Simplifying in silico protein evolution with minimal screening by unZipro.
TL;DR
UnZipro is presented, an efficient, scalable, and generalizable framework for zero-shot, in silico protein evolution, that integrates a compact, pre-trained inverse folding model with meta-learning to derive family-specific fitness landscapes.
Abstract
Despite the significant progress made in recent years, current AI-driven protein engineering methods often suffer from iterative experimental validation, limited success rates, modest improvements, and/or high computational costs. Moreover, their utility for evolving plant proteins remains unexplored. Here we present unZipro, an efficient, scalable, and generalizable framework for zero-shot, in silico protein evolution. unZipro integrates a compact, pre-trained inverse folding model with meta-learning to derive family-specific fitness landscapes. We demonstrate unZipro's ability to directly identify high-activity variants from minimal libraries (∼10 candidates) with an average success rate of 61% across nine diverse proteins, achieving up to a 28-fold increase in the gene-editing activity of T5E-CasΦ2 fusion variants. Leveraging unZipro, we developed state-of-the-art plant gene-editing nucleases, superior firefly luciferase variants, enhanced rice transcription factors, and barley-derived antiviral proteins with elevated potency. Overall, unZipro represents a new paradigm for cost-effective, widely accessible, and transformative protein engineering in biology, biotechnology, and agriculture.
