Free energies at QM accuracy from force fields via multimap targeted estimation

Area of Science:

• Computational Chemistry
• Molecular Modeling
• Drug Discovery

Background:

• Accurate prediction of ligand binding affinities is crucial for accelerating drug discovery.
• Current methods using quantum mechanics (QM) free energy calculations are computationally expensive and largely unfeasible for early drug discovery.
• The slow convergence of corrections from force fields (FFs) to QM potentials presents a significant bottleneck.

Purpose of the Study:

• To develop an efficient method for computing QM free energies using inexpensive reference potentials like FFs.
• To overcome the slow convergence issue in QM/FF free energy calculations.
• To accelerate the prediction of ligand binding affinities for drug discovery applications.

Main Methods:

• Generalization of targeted free energy methods using multiple maps implemented with normalizing flow neural networks (NNs).
• NNs are trained efficiently with a one-epoch learning policy to avoid overfitting and maximize distribution overlap.
• Integration with enhanced sampling strategies to address slow degrees of freedom and improve convergence.

Main Results:

• The method accelerates QM free energy calculations by three orders of magnitude compared to standard free energy perturbation.
• Achieved an eight-fold acceleration compared to previous nonequilibrium calculations on the HiPen dataset.
• Demonstrated efficient computation of QM free energies from simulations using FF reference potentials.

Conclusions:

• The developed method significantly reduces the computational cost of QM free energy calculations.
• This approach enables the use of QM accuracy in lead optimization campaigns and protein-ligand binding studies.
• The method offers a feasible pathway for integrating high-accuracy QM potentials into early-stage drug discovery.