#37 · Systems biology

Learned models and enzyme-constrained networks extend genome-scale metabolic prediction

The dominant thread is the use of genome-scale metabolic models (GEMs) as a mechanistic scaffold for learned components. HyTT pairs MARS regression with an enzyme-constrained metabolic model to predict proteome reallocation ab initio in S. cerevisiae, validated against chemostat data. AlphaGEM targets reconstruction itself, combining proteome-scale structural alignment and PLMSearch with deep-learning mining of nonhomologous 'dark' metabolic genes to build genome-scale models. ChatGEM addresses access rather than construction, using a multi-agent retrieval-augmented architecture with COBRApy so that simulations can be run via natural language, benchmarked on three tasks. Not every item shares the same learning component. The new element is the integration of learned models with metabolic network context, with protein language model search appearing only in AlphaGEM.