Limitations

Understanding what Pose Rescorer can and cannot do

Methodological Limitations

1. Single-Frame Approximation

   Rescore performs calculations on a single structure. It does not sample conformational space or account for protein-ligand dynamics. Real binding involves ensemble averaging over many conformations, which is not captured here.

   Impact: Scores can be sensitive to the input structure. Small changes in ligand pose or protein conformation can produce different energies.

2. No Entropy Calculation

   The binding free energy is ΔG = ΔH − TΔS. Rescore computes ΔH (enthalpy) using molecular mechanics but does not compute ΔS (entropy). The reported ΔG_bind implicitly assumes ΔS = 0.

   Impact: True binding free energies include both enthalpic and entropic contributions. Neglecting entropy can lead to large errors, especially for flexible ligands or systems with significant conformational changes upon binding.

3. Implicit Solvent Approximation

   Generalized Born (GB) and Poisson-Boltzmann (PB) models approximate solvent effects using continuum electrostatics. They do not represent explicit water molecules, hydrogen bonding networks, or desolvation effects accurately.

   Impact: Systems where water-mediated interactions are critical may be poorly represented. Explicit solvent MD simulations provide more accurate solvation modeling but are computationally expensive.

4. Force Field Limitations

   AMBER force fields (ff14SB, GAFF2) are parametrized for typical drug-like molecules and standard amino acids. They may not accurately represent:

   • Unusual ligands (e.g., organometallics, boron-containing compounds)
   • Modified amino acids or non-standard residues
   • Covalent inhibitors or transition-state analogs
   • Highly charged species or zwitterions

Structural Requirements

1. Protein Preparation

   • The receptor must be a complete protein structure (no missing heavy atoms)
   • Protonation states are inferred by pdb4amber—manual curation may be needed for histidines, cysteines, or other ionizable residues
   • Waters, ions, and cofactors are removed during preparation
   • Multi-chain complexes are supported but not validated extensively

2. Ligand Preparation

   • The ligand must be provided as a MOL2 file with 3D coordinates
   • Protonation state is taken from the input file—use tools like MolProbity or Epik to assign correct states
   • Ligand must be a single molecule (no multi-component ligands)
   • AM1-BCC charge calculation assumes neutral or singly charged molecules

3. Unsupported Systems

   Rescore cannot handle:

   • Metal cofactors (e.g., zinc fingers, heme groups)
   • Covalent ligands
   • DNA/RNA complexes
   • Membrane proteins (no lipid bilayer support)

Interpretation of Results

1. Relative Ranking Only

   ΔG_bind values are relative scores for ranking purposes. They should not be interpreted as experimental binding affinities (K_d, K_i, IC₅₀).

   Example: A ligand with ΔG = −30 kcal/mol is likely to bind more favorably than one with ΔG = −20 kcal/mol, but the absolute values do not correspond to measured affinities.

2. Lack of Error Bars

   Single-frame calculations provide no statistical uncertainty. There are no error bars or confidence intervals. Scores should be treated as point estimates with unknown precision.

3. Correlation with Experiment

   MM/GBSA scores do not reliably correlate with experimental binding affinities across diverse systems. The method works best for:

   • Comparing structurally similar ligands
   • Ranking poses for the same ligand
   • Filtering large virtual screening libraries

   It works poorly for:

   • Predicting absolute K_d or ΔG_exp
   • Comparing chemically diverse compounds
   • Systems with large conformational changes or allosteric effects

Computational Considerations

1. Performance

   A typical calculation for a ~300 residue protein and a small molecule takes on the order of minutes on modern hardware. Batch rescoring scales linearly with the number of ligands (no parallelization across ligands, but AmberTools parallelizes internal GB calculations).

2. Reproducibility

   Results are deterministic for the same input files. However, AM1-BCC charge calculations may have small numerical variations (±0.001 e) between runs, leading to ~1 kcal/mol differences in binding energy.

3. File Management

   Each ligand generates ~50-100 MB of output files (topologies, coordinates, logs). For large batches, ensure sufficient disk space.

Appropriate Use Cases

Alternative Methods

For more rigorous predictions, consider:

• Free energy perturbation (FEP): Alchemical method with explicit solvent and sampling (e.g., FEP+, OpenFE)
• Thermodynamic integration (TI): Similar to FEP but with different integration approach
• Metadynamics: Enhanced sampling for exploring free energy landscapes
• QM/MM calculations: Quantum mechanics for ligand binding site, MM for rest of protein

Best Practices

1. Always validate docking poses visually before rescoring
2. Use consistent protonation states across all ligands in a series
3. Compare scores within a congeneric series, not across diverse scaffolds
4. Report methods clearly if publishing: state that MM/GBSA is for relative ranking only
5. Do not over-interpret absolute values—focus on rank order

Further Reading

• Genheden & Ryde (2015) "The MM/PBSA and MM/GBSA methods to estimate ligand-binding affinities." Expert Opin. Drug Discov. 10(5):449-461.
• Wang et al. (2019) "End-Point Binding Free Energy Calculation with MM/PBSA and MM/GBSA: Strategies and Applications in Drug Design." Chem. Rev. 119(16):9478-9508.
• Tuccinardi (2021) "What is the current value of MM/PBSA and MM/GBSA methods in drug discovery?" Expert Opin. Drug Discov. 16(11):1233-1237.