Vincent Zaballa

[Keshigeyan Chandrasegaran @keshigeyan]

Excited to share Halcyon, our latest research on multimodal pretraining!

Building strong text-to-image generative models is simpler than we think!

Blog: https://t.co/HOrwIJIHTf
Models: https://t.co/xBPJkEAOPU
Paper: https://t.co/8kROIz1c7L

[Vincent Zaballa @VZaballa]

More information theoretic diffusion models ^^

[Sander Dieleman @sedielem]

This is a beautiful analysis of the order in which diffusion models resolve uncertainty about information at different abstraction levels in different representation spaces, with significant implications for latent diffusion.

https://t.co/2nzL1Gepb8

[Vincent Zaballa @VZaballa]

More structures coming soon! And thanks to @dinhscience for the structural biology feedback

[Vincent Zaballa @VZaballa]

Fun weekend project-there’s a gap between what protein structure models predict and what decades of structural biology have taught us

Found that known interface interactions can help guide protein complex assembly, but incorrect or overly rigid ones can steer us wrong

[Vincent Zaballa @VZaballa]

Fine-tuning on this kind of interface evidence can improve future predictions and reduce epistemic uncertainty around protein–protein interactions

Idea grew out of TAing @chang_c_liu class. More reason to get ML+bio folks together to identify these gaps

[Ryan Krueger @RyanKrue]

Over the past few weeks, we have been putting Gemini 4 Argon to the test on protein engineering workflows at Google DeepMind, and we have some interesting initial (zero shot!) results to share. 🧵👇

[Vincent Zaballa @VZaballa]

Hopefully this makes it easier for new entrants into SBI, especially those with design-dependent simulators. Open for collaborations and questions. Code coming soon

[Vincent Zaballa @VZaballa]

Excited to share the final paper of my PhD! Driven by challenges in modeling large systems biology models with diffusion transformers, we developed a framework to overcome difficulties in 1) compositional sampling, 2) hierarchical sampling and 3) the sim-2-real gap 🤖 ↔️ 🌍

[Vincent Zaballa @VZaballa]

This is a lot because it's the culmination of trying simpler things and realizing that the SBI toolbox didn't quite have everything I needed for my domain applications. eg what good is compositional sampling if your probabilistic model is misspecified?