Abstract
The intersection of machine learning and molecular biology is revolutionizing our understanding of genome organization and its role in disease. This abstract explores how advanced machine learning, particularly deep learning, can accelerate biological experiments and enhance our insights into 3D genomics. There is a need for more efficient, accurate in silico prediction of gene expression and its deregulation, which is often central to genetic diseases.
Starting with DNA sequencing data, we developed a deep feed-forward neural network based on eight state-of-the-art variant calling algorithms, significantly reducing the time and cost of variant discovery. Building on this, we leveraged the DNABERT architecture to predict spatial chromatin conformation, focusing on ChIA-PET loop discovery. However, to overcome limitations in accuracy and scale, we introduced the HiCDiffusion model, which combines encoder-decoder architecture with a transformer for context learning and diffusion-based enhancement, achieving results indistinguishable from experimental Hi-C data. Finally, we demonstrated how these AI models reveal key relationships between chromatin spatial organization and gene expression, highlighting how distal sequences in linear DNA can interact in 3D to regulate transcription.
We therefore provide a comprehensive AI-based framework for understanding the relationship between genome sequence variations and chromatin architecture, with implications for gene regulation, disease mechanisms, and the potential to accelerate genomic discoveries through machine learning. This journey, from DNA sequence variant discovery to spatial chromatin conformation prediction and gene expression modelling, underscores the transformative potential of Artificial Intelligence in 3D genomics.
Plewczynski is visitor to BiLS and the Department of Biosciences
There are a few opportunities to meet with professor Plewczynski. If you are interested, please contact the hosts: Rein Aasland and Jonas Paulsen