Discovery of synthetic lethal interactions from large-scale pan-cancer perturbation screens

Sumana Srivatsa, Hesam Montazeri, Gaia Bianco, Mairene Coto‐Llerena, Mattia Marinucci, Charlotte K.y. Ng, Salvatore Piscuoglio, Niko BeerenwinkelNature Communications 2022 · DOI

TL;DR

This paper introduces SLIDR, a new statistical framework for identifying synthetic lethal (SL) gene pairs from large-scale cancer cell line perturbation screens. SLIDR is more robust than existing methods, especially for rare cancer types with limited data, and has been experimentally validated in hepatocellular carcinoma, offering a promising tool for discovering new cancer drug targets.

What Problem Does This Paper Address?

Developing effective cancer therapies is hampered by the limited number of actionable drug targets. Synthetic lethality (SL) offers a promising avenue: it describes pairs of genes where the malfunction of either gene alone is tolerated by a cell, but the simultaneous malfunction of both leads to cell death. Identifying these SL pairs is crucial for developing targeted therapies that exploit cancer-specific genetic vulnerabilities. However, existing computational methods often struggle with the small sample sizes typical of rare cancer types and can generate a high number of false positives. This paper aims to address these limitations by developing a more powerful and accurate method for identifying SL interactions from large-scale perturbation screening data.

Methods

The core contribution is the development of SLIDR (Synthetic Lethal Identification in R), a novel statistical framework. Unlike previous methods that often rely on raw viability readouts, SLIDR utilizes the normalized ranks of viabilities across all perturbed genes for each cell line. This ranking approach is key to improving statistical power, particularly with small sample sizes.

SLIDR's workflow involves:

  1. Stratification: Dividing cell lines into "mutated" and "wild-type" groups based on the status of a "driver" gene.
  2. Irwin-Hall Tests: Applying one-sided Irwin-Hall tests to identify perturbed genes (potential SL partners) that cause significantly lower ranked viabilities in mutated cell lines compared to wild-type cell lines.
  3. Causal Inference: Employing matching-based potential outcome models to account for confounding factors arising from multiple co-occurring mutations in cancer cells, thereby prioritizing the most likely SL pairs.

The method is designed to be flexible, applicable to both pan-cancer analyses (across all cell lines) and cancer type-specific analyses.

Datasets

The primary dataset used is the Project DRIVE dataset, a large-scale shRNA perturbation screen of 7,837 genes across 398 cancer cell lines from the Cancer Cell Line Encyclopedia (CCLE). Mutation and copy number data from CCLE were also utilized.

For validation, the authors used:

  • PRISM dataset: A large-scale drug-response profile dataset to assess the drug-targetability of predicted SL interactions.
  • Experimental SL interaction data: Data from 17 in vitro SL screens and 10 combinatorial CRISPR screens to benchmark SLIDR's performance against experimentally validated SL pairs.
  • CRISPR data from the DepMap consortium (Project Achilles): Used to further validate SLIDR's predictions and assess its performance on a different type of perturbation screen.

Key Insights and Ideas

  • Ranking is More Powerful Than Raw Scores: The core insight is that using ranked viabilities, rather than raw viability scores, significantly boosts the statistical power of SL identification, especially in datasets with limited samples. This is a departure from many existing methods.
  • SLIDR Excels in Small Sample Sizes: The method's design, particularly the use of ranked viabilities and Irwin-Hall tests, makes it robust for identifying SL pairs even in rare cancer types where sample sizes are small.
  • Pan-Cancer vs. Cancer Type-Specific Signals: The study highlights that while pan-cancer analyses can identify broad SL interactions, many of these signals are driven by specific cancer types. This underscores the importance of performing both types of analyses for a comprehensive understanding.
  • Experimental Validation is Crucial: The successful experimental validation of ARID1A-TEAD1 and AXIN1-URI1 in hepatocellular carcinoma provides strong evidence for SLIDR's predictive power and its potential for identifying clinically relevant drug targets.

Key Results

  • SLIDR identified 151 pan-cancer SL pairs involving 84 driver genes.
  • After accounting for confounding mutations, 90 SL pairs across 42 driver genes were predicted.
  • The method identified 839 cancer type-specific SL pairs over 233 unique driver genes.
  • Validation on the PRISM drug-response dataset showed 13 SL pairs with significant drug response differences, with some remaining significant after multiple testing correction.
  • SLIDR recovered a significant fraction of experimentally identified SL interactions from both in vitro and CRISPR screens with high accuracy (~93%).
  • ARID1A-TEAD1 and AXIN1-URI1 were experimentally validated as SL pairs in hepatocellular carcinoma, demonstrating SLIDR's ability to predict actionable targets.

Limitations and Future Work

  • Data Integration: SLIDR currently focuses on perturbation screening data, copy number, and mutation data. The authors acknowledge that integrating expression and pathway data, as used by other methods, could further improve performance.
  • WT p-value Threshold Sensitivity: The method's performance can be sensitive to the p-value threshold used for filtering wild-type cell lines, suggesting a need for optimization or alternative filtering strategies.
  • Confounding Driver Mutations: While causal inference is used to address multiple mutations, the current approach can be prone to selection bias and may be underpowered in sparse mutation data. Future work could explore using other omics data to improve confounder selection.
  • Experimental Validation: While two HCC SL pairs were validated, extensive further experimental validation is needed for all predicted SL pairs.
  • Broader Application: The authors suggest extending SLIDR to incorporate multi-omics data and exploring its use in tandem with other methods for enhanced predictive power.

Summary updated

Abstract

The development of cancer therapies is limited by the availability of suitable drug targets. Potential candidate drug targets can be identified based on the concept of synthetic lethality (SL), which refers to pairs of genes for which an aberration in either gene alone is non-lethal, but co-occurrence of the aberrations is lethal to the cell. Here, we present SLIdR (Synthetic Lethal Identification in R), a statistical framework for identifying SL pairs from large-scale perturbation screens. SLIdR successfully predicts SL pairs even with small sample sizes while minimizing the number of false positive targets. We apply SLIdR to Project DRIVE data and find both established and potential pan-cancer and cancer type-specific SL pairs consistent with findings from literature and drug response screening data. We experimentally validate two predicted SL interactions (ARID1A-TEAD1 and AXIN1-URI1) in hepatocellular carcinoma, thus corroborating the ability of SLIdR to identify potential drug targets.
BibTeX
@inproceedings{srivatsa2022discovery,
  author = {Sumana Srivatsa and Hesam Montazeri and Gaia Bianco and Mairene Coto‐Llerena and Mattia Marinucci and Charlotte K.Y. Ng and Salvatore Piscuoglio and Niko Beerenwinkel},
  title = {{Discovery of synthetic lethal interactions from large-scale pan-cancer perturbation screens}},
  booktitle = {Nature Communications},
  year = {2022},
  doi = {10.1038/s41467-022-35378-z},
}
Figure