(639a) A Generative Model of S. Cerevisiae for the Design of Multiplexed knockout Mutants | AIChE

(639a) A Generative Model of S. Cerevisiae for the Design of Multiplexed knockout Mutants


Zhao, H., University of Illinois-Urbana
Metabolic engineering of yeasts allows for modification of industrial relevant yeast phenotypes including improved growth rates and overproduction of target metabolites. Traditional in silico design tools including constraint-based modeling kinetic modeling rely on mechanistic representations of the cell but lack flexibility including disparate data. We have developed a pretrained machine learning model for the generation of strain designs according to a user specified phenotype. The model is trained as an autoencoder for the generation of new strain designs and is supervised on multimodal data from S. cerevisiae multiplexed knockout mutants, including cell growth, gene expression, and metabolite concentrations for kinase knockouts. Additionally, metabolic states of the model are supervised by simulated data from genome scale model YEAST8 FBA simulations. The model architecture uses graph neural networks on a multigraph composed of flat networks including gene-gene interactions, protein-protein interactions, and the regulatory interaction network, and the hierarchical representation of the cell, gene ontology. We first do a benchmark predicting yeast fitness to show performance variation across three alternative representations of gene manipulations to prove graph neural networks can capture gene deletion effects. The model achieves improved performance compared to 3 separate prediction models of yeast fitness including DCell, Yeast8, and MMANN. As a demonstration, we predict and test the metabolic profiles from multiplexed kinase knockouts. The model is generalized to construct representations at node, edge, and global levels of the multigraph, allowing for predictions of previously uncharacterized interactions, gene annotations, and global cell states. The model is also modular, allowing users to inherit a pretrained model and fine tune it on a smaller custom dataset. We envision this model serving as a design tool for constructing multiplexed knockout mutants that can warn against strain designs with high-order synthetic lethality, for suggesting tuned genetic expression, and for the over production of target chemicals.