A systematic evaluation of protein allosteric site prediction tools with independent datasets

Allostery plays a critical role in protein dynamics and is essential for many biological functions. Over the past decade, various computational approaches have been proposed for predicting allosteric sites. However, the strengths and weaknesses of each method are not well understood. In this study, we created two independent datasets that had not been used in selected computational protocols: a CAPASP-General subset comprising holo state allosteric proteins and a CAPASP-Unbound subset comprising apo state allosteric proteins. We then systematically evaluated the accuracy of five allosteric site prediction tools across five dimensions: sensitivity, specificity, F1-score, MCC value and ranking capability. The results indicated that the machine learning models PASSer and APOP, which are based on protein physicochemical properties, not only achieved the highest success rate in sensitivity prediction but also lead in average F1-score and MCC value. However, these models performed better with the CAPASP-General subset than with the CAPASP-Unbound subset, suggesting that the prediction models require further improvement. These findings could facilitate the selection of appropriate prediction models for different allosteric proteins and enhance our understanding of protein function and regulatory mechanisms.