Unsupervised decomposition of massively parallel TraDIS-Xpress fitness data from Escherichia coli identifies distinct functional and expression modules of antimicrobial resistance determinants
Linking genotype to phenotype is a central goal in biology, one way to do this is using transposon insertion sequencing (TIS) where the fitness impacts of mutations across a genome are assessed in different conditions. Traditional analyses use comparison of each gene response independently in control versus test conditions, so a compendium of hundreds of test conditions returns hundreds of separate ranked gene lists. To unlock the potential information hidden within large data sets, we applied independent component (ICA) analysis to 754 Escherichia coli BW25113 TraDIS-Xpress samples. These samples included a series of antibiotic and other stress condition exposures. ICA is designed to identify hidden structure within complex data and to extract meaningfully linked components. In this analysis each annotated gene was resolved into a gene body and 198 bp 5' and 3' features so that expression effects driven by the transposon's outwards facing promoter enter the matrix separately from gene disruption effects. The decomposition returned 99 components, which we term fxModules, which were defined as containing elements with similar responses across the whole compendium rather than per any individual experiment. Of the 754 stress samples, the antibiotic arm included 110 samples derived from exposure to 14 separate agents at multiple concentrations. Many of these have been highly studied allowing us to explore the accuracy of predicted responsive fxModules. This analysis showed excellent predictive power of the approach and the fxModules responsive to individual drugs typically reproduced known mechanisms at a resolution superior to analysis of single gene sets. For example, exposure to fosfomycin identified responsive fxModules containing genes which are known to be involved in uptake of the drug, cAMP-CRP control and exposure to gentamicin identified oxidative phosphorylation split across three fxModules responding with different magnitudes and directions at different concentrations. This recapitulates recent work studying the landscape of E. coli genes sensitive to this drug with high precision without any prior knowledge of functional annotation. Similarly precise predictions were made for other drugs. Responsive modules also identified novel groups of genes and linkages beyond the boundaries of defined pathways as being responsive to various stresses demonstrating the ability to predict new genotype-phenotype links. Inclusion of the regulatory regions also resolved fxModules which were entirely contributed by promoter regions, for example for chloramphenicol genes responsible for efflux were identified by signals in intergenic control regions, and unidirectional insertions in gene bodies, a signal that a gene level matrix would not be able to predict with accuracy. These data show that ICA decomposition provides an efficient and accurate annotation independent approach to identifying sets of genes with related functional impacts in any given condition in the compendium. The functional modular association of genes also help in predicting gene functions of unknown genes based on the group function. This approach allows rapid identification of novel genotype-phenotype links and unlocks the power of multiple comparisons of transposon mutagenesis data.