Whole-genome phenotype prediction with machine learning: open problems in bacterial genomics.

Area of Science:

• Genomics
• Machine Learning
• Bacterial Genetics

Background:

• Accurate prediction of bacterial traits from genetic data is crucial but hindered by machine learning models identifying spurious correlations.
• Current methods struggle with high-dimensionality and noise in bacterial genomics, leading to unreliable identification of causal genetic variants.

Purpose of the Study:

• To define open problems in predicting bacterial phenotypes from whole-genome data.
• To extend phenotype prediction approaches to learning causal effects.
• To discuss challenges impacting machine decision-making reliability in bacterial genomics.

Main Methods:

• Analysis of genotype-to-phenotype mapping function injectivity.
• Identification of sources of non-injectivity: linkage disequilibrium, limited sampling, information loss, unmeasured confounders, and observational noise.
• Application to a dataset of 4,140 Staphylococcus aureus isolates.

Main Results:

• Major sources of non-injectivity in genotype-to-phenotype mapping were identified.
• Implications of these sources for machine learning applications in bacterial genomics were analyzed.
• Challenges in predicting bacterial phenotypes and learning causal effects were illustrated using Staphylococcus aureus data.

Conclusions:

• Machine learning models face significant challenges in accurately identifying causal genetic mechanisms for bacterial traits.
• Understanding sources of non-injectivity is critical for developing reliable predictive models in bacterial genomics.
• Further research is needed to overcome these hurdles for discovering high-risk bacterial genetic variants.