Benchmarks for interpretation of QSAR models

Abstract Interpretation of QSAR models is useful to understand the complex nature biological or physicochemical processes, guide structural optimization perform knowledge-based validation models. Highly predictive are usually and their interpretation non-trivial. This particularly true for modern neural networks. Various approaches these exist. However, it difficult evaluate compare performance applicability ever-emerging methods. Herein, we developed several benchmark data sets with end-points determined by pre-defined patterns. These purposed evaluation ability retrieve They represent tasks different complexity levels: from simple atom-based additive properties pharmacophore hypothesis. We proposed quantitative metrics performance. Applicability benchmarks was demonstrated on a set conventional end-to-end graph convolutional networks, interpreted previously suggested universal ML-agnostic approach interpretation. anticipate be in new investigation decision making “black box”