Geometric Intuition Enables Resource‐Efficient Screening for β ‐Barrel‐Like Folds in Viral Structural Dark Matter

ABSTRACT

Predicted protein structures now enable fold discovery beyond the reach of sequence homology, but scalable and interpretable methods remain limited. We present Cooper‐Beta, a resource‐efficient, chain‐resolved framework for detecting β ‐barrel‐like folds from strand connectivity. Cooper‐Beta converts DSSP‐assigned physical β ‐strands into an adjacency graph and evaluates four measurements using calibrated direct rules or logistic regression. On a grouped held‐out benchmark of 466 positive and 1063 negative chains, the direct classifier achieved 87.77% recall and 97.65% specificity, while the learned classifier reached 94.42% recall and 97.08% specificity. A separate evaluation of 114 manually classified CATH Class II non‐barrels yielded specificities of 95.61% and 92.98%, respectively. Both classifiers retained prediction‐flip rates below 3% through 1.0 Å Gaussian C α perturbation, with stronger learned‐model sensitivity at higher noise. Screening 351 242 structures from the Big Fantastic Virus Database yielded 5803 joint‐positive candidates; sequential Swiss‐Prot and UniRef50 searches identified 1568 sequence‐available candidates without a passing Swiss‐Prot or UniRef50 match. Cooper‐Beta provides an interpretable route from predicted structural repositories to β ‐barrel‐like candidates for experimental follow‐up.