Proteins carry out nearly all life processes by interacting with other molecules, particularly other proteins, through protein–protein interactions (PPIs).[1], [2] These interactions are essential for almost all aspects of cellular function, including signal transduction, immune responses, enzymatic regulation, and structural organization. At the molecular level, PPIs are governed by a range of forces – such as hydrogen bonding, hydrophobic effects, electrostatics, and van der Waals interactions – that drive specific and dynamic recognition between complementary surfaces. Understanding the principles underlying these interactions is crucial for dissecting biological mechanisms and designing therapeutic interventions.[3], [4], [5], [6]

The accurate determination of protein–protein complex structures is key to unlocking the roles of PPIs in health and disease.[7], [8] Experimental techniques such as nuclear magnetic resonance (NMR), X-ray crystallography, and cryo-electron microscopy (cryo-EM) have been instrumental in solving such structures.9 However, their high cost, long experimental timelines, and limited scalability have motivated the development and growing use of computational modeling as a complementary approach.

Traditionally, protein–protein docking has served as the primary computational strategy for modeling PPIs. Docking approaches fall into two broad categories: template-based and template-free (Table 1). Template-based docking relies on structural homologs available in the Protein Data Bank (PDB)10 and works well when close templates exist. In the absence of such templates, template-free docking explores binding modes by sampling conformational space and scoring predicted complexes. Despite decades of refinement, template-free methods often struggle with accuracy owing to the vast search space and limitations in scoring functions.

Recent breakthroughs in artificial intelligence (AI) and deep learning have fundamentally transformed the landscape of protein complex prediction.[11], [12] Unlike traditional pipelines that treat structure prediction and docking as separate tasks, modern end-to-end deep learning approaches, such as AlphaFold[13], [14] and its derivatives, can simultaneously predict the 3D structure of entire complexes.[15], [16] These methods leverage large data sets and neural networks to directly infer residue–residue contacts and structural configurations, bypassing the need for explicit docking steps and offering unprecedented predictive accuracy. Although end-to-end AI models dominate current practice, the prediction of larger assemblies often still follows a modular two-step workflow: first, predicting individual subunit structures, then assembling them computationally into full complexes.[17], [18] Fig. 1 provides a visual summary of this general approach and highlights that the process can also be supported by integrative modeling methods that incorporate experimental data, such as cryo-EM, X-ray diffraction (XRD), and cross-linking mass spectrometry (XL-MS).

In recent years, rapid advances in AI have reshaped the landscape of PPI modeling. Understanding these new capabilities – as well as the persisting challenges of modeling protein flexibility, large assemblies, and disordered regions – is crucial for both basic biological research and therapeutic innovation. In this review, we aim to provide an up-to-date perspective on computational strategies for PPI structure prediction, highlighting recent breakthroughs and pointing toward future opportunities for further improvement.