Computational alanine scanning with GBSA and interaction entropy: a review of methodologies and applications in protein–protein and protein–ligand binding

Abstract

Computational alanine scanning (CAS) has emerged as a powerful in silico strategy for identifying binding hot spots in protein–protein and protein–ligand complexes, addressing the limitations of experimental alanine scanning mutagenesis. This review provides a comprehensive overview of CAS methodologies, with particular focus on the integration of GBSA (molecular mechanics/generalized born surface area) and the Interaction Entropy (IE) method. Traditional CAS approaches have been hampered by the systematic neglect of entropic contributions or computationally prohibitive entropy calculations via normal-mode analysis. The IE method overcomes this bottleneck by enabling efficient and rigorous entropy calculation directly from a single molecular dynamics trajectory, using only fluctuations of intermolecular interaction energies. We present the theoretical foundations of the ASGBIE (Alanine Scanning with Generalized Born and Interaction Entropy) framework, which combines GBSA with IE to accurately compute residue-specific binding free energies. The methodology is systematically described for both protein–protein and protein–ligand systems, including practical considerations such as residue type-specific dielectric constants and the single-trajectory approximation. We survey diverse applications across biologically and pharmacologically relevant systems, including the p53/MDM2 complex, PD-1/PD-L1 immune checkpoint, SARS-CoV-2 variant escape from monoclonal antibodies, Nipah virus antibody synergy, and drug resistance mechanisms in kinase inhibitors. These case studies demonstrate the practical utility of ASGBIE across diverse protein-binding systems of biological and pharmaceutical relevance.

Computational alanine scanning (CAS) has emerged as a powerful in silico strategy for identifying binding hot spots in protein–protein and protein–ligand complexes, addressing the limitations of experimental alanine scanning mutagenesis.

1. Introduction

Specific, high affinity molecular recognition-whether between proteins, or between a protein and a small-molecule ligand underpins nearly every essential biological process, spanning signal transduction, immune recognition, enzymatic catalysis, transcriptional regulation, and viral entry.^1–5 In modern drug discovery, protein engineering, and computational structural biology, the ability to quantify binding energies and pinpoint functionally critical residues has become indispensable. Beyond simply measuring global binding affinity (ΔG), a mechanistic understanding demands residue-level resolution: identifying which amino acids dominate binding stability, govern specificity, drive resistance mutations, or represent actionable hot spots for therapeutic intervention. Such energetically privileged residues, widely defined as binding hot spots,^6–21 contribute disproportionately to the total binding free energy and serve as the functional “fingerprint” of a molecular interface.

Experimental alanine scanning mutagenesis, which quantifies changes in PPI binding affinity upon systematic substitution of interfacial residues with alanine, is the established gold standard for hot spot validation.²² By systematically mutating target residues to alanine and quantifying the resulting change in binding free energy (ΔΔG), experimentalists can rigorously classify residues as energetically essential (hot spots with ΔΔG > 2 kcal mol⁻¹) or negligible (null spots).⁸ Despite its proven value, experimental alanine scanning has inherent limitations: it is labor-intensive, time-consuming, costly, and challenging to scale to large complexes, multisite mutations, or difficult-to-express systems. Furthermore, it cannot readily dissect energetic contributions or isolate the individual roles of enthalpy and entropy, which are essential for understanding binding mechanisms.⁵

These challenges have driven the rapid development of computational alanine scanning (CAS),^11,23–33 an in silico framework that predicts residue-specific ΔΔG values rapidly, economically, and at full interface coverage.^{16,20,21,24,28,33–37} Computational alanine scanning (CAS) was first formalized by Massova and Kollman in 1999, who demonstrated that in silico alanine substitution could accurately predict hot spot residues in the MDM2-p53 complex with near-perfect agreement to experimental data.³³ A key advantage of the CAS method is that it requires only a single molecular dynamics (MD) simulation trajectory of the wild-type PPI complex, enabling high-throughput, genome-wide scanning of interfacial residues via post-processing calculations, without the need for independent simulations of each alanine mutant. This makes the CAS calculation for PPI extremely attractive. CAS methods relied heavily on molecular mechanics (MM) force fields combined with continuum solvation models, most prominently GBSA (and MM/PBSA),^38–42 to estimate binding free energies from static structures or short molecular dynamics (MD) trajectories.^{22,24,28,33,34,37,43} Although GBSA became widely adopted for its efficiency and favorable balance between speed and accuracy, conventional implementations suffer from a critical flaw: the systematic neglect of entropy contributions. Ignoring it leads to systematic overestimation of binding strength, false-positive hot spot predictions, and poor agreement with experimental mutagenesis data. Traditional entropy estimation via normal-mode analysis (NMA) is theoretically approximate and computationally prohibitive for large biological assemblies, creating a long-standing bottleneck in reliable CAS.^44,45

This bottleneck was decisively overcome by the pioneering development of the Interaction Entropy (IE) method.⁴⁶ IE enables computationally efficient calculation of binding entropy directly from a single conventional MD trajectory, using only the fluctuations of intermolecular interaction energies-no expensive normal mode calculations⁴⁵ or separate mutant simulations are required. By integrating IE into computational alanine scanning, the field has advanced to a robust, high-accuracy new paradigm: computational alanine scanning with GBSA and interaction entropy (ASGBIE).^47–49 ASGBIE refers to (Alanine Scanning with Generalized Born and Interaction Entropy), where GB denotes that the Generalized Born (GB) model is employed to compute the solvation free energy.^50–67 This combined framework rectifies the historical entropy deficit, markedly improves prediction precision, reduces false hot spots, and enables quantitative decomposition of binding into enthalpy and entropy components at per-residue resolution.

In this review, we present a comprehensive, theoretically grounded overview of computational alanine scanning using interaction entropy. We begin by establishing the theoretical foundations of binding free energy calculation, with particular emphasis on GBSA as the mainstream end-point method for solvation energy calculation. Next, we systematically introduce computational alanine scanning in both protein–protein and protein–ligand contexts, clarify the physical origins of conformational entropy in binding, and formally present the interaction entropy theory along with its numerical implementation. We further elaborate the unified ASGBIE pipeline as a complete, practical workflow for accurate residue-specific binding free energy calculation. Finally, we survey broad applications across diverse biological systems-including enzyme inhibition, antibody–antigen recognition, viral immune escape, antibiotic resistance, and drug optimization. Collectively, this review demonstrates that computational alanine scanning with interaction entropy has evolved into a mature, versatile, and high-impact tool for dissecting molecular recognition, guiding rational drug design, and uncovering the thermodynamic logic of biological assembly.

