Dissertation Title: A Geometric Deep Learning Framework for 3D Biomolecular Compression, Generation and Accelerated Simulation

Abstract: Biomolecules organize over vast scales of length and time, forming the bridge from physical principles to functional complexity. While this layered structure defies classical bioinformatics and traditional simulation approaches, breakthroughs in deep learning suggest a promising path for tackling the multi-resolution organization and dynamics of biomolecules. In this thesis, we advance these methods with geometric neural architectures and generative algorithms for learning 3D biomolecular data at coarser, more tractable and effective spatiotemporal scales. Leveraging the physical symmetries of 3D molecules, we draw on irreducible representations of the Euclidean group to construct coarse atomistic encodings at the protein or RNA residue level. From these representations, we build a suite of Euclidean-equivariant neural networks that enable generation, design, and accelerated simulation. First, we introduce autoencoders for deep protein coarse-graining and compression, uncovering structured latent spaces conducive to efficient generation. Then, we explore a multimodal model for co-generating sequence and all-atom structure of RNA, enabling efficient, simultaneous optimization in structural design. Finally, we train large-jump transfer operators for accelerated protein dynamics simulation, profiling quality tradeoffs. We show that the learned simulators generalize with kinetic accuracy across proteins, while achieving orders-of-magnitude speedup over traditional methods. Taken together, these efforts advance a paradigm grounded in physical symmetries for treating 3D biomolecular data, establishing foundations for efficient exploration of vast molecular design spaces and accessible simulation of long-timescale dynamics.

Committee members: 
Joseph Jacobson, Associate Professor at MIT Media Lab
Tess Smidt, Associate Professor at MIT RLE
Emine Kucukbenli, Research Manager at NVIDIA