AI-driven prediction of 14-3-3–intrinsically disordered protein interactions: From SLiM discovery to structural modeling of dynamic binding interfaces

Abstract

Artificial intelligence (AI) has revolutionized structural biology by enabling the prediction of protein structures, biomolecular complexes, and interaction interfaces with unprecedented accuracy. The field has evolved from sequence motif searches and conventional machine learning algorithms to deep learning architectures and protein models capable of learning structural and functional representations directly from protein sequences. Recent advances, including protein language models (PLMs), AlphaFold2/3, ESMFold, and Boltz-2, have transformed the prediction of protein–protein interactions by integrating sequence, evolutionary, and structural information while reducing reliance on experimental structural data. These methods have expanded the ability to identify binding interfaces, characterize biomolecular complexes, and investigate interactions involving intrinsically disordered proteins and short linear motifs (SLiMs), although accurately modeling dynamic and context-dependent interactions remains a significant challenge. In this review, we summarize the evolution of AI-driven approaches for predicting protein–protein interactions, from classical machine learning to modern foundation models, highlighting their strengths, limitations, and emerging applications. The 14-3-3 protein family serves as a representative example to illustrate how AI has progressively improved the identification of phosphorylation-dependent binding motifs and interaction partners. Finally, we discuss future perspectives for integrating protein language models, structural prediction, and experimental data to achieve a more comprehensive understanding of protein interaction networks and to accelerate the discovery of regulatory mechanisms and therapeutic targets.