2. Computational alanine scanning for protein systems

2.1. CAS for protein–protein binding

Protein–protein interactions (PPIs) govern the spatiotemporal organization of cellular pathways, and their dysregulation is a hallmark of human diseases including cancer, autoimmune disorders, and infectious diseases.^12,68 PPI interfaces are typically broad, shallow, and dynamically regulated, making them challenging targets for therapeutic intervention. A seminal observation in PPI biology is that binding affinity is not uniformly distributed across the interface: a small subset of “hot spot” residues contributes most of the binding free energy, forming cooperative interaction networks within structurally conserved, complementary interfacial pockets.^13,68 These hot spots represent the functional and pharmacological core of PPIs, making their identification the critical first step in both mechanistic studies and drug discovery programs.

For a protein–protein interaction system involving protein A and B, we assume that alanine scanning is performed on protein A with one of its residue x mutated to alanine a. Then the difference in binding free energy (ΔΔG_bind^x→a) between the alanine mutant and the wild type (WT) protein A to protein B is defined as

ΔΔG_bind^x→a = ΔG_bind^a − ΔG_bind^x 1

where ΔG_bind^a and ΔG_bind^x are the binding free energy between protein A and protein B before and after alanine mutation, respectively. These binding free energies can be calculated by decomposing into the gas-phase and solvation components,

2

and similarly,

3

In the above equations (eqn (1) and (2)), the superscript “x” denotes a specific residue of protein A at the protein–protein interface, and “a” denotes alanine that replaces residue “x” upon mutation. A(x) denotes the WT protein A and A(a) denotes the alanine mutant protein A. ΔG(A(x) − B) and ΔG(A(x) − B) represent, respectively, the binding free energy of the wide type and mutation type complex. Eqn (1) can be further expanded to give

4

In computing the above equation, we can make the following approximation,

5

Here an assumption is made that the difference in inter-protein gas-phase interaction energies between the wild type A(x) − B and the mutant A(a) − B complexes is equivalent to that between residue x and protein B and residue a and protein B. In the above equation, single trajectory approach was used and free energy change is calculated without the rearrangement of the surrounding environment.

Since the solvation part of free energy in eqn (4) is given by

ΔG_sol(A(x) − B) = G_sol(A(x)^B) − G_sol(A(x)) − G_sol(B), 6

where G_sol(A(x)^B) is the solvation energy of the A(x) − B complex, and similarly,

ΔG_sol(A(a) − B) = G_sol(A(a)^B) − G_sol(A(a)) − G_sol(B), 7

we have

8

2.2. GBSA calculation of solvation energy ΔG_sol

In CAS, the solvation free energy in PPI is typically calculated by using the popular MM/PBSA(GBSA) approach, in which the solvation free energy is split into two terms

ΔG_sol = ΔG_pb/gb + ΔG_np 9

where the first term in the above equationΔG_pb/gb is the electrostatic solvation free energy given by the analytical formula,^69,70

10

where ε_in is the dielectric constant of the solute, ε_out, the dielectric constant of the solvent (typically 80 for water at 298 K). The q_i and q_j are partial atomic charges, r_ij is the distance between atoms i and j, R_i and R_j are the effective Born radii of atoms i and j, respectively. The f_GB is a smooth function that interpolates between the Born (monatomic) and coulombic limits. A common form is:

f_GB(r_ij², R_i, R_j) = [r_ij² + R_iR_j exp(−r_ij²/4R_iR_j)]^1/2 11

The second termΔG_np in eqn (9) is the non-polar solvation free energy obtained by using an empirical solvent-accessible surface area (SASA) formula^71,72

ΔG_np = γSASA + β 12

The γ and β are given the standard values of 0.00542 kcal (mol⁻¹ Ų) and 0.92 kcal mol⁻¹, respectively (Fig. 2).

Fig. 2. Calculation of Interaction Entropy.

2.3. Interaction entropy

The “gas-phase” component of the binding free in eqn (5) is given by thermodynamic expression⁴⁶

13

and similarly, for ΔG_gas(a − B). In the above equation, we define the interaction entropy (IE)^46,73–78 as

14

Eqn (13) involves the natural log of an ensemble average of , which can be extracted along with MD simulation without extra computational cost. The interaction entropy of eqn (14) can be computed by numerical integration over simulation time, e.g.,

15

whereΔE^int = E^int − 〈E^int〉 and the average interaction energy is obtained by

16

Alternatively, one could also perform integration over the interaction energy distribution in eqn (14) by the integration⁴⁶

17

whereρis the distribution of the interaction energy. If the energy distribution is Gaussian, namely, then the above integral gives the analytical result

18

which yields the analytical result for IE

19

Here σ is the width of Gaussian distribution. This approximation should be validated against direct numerical integration (eqn (15)) when possible, as non-Gaussian tails are common in protein interactions.

However, extreme caution must be exercised when using the above formula. Unless the sampling is sufficiently converged-which is very challenging for large interaction energies, such as total protein–protein interaction energies-using eqn (17) can lead to substantial overestimation of the entropic term if the energy distribution is not converged. To avoid this issue, a cutoff based on energy fluctuation is often applied: all energy points with fluctuations exceeding a certain threshold are excluded from the calculation of eqn (15).

2.4. CAS for protein-ligand binding

Protein-ligand binding^79,80 is the fundamental basis of numerous biological processes, including enzymatic catalysis, signal transduction, and drug action. The identification of key amino acid residues (hot spots) that dominate the binding affinity and specificity between proteins and ligands is crucial for understanding molecular recognition mechanisms, guiding rational drug design, and interpreting drug resistance. Unlike protein–protein interactions, which involve large interfacial areas with dozens of interacting residues, protein-ligand binding typically occurs at a relatively compact active site, where a small number of residues form specific noncovalent interactions (e.g., hydrogen bonds, hydrophobic interactions, electrostatic interactions) with the ligand. This structural feature endows CAS for protein-ligand binding with unique methodological characteristics and application priorities, which are distinct from those for protein–protein binding.

