(586e) Structure-Guided Metabolic Modeling of Non-Model Organisms

Current metabolic pathway designs for genomic variants rely on naïve statistics from only a small fraction of metagenomic data. Protein structure predictions using deep learning techniques such as AlphaFold2 or RosettaFold, on the other hand, exploit any available metagenomic information to construct deep family sequence alignments (or multiple sequence alignments in general) - MSAs. This helps us to create more reliable, biophysically grounded enzyme activity (kcat, k_on, k_off) estimators for whole pathways. Also, latent space encodings derived from ligand-enzyme binding (ESM1b or AminoBERT-type representations) help us alleviate distance/ energy approximations that arise from molecular mechanical models. This, in principle, provides a mechanism to construct a high fidelity de novo pathway design workflow.

Combining structural information (of each and every enzyme per biocatalytic conversion), and more comprehensive utilization of metagenomic data, with metabolic pathway analyses – provides us with the right set of reaction cascades and enzymes’ combinations that aid intended overall bioconversions in vivo in engineered strains. This lends into a future generation of synthetic biology and more efficient synthesis of precursors for key pharmaceutical concoctions, biodegradable polymers, agriculturally relevant toxins, food-flavoring/ preserving agents, and even biofuel.

Our learning algorithms glean rules for energetically-favorable (ΔG>0) conversion routes for synthesis of a desired biomolecule – by optimizing protein cost, translation efficiency, and ATP usage etc. We additionally explore the limits of promiscuity of natural enzymes (from organisms phylogenetically related to the host) that show any activity for a desired bioconversion by compromising its native activity.