Consistency of in vitro drug sensitivities within pharmacological classes

C. Suk-Yee Hon, Sisira Kadambat Nair, Petr Smirnov, Hossein Sharifi-Noghabi, Nikta Feizi, Shaun Shepherd, Benjamin Haibe‐KainsJournal of Undergraduate Life Sciences 2021 · DOI

Abstract

Multiple comparative analyses between the common drugs and cell lines of the Genomics of Drug Sensitivity in Cancer (GDSC) and the Cancer Therapeutics Response Portal (CTRP) have previously shown low consistency between the in vitro phenotypic measures of a drug in one study with the other. While several potential sources of inconsistency have been tested, the similar targets of tested compounds has yet to be tested as a contributing factor of discrepancy. This analysis includes two methods of reclassifying drugs into classes based on their targets to identify the truer set of consistent cell lines, showing an increased correlation between the two pharmacogenomic studies.

Introduction

The advancement of personalized oncology relies heavily on large-scale pharmacogenomic databases, such as the Genomics of Drug Sensitivity in Cancer (GDSC) and the Cancer Therapeutics Response Portal (CTRP). These resources provide maps of drug responses across hundreds of cancer cell lines, aiming to identify predictive biomarkers. However, a significant "reproducibility crisis" persists: drug sensitivity measurements for the same drug-cell line pairs often show poor concordance across different studies. While factors like tissue type and statistical methodology have been explored, this study investigates whether grouping drugs by their biological targets or Mechanisms of Action (MOA) can help filter out inconsistent data and improve overall reliability.

Methodology

Data from GDSC and CTRP were extracted and analyzed using the PharmacoGx R package. The study focused on an intersection of 619 cell lines and 71 drugs common to both databases.

Drugs were reclassified using two distinct approaches:

  1. Binary Classification: Compounds were categorized as either "Broad-spectrum" (e.g., traditional chemotherapy) or "Targeted" therapies using PubChem data.
  2. Pharmacological Classes (PCLs): Drugs were grouped by specific molecular targets (such as PI3K or Tubulin inhibitors) based on Connectivity Map (CMAP) data.

A filtering step was implemented within these classes to identify and remove cell lines exhibiting poor consistency (Pearson correlation $r < 0.5$) between the two datasets. The primary metrics for evaluation were the half-maximal inhibitory concentration (IC50) and the Area Above the Curve (AAC).

Results

The analysis confirmed that AAC is a significantly more stable and concordant metric than IC50. AAC's robustness stems from its ability to provide a value even when 50% inhibition is not reached, whereas IC50 often results in noisy or null values in such scenarios. By integrating the response over the entire dose-response curve, AAC captures more nuanced information about drug potency and efficacy, making it a superior metric for cross-study comparisons.

Applying the class-based filtering significantly improved data concordance. After filtering based on binary classification, the concordance index for AAC reached 0.939, indicating a very high level of agreement between GDSC and CTRP. IC50 concordance also saw improvements, rising to 0.717. On an individual drug level, compounds such as Bms-754807 and Venetoclax demonstrated substantial increases in correlation following the application of class-based filtering.

Discussion and Conclusion

The results suggest that biological mechanism is a critical factor in understanding data discrepancy. Grouping drugs by target allows for the identification of "true" consistent cell line responses; if a cell line responds inconsistently to an entire pharmacological class across different laboratories, it likely represents an experimental outlier. Leveraging shared biological targets provides a context for identifying experimental noise that is invisible when looking at drugs in isolation.

A notable limitation was the sparsity of specific PCLs; only four classes contained the minimum requirement of three drugs for robust analysis. This highlights the need for broader drug testing across platforms to fully leverage PCL-based filtering. While target-based reclassification improves consistency, it is not the sole solution. Future efforts must focus on standardizing laboratory assays and refining classification rules that integrate chemical structure with biological mechanism to further mitigate noise in pharmacogenomic data.

Summary updated

Abstract

Multiple comparative analyses between the common drugs and cell lines of the Genomics of Drug Sensitivity in Cancer (GDSC) and the Cancer Therapeutics Response Portal (CTRP) have previously shown low consistency between the in vitro phenotypic measures of a drug in one study with the other. While several potential sources of inconsistency have been tested, the similar targets of tested compounds has yet to be tested as a contributing factor of discrepancy. This analysis includes two methods of reclassifying drugs into classes based on their targets to identify the truer set of consistent cell lines, showing an increased correlation between the two pharmacogenomic studies.
BibTeX
@article{hon2021consistency,
  author = {C. Suk-Yee Hon and Sisira Kadambat Nair and Petr Smirnov and Hossein Sharifi-Noghabi and Nikta Feizi and Shaun Shepherd and Benjamin Haibe‐Kains},
  title = {{Consistency of in vitro drug sensitivities within pharmacological classes}},
  journal = {Journal of Undergraduate Life Sciences},
  year = {2021},
  doi = {10.33137/juls.v15i1.37046},
}
Figure