Search

It is widely understood that QSAR models greatly improve if more data are used. However, irrespective of model quality, once chemical structures diverge too far from the initial data set, the predictive performance of a model degrades quickly. To increase the applicability domain we need to increase the diversity of the training set. This can be achieved by combining data from diverse sources. Public data can be easily included; however, proprietary data may be more difficult to add due to intellectual property concerns. In this contribution, we will present a method for the collaborative development of linear regression models that addresses this problem. The method differs from other past approaches, because data are only shared in an aggregated form. This prohibits access to individual data points and therefore avoids the disclosure of confidential structural information. The final models are equivalent to models that were built with combined data sets.

The ionization state of drugs influences many pharmaceutical properties such as their solubility, permeability, and biological activity. It is therefore important to understand the structure property relationship for the acid-base dissociation constant pKa during the lead optimization process to make better-informed design decisions. Computational approaches, such as implemented in MoKa, can help with this; however, they often predict with too large error especially for proprietary compounds. In this contribution, we look at how retraining helps to greatly improve prediction error. Using a longitudinal study with data measured over 15 years in a drug discovery environment, we assess the impact of model training on prediction accuracy and look at model degradation over time. Using the MoKa software, we will demonstrate that regular retraining is required to address changes in chemical space leading to model degradation over six to nine months.