Cross-family learning transfers thermal adaptation between enzymes

THOR enables cross-family transfer of thermostability to enzymes without thermophilic homolog exposure, increasing apparent melting temperature by up to 15°C and up to 36°C after fine-tuning.

• Why it matters: It addresses whether thermal adaptation information can be transferred between evolutionarily unrelated protein families, beyond lineage-specific evolutionary ancestry.
• What they did: THOR learns background-specific sequence transformations between unpaired mesophilic and thermophilic populations while preserving the starting protein identity, then applies zero-shot designs across divergent enzyme families.
• The result: In malate dehydrogenase excluded from cross-family training, zero-shot designs raised apparent melting temperature by up to 15°C, with gains reaching 36°C after limited family-specific fine-tuning, and selected variants showed nearly threefold higher catalytic turnover.