OpenSplitProtein: Interpretable Split-site Prediction for Split Tolerance and Fragment Reassembly

Split-protein engineering divides a folded protein into inactive fragments that regain activity upon reassembly, providing spatial and temporal control over cellular function. However, choosing a split site remains a trial-and-error process, and structure-based predictors have not been benchmarked against the positional baseline, defined as how far the split lies from the nearest chain terminus. Here we show that this positional baseline, rather than the structure around it, explains most of the marginal association between cross-cut contact counts and both backbone insertion cost and intein reassembly. Across 13,622 backbone positions in 293 domains and 129 clusters, the within-domain partial rank correlation of the cross-cut contact count reverses from 0.174 to -0.108 [-0.156, -0.036] once the length of the shorter fragment is held fixed. By contrast, local packing density keeps a partial rank correlation of 0.409 [0.378, 0.446] and raises the out-of-fold rank correlation from 0.188 to 0.485, so it carries structural information that the positional baseline does not. OpenSplitProtein ranks 21 experimentally validated split sites at a median percentile of 20.6, compared to 23.5, 33.0, and 33.0 for three existing models. By providing an interpretable, baseline-calibrated split-site predictor, OpenSplitProtein enables the systematic design of dimerization-free AND logic gates, cellular biosensors, and modular conditional therapeutics.