In computational alanine scanning for protein–ligand binding, a specific residue is mutated to alanine, and the difference in binding free energies before and after the mutation is computed, as schematically illustrated in Fig. 1. It is generally assumed that the mutated alanine does not make a measurable contribution to the free energy of protein–ligand binding. Therefore, the difference in binding free energies directly provides a quantitative measure of that residue's contribution to the overall binding free energy. Under this assumption, eqn (5), originally derived for protein–protein interactions, can be directly adapted for protein–ligand binding,

ΔΔG_gas^x→a ≈ ΔG_gas^a(a − L) − ΔG_gas^x(x − L) 20

where protein B is simply replaced by the ligand L. However, one should keep in mind that, unlike in protein–protein systems, an alanine mutation in a protein–ligand system may induce small conformational changes in the ligand, which can complicate the analysis.

Fig. 1. Binding free energy difference in alanine scanning mutagenesis in (A) protein–protein and (B) protein–ligand bindings.

2.5. Use of residue type-specific dielectric constant

The evaluation of protein solvation energy using an implicit model such as PBSA and GBSA are generally sensitive to the value of the internal dielectric constant assigned to the protein.^42,81,82 Studies have shown that different types of residues prefer different internal dielectric constants in computational alanine scanning.^83–85 To account for this, we classify the 20 standard amino acids into four groups based on side-chain structure: nonpolar (phenylalanine, alanine, leucine, methionine, isoleucine, tryptophan, proline, valine); polar (cysteine, glycine, glutamine, asparagine, serine, tyrosine, threonine); positively charged (histidine, arginine, lysine); and negatively charged (aspartic acid, glutamic acid). As demonstrated in ref. 47. (Yan et al., JCIM), when using a GB model to calculate solvation free energies, the dielectric constants that yield minimal mean absolute errors (MAE) are approximately 3 for polar residues, 1 for hydrophobic residues, and 5 or above for charged residues. For charged residues, using dielectric constants of 5, 6, or higher results in little difference when the GB model is employed. Therefore, when any residue X is mutated to alanine, a corresponding dielectric constant is used in both gas-phase and solvation energy calculations involving that residue.

2.6. ASGBIE for total binding free energy

Once the residue-specific relative binding free energies have been obtained for all contributing residues x using the methods described above, the total binding free energy can be approximated by the following relation⁴⁸,

21

where the summation runs over all contributing (non-alanine) residues. As will be shown later, eqn (21) offers an alternative approach to computing total protein–protein and protein-ligand binding free energies, providing a clearer physical interpretation directly from the computational alanine scanning (CAS) calculation.

The implementation of the ASGBIE method involves four key steps, applicable to both protein–protein and protein–ligand systems with minor adjustments depending on the system characteristics. The standard workflow is as follows:

(1) System preparation: build the wild-type protein–protein or protein-ligand complex, perform energy minimization to remove steric clashes, and run molecular dynamics (MD) simulations to obtain an equilibrium conformational trajectory.

(2) Alanine mutation and trajectory generation: for each target residue at the binding interface-i.e., interfacial residues for protein–protein systems or active-site residues for protein-ligand systems-generate the alanine mutant complex while preserving the backbone conformation of the wild type (or allowing minor adjustments to avoid steric clashes).

To improve computational efficiency, the ASGBIE method can reuse the wild-type MD trajectory to calculate the binding free energy of the mutant, which significantly reduces the computational cost without sacrificing much accuracy.

(3) Binding free energy decomposition: for both wild-type and mutant complexes, calculate the enthalpy contribution using the MMGBSA method, including MM energy (van der Waals and electrostatic interactions) and GB solvation free energy (polar and nonpolar components); calculate the entropy contribution using the IE method based on the MD trajectory.

(4) ΔΔG calculation and hot spot identification: Calculate the binding free energy of the wild-type and mutant complexes, and obtain the ΔΔG value (ΔΔG = ΔG_wt − ΔG_mut) to evaluate the contribution of each residue to binding free energy. Residues with ΔΔG > 2 kcal mol⁻¹ are identified as hot spots, which are critical for maintaining binding affinity.

2.7. Molecular dynamics simulation

MD simulations are performed starting from the complex structures of protein–protein or protein–ligand systems. Notably, only a single MD trajectory of the complex is required to generate alanine scanning results for all residues of interest, making the CAS approach highly efficient for hot spots identification. Typically, about 1000 configurations (snapshots) are extracted from the MD trajectory for computational alanine scanning calculations. Solvation energies are computed using the GB model, and the interaction entropy is calculated by time-averaging over all selected configurations according to eqn (13). The single-trajectory approach is employed to compute binding free energy differences, where error cancellation helps mitigate insufficient conformational sampling.^22,33,37 Mutant trajectories are obtained by a single truncation of the amino acid side chain: the C_γ atom is replaced with a hydrogen atom, and the C_β–H bond direction is set to match the original C_β–C_γ orientation from the wild-type trajectory.The single-trajectory approximation employed in this study assumes that the structural perturbations induced by mutations can be adequately captured within a single molecular dynamics (MD) simulation. This approximation remains valid when mutations cause only localized or moderate conformational changes that do not substantially alter the global architecture of the protein complex. However, this method becomes unreliable when target mutations induce substantial global conformational rearrangements, where distinct structural basins cannot be adequately sampled within a single MD trajectory. In such cases, the free energy landscape may feature multiple minima separated by high barriers, and a single trajectory risks becoming trapped in a local basin, leading to incomplete sampling and biased estimates. For highly flexible systems or those anticipated to undergo large-scale domain motions upon mutation, practitioners are advised to cross-check results with independent multi-trajectory simulations to eliminate sampling bias and ensure robustness of the predicted thermodynamic quantities.

To ensure the reliability of thermodynamic quantities derived from molecular dynamics simulations, rigorous assessment of sampling convergence is essential. In this work, convergence was evaluated using the following criteria: the root-mean-square deviation (RMSD) of the protein backbone relative to the initial structure was monitored throughout each independent simulation, and convergence was deemed satisfactory when the RMSD profile reached a stable plateau without systematic drift over the production phase. Additionally, for each mutant system, multiple independent MD trajectories were generated with different initial velocity distributions, and the calculated ΔΔG values were averaged across these replicas to improve statistical reliability and reduce random noise. The standard deviation across independent runs was used to estimate the uncertainty of the reported ΔΔG values. For systems exhibiting high flexibility or slow conformational transitions-such as protein–protein complexes with extensive interfacial dynamics-extended simulations beyond 100 ns per trajectory are strongly recommended to achieve sufficient sampling of the relevant conformational ensemble for interaction entropy calculations.

