Predicting Activity Cliffs with Free-Energy Perturbation

Abstract

Activity cliffs (ACs) are an important type of structure−activity relationship in medicinal chemistry where small structural changes result in unexpectedly large differences in biological activity. Being able to predict these changes would have a profound impact on lead optimization of drug candidates. Free-energy perturbation is an ideal tool for predicting relative binding energy differences for small structural modifications, but its performance for ACs is unknown. Here, we show that FEP can on average predict ACs to within 1.39 kcal/mol of experiment (∼1 log unit of activity). We performed FEP calculations with two different software methods: Schrodinger−Desmond FEP+ and GROMACS implementations. There was qualitative agreement in the results from the two methods, and quantitatively the error for one data set was identical, 1.43 kcal/mol, but FEP+ performed better in the second, with errors of 1.17 versus 1.90 kcal/mol. The results have far-reaching implications, suggesting well-implemented FEP calculations can have a major impact on computational drug design.

Rapid, accurate, precise and reproducible ligand–protein binding free energy prediction

“…This may or may not be valid, depending on each and every case under study; when invalid, however, its predictions will fail badly. Thus, for example, when the free energy changes more or less discontinuously with molecular structure, as it does for the case of free energy cliffs [94], there is no way such an ML algorithm will in general be able to spot such phenomena. It could only do so if the coverage of the state space were exceptionally dense, implying that the data upon which it has been trained would need to be enormous.…”

Large scale relative protein ligand binding affinities using non-equilibrium alchemy

“…In turn, application of FEP to a vast range of protein-ligand systems revealed that the method can indeed deliver accurate relative binding affinity predictions with an error of <1 kcal mol À1 with respect to experiment. [23][24][25][26][27][28][29][30][31][32][33][34][35][36] However, the application of FEP using most MD soware remains challenging, preventing its widescale uptake.…”

Automated Assessment of Binding Affinity via Alchemical Free Energy Calculations

“…For comparison, it has been reported that FEP+ takes 86 minutes to sample a 1ns perturbation for 12 windows on 4 NVIDIA Tesla K80 for BACE1 69 complexes (containing 401 amino acids) and 34 minutes for the corresponding simulations of the solvated ligands. 70 This corresponds to ~29 min/ns for the complex and ~11 min/ns for the ligand-only simulations, respectively. The Gromacs-FEP implementation was reported to be 3 to 6 times slower than FEP+.…”

“…However, there is evidence that shorter calculations times (usually 1 ns) may be sufficient to obtain a reasonable free energy estimate for certain classes of perturbations. 23,70 Thus, there may be scope to further…”