AI-assisted CRISPR screening for metabolic pathway engineering in microbes

Thesis Abstract

Advances in genome editing and machine learning present an opportunity to accelerate microbial metabolic pathway optimization through AI-guided CRISPR screens that map genotype–phenotype relationships with high resolution. The study addresses the persistent challenge of efficiently identifying genetically actionable edits that enhance production of target metabolites while preserving cellular fitness, enabling scalable and tunable pathway engineering in microbial hosts. The aim is to develop an integrated AI-assisted CRISPR screening framework that prioritizes gene targets and edit combinations to maximize pathway flux and product yield under defined cultivation conditions. Specific objectives include (1) to curate a comprehensive CRISPR perturbation library targeting key enzymes, transporters, and regulatory nodes across a model solventogenic bacterium and a budding yeast chassis; (2) to implement high-throughput single-guide RNA (sgRNA) screens coupled with multiplexed barcoded sequencing to quantify fitness and production phenotypes; (3) to train and validate machine learning models that correlate genotypic perturbations with phenotypic outputs, incorporating mechanistic priors from metabolic network analysis; (4) to compare AI-guided recommendations with conventional rational design in iterative rounds of wet-lab validation; and (5) to assess robustness of the framework across variable fermentation conditions and scale-up scenarios. The methodology adopts a mixed-methods research design combining experimental CRISPR screening and quantitative modeling. The population comprises two microbial platforms Escherichia coli engineered for acetate-to-product conversion and Saccharomyces cerevisiae tailored for high-value metabolite synthesis. A combinatorial CRISPR library of approximately 3,000 sgRNA perturbations per chassis is constructed, with each perturbation represented by at least five replicates to enable statistical power. Data collection employs high-throughput sequencing for sgRNA abundance, metabolite assays via LC-MS/MS for product titers, and growth metrics from automated plate readers. Instrumental data also include time-series transcriptomic and proteomic measurements for a subset of 300 perturbations to inform mechanistic interpretation. The AI framework integrates supervised learning algorithms (random forests, gradient boosting, and deep neural networks) with Bayesian regularization and physics-informed constraints from flux balance analysis (FBA). Feature engineering leverages pathway topology, gene essentiality, expression levels, and historical perturbation outcomes. Model performance is evaluated using cross-validation, with performance metrics including R-squared, RMSE for yield predictions, and area under the precision–recall curve for hit identification. Validity and reliability procedures include technical replicates, spike-in controls for sequencing, and preregistered analysis pipelines to ensure reproducibility. Ethical considerations focus on biosafety compliance and responsible genome editing practices, with risk assessment and containment strategies reviewed by an institutional biosafety committee. Anticipated results include identification of non-obvious gene combinations that synergistically enhance product yield by rerouting carbon flux, alongside resilient edit strategies that maintain growth under stress. Comparative analyses are expected to show AI-guided design achieving 20–35% higher yield and 15–25% improved productivity relative to traditional rational design approaches, with reduced experimental iterations. The study contributes to knowledge by (a) demonstrating the feasibility and advantages of an integrated AI-guided CRISPR screening pipeline for metabolic engineering, (b) providing a transferable modeling framework that links genotype to phenotype in microbial systems, and (c) delivering a curated, scalable perturbation library and associated analytical tools for future strain optimization. The main conclusion posits that AI-assisted CRISPR screening can systematically uncover high-impact regulatory nodes and synergistic edit sets to optimize metabolic pathways beyond conventional intuition, underpinned by robust predictive models and mechanistic insights. Practical recommendations include extending the framework to additional chassis and product classes, integrating real-time sensing to close the loop between production and regulation, and developing standardized data formats to enhance cross-study comparability.