Mihir, a senior studying Computer Science at Georgia Tech in the BSMS program specializing in Machine Learning and Theory threads. His research interests lie at the intersection of machine learning and structural biology, specifically with generative modeling. Additionally, he continues to be excited by advances in single-cell genomics and how graph representation learning can be used to make inferences.
Since the Human Genome was sequenced in 2001, countless studies have utilized bulk sequencing. You can't ever reach perfection, but you can believe in an asymptote toward which you are ceaselessly striving.” - Paul Kalanithi However, this only captures the linear 1D structure of the genome. In reality, the genome is organized in this complex 3D structure that makes up the chromatin, and hidden away in these folds are possible enhancer-promoter interactions that may affect gene expression. We use single-cell Hi-C sequencing, a proximity-ligation-based method where the contacts are stored in these contact map representations. However, these contact maps are plagued with sparsity at a single-cell level. We seek to use modern generative deep learning to mitigate this issue. Specifically, we utilize diffusion models, which have been shown to outperform GANs on image synthesis tasks. And, as contact maps and images are represented the same computationally, we find diffusion to be a simple extension of our task. Thus, we propose Daifuku: a diffusion denoising model that uses bulk Hi-C to inform the generative imputation of sparse single-cell Hi-C maps. We find that Daifuku imputation mitigates sparsity issues while preserving the single-cell heterogeneity between contact maps. With these initial promising results, we look forward to validating our method on other real datasets to gain more biological insights and elucidate the true 3D structure of the genome.
Project: Daifuku: elucidating 3D genome structure via diffusion-denoising imputation of scHi-C contact maps
Mentor: Ruochi Zhang
PI: Bonnie Berger, MIT