Doctoral Researcher · University of Miami
I build AI and deep learning for the three-dimensional architecture of life, from the folding of the genome to single cells and the molecules they build.
The through-line
Graph transformers and geometric deep learning: the same structural way of thinking, from the folding of whole genomes down to the surfaces where proteins meet. Scroll to travel between them.
3D chromatin folding
How the two meters of DNA in every cell fold into loops and domains, and what it means when that folding changes across time or disease.
I built it to flag exactly where and when the folding goes wrong across time.
Single-cell genomics
Reading one cell at a time: linking a single cell's 3D genome structure to the genes it actually switches on.
It reads one cell's 3D genome and predicts the genes that cell switches on.
Biomolecular structure
Scoring and understanding the structures of proteins and protein-RNA complexes, the machinery the genome builds.
It scores how accurate a predicted protein interface really is.
An unsupervised deep-learning tool that flags exactly where and when the 3D folding of the genome goes wrong across time.
A graph transformer that predicts a single cell's gene expression directly from that same cell's 3D genome structure.
Status & scope
An ensemble of graph neural networks that scores how accurate a predicted protein interface really is.
I started in computer engineering in Kathmandu, drawn by a restless curiosity for how things work. That curiosity turned into a commitment to biological discovery through computation — computational biology and bioinformatics: from graph neural networks for genomic data to systems that catch the moment genome structure shifts during development or disease.
I conceived and led HiC4D-SPOT, scHiGex, and ARC as first author, and I work embedded with wet-lab teams: for the CITED2 study in EMBO Molecular Medicine, a druggable-target paper, I ran the computational analysis.
Research program
One way of thinking should read biological structure across every scale. Three claims I am working to prove:
Detect structural failure across the genome's time axis.
HiC4D-SPOT shows 4D folding drift is detectable without labeled disease states; the same unsupervised idea should generalize across chromatin transitions.
Treat 3D structure as a sufficient statistic for expression.
scHiGex recovers cell types from single-cell Hi-C alone (ARI 1.00 on HiRES); next is predicting how expression shifts when structure is perturbed.
Score interface accuracy before an experiment confirms it.
ARC, my CASP16 entry, estimates per-residue quality where proteins meet — and extends to protein-RNA and larger complexes.
Research is where my focus sits. These are things I build for the craft of it — when an idea is worth making real.
A post-hoc reliability layer for single-cell perturbation predictors. Rather than build a better predictor, it scores which individual predictions to trust, giving each a calibrated error and a 90% conformal interval over any model's outputs. The scoring core is ported to JavaScript and runs entirely in the browser, so a lab's predictions never leave their machine.
An interactive study of associative memory: store patterns as memories, corrupt them, and watch a Hopfield network recall the original. It walks from the classical 0.138·N capacity wall to the modern continuous model — one-step softmax retrieval that is, exactly, the transformer's attention.
A living 3D map of the Kathmandu Valley, built from OpenStreetMap geometry and real terrain, with eight monuments procedurally modelled from surveyed dimensions. Time of day, seasons, weather and night are driven by real astronomy (Meeus-style sun and moon) and WMO climate normals; real aircraft stream in over ADS-B.
Graduate Teaching Assistant, Spring 2025, Spring 2024
Visiting Lecturer, Spring 2025, Fall 2024 (2 sections)
Graduate Teaching Assistant, Spring 2026
Graduate Teaching Assistant, Fall 2023
ARC: Assessment of Interface Residue Conformation using an Ensemble of Graph Neural Networks, Computing Day, University of Miami (May 2026).
Co-authored “Inferring the qualities of protein-RNA models with graph transformers,” extending structure-quality estimation beyond proteins to multi-biomolecule systems.
Co-authored SCW, which reconstructs whole-genome 3D structures from extremely sparse single-cell Hi-C data, benchmarked across multiple cell types.
Read the paperSubmitted ARC, an ensemble of graph neural networks for interface-residue conformation assessment, my first-author project from participating in CASP16.
Co-authored work identifying CITED2 as a druggable epigenetic switch that promotes axonal growth and functional recovery via the HDAC inhibitor Panobinostat.
Read the paperServing as a reviewer for ISMB 2026, the flagship conference of the International Society for Computational Biology.
An unsupervised ConvLSTM autoencoder that detects structural anomalies in spatiotemporal 3D chromatin data, TAD disruptions, loop perturbations, and cohesin-mediated remodeling.
Read the paperHiC4D-SPOT: a spatiotemporal outlier detection tool for Hi-C data, Computing Day, University of Miami (April 2025).
My first PhD paper, scHiGex, published in NAR Genomics and Bioinformatics, a graph transformer bridging single-cell Hi-C data and gene-expression prediction.
Read the paperPredicting single-cell gene expression from single-cell Hi-C using a graph transformer, Computing Day (April 2024) & Sylvester Comprehensive Cancer Center Annual Retreat (Oct 2024).
Began my PhD in Computer Science under Dr. Zheng Wang at the University of Miami, Coral Gables, Florida.
I’m glad to talk — about research collaborations, a faculty-track opening, or a machine-learning team building in biology. Email reaches me fastest.