3. Applications of ASGBIE

3.1. Applications to protein–protein systems

The accurate and efficient computation of protein-ligand and protein–protein binding free energies constitutes one of the grand challenges in computational biology and biophysics. Protein–protein interactions involve large interfacial areas, with dozens of residues participating in binding, and the binding free energy is the result of the synergistic effect of multiple residues. The ASGBIE method can accurately quantify the contribution of each interfacial residue to binding, effectively identifying hot spot residues, and interpreting the synergistic mechanism between residues.

3.1.1. P53/MDM2 system

This study applies the ASGBIE method to identify hot spots (ΔΔG > 2 kcal mol⁻¹) and warm spots (ΔΔG ∼1–2 kcal mol⁻¹) in four protein–protein interaction (PPI) systems: p53-MDM2, p53-MDMX, PMI-MDM2, and PMI-MDMX. The work aims to quantitatively understand why the peptide PMI binds MDM2 and MDMX ∼100-fold more strongly than the native p53 peptide, despite sharing a similar binding mode.⁸⁶

Traditional GBSA often neglects entropy, leading to overestimated binding free energies. The IE method efficiently calculates entropic contributions from interaction energy fluctuations along MD trajectories. Combining IE with GBSA yields more accurate ΔΔG predictions at low computational cost, avoiding normal-mode analysis. Computational alanine scanning of p53 and PMI residues was compared with published experimental data. GBSAIE achieved a correlation coefficient of 0.89 with experiment, significantly better than GBSA alone (0.67). Error metrics (MAE: 0.90 kcal mol⁻¹) and statistical scores (F1 = 0.95) confirm high accuracy. The method produced only one false positive (p53 L22A in MDMX), whereas GBSA gave four false positives (Fig. 3).

Fig. 3. (A) Crystal structure of p53-MDM2 (PDB ID: 1YCR), p53-MDMX (PDB ID: 3DAB), PMI-MDM2 (PDB ID: 3EQS), and PMI-MDMX (PDB ID: 3EQY) interactions. MDM2 and MDMX are shown in surface and p53 and PMI are shown in cartoon and colored in cyan. Hot spots (ΔΔG_exp ≥ 2.0 kcal mol⁻¹) and warm spots (2.0 kcal mol⁻¹ ≥ ΔΔG_exp ≥ 1.0 kcal mol⁻¹) determined by experiments are shown in red and yellow sticks, respectively. (B) Interaction details of hot spots and warm spots in the four PPIs: p53-MDM2 (PDB ID: 1YCR), p53-MDMX (PDB ID: 3DAB), PMI-MDM2 (PDB ID: 3EQS), and PMI-MDMX (PDB ID: 3EQY) interactions. p53/PMI and MDM2/MDMX are shown in cartoon and colored in green and blue respectively. Hot spots and warm spots are shown in sticks while the CA atoms of the hot spots and warm spots of p53/PMI are colored in green and the CA atoms of the hot spots and warm spots of MDM2/MDMX are colored in yellow. The hot spots and warm spots of MDM2/MDMX are labeled in red and black font respectively. The hydrogen bonds that conserved in the four interactions are shown in dashed lines with their distances labeled in blue font. (C) Average error estimation of the GBSA and ASGBIE results. Blue bars represent the GBSA (without entropy contribution) results and red bars represent the ASGBIE results. (D) Comparison of alanine scanning of PMI in PMI/MDM2 and PMI/MDMX by ASGBIE calculations and experiments. Blue bars represent the ASGBIE calculation results and red bars represent the experimental results. Reproduced from ref. 86 with permission from [John Wiley and Sons], copyright [2019].

Via residue-by-residue energetic decomposition, ASGBIE uncovers why the PMI peptide binds tighter than wild-type p53. PMI forms stronger van der Waals contacts at hot spot residues, and its preorganized side-chain conformation minimizes entropic penalty upon binding. Such enthalpy-entropy decomposition cannot be obtained via experimental alanine scanning.

3.1.2. PD-1/PD-L1 system

This study applies the ASGBIE method to quantitatively identify hot spots (ΔΔG > 2 kcal mol⁻¹) in the PD-1/PD-L1 immune checkpoint complex. Understanding the binding mechanism of this interaction is crucial for developing cancer immunotherapies, as tumors often overexpress PD-L1 to evade T-cell responses.⁸⁷ We performed molecular dynamics simulations on the PD-1/PD-L1 crystal structure (PDB: 4ZQK) for 20 ns, followed by computational alanine scanning using the single-trajectory approach with the ASGBIE method, which incorporates entropic contributions via interaction entropy. They validated predictions against three knowledge-based methods (Robetta, KFC, HotRegion).

The ASGBIE predictions largely agreed with Robetta, KFC, and HotRegion, though discrepancies arose for PD-1 E84 (only ASGBIE predicted hot spots)-likely because MD captures induced conformational changes (e.g., shortened hydrogen-bond distances) not visible in static crystal structures. The authors propose that combining MD-based and structure-based methods increases confidence in hot spot identification. The PD-1/PD-L1 interaction is stabilized by a hydrophobic core (primarily I134, I126, L128, M115, Y123, Y56) and a surrounding hydrogen-bond network involving charged and polar hot spots (Q75, R125, E84, K78, etc.). These hot spots represent promising targets for designing small-molecule or peptide inhibitors that disrupt the PD-1/PD-L1 axis for cancer immunotherapy. The study also highlights that MD-based methods can reveal dynamic features (e.g., induced side-chain rearrangements) that static structure-based methods may miss (Fig. 4).

Fig. 4. (A) Close-up view of the hot and warm spots in the PD-1/PD-L1 interaction that predicted by the ASGBIE method. The hot spots of PD-1(blue cartoon) and PD-L1(green cartoon) are shown in sticks with carbon atoms colored in green and orange, respectively. (B) Surface representation of the hydrophobic core that formed by the hot and warm spots in the PD-1/PD-L1 interaction. The hot and warm spots that compose the hydrophobic core are colored surface and labeled in green font. (C) Hot and warm spots in the PD-1/PD-L1 interaction predicted by the GBSAIE method. (D) RMSF of the Ca atoms calculated from the last 20 ns MD trajectory. The hot and warm spots predicted by the ASGBIE method are labeled. Reproduced from ref. 87 with permission from [Royal Society of Chemistry], copyright [2019].

