gemiz
A tool for reconstructing genome-scale metabolic models. I wanted to see whether protein-language-model embeddings could help with reaction-gene mapping, so I added ESM C 600M scoring and compared it with a no-ESM baseline.
What I did
I built the reconstruction pipeline and the benchmark setup.
Why I built it
Most GEM reconstruction tools lean heavily on homology. That works well in many cases, but I wanted to test whether protein embeddings could add another useful signal.
How it works
gemiz combines sequence search, reaction evidence, COBRApy model building, and an ESM C embedding score. Then it solves and checks the model instead of stopping at annotation.
What came out of it
- Benchmarked the approach against gold-standard E. coli models like iML1515 and iJO1366.
- Added a no-ESM mode so the embedding part can be tested honestly.
- Made a base pipeline I can keep improving for different organisms.
Main pieces
- DIAMOND and BioPython for the sequence side.
- ESM C 600M embeddings for an extra reaction-gene score.
- COBRApy and Gurobi for model assembly, gap filling, and checks.
What I would improve
Embeddings make the pipeline heavier. For me, the point was not to make the cheapest version, but to test whether the extra signal is worth it.