncology.
3. Engineering and analysis of proteins for biotechnology and biomedicine.
4. Development of software tools for protein engineering.