Towards context-aware similarity metrics

Comparing molecular structures is often done based on relatively high-dimensional, binary fingerprints that represent absence or presence of certain structural elements. The then commonly used Tanimoto distance considers all those bits similar, although some of the underlying structural elements may be quite closely related. We propose a weighting scheme for such related bits which softens the influence of differences in bit patterns that represent similar contexts. We demonstrate the usefulness of this context aware similarity metric using an automated bit-weighting mechanism and a set of similarity weights based on Topomer shape similarity between the structural elements encoded in the finger print.