Description
The models will be tested against 3 types of data. There are 2 types of outcomes for each type of data. First, is when model predicts growth or secretion, but experimental data tells otherwise. This situation is considered as false positive prediction. Generally, this requires a removal of a reaction to make the model consistent with the data. Second, is the opposite, when model does not predict growth or secretion, but it does happen. This is considered false negative. To address this it typically requires adding a reaction.
In summary:
false positive -> removing reaction
false negative -> adding reaction

A special case is when a metabolite is a byproduct of growth. If it must be produced (according to the model), but not present in the experimental data of the secreted metabolites. It may be missing (false positive) for two reasons. First - metabolomics experimental platform can't detect it (e.g. ethanol or ammonia). In this case we can't make any conclusions. Second - it is detectable, but missing. In this situation, the model requires more complicated update, rather than removing a reaction.

Approaches for adding/removing reactions:

• gapfill
• crop
  The probabilistic approaches take together 1) experimental, 2) computational, 3) pathway context and 4) thermodynamic evidences to come up with a probability for a given reaction that could be added.

Deliverable
Refined models pass more tests (less FP and FN) than previously.