ASGBIE identifies functionally critical hot spot residues that only emerge after conformational rearrangements captured in molecular dynamics (MD) simulations. Static crystal structure-based computational methods fail to pinpoint these dynamic hot spots, directly demonstrating the indispensable role of protein conformational dynamics in regulating immune checkpoint binding affinity.

3.1.3. SARS-CoV-2 variants systems

As another important application to protein–protein systems, ASGBIE has recently been extended to investigate viral immune escape, particularly for SARS-CoV-2 variants against therapeutic monoclonal antibodies. A representative study by Xiong and colleagues applied MD simulations and the ASGBIE method to explore the binding of Delta and Omicron variants to two monoclonal antibodies that received emergency use authorization, Bamlanivimab and Etesevimab (Fig. 5(A–C)).⁸⁸ GBSA combined with the IE method revealed that the Omicron variant significantly weakened binding to both antibodies, whereas the Delta variant only reduced the binding free energy to Bamlanivimab and retained binding to Etesevimab comparable to the wild type (Fig. 5(B)). Energy decomposition traced these binding free energy losses to the introduction of multiple positively charged residues, such as L452R and T478K in Delta, and Q493R, Q498R, and Y505H in Omicron, which severely impaired electrostatic interactions. Omicron additionally suffered from conformational deflections that reduced vdW contacts and stable hydrogen bonds. To pinpoint the residue-level origins of these changes, the authors employed ASGBIE, which calculates the alanine-scanning free energy change for each residue without empirical entropy approximations by incorporating the interaction entropy method (Fig. 5(D)). ASGBIE identified hot spot residues whose mutation to alanine would significantly alter binding. In the Bamlanivimab–Delta complex, hot spot residues included the mutated T478K as well as Q493 and S494, all of which lie near an extended region of the antibody's heavy chain. In the Bamlanivimab–Omicron complex, more hot spot residues emerged, including T478K, E484A, Q493R, Q498R, and non-mutated residues V483 and F490, all contributing through weakened electrostatics or vdW interactions. For Etesevimab–Delta, no major hot spot residues were found, confirming that the binding mode remained similar to wild type, which explains why this antibody retains efficacy against Delta. In contrast, the Etesevimab–Omicron complex showed hot spot residues at K417N, T478K, and nearby residues E406, R408, Y421, and N460, indicating that large conformational changes in that region destabilized the interaction. Overall, this work demonstrates that ASGBIE, by rigorously including entropic contributions, can accurately identify key residues driving mutation-induced binding loss in viral protein–antibody complexes, providing timely molecular insights for developing next generation therapeutics against emerging variants.

Fig. 5. (A) The mutation residues of the SARS-CoV-2 Delta and Omicron variants in the RBD are green and orange, respectively, and the binding sites of mAbs to the RBD are marked with dashed ellipses; (B) binding free energy calculated by the GBSA and IE methods; (C) stable hydrogen bonds of each complex; (D) binding free energy contribution of Bamlanivimab–RBD and Etesevimab–RBD complexes obtained by the ASGBIE method. Reproduced from ref. 88 with permission from [American Chemical Society], copyright [2022].

ASGBIE quantitatively traces the molecular origin of bamlanivimab evasion in Omicron variants. Immune escape arises not merely from the mutated residues themselves, but from cascading electrostatic perturbations and global conformational shifts that propagate to neighboring unmutated hot spots, collectively weakening antibody–antigen interactions. This multi-layered allosteric mechanism cannot be deduced solely from bulk experimental binding affinity measurements.

3.1.4. Nipah virus systems

The ASGBIE method has also been applied to investigate antibody escape and synergy in Nipah virus, another emerging pathogen with pandemic potential. For the Nipah virus system, Liu et al. used the ASGBIE method together with MD simulations to study three monoclonal antibodies (Fig. 6), m102.4, 1 × 10⁵, and n425, against the V507I mutant of the attachement Glycoprotein (G protein).⁸⁹ The V507I mutation in the G protein is known to compromise m102.4 neutralization, but the molecular basis and potential synergy among antibodies remained unclear. Using ASGBIE, the authors quantitatively computed binding free energies and identified hot spot residues with large contribution changes (Fig. 6(C)). The results showed that the V507I mutation significantly reduced the binding free energy of m102.4, consistent with experimental immune escape. In contrast, 1 × 10⁵ and n425 retained near wild type binding free energies, indicating continued efficacy against the mutant. Energy decomposition traced these losses primarily to weakened electrostatic interactions and disrupted hydrogen bond networks. The loss for m102.4 arose from disrupted electrostatic interactions and a destabilization of hydrogen bonds in a critical loop region of the antibody, causing key residues such as T281 to move away from the binding interface (Fig. 6(B)). For 1 × 10⁵, although the mutation also reduced electrostatic interactions, compensatory enhancement of vdW energy from residues maintained overall binding free energy. Importantly, the critical loop region of 1 × 10⁵ retained several high occupancy hydrogen bonds via residues G107 and S108, preserving stable binding. Beyond individual antibody evaluation, the study uncovered a strong synergistic effect between n425 and m102.4. The total binding free energy of the ternary complex n425-m102.4-wild-type G protein is much stronger than the sum of individual antibodies, confirming a significant synergistic effect (Fig. 6(D–E)). Energy analysis showed that m102.4 binding expanded the n425 binding interface, generating additional hot spot residues such as K560, E579, and I588. Reciprocally, n425 binding stabilized the critical loop region of m102.4, reducing entropy loss of key residues and increasing the number of high occupancy hydrogen bonds. Overall, this work demonstrates that ASGBIE can accurately map mutation induced binding changes in viral glycoprotein antibody interactions and quantify antibody synergy, providing theoretical guidance for developing combination therapies against Nipah virus.

Fig. 6. (A) Binding mode of the G protein (light pink) with n425 (green) and m102.4 (orange) and binding structure of the G protein with three mAbs and the V507I mutation site; (B) key regions of the m102.4 and 1 × 10⁵, along with key residues in their vicinity; (C) scatter plots showing the energy contribution of each hot-spot residue, as determined by the ASGBIE method; (D) heat map of the binding free energy of key residues in n425/m102.4-WT and n425-m102.4-WT systems; (E) the energy contributions of each residue near the binding interface. Reproduced from ref. 89 with permission from [American Chemical Society], copyright [2026].

