Navigating Protein Fitness Landscapes Through Simulated Evolutionary Jumps

Natural and directed evolution are powerful for enhancing protein function, but their utility is limited by local exploration, as each mutation must maintain function. Here we introduce SPIN-JEvo (Sequence-based Prediction with Integrated Neural Network by Jump and Evolution), a computational framework that decouples exploration from functional constraint by initiating from functional sequences and jumping to likely nonfunctional variants with ~20% random substitutions. These sequences then evolved using a genetic algorithm to reach remote homologs. Despite training on only a few binary-labeled sequences and starting from a cluster of close sequence neighbors, SPIN-JEvo identified previously inaccessible remote homologs in minutes for both a structured enzyme (tRNA-specific adenosine deaminase) and an intrinsically disordered antitoxin (CcdA)—without requiring structural information during model training or sequence generation. SPIN-JEvo shifts protein engineering from incremental local optimization to global navigation of sequence space, enabling the discovery of novel functional families beyond the reach of traditional methods.