Generative and geometric deep learning for 3D genome organization and biomolecular interaction modeling

The three-dimensional organization of the genome and the interactions between biomolecules are fundamental to understanding cellular function and disease. Recent advances in sequencing technologies and structural biology have generated unprecedented amounts of biological data, creating new opportunities to study these processes. However, these data are often sparse, noisy, and highly complex, making it difficult to accurately model biological systems using conventional computational methods. This dissertation explores how generative and geometric deep learning can address these challenges by learning meaningful biological representations from high-dimensional genomic and structural data. The first part of this dissertation focuses on modeling three-dimensional genome organization from single-cell Hi-C data. Because single-cell Hi-C experiments capture only a small fraction of chromatin interactions, computational methods are essential for recovering missing information and revealing genome architecture. I first developed a deep residual generative adversarial network to enhance sparse single-cell Hi-C contact maps by recovering biologically meaningful chromatin interactions. Building on this work, I proposed HiCDiff, a diffusion model that iteratively denoises chromatin contact maps and produces more accurate representations of genome organization. These improved contact maps provide a stronger foundation for downstream analyses and structural modeling. To further characterize genome architecture, I developed an SO(3)-equivariant graph neural network that reconstructs three-dimensional chromosome structures directly from single-cell Hi-C data. By incorporating geometric equivariance into the learning process, the model captures the spatial relationships among genomic loci and reconstructs chromosome conformations with improved accuracy and biological consistency. The second part of this dissertation extends geometric deep learning to biomolecular interaction modeling. I developed MPBind, a multitask framework for predicting protein binding sites by integrating protein language model embeddings with equivariant graph neural networks. This approach combines evolutionary information encoded in protein sequences with three-dimensional structural features, enabling accurate prediction of diverse protein interactions while achieving strong generalization across previously unseen proteins. Although these studies address different biological problems, they share a common theme: combining generative models with geometry-aware neural networks to learn from complex biological data. Generative deep learning improves the quality of sparse and noisy experimental measurements, while geometric deep learning enables models to capture the structural principles underlying biological systems. Together, the methods presented in this dissertation provide a unified framework for studying genome organization and biomolecular interactions, demonstrating how modern artificial intelligence can advance computational genomics and structural biology and support future discoveries in biology and precision medicine.