ASGBIE is the first tool to elucidate the synergistic binding mechanism between antibodies n425 and m102.4. Binding of m102.4 expands the accessible binding interface of n425 and generates novel energetic hot spots; reciprocally, n425 stabilizes the flexible critical loop of m102.4 and reduces unfavorable entropic loss. This mechanistic finding provides direct theoretical guidance for the rational design of combinatorial antibody therapeutics.

3.2. Applications to protein-ligand systems

Protein-ligand binding^6,23,90–93 occurs at a compact active site, and the binding affinity is mainly determined by the specific interactions between a small number of active site residues and the ligand. The ASGBIE method focuses on the local environment of the active site, accurately quantifying the contribution of each active site residue to ligand binding, and providing important guidance for rational drug design. For protein-ligand approximation, the mutation of a large residue to a smaller alanine could likely induces a small conformational change of the ligand, if the ligand is loosely bind to the protein. This effect can cause additional error in the computational alanine scanning calculation. On the other hand, if the ligand is bind quite tightly to the protein, this conformational change of the ligand should be small.

3.2.1. Hot spot predictions and binding free energy calculations

He et al. applied ASGBIE method to investigate the binding mechanism between mitochondrial serine hydroxymethyl transferase2 (SHMT2) and 28 structurally similar inhibitors, including representative compound 1(Fig. 7(A)) (Fig. 7(D)); they quantified residue-specific binding free energies, identified critical hot-spot residues such as Tyr105′, Leu166, Arg425, and Leu172 that contribute dominantly through van der Waals interactions, and obtained calculated binding affinities with strong correlation (R = 0.8) and low errors relative to experimental values, outperforming the conventional GBSA approach and providing a quantitative basis for rational design of high-potency SHMT2 inhibitors⁹⁴(Fig. 7(C)).To characterize the interaction between a pocket residue and different ligands and identify the key residues that determine the binding strength, they classified the 28 complexes into five different groups according to their magnitude of binding affinities: groups 1–5 consist of compounds 23–28 (IC50 ≥ 10 000 nM), 17–22 (100 nM ≤ IC50 < 1000 nM), 11–16 (10 nM ≤ IC50 < 100 nM), 5–10 (1.0 nM ≤ IC50 < 10 nM), and 1–4 (IC50 < 1.0 nM), respectively (Table S1). The binding free energy of hot-spot residues Tyr105′, Leu166, Leu172, Tyr176, His171, and Ile183 with five different groups increases with the increase of binding affinity from groups 1–5, suggesting that these residues are the major determinants of the binding strength. The variation of the energy with different groups mostly comes from different substituents (Fig. 7(B)).

Fig. 7. (A) Cocrystal structure of SHMT2 and compound 16(pdbid 5V7I). The receptor SHMT2 consists of two chains, A (orange) and B (green). The red, yellow, and blue lines represent the pi–pi, pi–alkyl, and van der Waals interactions⁹⁴. (B) Average residue specific contributions to the total binding free energy in the five groups of inhibitors with different magnitudes of binding affinities. (C) Comparison of experimental and calculated total binding free energies of 28 inhibitors from the ASGBIE and the standard GBSA with the normal mode method. (D) Residue-specific interactions between pocket residues and compound 1 (pdbid 5V7I). Average contributions of the pocket residues in SHMT2 to compound1. Only residues with ΔΔG > 0.6 kcal mol⁻¹ are plotted. Interaction diagram of pocket residues with compound1.⁹⁴ Reproduced from ref. 94 with permission from [American Chemical Society], copyright [2019].

Shao et al. employed the ASGBIE method to systematically investigate the residue-specific binding free energy contributions of the androgen receptor (AR) ligand-binding domain with 15 structurally diverse ligands.The study quantitatively identified key hydrophobic hot spot residues including Leu704, Leu873, Met745, and Phe764, which stabilize AR ligand complexes mainly through van der Waals interactions. By summing residue-level contributions, ASGBIE reproduced experimental binding affinities with a high correlation coefficient (R = 0.85) and much lower errors than conventional GBSA. The method further revealed distinct binding patterns across four ligand classes and linked residue contributions to ligand binding modes and structural features. Overall, this work demonstrates that ASGBIE serves as a reliable computational tool to quantify energetic contributions, identify hot spots, and elucidate the molecular binding mechanism of the androgen receptor.⁹⁵

3.2.2. Prediction of potential drug resistance mutations

Liu et al. first extended ASGBIE method for protein–ligand interactions. ASGBIE can accurately identify key hot spot residues at the binding interface of the anaplastic lymphoma kinase (ALK) protein–ligand complex.⁴⁸ Through computational analysis of two generations of inhibitors, residues Leu1256 and Leu1122 are identified to be the two most important residues that contribute to the binding free energy. Furthermore, by quantitatively calculating the contribution of each residue to the binding free energy, the method reveals the molecular mechanism of drug resistance induced by residue mutations at the atomic level. In the ALK crizotinib (first generation inhibitor) system, ASGBIE calculations clearly identify Leu1122 as the core hot spot residue, which contributes up to 2.46 kcal mol⁻¹ to the binding free energy(Fig. 8(A)). The resistance mutation L1196M does not directly weaken interactions at the mutation site itself. Instead, it exerts a steric effect that increases the distance between the adjacent hot spot residue Leu1122 and the R1 ring of the ligand from 4.3 Å in the wild type protein to 4.7 Å in the L1196M mutant(Fig. 8(B)), significantly reducing their intermolecular interaction strength and ultimately leading to decreased binding free energy and drug resistance.These results demonstrate that the ASGBIE method can not only efficiently locate protein–ligand binding hot spots but also accurately elucidate mutation induced drug resistance mechanisms through quantitative residue contribution and conformational change analysis, providing critical theoretical support for the rational design of anti-resistance drugs.⁴⁸

Fig. 8. (A) Main contributing residues from ASGBIE calculation for crizotinib with WT ALK and its L1196M mutant. (B) Co-crystal structures of the first generation inhibitor crizotinib with (left) WT ALK (pdbid 2xp2) and (right) L1196M mutant (pdbid 2yfx).⁴⁸ Reproduced from ref. 48 with permission from [American Chemical Society], copyright [2018].

