Ponente
Descripción
Genome-scale metabolic models (GSMMs) are fundamental tools for studying metabolism, microbial ecology, and biotechnology. However, the reliable reconstruction of GSMMs remains a major challenge for non-model organisms and microbial communities, where sparse annotations, unknown gene functions, and uncertain biological objectives limit predictive accuracy. We propose a unified computational framework that begins from protein FASTA sequences, either from a single organism or an entire microbial community. Draft GSMMs are first reconstructed using automated platforms such as CarveMe or ModelSEED and subsequently refined through an artificial intelligence (AI)-assisted curation layer.
The proposed AI framework integrates multiple sources of evidence, including sequence homology (BLASTp, DIAMOND, UniProt, NCBI), functional annotation (eggNOG, BioCyc), and structural biology. For proteins lacking significant sequence homologs, three-dimensional structures are predicted using AlphaFold3 or RoseTTAFold and evaluated through pLDDT confidence scores together with geometric quality metrics such as PROCHECK, MolProbity, and QMEAN4. Structural homology searches against repositories and fold-classification resources provide complementary functional evidence, enabling the inference of candidate metabolic reactions that would otherwise remain undetected. Rather than acting as an automatic annotation engine, the AI system aggregates heterogeneous evidence and assigns confidence scores to proposed metabolic functions, while maintaining a human-in-the-loop curation process.
The resulting evidence-supported metabolic network is subjected to AI-guided gapfilling, where reactions are prioritized according to biological plausibility rather than solely topological criteria. Model quality is then evaluated using MEMOTE, generating a quantitative reliability score that determines whether further reconstruction cycles are required. This stage yields a curated GSMM together with explicit confidence estimates associated with its underlying biological assumptions.
Beyond reconstruction, the central open problem concerns the inference of biological objectives. Classical flux balance analysis assumes that organisms maximize biomass production,
$$ \max v_{\text{biomass}}, $$ an assumption that is often appropriate for model organisms but may not hold for many environmental species. Non-model organisms may instead optimize survival, osmoadaptation, stress tolerance, resource efficiency, or ecological persistence. Consequently, a more general formulation may involve multiple competing objectives, $$ \max \left(F_{\text{growth}}, F_{\text{survival}}, F_{\text{stress}}, \ldots \right), $$
whose relative importance remains largely unknown. We therefore identify the inference of organism-specific objective functions from genomic, structural, metabolic, and ecological evidence as a fundamental open problem.
The framework naturally extends to microbial communities, where multiple GSMMs are reconstructed simultaneously and integrated into ecosystem-scale models to investigate cross-feeding, competition, mutualism, and emergent community objectives. The overarching challenge is thus to develop a mathematically rigorous framework capable of reconstructing reliable metabolic networks, quantifying uncertainty, and inferring biological objectives across multiple organizational scales. Such a framework would establish a new foundation for AI-assisted mathematical biology, bridging systems biology, machine learning, structural biology, and ecological modeling.