Time-series QSAR analysis of human plasma protein binding data

Quantitative Structure-Activity Relationship (QSAR) models are built using a wide variety of methods and descriptors in order to produce the optimal model for a given property or activity. Once built, models tend to be static or very rarely updated, yet chemistry progresses with time. Any QSAR model is limited by the data used to create it, which in turn is dependent on the data available at the time of building. This study investigates the time dependence of the predictive ability of QSAR models. This is achieved by examining the performance of a human plasma protein binding model over time. A number of QSAR models were built using Partial Least-Squares (PLS) and tested using data collected from experiments conducted at specific time points over a two-year period. The ability of these updated models to make predictions was compared with that of a model built with data from the beginning of the time period. In general, the updated models are stable over time and predictions vary only to a small degree from one model to the next. However, for both current and future test data, the updated models are overall more effective at making predictions than the static model. This highlights the benefits of using more current data for making predictions and thus the need for autoupdating QSAR models.