Rapid, economical diagnostic classification of ATRT molecular subgroup using NanoString nCounter platform

Beng‐Choon Ho, Anthony Arnoldo, Yunan Zhong, Mei‐Yeh Jade Lu, James A. Torchia, Fangzhou Yao, Cynthia Hawkins, Annie HuangNeuro-Oncology 2024 · DOI

TL;DR

This paper introduces "AClass," a rapid and cost-effective assay for classifying molecular subgroups of atypical teratoid rhabdoid tumors (ATRTs) using the NanoString nCounter platform. This new method achieves high accuracy and offers a practical solution for clinical settings, addressing the current challenges in ATRT subgrouping.

What Problem Does This Paper Address?

Atypical teratoid rhabdoid tumors (ATRTs) are aggressive pediatric brain tumors that, despite their genomic simplicity, exhibit significant molecular and clinical heterogeneity. Recent research has identified at least three major ATRT subgroups (ATRT-SHH, ATRT-TYR, and ATRT-MYC), but reliably classifying these subgroups in a clinical setting remains a significant challenge. Current methods, such as DNA methylation or transcriptomic array-based classifications, are often expensive, time-consuming, require large sample sizes, and involve complex sample preparation. This lack of a standardized, accessible, and accurate classification method hinders personalized treatment strategies and the development of targeted therapies.

Methods

The researchers developed a novel ATRT subgroup predictor assay by combining a custom gene panel, "CodeSet30," with a flexible machine learning classifier package.

  • CodeSet30 Panel Design: They identified 30 highly differentially expressed genes (≥3-fold change, adj. P < .005) from a previously published Illumina HT12 v4 array dataset of 90 ATRTs. These genes were selected to be representative of the three ATRT subgroups and to minimize platform bias. Six housekeeping genes were included for normalization.
  • Ensemble Machine Learning Classifier (AClass): To build the classifier, they employed an ensemble strategy, integrating predictions from eight different algorithms (glmnet, knn, rf, pam, nb, rpart, svmLinear, Ida). This approach aims to improve accuracy and robustness by aggregating multiple predictions, thus mitigating the biases and variances of single algorithms. The final classifier selects the subgroup favored by the majority of the 55 integrated models.
  • NanoString nCounter Platform: The assay utilizes the NanoString nCounter platform, known for its efficiency and ability to work with limited RNA input and formalin-fixed paraffin-embedded (FFPE) samples.

Datasets

  • Training and Initial Validation: The primary dataset used for developing the CodeSet30 panel and training the AClass classifier was an Illumina HT12 v4 array dataset of 90 primary ATRTs from Torchia et al. (2015). This dataset was split into training (60%) and testing (40%) cohorts for model development and validation.
  • Independent Validation: The AClass classifier was further validated using an independent gene expression array dataset from 112 ATRT samples.
  • Extra-cranial Rhabdoid Tumors (MRTs): 14 MRT samples were analyzed to assess the classifier's applicability to related tumor types.
  • Comparison Datasets: The study also compared its results with DNA methylation classification data from Capper et al. (2018) and a previously published NanoString gene set from Leruste et al.

Key Insights and Ideas

  • The Power of Ensemble Learning: The study highlights how combining multiple machine learning algorithms (ensemble learning) can significantly improve classification accuracy and robustness, especially when dealing with complex biological data and limited sample sizes.
  • NanoString as a Clinical Tool: The research demonstrates the NanoString nCounter platform's suitability for rapid and cost-effective molecular diagnostics in oncology, particularly for FFPE samples, which are common in clinical archives.
  • CodeSet30's Robustness: The identified 30-gene signature (CodeSet30) proved to be a stable and reliable marker for ATRT subgroup classification across different platforms (gene expression array, RNASeq, and NanoString).
  • MYC Subgroup Heterogeneity: The study suggests that the MYC subgroup, which encompasses a broader range of tumor locations and potentially more unstable genomes, is inherently more heterogeneous, leading to less distinct signature genes and potentially more misclassifications.

Key Results

  • High Accuracy: The AClass classifier achieved an average accuracy of 93.6% on the validation dataset. For high-prediction score samples, the accuracy reached an impressive 98.9%.
  • Concordance with DNA Methylation: The AClass classifier showed high agreement with DNA methylation classification, with 84.5% concordance overall and up to 95.8% concordance for high-confidence predictions.
  • Rapid and Economical: The NanoString workflow is significantly faster (days vs. weeks for DNA methylation) and more cost-effective (approximately one-tenth the cost) than traditional DNA methylation array methods.
  • FFPE Sample Compatibility: The assay is effective for both fresh frozen and archival FFPE samples, a crucial advantage for clinical applications.
  • MRT Classification: All analyzed extra-cranial rhabdoid tumors (MRTs) were classified as the MYC subgroup by AClass.

Limitations and Future Work

  • ATRT Specificity: The current AClass workflow is specifically designed for ATRT classification and does not infer copy number variants or differentiate ATRT from other tumor entities.
  • MYC Subgroup Definition: The study acknowledges that the MYC subgroup is less well-defined due to its heterogeneity, which can impact classification specificity.
  • Sample Quality Impact: While the assay is robust, sample quality (especially RNA integrity in FFPE samples) can influence prediction scores and agreement. The authors suggest that adjustments to acceptance criteria for FFPE samples might be necessary.
  • Future Directions: The authors suggest that the CodeSet30 panel could be expanded with additional samples to improve its performance and potentially differentiate ATRT from other tumor types. Further validation of the prognostic value of ATRT molecular subgroups in clinical trials is also warranted.

Summary updated

Abstract

Here we present a rapid, cost-effective, and accurate ATRT subgrouping assay applicable for clinical use.
BibTeX
@inproceedings{ho2024rapid,
  author = {Beng‐Choon Ho and Anthony Arnoldo and Yunan Zhong and Mei‐Yeh Jade Lu and James A. Torchia and Fangzhou Yao and Cynthia Hawkins and Annie Huang},
  title = {{Rapid, economical diagnostic classification of ATRT molecular subgroup using NanoString nCounter platform}},
  booktitle = {Neuro-Oncology},
  year = {2024},
  doi = {10.1093/noajnl/vdae004},
}
Figure