The Science
Computational evolution shaped by fermentation demands.
We capture the biology your fermentation team knows, then search sequence space beyond manual exploration.
Genome mapping
Not a black box. Each proposal links to a flux prediction.
The platform combines a genome-scale metabolic network (GSM) and sequence-level transformer scoring. The GSM represents organism-specific reaction stoichiometry and flux limits. The transformer scores sequence changes by their predicted effects on those limits.
Each proposed modification connects to a predicted pathway flux change. You can see why a candidate ranked first, which pathway interaction it targets, and what the model forecasts at the metabolite level.
The model uses data from your organism, rather than another host. A citric acid model for your specific Aspergillus niger production strain differs from a model for a published lab strain.
Architecture
Transformer and graph neural network for fermentation.
In the metabolic graph, reactions and metabolites are nodes, while flux constraints form edges. A graph neural network spreads modification signals through the network to predict systemic flux changes from targeted gene edits.
A transformer layer compares candidate sequences with predicted flux outcomes, ranking modifications by expected effects on yield and tolerance. Training combines public genome databases with proprietary fermentation outcome data from early-access pilot partners.
Validating the predictions.
Early validation markers, scope stated plainly.
Pilot feedback loop
Lab results from 4 early-access partners refine predictions through each iteration. Each partner's results update the model's flux assumptions for its organism-molecule pair.
Retrospective review
We test published fermentation datasets covering 8 organism-molecule pairs from the literature. Retrospective rankings of published modifications check prediction direction and magnitude.
Internal check
12 characterized organism and target molecule combinations serve as an internal standard for model quality checks. Includes common industrial chassis for citric acid, lactic acid, and terpene production.
Research foundations for fermentation reality.
StrainX Works draws on genome-scale metabolic modeling research and newer protein language models for functional sequence design. Its genome-scale metabolic modeling approach follows the COBRA toolbox research lineage developed through two decades of computational systems biology.
We apply these research directions to industrial fermentation constraints: low-cost feedstocks, scale-up reliability, and regulatory-accepted host organisms. The challenge is more than prediction accuracy. Models must align with commercial fermentation reality, since performance on published benchmarks is not useful to process development if physiological scale-up realities are overlooked.
Our work addresses the gap between published genome-scale modeling methods and industrial strain development needs: non-model organism support, scale-dependent constraint handling, and iteration speeds that academic models were not designed to provide.
Our founding team combines experience in metabolic engineering, ML for biological sequences, and industrial fermentation process development.
Meet the teamReady to apply it to your organism?
Join the early-access program or book a call with our scientific team.