Sampling out-of-distribution chemical spaces via Bayesian flow

Nianze Tao, Minori Abe · 2026

Read Sampling out-of-distribution chemical spaces via Bayesian flow by Nianze Tao, Minori Abe (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/PMC13580001.txt.

Description

Nianze Tao and Minori Abe show that a Bayesian flow model can dream up molecules better than any it was trained on, in a 2026 study. Generators that learn a data distribution tend to copy it, which caps drug hunting at known chemistry. Their model builds molecules through a flow of probability updates rather than noise removal, a design that turns out to wander productively beyond the training fold.

A reward driven training add on steers the dreaming, and a solver style rollout speeds it up. A part by part building trick, added in training and sampling alike, lifted quality further. Tests on small molecules and proteins beat leading rivals at finding strong out of family candidates. Theory in the paper backs why the wandering works.

Catalog record

CREATORNianze Tao, Minori Abe
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE81238
LICENSEhttps://creativecommons.org/licenses/by/4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/article/MED/PMC13580001

Browse the full library · About the Tower of Babel