Master's Thesis (Incoming)
Incoming EPFL Master's thesis, hosted at ETH Zurich's Laboratory for Systems and Synthetic Immunology (Prof. Sai Reddy). A 26-week research project on diffusion language models for antibody design, starting October 2026.
- ▸EPFL Master's thesis hosted at ETH Zurich, in the Laboratory for Systems and Synthetic Immunology (Prof. Sai Reddy)
- ▸Research direction: diffusion language models for antibody sequence generation, including CDR-H3 generation conditioned on the antibody framework
- ▸Planned work on comparing guidance methods for generation and on structural modeling and validation of generated candidates
- ▸26-week project, starting October 2026
Overview
This is my EPFL Master's thesis, hosted at ETH Zurich in the Laboratory for Systems and Synthetic Immunology, supervised by Prof. Sai Reddy. It counts toward my EPFL degree; it is not a separate ETH Zurich degree.
Incoming
The thesis starts in October 2026 and runs for 26 weeks. This page describes the planned direction; it will be updated with concrete methods and results once the work is underway.
Planned Direction
The topic is diffusion language models for antibody design. Antibodies bind their targets largely through a small number of highly variable loops, and the third loop of the heavy chain (CDR-H3) is the hardest to design because it is the most diverse. The plan is to use diffusion language models to generate antibody sequences, with a focus on generating CDR-H3 conditioned on the surrounding framework, or scaffold.
The directions I expect to explore include:
- diffusion language models for antibody sequence generation
- CDR-H3 generation conditioned on the antibody framework
- comparing and investigating guidance processes for generation
- studying antibody repertoire embeddings
- developing and evaluating model architectures
- structural modeling and validation of generated candidates
Why This
It brings together the two threads I have been working on: language models and the systems and engineering side of training them. Applying generative models to a real biological design problem, where a generated sequence has to be evaluated against structure rather than just a held-out loss, is the part I am most looking forward to.