Residue-resolved energetic decomposition by ASGBIE clarifies the L1196M resistance mechanism. This mutation does not directly disrupt intermolecular interactions at the mutated site; instead, it introduces steric hindrance that widens the distance between adjacent hot spot residues and the inhibitor from 4.3 Å to 4.7 Å, eroding core favorable contacts. This mechanism explains why the L1196M mutation confers drug resistance, eliminating the need for experimental screening of every possible mutation.

Liu et al. employed the alanine scanning with generalized Born and interaction entropy (ASGBIE) approach to provide an all-atom understanding of the inhibition mechanism of DENV-3 methyltransferase (MTase), and further revealed why inhibitor 36A exhibits stronger binding and more potent inhibition than SAH.Through quantitative residue-specific binding free energy calculations, Liu et al. identified Ile147, Arg163, and Met167 as the key interfacial residues dominating the binding affinity.Compared with SAH, 36A forms significantly stronger π-aromatic interactions with Ile147 and more favorable electrostatic contacts with Arg163, leading to a clear improvement in binding enthalpy. The experimental binding free energies further validate this trend: the binding affinity of 36A reaches −9.2 kcal mol⁻¹, stronger than that of SAH at −8.5 kcal mol⁻¹. Notably, Liu et al. also discovered that the enhanced inhibition of 36A is mainly driven by enthalpy contributions, while the entropic penalty remains negligible. Moreover, their work highlights that the ASGBIE method can accurately distinguish binding differences among analogues by quantifying hotspot contributions, rather than relying on global protein contributions (Fig. 9(A)). They used ASGBIE calculations to identify the hot spot residues of DENV3-SAH and DENV3-36A, with specific energy contributions quantified for each residue, including van der Waals, electrostatic interaction, and solvation free energy components et al (Fig. 9(B)). This study by Liu et al. demonstrates that ASGBIE serves as a powerful tool to uncover non selective or selective inhibition mechanisms at the atomic level and provides critical structural and energetic insights for the rational design of more potent anti-dengue virus inhibitors.⁹⁶

Fig. 9. (A) Component-specific contributions of the protein–ligand binding affinities for the WT and F133A mutant. Structural comparison between the mutants and the comparison between the computed values with and without simulation relaxation. (B) Component-specific contributions for interfacial residues scanned in ASGBIE calculations.⁹⁶ Reproduced from ref. 96 with permission from [American Chemical Society], copyright [2024].

ASGBIE can also predict potential drug-resistant mutations.^96,97 With solid numerical statistics supporting its accuracy, ASGBIE can perform mutational scans under a relaxed regime (mutating target sequences, conducting simulation relaxation, and postprocessing snapshots). Taking the DENV-3 MTase protein–igand system as an example, ASGBIE was used to scan three critical residues (110th, 147th, and 163rd) related to MTase inhibitor interactions, mutating each original residue to all other naturally occurring residues and performing 100 ns simulations for each mutant. Through analyzing affinity differences, it successfully identified mutations (e.g., H110G, I147L, R163D) that could escape inhibitor inhibition (affinity drops larger than 1 kcal mol⁻¹), providing critical hints for potentially threatening drug-resistant mutations and revealing that critical sites in protein–ligand coordination may not strongly correlate with drug resistance⁹⁶ (Fig. 10). The ASGBIE method has also been applied to elucidate the protein–drug interaction mechanisms of various targets, including JAK2,⁹⁸ CDK2,⁵⁵ ZIKV,⁹⁹ PARP1,¹⁰⁰ and SHMT2,¹⁰¹ HIV1/HIV2 (ref. 63) et al.

Fig. 10. (A) Binding affinities of MTase-SAH and MTase-36A complexes in saturation mutagenesis computed with ASGBIE and the (B) mutation induced drug-resistance thermodynamics estimated by the 36A-SAH affinity difference of mutants offset by the WT reference,(ΔΔG^{mutant,36A−SAH}_bind − ΔΔG^{wild,36A−SAH}_bind).⁹⁶ Reproduced from ref. 96 with permission from [American Chemical Society], copyright [2024].

4. Advantages and limitations of various binding free energy calculation methods

A comprehensive comparison of widely adopted binding free energy calculation approaches is summarized in Table 1, covering computational overhead, entropy calculation schemes, core strengths and inherent drawbacks. Conventional high-precision methods such as FEP and TI rely on exhaustive conformational sampling and deliver the most rigorous thermodynamic predictions with few simplifying approximations, yet their extremely high computational cost restricts their application to only a small number of candidate molecules rather than large-scale screening workflows. The MM/GBSA framework is computationally cheap and widely accessible; however, the standard version completely ignores entropic effects and suffers systematic over prediction of binding strength, generating numerous false hot spot residues. When combined with normal mode analysis (NMA) to recover vibrational entropy, MM/GBSA substantially elevates computing expenses and becomes impractical for large protein complexes.^39,102–104 Quasi-harmonic analysis achieves moderate cost and captures conformational entropy via covariance matrices, but it exhibits very slow convergence for biomolecules undergoing large conformational transitions.¹⁰⁵ Knowledge-based and machine learning models achieve ultra-fast inference suitable for high-throughput screening, yet their predictions lack fundamental physical interpretability due to empirical fitting.^106–109

Table 1. Advantages and limitations of various binding free energy calculation methods.

Method	Computational cost	Entropy treatment	Core advantages	Major limitations
ASGBIE	Low	Calculated via interaction entropy (IE) from energy fluctuations	High efficiency; residue-resolved decomposition; entropic contribution included	Reliant on single-trajectory approximation; entropy calculation is sensitive to energy cut off thresholds
MM/GBSA (no entropy)	Low	Entropic terms omitted entirely	Fast computation speed	Systematically overestimates binding affinity; prone to false positive hot spots
MM/GBSA + NMA	High	Normal mode analysis (NMA)	Estimates entropy contributions	Computationally intractable for large biomolecular systems
FEP/TI	Extremely high	Full conformational sampling	Theoretically rigorous	Very high computational cost; unsuitable for routine high-throughput screening
Quasi-harmonic analysis	Moderate	Entropy derived from covariance matrices	Quantify conformational entropy	Slow convergence
Knowledge-based/machine learning	Low	Empirical fitting	Ultra-fast; compatible with high-throughput virtual screening	Lacks clear physical interpretability of prediction outputs

By contrast, the ASGBIE method proposed in this work balances calculation efficiency and thermodynamic completeness. It only requires a single molecular dynamics trajectory and quantifies entropic contributions via the interaction entropy (IE) formalism derived from energy fluctuations, enabling residue-by-residue energetic decomposition to pinpoint binding hot spots. Its primary caveats stem from the single-trajectory no-relaxation approximation tailored for alanine scanning, alongside performance fluctuations caused by statistical cutoff filtering of abnormal energy values. Overall, ASGBIE provides a well-balanced solution for residue-level mechanistic analysis and medium-throughput mutational scanning of protein–ligand and protein–protein complexes.

