Computational approaches are increasingly used to predict toxicity due, in part, to pressures to find alternatives to animal testing. Read-across is the "new paradigm" which aims to predict toxicity by identifying similar, data rich, source compounds. This assumes that similar molecules tend to exhibit similar activities i.e. molecular similarity is integral to read-across. Various of molecular fingerprints and similarity measures may be used to calculate molecular similarity. This study investigated the value and concordance of the Tanimoto similarity values calculated using six widely used fingerprints within six toxicological datasets. There was considerable variability in the similarity values calculated from the various molecular fingerprints for diverse compounds, although they were reasonably concordant for homologous series acting via a common mechanism. The results suggest generic fingerprint-derived similarities are likely to be optimally predictive for local datasets, i.e. following sub-categorisation. Thus, for read-across, generic fingerprint-derived similarities are likely to be most predictive after chemicals are placed into categories (or groups), then similarity is calculated within those categories, rather than for a whole chemically diverse dataset.