Traditional structural biology often treats water as a homogeneous, passive background fluid. When visualizing a protein structure, the surrounding solvent is frequently stripped away to reveal a clean macromolecular surface. However, this "dry" representation omits a fundamental component of molecular biophysics: highly ordered, immobilized water molecules trapped within protein cores or at binding interfaces.
Rather than acting as inert fillers, these structural water molecules serve as active components of a protein's architecture. By explicitly integrating structural water into computational models, researchers have shifted from a binary view of biomolecular interactions to a tripartite framework (Protein–Water–Ligand or Protein–Water–Protein). This paradigm shift reveals how water acts as both a physical glue for molecular recognition and a critical node for long-range allosteric communication.
Water as Structural Glue in Molecular Recognition
Why does omitting water molecules lead to massive inaccuracies in drug discovery and binding affinity predictions? The answer lies in the geometric and energetic versatility of the water molecule. With its ability to act simultaneously as a hydrogen-bond donor and acceptor, a single trapped water molecule can bridge gaps between chemical groups that are otherwise too far apart to interact directly.
In Rapid and Accurate Prediction and Scoring of Water Molecules in Protein Binding Sites, Gregory A. Ross et al. note that approximately 85% of protein–ligand complexes rely on these water-mediated structural bridges. When a drug candidate enters a binding pocket, it does not simply displace all water molecules; instead, it often traps specific, high-affinity waters to complete its binding interface. As Nagasuma R. Chandra et al. demonstrate in PLIC: protein-ligand interaction clusters, failing to account for these interactions causes models to mischaracterize the steric and electrostatic landscape of the binding site. Explicitly modeling these interfacial waters is essential for capturing both the enthalpic gains of targeted hydrogen bonding and the entropic costs of water immobilization, as highlighted by Vittorio Limongelli et al. in Sampling protein motion and solvent effect during ligand binding.
Standard Binary Model:
[ Protein Residue ] - - - (Steric Gap / Weak Attraction) - - - [ Ligand ]
Tripartite Integrated Model:
[ Protein Residue ] <--- H-Bond ---> [ Ordered Water ] <--- H-Bond ---> [ Ligand ]
This structural stabilization is equally vital at protein–protein interfaces. In Progressive dry-core-wet-rim hydration trend in a nested-ring topology of protein binding interfaces, Ying He et al. map the spatial distribution of these molecules, revealing a "dry-core-wet-rim" topology. While the center of a protein-protein interface excludes water to form a hydrophobic core, the outer rim organizes a dense network of immobilized water molecules. This tripartite boundary layer provides a smooth energetic transition to the bulk solvent and seals the interface.
A stark example of this mechanism is found in host-pathogen interactions. In Computational simulations reveal the binding dynamics between human ACE2 and the receptor binding domain of SARS-CoV-2 spike protein, Cecylia S. Lupala et al. show that persistent water molecules trapped at the contact surface between the viral spike protein and the human ACE2 receptor act as a literal "molecular glue." These solvent-mediated hydrogen bonds stabilize the complex in ways that direct amino acid contacts alone cannot, directly influencing the virus's high binding affinity.
Completing the Graph for Allosteric Communication
Beyond stabilizing static interfaces, structural water molecules function as active transmission lines for allostery—the process by which a protein senses a stimulus at one site and transmits a structural signal to a distant functional site.
To model allostery computationally, researchers often represent proteins as mathematical graphs, where amino acid residues are nodes and non-covalent interactions are edges. If these graphs are built using only direct residue-to-residue contacts, the network appears fragmented. As Gaetano Invernizzi et al. argue in Communication Routes in ARID Domains between Distal Residues in Helix 5 and the DNA-Binding Loops, omitting the surrounding solvent leaves out critical pathways through which distal domains communicate.
Disconnected Graph (No Water):
[Residue A] ---- [Residue B] [Residue C] ---- [Residue D]
^
(Artificial Gap)
Connected Graph (With Structural Water Node):
[Residue A] ---- [Residue B] ---- [Water Node] ---- [Residue C] ---- [Residue D]
By integrating structurally conserved water molecules directly into the interaction graph as functional nodes, the true connectivity of the protein is restored. In Prediction of allosteric sites and mediating interactions through bond-to-bond propensities, Benjamin Amor et al. demonstrate that tracking how energy propagates through these hybrid residue-water networks allows for the highly accurate prediction of remote allosteric switches. Similarly, Nan Wu et al. in Prediction of protein allosteric signalling pathways and functional residues through paths of optimised propensity confirm that these ordered water networks are essential for maintaining the continuous pathways required for long-range signaling.
This network integration is particularly evident in transmembrane receptors like G protein-coupled receptors (GPCRs), which must transmit signals across the cell membrane. Basak Isin et al. in Identifying Ligand Binding Conformations of the β2-Adrenergic Receptor by Using Its Agonists as Computational Probes show that internal water networks act as structural stabilizers that bridge distant helical domains, allowing the receptor to transition smoothly between functional states.
More recently, high-resolution structural insights from Chongzhao You et al. in Molecular mechanism of pH sensing and activation in GPR4 reveals proton-mediated GPCR signaling mapped an intricate network of ordered water molecules within the core of the GPR4 receptor. These waters form a continuous hydrogen-bond network that stabilizes the active conformation of the receptor, acting as a structural relay that transmits extracellular proton-sensing events down to intracellular signaling partners.
Machine Learning and the Hydration Landscape
Because structural waters are highly dynamic and difficult to resolve experimentally, the field has increasingly turned to advanced computational frameworks to predict their locations and exploit them for design.
Early methods repurposed physics-based docking algorithms to identify stable hydration sites. However, modern approaches leverage deep learning to capture the complex, non-local dependencies of the tripartite interface. In Incorporating Water Molecules into Highly Accurate Binding Affinity Prediction for Proteins and Ligands, Diya Zhang et al. introduced GraphWater-Net, a model that explicitly represents protein-ligand-water complexes as unified graphs. By using a Graphormer architecture to process these explicit water nodes, the model achieves state-of-the-art accuracy in binding affinity predictions, proving that water's energetic contributions cannot be fully captured by implicit solvent approximations.
To scale this capability across the proteome, generative artificial intelligence has been deployed to map the complete hydration landscape of uncharacterized proteins. In Superwater as a generative AI framework to predict water molecule positions on protein structures, Xiaohan Kuang et al. developed a framework combining score-based diffusion models with equivariant graph neural networks. This approach allows for the rapid, high-throughput prediction of 3D structural water coordinates directly onto dry protein structures.
By treating water as an explicit, structural coordinate rather than a statistical background, these computational frameworks enable researchers to design drugs that intentionally displace or conserve specific water molecules, unlocking a higher level of precision in structure-based design.
Key Foundational Work
Sampling protein motion and solvent effect during ligand binding (2012) — contributed 1 concept elements to this direction.
Rapid and Accurate Prediction and Scoring of Water Molecules in Protein Binding Sites (2012) — contributed 1 concept elements to this direction.
Identifying Ligand Binding Conformations of the β2-Adrenergic Receptor by Using Its Agonists as Computational Probes (2012) — contributed 1 concept elements to this direction.