MicroGrowAgents: An Agentic AI System for Microbial Cultivation Engineering
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Microbial cultivation optimization remains labor-intensive and inefficient, requiring extensive experimental screening to identify suitable growth conditions. Traditional one-factor-at-a-time approaches are particularly ineffective for exploring complex, multidimensional nutrient parameter spaces. We present MicroGrowAgents, an agent-assisted system for auditable design of candidate growth media through integration of knowledge graphs, metabolic modeling, and optimal experimental design. The system comprises 29 documented agents implementing 58 skills across seven functional categories that query structured biological knowledge (KG-Microbe: 864,363 validated species), Bakta-annotated genomes (667,000+ annotated features), curated organism-specific FACTS sheets, and a DOI-linked publication corpus, and apply this evidence to three ends: recommending candidate ingredients and concentration ranges, specifying the factors of a statistically optimal MaxPro experimental design, and interpreting cultivation outcomes against known biochemistry. We applied the approach to Methylorubrum extorquens AM1 (formerly Methylobacterium extorquens; reclassified per LPSN) by cultivating 69 designed media plus a default media baseline (70 total tested conditions) in quadruplicate and assessing two concurrent objectives: biomass turbidity (740 nm) and apparent residual-Nd depletion (residual Nd by arsenazo III). Monte-Carlo resampling of the replicate-level uncertainty (1000 iterations) identified MPOB_058 as the single MC-stable Pareto-optimal condition (membership frequency 0.922); paired-control biology analysis flags MPOB_058 as chemistry-confounded and nominates MPOB_008 as the cleaner biological-signal anchor (its lower abiotic-drift contribution to the residual-Nd measurement makes it the better candidate for confirming biological lanthanide handling in a follow-up round), with MPOB_019 borderline-stable, providing a prioritized anchor set for confirmation in subsequent rounds rather than a single declared optimum. The integration of chemical similarity search (208,000+ embeddings), metabolic gap analysis, and multi-modal reasoning enables evidence-based hypothesis generation that reduces experimental burden while accelerating discovery of growth-promoting conditions. On the Biolog Odin platform, the MC-stable composite candidate MPOB_058 grew 79% more integrated biomass (area under the 740 nm growth curve) and 46% faster (maximum specific growth rate μ_max; Gompertz fits, R² > 0.99) than the standard MP base medium for this organism (the unsupplemented starting-point recipe carried as the on-plate control). This biomass and growth-rate advantage is a direct kinetic measurement; the condition’s apparent residual-Nd depletion ranking, by contrast, awaits confirmation in a subsequent round, because MPOB_058’s apparent Nd depletion is partly abiotic. Unlike general-purpose AI co-scientists, MicroGrowAgents grounds every recommendation in inspectable evidence — structured provenance manifests with per-session input checksums, schema- and ontology-validated outputs, and 90.5% literature citation coverage (143 of 158 curated DOIs with evidence-supports-claim verification) — separating deterministic design and analysis from agentic interpretation so that recommendations are transparent, explainable, and auditable, while passing 7 of 9 bbop-skills agentic-system criteria in a Claude-Code self-audit (the two unmet criteria are MCP-standardized tool exposure and full input-data cryptographic hashing, both tracked roadmap items).
The Bigger Picture
Identifying growth conditions for poorly characterized microbes is one specific instance of a recurring problem in the natural sciences: how to integrate heterogeneous prior knowledge — structured databases, published literature, mechanistic models — with statistically efficient experimental design so that limited wet-lab time is spent where it most reduces the search. We show that a multi-agent AI system can perform that integration from organism input to wet-lab design, with every recommendation traceable from its source evidence through to the experimental design that tested it. The same architecture (specialized agents over structured knowledge bases, mechanistic models, and statistical design, with checksummed provenance recorded as the unit of evidence) is in principle transferable to other DBTL loops against under-explored search spaces such as catalyst discovery, materials formulation, cell-line engineering, drug-combination screens, or environmental remediation. We demonstrate it here only for microbial cultivation and flag cross-domain transfer as future work rather than a demonstrated capability. Our cultivation results on Methylorubrum extorquens AM1 are a concrete demonstration that this pattern works on real laboratory data: an uncertainty-aware shortlist of candidate media emerged from Monte-Carlo stability analysis of two competing objectives, with the leading composite candidate growing 79% more integrated biomass and 46% faster than the standard MP base medium for this organism; this biomass and growth-rate advantage is directly measured, while its apparent residual-Nd depletion advantage awaits confirmation in a subsequent round, since the apparent Nd depletion is partly abiotic. But the lesson we want to port across research domains is the architecture and its provenance discipline, not the specific organism. MicroGrowAgents is released under BSD-3 with checksummed YAML provenance manifests so that any group adopting these agents inherits a workflow designed to be re-runnable from its recorded inputs.
