New methods for alchemical absolute binding free energy calculations

Abstract

Alchemical absolute binding free energy (ABFE) calculations are of increasing interest in drug discovery. They can predict the binding affinities of structurally dissimilar ligands to their targets and offer higher accuracy than alternative methods. However, their widespread application is still limited by high computational cost, lack of automation, and inaccuracy. This work investigates methods to improve the speed, accuracy, and automation of ABFE calculations.

Receptor-ligand restraint schemes were compared. A new scheme based on multiple distance restraints was proposed which avoids inherent instabilities and may provide convergence benefits. This produced results comparable to the common Boresch scheme, while omitting orientational restraints led to large errors.

A fully automated ABFE workflow was developed, including automated lambda-window selection, the ensemble-based detection of equilibration, and the adaptive allocation of sampling time based on inter-replicate statistics. The workflow produced equivalent results to a nonadaptive scheme over several test systems, while often accelerating equilibration.

White’s marginal standard error rule was reformulated to provide a spectrum of equilibration detection heuristics applicable to single simulations. These were tested on ensembles of synthetic time series modelled on free energy change estimates from long ABFE calculations. Methods that more thoroughly accounted for autocorrelation often showed late and variable truncation times, while methods that less thoroughly accounted for autocorrelation often showed early truncation, relative to the optimal truncation point. A method was identified which achieved robust performance across test sets by balancing these extremes.

The performance of extremely fast ABFE calculations was investigated over a range of test systems and compared to "standard-length" calculations. Short ABFE calculations showed large absolute errors, but often retained similar ranking performance to "standard" calculations.