ChemGAPP: a tool for chemical genomics analysis and phenotypic profiling

Abstract

Motivation: High-throughput chemical genomic screens produce informative datasets, providing valuable insights into unknown gene function on a genome-wide level. However, there is currently no comprehensive analytic package publicly available. We developed ChemGAPP to bridge this gap. ChemGAPP integrates various steps in a streamlined and user-friendly format, including rigorous quality control measures to curate screening data.

Results: ChemGAPP provides three sub-packages for different chemical-genomic screens: ChemGAPP Big for large-scale screens; ChemGAPP Small for small-scale screens; and ChemGAPP GI for genetic interaction screens. ChemGAPP Big, tested against the Escherichiacoli KEIO collection, revealed reliable fitness scores which displayed biologically relevant phenotypes. ChemGAPP Small demonstrated significant changes in phenotype in a small-scale screen. ChemGAPP GI was benchmarked against three sets of genes with known epistasis types and successfully reproduced each interaction type.

Availability and implementation: ChemGAPP is available at https://github.com/HannahMDoherty/ChemGAPP, as a standalone Python package as well as Streamlit applications.

Figure 1

Workflow of the ChemGAPP packages. ChemGAPP Big analyses large scale chemical genomic screens. ChemGAPP Small analyses small scale chemical genomic screens. ChemGAPP GI analyses small scale genetic interaction screens.

Figure 2

ChemGAPP Big highlights plates with errors common to chemical genomic screens. (A) Plate matrix depicting the colony sizes within replicate plates of the condition 20C (cold shock 20°C) and the percentage normality determined by the Z-score test. (B) Density of colony sizes for replicates A, B, and C for condition Rif2 (Rifampicin 2 µg/ml), Plate 6, Batch1. Difference in the distribution of C versus A and B is statistically significant by Mann–Whitney test. **0.001 < P-value ≤0.01; ns, non-significant.

Figure 3

ChemGAPP produces robust and accurate S-scores. (A and B) Joint scatter and density plot depicting the difference between (A) the original non-curated S-scores versus the mean bootstrapped S-scores for each mutant within each condition of the KEIO dataset. (B) The curated S-scores versus the mean bootstrapped S-scores for each mutant within each condition of the KEIO curated dataset. Density plots show distribution of hits with various S-scores. For ease of visualization, outlier bootstrapped S-scores >125 were excluded, representing a negligible 0.000018% (A) and 0.000026% (B) of all values (see Supplementary Additional File S2). Outlier curated S-scores >125 were also excluded, representing a negligible percentage of all values (0.00056%). MAE, mean absolute error; % of std, percentage of the standard deviation for the original dataset that the MAE constitutes. (C) Receiver operating characteristic curve with AUC values for the KEIO ChemGAPP Big non-curated dataset (red line), the KEIO ChemGAPP Big curated dataset (blue line), and the KEIO dataset from Nichols et al. (2011) (yellow line).

Figure 4

Clustered heatmaps displaying the S-scores for various single gene knockout mutants in different conditions. (A) The GCV system in sulfonamide drugs (Sulfa), from left to right: Sulfamonomethoxine 100 µg/ml; Sulfamethoxazole 100 µg/ml; Sulfamethoxazole 200 µg/ml; Sulfamethoxazole 300 µg/ml; Sulfamonomethoxine 50 µg/ml. (B) AcrAB-TolC system mutants in presence of AcrAB-TolC substrates. ns, non-significant.

Figure 5

ChemGAPP Small produces informative fitness ratios. ChemGAPP Small provides the option of two different output plot types. (A) Bar plot output of ChemGAPP Small for ΔenvC in LB, and 0.25% SDS + 0.25 mM EDTA, error bars represent 95% confidence intervals. (B) Swarm plot output of ChemGAPP small for ΔenvC in LB, and 0.25% SDS + 0.25 mM EDTA. ns = P > 0.05; ***0.0001 < P ≤ 0.001.

Figure 6

ChemGAPP GI can accurately predict all types of genetic interactions. (A) Nonsignificant difference between the double expected and double observed fitness ratios, therefore indicating no epistasis between dacB and nlpI. (B) Double observed significantly lower than double expected, showing negative epistasis between mrcB and nlpI. (C) Double observed significantly fitter than double expected. Showing positive epistasis between diaA and bamB. ns = P > 0.05; ***0.0001 < P ≤ 0.001.