Perturbation-specific transcriptional mapping for unbiased target elucidation of antibiotics

Abstract

The rising prevalence of antibiotic resistance threatens human health. While more sophisticated strategies for antibiotic discovery are being developed, target elucidation of new chemical entities remains challenging. In the postgenomic era, expression profiling can play an important role in mechanism-of-action (MOA) prediction by reporting on the cellular response to perturbation. However, the broad application of transcriptomics has yet to fulfill its promise of transforming target elucidation due to challenges in identifying the most relevant, direct responses to target inhibition. We developed an unbiased strategy for MOA prediction, called perturbation-specific transcriptional mapping (PerSpecTM), in which large-throughput expression profiling of wild-type or hypomorphic mutants, depleted for essential targets, enables a computational strategy to address this challenge. We applied PerSpecTM to perform reference-based MOA prediction based on the principle that similar perturbations, whether chemical or genetic, will elicit similar transcriptional responses. Using this approach, we elucidated the MOAs of three molecules with activity against Pseudomonas aeruginosa by comparing their expression profiles to those of a reference set of antimicrobial compounds with known MOAs. We also show that transcriptional responses to small-molecule inhibition resemble those resulting from genetic depletion of essential targets by clustered regularly interspaced short palindromic repeats interference (CRISPRi) by PerSpecTM, demonstrating proof of concept that correlations between expression profiles of small-molecule and genetic perturbations can facilitate MOA prediction when no chemical entities exist to serve as a reference. Empowered by PerSpecTM, this work lays the foundation for an unbiased, readily scalable, systematic reference-based strategy for MOA elucidation that could transform antibiotic discovery efforts.

Keywords: RNAseq; antibiotics; gene expression; mechanism-of-action; transcriptomics.

Fig. 1.

Correlating antimicrobial gene expression profiles of the antimicrobial reference set. (A) Heat map of Pearson correlations of the replicate-collapsed, z-score expression profiles of PA14 treated with all 37 compounds in the antimicrobial reference set at the highest concentration at 90 min. Rows and columns are organized and color-coded by MOA. All correlation values below zero were set to zero for visualization purposes. (B) UMAP visualization of all compound treatments (at the highest dose) of wild-type PA14 and seven engineered hypomorphs at 90 min. Colors correspond to antibiotic MOA classes: DNA synthesis (red), membrane integrity (blue), cell wall synthesis (green), and protein synthesis (magenta). Closed circles indicate wild-type PA14 treatments, whereas all other markers indicate treatment of a hypomorphic strain, as indicated. Antimicrobial compounds with robust activity at the experimental inoculum group by MOA, agnostic of the genetic strain. Protein synthesis inhibitors form subgroups that separate tetracyclines, macrolides, and aminoglycosides. Except for novobiocin, compounds that are not active against PA14, such as fosmidomycin, piperacillin, rifampicin, doxorubicin, and hydroxyurea, appear close to DMSO vehicle. Treatments that were deemed transcriptionally weak based on DESeq2 analysis were not included in the visualization. (C) Confusion matrix showing the number of compounds in each query target category that were predicted to have indicated targets in the leave-one-out cross-validation (LOOCV), using the method of calculating the target correlation prediction score (r_predict). Compounds with correctly predicted target fall along the diagonal, while those with correctly predicted MOA fall within the colored boxes. Note that query target categories that only contain one compound (the column sums to one) could not be predicted correctly, due to the LOO approach.

Fig. 2.

PerSpecTM analysis of external dataset by mapping to the antimicrobial reference set. (A) Target correlation scores (${\overline{r}}_{\mathit{max}}$) of the top six external query compounds against the target categories indicated in the antimicrobial reference set. The highest ${\overline{r}}_{\mathit{max}}$ score is the target correlation prediction score (r_predict) and identifies the predicted target. Colors correspond to MOA classes: DNA synthesis (red), membrane integrity (blue), cell wall synthesis (green), and protein synthesis (magenta). (B and C) Distribution of r_predict for all compounds in B the internal reference set in the LOOCV compared to the external query set using z-scores and (C) the external query set using z-scores compared to nz-scores. Each point represents a compound. Points of the same color identify the same compound in both the reference and external query sets, as shown. Dotted lines indicate three r_predict thresholds, labeled with approximate target PPVs determined from the LOOCV z-score analysis (with MOA PPVs in parentheses). Target prediction scores are not indicated for compounds that were most correlated to negative control.

Fig. 3.

MOA determination of antipseudomonal compounds. (A) Target correlation scores (${\overline{r}}_{\mathit{max}}$) of PA14 treated with PA-0918 against the target categories indicated in the antimicrobial reference set. PA-0918 is predicted to be a direct LPS binder, similar to colistin, based on a target correlation prediction score (r_predict) of 0.74 (target PPV of 100%). (B) Ethidium bromide uptake in cells treated with PA-0918 and colistin, with ciprofloxacin as a negative control. PA-0918 treatment phenocopies colistin with rapid uptake of ethidium bromide suggestive of membrane disruption. In all experiments, error bars represent SEM of three biological replicates (n = 3). (C) ${\overline{r}}_{\mathit{max}}$ scores of oprL-hypomorph treated with PA-69180 against the target categories indicated in the antimicrobial reference set. PA-69180 is predicted to be an MreB inhibitor, similar to A22, based on an r_predict of 0.38 (target PPV of 68%). (D) MG1655 overexpressing ftsZ (MG1655/ FtsZup) is resistant to both A22 and PA-69180 and is rescued from mecillinam killing (mec- and mec+ denote 0 μg/mL and 2.5 μg/mL mecillinam, respectively) by both compounds, implicating MreB as the target. (E) ${\overline{r}}_{\mathit{max}}$ scores of PA14 treated with PA-5750 against the target categories indicated in the antimicrobial reference set. PA-5750 is predicted to be a GyrA/ParC inhibitor, similar to ciprofloxacin, based on a r_predict of 0.56 (target PPV 88%). (F) Two ciprofloxacin-resistant E. coli J53 mutant strains, each containing single point mutations (G81D or S83L) in GyrA, are cross-resistant to PA-5750, implicating DNA gyrase at the target.

Fig. 4.

PerSpecTM analysis of antimicrobial treatments queried against the genetic depletion reference set. Heatmap intensities show target correlation scores (${\overline{r}}_{\mathit{max}}$) of 90-min antibiotic treatments when queried against CRISPRi strains in the genetic depletion reference set treated with 0.125% or 0.5% arabinose to induce dCas9 expression. PA14 was treated with all query compounds except fosmidomycin, which lacked wild-type activity, and thus the dxr-hypomorph was used. White dots indicate the target correlation prediction score (r_predict) for each query compound. Colored boxes indicate the MOA relating to each target gene or query compound. Note that from the genetic depletion reference set, only those that have corresponding compound treatments targeting the same gene targets, or the associated ribosomal/protein complexes, are shown.