Yes we have recommended settings for the use of Sarah.
The user is able to change the settings within Sarah. This gives the user some flexibility in how the predictions are derived in order to best match their use-case. For example if harsh triaging was desired to answer the question “do any of my molecules have potentially activating features?”, then the user could select ‘conservative’ whereby any activating fragments would trigger a positive prediction irrespective of the presence of deactivating features. Similarly if only very confident predictions are wanted, then the user can raise the equivocal threshold (level below which the weight of evidence is not high enough to make a call).
However, for typical use, and particularly for regulatory submission, we recommend that the settings for both equivocal threshold and sensitivity are both set at 8% and the reasoning is set at weighted. This then provides the best compromise between sensitivity and specificity and also provides the best overall accuracy. We have done extensive analyses using a number of proprietary datasets to derive these settings and these were also recently confirmed by the FDA and presented in their poster shown at the Society of Toxicology (SOT) meeting in Phoenix, March 2014.
Further details of the performance under these settings is described in a poster we presented in collaboration with the FDA at the SOT meeting.
In addition, details of our analysis were presented at the 2014 SOT by Chris Barber, Director of Science at Lhasa Limited SOT.