Membrane pan-assay interference compounds (PAINS) are a class of molecules that interact nonspecifically with lipid bilayers and alter their physicochemical properties. An early identification of these compounds avoids chasing false leads and the needless waste of time and resources in drug discovery campaigns. In this work, we optimized an in silico protocol on the basis of umbrella sampling (US)/molecular dynamics (MD) simulations to discriminate between compounds with different membrane PAINS behavior. We showed that the method is quite sensitive to membrane thickness fluctuations, which was mitigated by changing the US reference position to the phosphate atoms of the closest interacting monolayer. The computational efficiency was improved further by decreasing the number of umbrellas and adjusting their strength and position in our US scheme. The inhomogeneous solubility-diffusion model (ISDM) used to calculate the membrane permeability coefficients confirmed that resveratrol and curcumin have distinct membrane PAINS characteristics and indicated a misclassification of nothofagin in a previous work. Overall, we have presented here a promising in silico protocol that can be adopted as a future reference method to identify membrane PAINS.