Zero-shot deep learning with multi-objective optimization improves thermostability of zearalenone hydrolase and xylanase

Enhancing enzyme thermostability is crucial for industrial applications requiring robust performance under extreme conditions. Structure-based protein design models excel at improving thermal stability but often compromise enzymatic activity, while sequence-based models better preserve activity but struggle to enhance thermostability. It is challenging to efficiently generate multi-site mutants with both improved thermostability and intact activity using minimal experimental effort. Here, we used zearalenone hydrolase (RmZHD) and xylanase as model systems to evaluate different strategies for multi-site mutation design: (i) structure-based design with the ABACUS-R model, (ii) sequence-based design with the ProGen2 or MSA Transformer, (iii) integrated approaches combining either ProGen2 or MSA Transformer with ABACUS-R via Markov Chain Monte Carlo sampling with multi-objective scoring. Results showed that designing with ABACUS-R increased thermal stability by ~15°C but caused 80–100% activity loss. Sequence-based designs retained ~19% wild-type activity but failed to improve thermostability. Notably, zero-shot designs from integrating ABACUS-R with MSA Transformer achieved significant thermostability gains ( ΔT m ~8°C) while preserving >95% wild-type activity. This highlights the potential of combining sequence-and structure-based deep learning models for developing industrially relevant thermostable enzymes. • Reveals limits of sequence and structure models in thermostable enzyme design. • Multi-objective optimization yields thermostable multi-site enzyme variants. • Achieves thermostability improvement without experimental screening. • Combines sequence and structure-based deep learning models for enzyme engineering.