Improving Cell-type-specific 3D Genome Architectures Prediction Leveraging Graph Neural Networks

The mammalian genome organizes into complex three-dimensional structures, where interactions among chromatin regulatory elements play a pivotal role in mediating biological functions, highlighting the significance of genomic region interactions in biological research. Traditional biological sequencing techniques like HiC and MicroC, commonly employed to estimate these interactions, are resource-intensive and time-consuming, especially given the vast array of cell lines and tissues involved. With the advent of advanced machine learning (ML) methodologies, there has been a push towards developing ML models to predict genomic interactions. However, while these models excel in predicting interactions for cell lines similar to their training data, they often fail to generalize across distantly related cell lines or accurately predict interactions specific to certain cell lines. Identifying the potential oversight of excluding example genomic region interaction information from model inputs as a fundamental limitation, this paper introduces GRACHIP, a model rooted in graph neural network technology aiming to address this issue by incorporating detailed interaction information as a hint. Through extensive testing across various cell lines, GRACHIP not only demonstrates exceptional accuracy in predicting chromatin interaction intensity but showcases remarkable generalizability to cell lines not encountered during training. Consequently, GRACHIP emerges as a potent research tool, offering a viable alternative to conventional sequencing methods for analyzing the interactions and three-dimensional organization of mammalian genomes, thus alleviating the dependency on expensive and time-consuming biological sequencing techniques. It also offers an alternative way for researchers to investigate 3D chromatin interactions and simulate their changes in model systems to test their hypotheses.