Session 1: Systems that predict metabolism
Lhasa Limited, United Kingdom
ABSTRACT CONTENT / DETAILS:
There are problems with assessing the performance of systems that predict metabolism. Measures such as the Cooper statistics, selectivity, sensitivity and concordance, and receiver operating characteristic (ROC) curves are primarily suited to binary predictions. Some metabolism prediction systems attach varying confidence to predictions – i.e. they do not deliver simple binary predictions.
Standard metrics require figures for false positive predictions. Failing to observe, or record, the presence of a metabolite is not the same as confirming its absence and so there is uncertainty about false positive predictions.
Possible approaches to improving the assessment of performance of metabolism predictions will be discussed, including the use of a new metric, “veracity”, that takes into account how well systems judge levels of confidence in their predictions.