Selected publications
* Equal contribution. Selected publications are listed below; see Google Scholar for more.
Programmable design of functional proteins from natural language.
Nature, 2026
Pinal is a 16-billion-parameter generative model pretrained on 1.7 billion protein–text pairs for designing proteins from natural language. Experiments confirmed the intended functions across four design tasks, including fluorescence and enzyme catalysis, without iterative experimental optimization.
DOMINO: Learning Domain Co-occurrence for Multidomain Protein Design.
bioRxiv, 2026
DOMINO learns domain co-occurrence patterns in natural proteins and uses these patterns to guide domain recombination for multidomain protein design.
Democratizing protein language model training, sharing and collaboration.
Nature Biotechnology, 2025
SaprotHub enables protein language model training, prediction, and sharing without coding through a collaborative Google Colab platform. I trained the 1.3-billion-parameter SaProt model for this work.
Decoding the Molecular Language of Proteins with Evolla.
bioRxiv, 2025
Evolla is an interactive protein-language model trained on 546 million protein–text pairs and refined through direct preference optimization. It supports natural-language queries about protein function, with zero-shot performance comparable to state-of-the-art supervised prediction.