Universal annotation of regulatory regions with HMMs
Project Description
There’s a lot of work on explainable AI to understand what machine learning models deem important. The specialty for models trained on genomic data is that the input features can be continuous, as they are located next to each other on the DNA. We built Hidden Markov Models (HMM) that take the feature importance of already-trained machine learning models to merge them together into continuous regions. The models at hand were trained on H3K27ac ChIP-seq data to predict gene expression. The current HMMs work on sample-specific feature importance, which means that they return regions for each sample. For many downstream tasks however, it is necessary to have one unified feature set.
Research Goals
- Try different approaches to unify the sample-specific HMM regions into one set of regions.
- Explore how meaningful the unified regions are with external data.
Prerequisites
- R and/or Python programming
- Ideally some knowledge about epigenomics and HMMs
How to Apply
Send a short Email with a short motivation to the corresponding supervisor Ashrafiyan|med.uni-frankfurt.de.