Practical guidelines for optimising free energy calculations using thermodynamic integration

Free energy calculations are valuable in structure-based drug design, but their accuracy and reliability remain challenges. We present an automated workflow for estimation of protein–ligand binding affinity built with AMBER20, alchemlyb, and open-source cycle closure algorithm. Evaluated on 178 perturbations across four datasets, the short sub-nanosecond simulations performed comparably or better than prior studies for the MCL1, BACE, and CDK2 datasets, while the TYK2 dataset required a longer equilibration time ($\sim 2\mathrm{ns}$). Perturbations with $\left|\Delta \Delta G\right|>2.0$ kcal/mol exhibited higher errors, suggesting such perturbations are unreliable, hence providing a practical guideline for improving thermodynamic integration simulations.