The discovery of valuable medical information from biobanks is fundamental to the development of new personalised medicine. The project “Scalable Learning for Reproducibility in High-Dimensional Biomedical Signal Processing: A Robust Data Science Framework (ScReeningData)” provides researchers from biomedical disciplines with methods and reproducible findings for their research. The gist: a quality check is already built in.
Reproducibility of the results and statistical robustness of the methods are mathematically quantified and proven in the “ScReeningData” project. This is important, because the hypotheses from biomedical data must be tested in elaborate experiments and clinical trials. Without statistical guarantees of reproducibility, valuable time is spent studying relationships that in reality may not even exist. In contrast, a high rate of new reproducible discoveries accelerates and improves e.g. the development of individualised diagnostics and the therapy for diseases such as cancer, diabetes and heart failure.
Complex calculations possible in a few days
The “ScReeningData” methods can distinguish reproducible biomarkers from random patterns. They are also robust against outliers in the data. Furthermore they are scalable to highly complex problems such as the analysis of genetic data. In the future, it will be possible to carry out calculations that today take many years even using state-of-the-art high-performance computers in just a few days. This means that “ScReeningData” enables the systematic exploration of large biobanks.
The underlying concept of computer-aided learning behind “ScReeningData” only recently developed by Muma and his research team is called “Terminating-Knockoff (T-Knock)”. The idea is similar to a placebo-controlled trial in drug research. Randomised controlled experiments are systematically calculated on the computer and mathematically modelled. Biomarkers are only declared as reproducible discoveries if they sufficiently prevail over computer-generated placebo features (“knockoffs”). The speed advantage over existing methods comes from the fact that learning is stopped early (termination) when knockoffs are selected.
Michael Muma studied and received his doctorate at TU Darmstadt. Since 2017 he has been a Postdoctoral Research Fellow, lecturer and junior research group leader ( ) with the Signal Processing Group at the Department of Electrical Engineering and Information Technology (etit) at TU Darmstadt. He also conducts research at the LOEWE centre emergenCITY. His work and publications have received numerous awards, including the Early Career Award of the European Association For Signal Processing (EURASIP). Athene Young Investigator
ERC Starting Grants are awarded by the European Research Council to researchers from all disciplines up to seven years after their doctorate. In this way, the European Union promotes outstanding research and at the same time early-career scientists. The Starting Grant is aimed at researchers, who can show evidence of excellent work, are now at an early stage of their careers and would like to expand their independent research or set up their own research group. In the current round, TU Darmstadt has acquired three ERC Starting Grants, as the were successful in addition to the “ScReeningData” project. “DaVinci Switches” and “MotLang” projects