Genome files are just the beginning. With gemiz, I’m developing a Python pipeline that takes bacterial genomes and builds genome-scale metabolic models. It generates an SBML model and keeps the supporting evidence traceable.
The pipeline calls genes with pyrodigal, scores reaction evidence with MMseqs2, and builds the network with COBRApy and HiGHS. Each step has a different task. Keeping those steps separate makes it easier to track down an unexpected result.
An SBML model, by itself, does not explain why a reaction was included. That is why gemiz stores the evidence, quality-control artifacts, and exported model together. The benchmark setup allows comparisons with CarveMe and reference models. Those comparisons help describe the reconstruction’s limits; they do not establish that every prediction is correct.
The public alpha focuses on bacterial genomes. Eukaryotic support and experimental embedding-based scoring are outside the supported workflow. For now, the goal is a pipeline whose outputs can be checked, questioned, and reproduced.