I have the pleasure of being a PhD student in mathematics (de jure) and biochemistry (de facto) at Stanford, advised by Rhiju Das (Biochemistry) and George Papanicolaou (Mathematics). My research develops computational methods for RNA structural biology, connecting RNA 3D structure to experimental observables—and using that connection in both directions.
Structure ↔ reactivity. RNA chemical probing experiments (SHAPE, DMS) measure per-nucleotide flexibility. I build models that predict these profiles from 3D structure, enabling computational validation of predicted structures against cheap experimental data. This includes an SO(3)-equivariant transformer and a learnable Gaussian network model.
Reactivity-guided structure prediction. Structure prediction models like AlphaFold3 don't use experimental probing data. I fine-tune Protenix (an open AlphaFold3 reproduction) with LoRA adapters, using chemical mapping profiles as a supervisory signal via reward-weighted denoising.
Conditional structure generation. I also develop lightweight, trainable-from-scratch models for RNA 3D structure prediction as an alternative to fine-tuning large pretrained models, exploring flow matching and variational approaches with equivariant architectures.
Energy models from structure. Learning molecular energy functions directly from structural ensembles via score matching, bypassing hand-tuned force fields and quantum chemistry.
Cryo-EM reconstruction. Small RNA molecules produce sparse, noisy cryo-EM data. I develop autocorrelation-based reconstruction methods that could enable structure determination for molecules currently too small for conventional cryo-EM.
Read depth bias in MaP-seq. Transcriptome-scale MaP-seq libraries suffer from extreme read depth biases. I develop predictive models that enable library rebalancing, reducing zero-read transcripts from 16% to 2.6% in blind tests.