Vincent D. Zaballa
I work on machine learning methods in biology with a strong interest in integrating systems and structural biology with applications in drug discovery. Given the complexity of biological data, I seek to combine biological data into more robust machine learning models to reduce entropy of biological systems. A key application of this is in Bayesian optimal experimental design, which helps scientists decipher between competing biological hypotheses with maximum efficiency.
So far, the main software from my Ph.D. work is the JAX-based conditional normalizing flow library LFIAX with experimental design loss functions. I also contribute to open-source software, including an entropy-bounded sampler for discrete diffusion language models at Radical Numerics and reliable optimizer implementations in Optax.
selected publications
- Scalable Diffusion SBI for Compositional Inference under Simulator Misspecification2026