In drug discovery and protein engineering, the most common metric of success is binding affinity—a single number representing how tightly two molecules stick together. However, this aggregate value is a "black box" that obscures the underlying physical mechanisms. Knowing that a drug fails against a mutant protein is useful; knowing why it fails—whether due to a lost hydrogen bond, a change in surface charge, or increased molecular "floppiness"—is actionable.
Research across molecular dynamics and machine learning has converged on energetic decomposition: the practice of breaking down the total free energy of binding into its constituent physical forces and spatial contributions. By shifting focus from the what to the how, researchers can pinpoint "hotspots" of stability and predict resistance before it emerges in the clinic.
Beyond the Aggregate Score
Traditional methods often treat binding energy as a monolithic value. However, Debby D. Wang and colleagues demonstrated in Personalized prediction of EGFR mutation-induced drug resistance in lung cancer that aggregate energy is often too noisy for precise prediction. By decomposing energy into specific physical components—such as electrostatic forces, Van der Waals interactions, and solvation terms—they created features for machine learning that significantly outperformed models based on genetics alone.
This granular approach allows researchers to see the "invisible" drivers of resistance. For example, a mutation might not change the total binding energy significantly, but it might "rewire" the receptor to bond more strongly with neighboring proteins rather than the drug. Mengxu ZHU and Debby D. Wang used this logic to show how lung cancer mutations cause "amplified crosstalk" between receptors, allowing cancer cells to bypass targeted therapies (Genotype-determined EGFR-RTK heterodimerization and its effects on drug resistance in lung Cancer treatment revealed by molecular dynamics simulations).
Spatial Mapping and Residue Hotspots
If physical decomposition tells us what forces are at play, spatial decomposition tells us where they are acting. Techniques like Molecular Mechanics/Generalized Born Surface Area (MM/GBSA) allow researchers to calculate the energy contribution of every single amino acid in a protein.
This is particularly powerful for distinguishing between different types of binders. Matthew I. J. Raybould and Sandeep Kumar used per-residue decomposition to profile T-cell receptor-mimetic antibodies. They discovered that while these antibodies look like natural receptors, they often derive their stability from the surrounding structural framework rather than the central peptide ligand, explaining why they sometimes lack the necessary specificity for safe clinical use (Computationally profiling peptide:MHC recognition by T-cell receptors and T-cell receptor-mimetic antibodies).
In protein engineering, this spatial map acts as a guide for rational design. Instead of random "trial-and-error" mutagenesis, Yue Pan used MM/GBSA to identify specific residues that stabilize substrates in the desired orientation, enabling the targeted design of more selective enzymes (Rational Design of P450 aMOx for Improving Anti-Markovnikov Selectivity Based on the “Butterfly” Model).
Enthalpy, Entropy, and Molecular Floppiness
Decomposition also extends to the thermodynamic nature of the bond. Binding energy is a balance between enthalpy (the strength of chemical bonds) and entropy (the degree of molecular disorder).
A drug might lose efficacy not because it can no longer form bonds, but because the mutation makes the binding pocket too "floppy." Xiaotian Kong and Peichen Pan explored this in How Does the L884P Mutation Confer Resistance to Type-II Inhibitors of JAK2 Kinase, showing that a specific mutation increases the conformational freedom of the protein, making it energetically expensive for the inhibitor to stay locked in place. Similarly, Daniel Tapken used energy decomposition to explain why certain receptors have puzzlingly low affinities, identifying energetic penalties associated with the structural "hinge" movements required for binding (The low binding affinity of D-serine at the ionotropic glutamate receptor GluD2 can be attributed to the hinge region).
Interpretable AI through Physics
The push for decomposition is also transforming how we build artificial intelligence for drug discovery. Modern deep learning models, while fast, are often criticized as "black boxes." New frameworks are bridging this gap by forcing neural networks to speak the language of physics.
David Ryan Koes and colleagues introduced CENsible, a network that predicts weights for traditional physics terms rather than just a final affinity score. This allows chemists to visualize exactly which atom-pair interactions are driving the AI's prediction (CENsible: Interpretable Insights into Small-Molecule Binding with Context Explanation Networks). Similarly, the LumiNet framework by Qun Su integrates classical force-field equations directly into subgraph transformers, achieving high-precision absolute binding free energy calculations while remaining interpretable (Robust protein–ligand interaction modeling through integrating physical laws and geometric knowledge for absolute binding free energy calculation).
By breaking binding down into its fundamental parts, these researchers have turned a single, opaque number into a rich, multidimensional map of molecular interaction.