Repurposing Protein Language Models for Latent Flow–Based Fitness Optimization

Amaru Caceres Arroyo, Lea Bogensperger, Ahmed Allam, Michael Krauthammer, Konrad Schindler, Dominik Narnhofer · 2026

Read Repurposing Protein Language Models for Latent Flow–Based Fitness Optimization by Amaru Caceres Arroyo, Lea Bogensperger, Ahmed Allam, Michael Krauthammer, Konrad Schindler, Dominik Narnhofer (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/arxiv-2602-02425.txt.

Description

Amaru Caceres Arroyo and 5 co-authors repurpose protein language models to design fitter proteins without expensive sampling, in a 2026 arXiv paper. Protein fitness optimization searches a vast space of variants where high scorers are extremely rare. Many methods underperform or need costly gradient-based sampling at design time.

Their framework, CHASE, compresses embeddings from a pretrained protein language model into a compact latent space. Those embeddings carry evolutionary knowledge learned from natural sequences. A conditional flow-matching model, trained with classifier-free guidance, then generates high-fitness variants directly, with no predictor steering the sampling steps. CHASE reached the best reported performance on AAV and GFP design benchmarks, two standard protein engineering tests. Bootstrapping with synthetic data improved results further when real data was scarce.

Catalog record

CREATORAmaru Caceres Arroyo, Lea Bogensperger, Ahmed Allam, Michael Krauthammer, Konrad Schindler, Dominik Narnhofer
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE44 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-02425

Browse the full library · About the Tower of Babel