neoDL: a novel neoantigen intrinsic feature-based deep learning model identifies IDH wild-type glioblastomas with the longest survival

Ting Sun, Yufei He, Wendong Li, Guang Liu, Lin Li, Lu Wang, Zixuan Xiao, Xiaohan Han, Hao Wen, Yong Liu, Yifan Chen, Haoyu Wang, Jing Li, Yubo Fan, Wei Zhang, Jing ZhangBMC Bioinformatics 2021 · DOI

TL;DR

Researchers have developed "neoDL," a novel deep learning model that leverages intrinsic features of neoantigens to accurately stratify IDH wild-type glioblastomas (GBMs) into distinct survival groups. This advancement promises to improve personalized immunotherapy by identifying patients most likely to benefit and by highlighting key neoantigen characteristics that drive anti-tumor immune responses.

What Problem Does This Paper Address?

Glioblastoma (GBM) is a highly aggressive brain tumor with a poor prognosis, even with advanced treatments. While neoantigen-based immunotherapies have shown promise in other cancers, their effectiveness in IDH wild-type GBMs and the role of neoantigen intrinsic features in predicting prognosis have remained unclear. Existing methods often struggle with the limited number of high-quality neoantigens, hindering clinical application. This paper aims to bridge this gap by developing a model that uses deep learning to analyze neoantigen intrinsic features for better prognostic stratification in IDH wild-type GBMs.

Methods

The core of this research is the development of neoDL, a deep learning model. The process involved:

  • Neoantigen Feature Extraction: A comprehensive set of 2928 features was extracted for each neoantigen and its wild-type counterpart. These features encompassed physical-chemical properties, amino acid (AA) composition, AA descriptors at specific positions, and sequence-related characteristics.
  • Prognostic Feature Selection: Univariate Cox regression was used to identify "valid features" (189 in total) that were significantly associated with overall survival in IDH wild-type GBMs from the TCGA cohort.
  • Deep Learning Model Construction: A deep learning architecture was built using three hidden layers: two LSTM (Long Short-Term Memory) layers and one fully connected layer. This architecture was chosen for its ability to handle sequential data and avoid gradient vanishing issues. The model was trained using the selected valid features from the TCGA cohort.
  • Hierarchical K-Means Clustering: This was used to initially stratify patients into short-term and long-term survival groups, providing labels for training the deep learning model.
  • Validation: The model's performance was rigorously validated using both internal cross-validation (LOOCV) on the TCGA dataset and external validation on an independent Asian patient cohort (Pri cohort).

Datasets

The study utilized two primary datasets:

  • TCGA Cohort: This dataset, comprising 262 IDH wild-type GBM samples, was used for training and internal validation of the neoDL model. It included mutation data, gene expression data, and clinical information.
  • Pri Cohort: This independent dataset, consisting of 42 IDH wild-type GBM samples from an Asian population, served as the external validation set to assess the model's generalizability.

Key Insights and Ideas

  • Neoantigen Intrinsic Features Matter: Unlike tumor mutational burden or the sheer quantity of neoantigens, specific intrinsic features of neoantigens are crucial for predicting survival in IDH wild-type GBMs.
  • Deep Learning Unlocks Complex Patterns: Deep learning, particularly LSTM networks, proved highly effective in learning complex, high-level representations from the rich feature space of neoantigens, outperforming traditional methods.
  • Protective Features Identified: The study pinpointed 12 specific neoantigen intrinsic features (related to molecular weight, size, and electrostatic potential of dipeptides and tripeptides) that are consistently associated with longer survival across different cohorts.
  • Biological Pathways Revealed: Patients stratified into longer-survival groups by neoDL showed enrichment in biological processes related to "development" and "cell cycle," suggesting these pathways are linked to better outcomes in this GBM subtype.
  • Clinical Utility for Immunotherapy: The neoDL model has the potential to guide personalized immunotherapy by identifying GBM patients who are more likely to respond to neoantigen-based treatments.

Key Results

  • High Predictive Accuracy: The neoDL model achieved a remarkable AUC of 0.988 (p < 0.0001) for stratifying IDH wild-type GBMs in the TCGA cohort.
  • Successful External Validation: The model also demonstrated significant predictive power in the independent Pri cohort, with a p-value of 0.037.
  • Identification of 12 Protective Features: 12 features, primarily related to the physical-chemical properties of dipeptides and tripeptides at specific positions of neoantigens, were identified as protective factors (HR < 1) for survival.
  • Enrichment in Development and Cell Cycle: Gene Set Enrichment Analysis (GSEA) revealed that long-term survivors were enriched in GO terms related to development and cell cycle.
  • Superior Performance: The neoDL model outperformed other established deep learning models like DeepLearningModel and PASNet in predicting GBM patient survival.

Limitations and Future Work

  • Focus on Sequence Structure: The current study primarily focused on sequence structure and did not incorporate secondary or tertiary protein structure features. The authors suggest that integrating these could further improve the model's predictive power.
  • Generalizability Testing: While validated on two cohorts, the authors acknowledge the need for more extensive generalizability testing across diverse patient populations and clinical settings.
  • Therapeutic Exploitation: The paper highlights the potential for therapeutic exploitation of the identified prognostic features and the neoDL model itself, suggesting future work could involve designing neoantigen-based therapies informed by these findings.
  • Integration with Other Data: The authors propose that incorporating additional data modalities (e.g., medical images, gene expression) alongside neoantigen features could further enhance prognostic evaluation and decision-making in glioma.

Summary updated

BibTeX
@article{sun2021neodl,
  author = {Ting Sun and Yufei He and Wendong Li and Guang Liu and Lin Li and Lu Wang and Zixuan Xiao and Xiaohan Han and Hao Wen and Yong Liu and Yifan Chen and Haoyu Wang and Jing Li and Yubo Fan and Wei Zhang and Jing Zhang},
  title = {{neoDL: a novel neoantigen intrinsic feature-based deep learning model identifies IDH wild-type glioblastomas with the longest survival}},
  journal = {BMC Bioinformatics},
  year = {2021},
  doi = {10.1186/s12859-021-04301-6},
}
Figure