Learned mappings for targeted free energy perturbation between peptide conformations

Area of Science:

• Computational chemistry
• Statistical mechanics
• Machine learning applications

Background:

• Targeted free energy perturbation methods rely on invertible mappings for accurate free energy estimates.
• Developing effective mappings for complex systems remains a significant challenge in computational chemistry.
• Prior work demonstrated machine learning's potential for mapping between thermodynamic states.

Purpose of the Study:

• To adapt a machine learning approach for calculating free energy differences of a flexible molecule.
• To investigate the efficacy of neural network mappings with harmonic biases and varying spring centers.
• To assess the accuracy and limitations of this method for different thermodynamic state separations.

Main Methods:

• Utilized a deep neural network approach to train mappings between Boltzmann distributions of different thermodynamic states.
• Applied harmonic biases with adjustable spring centers to the deca-alanine molecule.
• Employed "early stopping" criteria during neural network training based on test set loss.
• Calculated free energy differences and compared results with reference methods.

Main Results:

• Accurate free energy differences were obtained for thermodynamic states with spring centers separated by up to 2 Å.
• The neural network mapping successfully promoted configuration space overlap for closely spaced states.
• For more distant states, the mapping failed to generate representative structures, limiting accuracy.

Conclusions:

• Machine learning-based mappings offer a promising avenue for improving free energy calculations in molecular systems.
• "Early stopping" is a crucial technique for preventing overfitting and ensuring reliable results.
• The method's applicability is dependent on the separation between thermodynamic states, with limitations for highly divergent states.