Austin Cheng

Excited to share Quetzal, a simple but scalable model for building 3D molecules atom-by-atom. 🐉 Named after Quetzalcoatl, the Aztec god of creation We equip a standard causal transformer with a per-atom diffusion MLP to model the continuous 3D position of the next atom. [1/3]

tagging folks who might be interested: Emily Jin, @andrei_nica @ChengHaoLiu1 @jarridrb @bose_joey @AlexanderTong7 @mmbronstein @ccdc_cambridge @wellingmax @xie_tian @k_neklyudov @lazar_atan @martoskreto @ADuvalinho @rishabh16_ @smnlssn @danyalrehman17 @FrankNoeBerlin @DidiKieran

Made possible by an awesome team: Alston Lo @alston1o, Luka Mucko @srboljubbosanac, Andy Cai, Alastair Price @Albd90, Wojciech Matusik, and Alán Aspuru-Guzik @A_Aspuru_Guzik

💠Introducing Clari, a generative model to predict organic crystal structure in seconds. Clari’s speed makes it feasible to do virtual screening of organic molecules on the basis of solid-state properties. co-first with Alston Lo @alston1o and Luka Mucko @srboljubbosanac

Code is now available: github.com/aspuru-guzik-g… Try it out now in Colab: colab.research.google.com/github/aspuru-…

Excited to share Quetzal, a simple but scalable model for building 3D molecules atom-by-atom. 🐉 Named after Quetzalcoatl, the Aztec god of creation We equip a standard causal transformer with a per-atom diffusion MLP to model the continuous 3D position of the next atom. [1/3]

🧵(1/6) Delighted to share our @icmlconf 2025 spotlight paper: the Feynman-Kac Correctors (FKCs) in Diffusion Picture this: it’s inference time and we want to generate new samples from our diffusion model. But we don’t want to just copy the training data – we may want to sample from an annealed distribution (e.g. to find modes), or combine two experts by sampling from the product of conditional densities (e.g. to design molecules that bind to two protein pockets). FKCs let us do this!✨But how? We explain in the 🧵 below. (work with an amazing amazing team -- @tara_aksa @OhanesianViktor Roberto Bondesan @A_Aspuru_Guzik @ArnaudDoucet1 @brekelmaniac @AlexanderTong7 and @k_neklyudov) 📃arxiv.org/abs/2503.02819 🧑‍💻github.com/martaskrt/fkc-…

Collaboration with the amazing @ChongSun20 and @A_Aspuru_Guzik arxiv.org/abs/2505.13791 [3/3]

This scalable autoregressive approach enables: - Fast generation with quality competitive with all-atom diffusion - Tractable, exact log-likelihood computation - Arbitrary-size generation for hydrogen decoration and scaffold completion [2/3]

If you know a molecule's molecular formula and moments of inertia, can you predict its all-atom 3D structure? I'm at ICLR presenting Stiefel Flow Matching for Moment-Constrained Structure Elucidation! Poster #1 today at 3 pm. arxiv.org/abs/2412.12540 github.com/aspuru-guzik-g…