Investigating Errors in Alchemical Free Energy Predictions Using Random Forest Models and GaMD

Area of Science:

• Computational chemistry
• Biophysics
• Pharmacology

Background:

• Accurate prediction of binding free energy changes (ΔΔG) is vital for antibody-antigen interactions.
• Current in silico methods (±1 kcal/mol accuracy) are insufficient for assessing post-translational modification (PTM) impacts critical for clinical development.
• PTMs affecting binding by >50% (ΔΔG of +0.5 kcal/mol) require higher accuracy predictions (±0.5 kcal/mol) for risk assessment.

Purpose of the Study:

• To improve the accuracy of in silico ΔΔG predictions for antibody-antigen complexes, specifically for assessing PTM impacts.
• To develop an error analysis approach for molecular dynamics simulations to identify sources of inaccuracy.
• To apply error corrections to enhance the practical utility of these predictions in clinical development.

Main Methods:

• Conventional molecular dynamics thermodynamic integration (cMD-TI) was used for ΔΔG predictions.
• Random forest (RF) models and Gaussian accelerated molecular dynamics (GaMD) were employed for error analysis and correction.
• Analysis focused on identifying inadequate sampling and violations of energetic interactions as error sources.

Main Results:

• The developed error analysis approach provided insights into simulation sampling limitations.
• GaMD-based error corrections improved accuracy by over 1 kcal/mol in challenging cases.
• The root-mean-square error (RMSE) for 13 predictions was reduced from 1.06 ± 0.22 to 0.70 ± 0.18 kcal/mol.

Conclusions:

• Alchemical free energy predictions can be applied to estimate PTM impacts on bioactivity.
• The study identifies key error sources limiting the practical application of these predictions in clinical development.
• Error correction strategies, particularly using GaMD, significantly enhance prediction accuracy for PTM impact assessment.