Darin Tsui

DarinTsui.png

Hi there! My name is Darin. I’m a Ph.D. candidate and NSF GRFP fellow at the Georgia Institute of Technology studying Electrical and Computer Engineering and Bioengineering. I previously received my BS degree at the University of California San Diego in Bioengineering.

I currently work with Dr. Amirali Aghazadeh in the AI ML & Information Group, where I develop scalable tools for generative models in the biosciences. I am currently supported by the NSF Graduate Research Fellowship. My research focuses on:

  • Leveraging mechanistic interpretability frameworks, such as cross-layer transcoders and sparse autoencoders, for circuit discovery, protein design, and extracting biological mechanisms from protein language models.

  • Extracting higher-order interactions in models 1000-fold quicker than state-of-the-art methods by developing algorithms at the intersection of signal processing, coding theory, and combinatorics.

  • Developing principled generative models for high-throughput cryo-electron microscopy (cryo-EM), autoregressive protein generation, and variant effect prediction.

My background is highlighly interdisciplinary. I previously validated deep learning models at Surgalign (acquired by Xtant Medical), developed an optical surgical platform in the Talke Biomedical Device Lab, and designed a bioelectronic COVID-19 test in the Integrated Systems Neuroengineering Laboratory.

Outside of research, I co-chair Bioengineering Day at Georgia Tech, which aims to highlight interdisciplinary research. Previously, I served as President of the IEEE student branch and helped found the IEEE EMBS chapter at UC San Diego.

News

Oct 02, 2025 I gave a guest lecture in ECE 8803–Generative and Geometric Deep Learning–on explainability and mechanistic interpretability in AI. You can find the recording of the lecture here.
Sep 24, 2025 My work on adapting sparse autoencoders for low-N protein engineering into the NeurIPS AI4Science workshop.
Sep 18, 2025 My work on explaining biological sequence models at scale has been accepted into NeurIPS 2025!
Sep 18, 2025 Our work on adapting speculative decoding for protein generation was accepted to NeurIPS 2025 as a spotlight!
Jul 16, 2025 Our work on assessing and explaining mutations in myocilin was accepted to the Machine Learning in Computational Biology (MLCB) meeting.

Selected Publications

  1. ProtoMech.png
    Protein Circuit Tracing via Cross-layer Transcoders
    Darin Tsui, Kunal Talreja , Daniel Saeedi , and 1 more author
    arXiv preprint arXiv:2602.12026, 2026
  2. ddpm.gif
    cryoSENSE: Compressive Sensing Enables High-throughput Microscopy with Sparse and Generative Priors on the Protein Cryo-EM Image Manifold
    Zain Shabeeb , Daniel Saeedi , Darin Tsui, and 2 more authors
    arXiv preprint arXiv:2511.12931, 2025
  3. SHAPzero.png
    SHAP zero explains biological sequence models with near-zero marginal cost for future queries
    Darin Tsui, Aryan Musharaf , Yigit Efe Erginbas , and 2 more authors
    In Advances in Neural Information Processing Systems (NeurIPS) , 2025
  4. SpecMER.png
    SpecMER: Fast protein generation with K-mer guided speculative decoding
    Thomas Walton , Darin Tsui, Aryan Musharaf , and 1 more author
    In NeurIPS (Spotlight) , 2025