Interpretable genotype-to-phenotype classifiers with performance guarantees

Abstract

Understanding the relationship between the genome of a cell and its phenotype is a central problem in precision medicine. Nonetheless, genotype-to-phenotype prediction comes with great challenges for machine learning algorithms that limit their use in this setting. The high dimensionality of the data tends to hinder generalization and challenges the scalability of most learning algorithms. Additionally, most algorithms produce models that are complex and difficult to interpret. We alleviate these limitations by proposing strong performance guarantees, based on sample compression theory, for rule-based learning algorithms that produce highly interpretable models. We show that these guarantees can be leveraged to accelerate learning and improve model interpretability. Our approach is validated through an application to the genomic prediction of antimicrobial resistance, an important public health concern. Highly accurate models were obtained for 12 species and 56 antibiotics, and their interpretation revealed known resistance mechanisms, as well as some potentially new ones. An open-source disk-based implementation that is both memory and computationally efficient is provided with this work. The implementation is turnkey, requires no prior knowledge of machine learning, and is complemented by comprehensive tutorials.

Predicting Phenotypic Polymyxin Resistance in Klebsiella pneumoniae through Machine Learning Analysis of Genomic Data

“…First, our reference-based approach had high performance (AUROC, 0.885 to 0.933) and outperformed the rule-based approach that we used as a benchmark. Despite PR being more biologically complex than other forms of antimicrobial resistance in K. pneumoniae, it was encouraging that our results were similar to those obtained for other antimicrobials with less complex mechanisms of resistance (e.g., carbapenem resistance) tested with ML approaches (14,15,(34)(35)(36). The performance of this approach falls below the FDA cutoffs for AST tests (37).…”

“…Impact of feature engineering with GWAS filtering and polymyxin exposure data on performance of ML-based prediction. A key challenge in AST genotypephenotype prediction is the sparsity of the input genomic data sets due to relatively few genomes in data sets compared to the number of genomic features (14). Appropriate feature selection prior to training ML algorithms is a potential solution to this…”

“…Second, these rule-based models struggle to account for complex interactions between variants in multiple loci. In order to move beyond these limitations, machine learning (ML) methods have been used to predict antimicrobial susceptibility (14)(15)(16)(17). Given the incomplete identification of contributing PR mutations and the possible polygenic nature of PR, we hypothesize that ML approaches may be well suited to AST genotype-phenotype prediction in this setting (18) and may ultimately be used to help identify isolates for confirmatory phenotypic testing.…”

Improved Prediction of Bacterial Genotype-Phenotype Associations Using Interpretable Pangenome-Spanning Regressions

“…They are potentially broadly applicable to any problem with vast amounts of data, though they perform best when the number of data points exceeds the number of dimensions. Unsurprisingly their uptake in sequence analysis (19,20) and bacterial genomics specifically has been rapid (4,5,21). In general, the predictive variants in the identified sets may not fully overlap those significant by themselves, and the mapping does not necessarily lend itself readily to the same interpretation as P values in a GWAS.…”

Bioinformatics Approaches to the Understanding of Molecular Mechanisms in Antimicrobial Resistance

“…Such additional insight may help one to better understand the AMR causality of alleles in future studies. The latest and most comprehensive work within this approach was published by Drouin and colleagues and covers 12 bacterial species and 56 drugs [13]. In addition to generating the collections of k-mers associated with susceptible or resistant strains, they applied rule-based ML algorithms-Classification and Regression Trees (CART) and Set Covering Machines (SCM).…”