5. Challenges and future perspectives

Despite the transformative advances in CAS for PPI systems, several critical challenges remain to be addressed. First, the fundamental trade-off between accuracy and throughput has not been fully resolved: rigorous free energy methods offer high precision but are incompatible with high-throughput screening, while rapid ML and end-state methods have reduced accuracy for PPI systems with large conformational rearrangements upon alanine substitution. Second, current CAS methods have limited ability to model physiological regulators of PPI binding, including interfacial water dynamics, post-translational modifications, and the conformational flexibility of intrinsically disordered regions, which are prevalent in signaling PPIs. Third, most CAS frameworks are optimized for tight, stable PPIs, with poor performance for the weak, transient PPIs that mediate most cellular signal transduction events. Fourth, a key simplification in most current computational CAS approaches is the neglect of entropic contributions to binding free energy changes; entropic effects upon alanine mutation, are often either excluded entirely or roughly approximated as uniform background, rather than being calculated explicitly on a per-residue basis. This omission introduces potential biases in the estimation of ΔΔG_bind and the identification of hotspot residues, as entropy can significantly modulate the energetic contribution of individual interfacial residues.

Among mainstream single-trajectory CAS algorithms, ASGBIE is widely adopted for its low computational overhead. Nevertheless, this method carries inherent physical defects that directly undermine its prediction reliability for flexible biomolecular systems, especially those with pronounced non-harmonic conformational motions. Three core physical limitations of ASGBIE are elaborated below: (1) conformational sampling and single-trajectory approximation: ASGBIE relies exclusively on a single trajectory molecular dynamics simulation generated from the wild-type protein–ligand complex. The single-trajectory approximations could introduce substantial errors if the alanine mutation triggers significant conformational change. (2) Non-Gaussian distribution of interaction energy fluctuation: for systems characterized by highly anharmonic energy landscapes, multiple metastable conformational states or slow conformational exchange kinetics, the distribution of pairwise interaction energies deviates markedly from a Gaussian profile. Such deviation yields unreliable entropy estimations unless extremely lengthy MD simulations are performed to gather sufficient sampling data. (3) Intrinsic limitations of the generalized born (gb) solvation model: the GB implicit solvent model serves as a computationally cheaper approximation to the more rigorous Poisson–Boltzmann (PB) model. For binding interfaces with dense charged groups or buried charged residues, the GB framework tends to under- or overestimate solvation free energies, thereby compromising the quantitative accuracy of predicted ΔΔG_bind values.

Looking forward, three key directions will define the next generation of CAS technology for PPI systems. First, deep integration with artificial intelligence will address many current limitations: the combination of CAS with AlphaFold and other structure prediction models will enable hot spot prediction for PPIs without experimental crystal structures, while physics-informed deep learning models will deliver end-to-end ΔΔG_bind prediction with the accuracy of rigorous free energy methods and the speed of knowledge-based approaches. Second, the integration of multiscale modeling-combining all-atom MD, coarse-grained simulations, and quantum chemical calculations-will enable more accurate modeling of dynamic PPI systems and the electronic effects that drive interfacial binding. Third, the development of fully integrated “CAS prediction-experimental validation” closed-loop platforms will accelerate the translation of in silico hot spot predictions into functional PPI modulators, bridging the gap between computational prediction and biological validation. Fourth, it is necessary to develop targeted computational strategies to explicitly calculate entropic changes upon alanine mutation on a per-residue basis, replacing the current rough approximation or complete exclusion of entropic contributions. This will reduce biases in ΔΔG_bind estimation and hot spot residue identification, providing a more comprehensive and accurate description of PPI binding energetics.

To further demonstrate the practical compatibility between AlphaFold-derived structures, alanine mutation analysis and machine learning pipelines, representative recent research works are summarized as supporting evidence for the above prospects:

First, AlphaFold2 can be combined with randomized alanine sequence scanning to characterize mutation-triggered allosteric conformational rearrangements of ABL kinase mutants, enabling rapid analysis of mutational conformational landscapes without high-resolution crystal templates.¹¹⁰ Second, the PFMI3DSC pipeline leverages AlphaFold -predicted 3D protein architectures to screen functional mutational hot spots via cross-family structural alignment.¹¹¹ Benchmarked on 288 protein systems, PFMI3DSC identifies more authentic hot spot residues than traditional algorithms including MutClustSW and Mutation Aligner, with most predicted hot spots localized to core functional domains and intermolecular binding interfaces. Third, the protein language model ProBASS, built upon ESM2 and ESM-IF1 encoders and fine-tuned on datasets with experimental ΔΔG measurements, delivers quantitative predictions of mutation-induced binding affinity shifts.¹¹² Fourth, graph deep learning framework DDMut-PPI adopts graph convolutional architectures, where residue embeddings extracted from ProtT5 serve as node features and various noncovalent molecular interactions are defined as edge attributes, to quantify mutational effects on protein–protein binding.¹¹³ In addition, multiple benchmarking works evaluate the thermodynamic applicability of AlphaFold3 complexes. One study utilizes the SKEMPI 2.0 database to test the reliability of AlphaFold3 in predicting mutation-caused binding free energy alterations.¹¹⁴

Author contributions

Xiao Liu: writing – original draft, visualisation. Xiaole Xu: writing – original draft, visualisation. Lili Duan: conceptualisation, project administration, writing – review and editing. John Z. H. Zhang: conceptualisation, project administration, writing – review and editing.

Conflicts of interest

The authors declare no competing financial interest.

Data availability

The authors confirm that the data supporting the findings of this study are available within the article and its referenced publications.