Overview
We evaluated Odinn on datasets from the largest human metabolic pathway database, Reactome, and on 7 organismal datasets from Metexplore (B. Aphidicola, B. Cicadellinicola, C. Rudii, E. Coli, H. Sapiens, S. Cerevisiae, S. Muelleri).
Reactome has 20,000 molecules and 12,000 reactions, and the organismal datasets range from 263 to 1,877 molecules and from 229 to 2,999 reactions.
These respectively map to the vertices and hyperedges in our hypergraphs, resulting in a hypergraph built from Reactome with 20,000 vertices and 12,000 hyperedges.
Any vertex without in-edges is treated as a source, and any vertex without out-edges is treated as a target.
An instance of the problem is to find the shortest hyperpath from all of the sources to a single target, resulting in 5000 Reactome instances, and between 44 and 344 instances for the organismal datasets.
Negative regulation
We found 83 instances from Reactome where all optimal factories were invalid under first-order negative regulation, and 11 instances where all first-order factories were invalid under second-order negative regulation.
Running time
Odinn has a median running time of just 3 seconds on all instances.
The maximum running time for computing 0th- or 1st-order factories is 61 seconds and for computing 2nd-order factories is nearly 15 hours, with only 17 instances running over an hour.
Source code
The MILP including negative regulation is available on GitHub.