Evolution‐informed modeling improves outcome prediction for cancers

Li Liu, Yung Chang, Tao Yang, David P. Noren, Byron L. Long, Steven M. Kornblau, Amina A. Qutub, Jieping YeEvolutionary 2016 · DOI

TL;DR

This paper introduces a novel computational method that leverages evolutionary conservation to identify more reliable cancer biomarkers. By combining evolutionary insights with statistical associations, their approach significantly improves prediction accuracy for cancer treatment outcomes and metastasis, while also reducing model complexity.

What Problem Does This Paper Address?

Despite advances in high-throughput "omics" technologies, biomarkers discovered for cancer prediction often perform poorly in real-world applications. This is largely because these methods tend to rely heavily on statistical associations, which can lead to overfitting and the inclusion of spurious markers that lack true biological relevance. Cancer is fundamentally an evolutionary disease, yet evolutionary principles have been largely overlooked in biomarker discovery. The authors aim to bridge this gap by integrating evolutionary conservation as a crucial prior knowledge source to filter for functionally important genes, thereby improving the robustness and interpretability of predictive models.

Methods

The core of the paper is a novel "evolution-informed modeling" framework. The key steps are:

  1. Composite Weighting Schema: Each feature (gene, protein, or clinical covariate) is assigned a composite weight (W) that combines two components:
  2. Evolutionary Weight (W<sub>E</sub>): This is the reciprocal of the evolutionary rate (R) of a gene/protein, reflecting how conserved it is across species. Higher conservation implies higher functional importance.
  3. Statistical Weight (W<sub>S</sub>): This is derived from the statistical significance (e.g., p-value from a t-test) of the feature's association with the clinical outcome.
  4. Clinical covariates are given a maximum evolutionary weight, acknowledging their known predictive power.
  5. Feature Selection:
    • Features are normalized (z-scores).
  6. The composite weights are applied to the normalized feature matrix, effectively prioritizing evolutionarily conserved and statistically significant features.
  7. L<sub>1</sub>-norm regularized logistic regression is used for feature selection. This method inherently promotes sparsity, selecting a subset of features.
  8. Stability selection (using 100 bootstraps) is employed to ensure robustness and reduce dependence on the regularization parameter. Features consistently selected across multiple runs are deemed informative.
  9. Classification:
    • A random forest classifier is trained using the unweighted values of the selected features to avoid bias introduced by the weighting scheme during prediction.
  10. To handle class imbalance, a bootstrapping approach is used, creating an ensemble of 100 submodels. Each submodel is trained on a bootstrapped subset of the data.
  11. Prediction: For an unknown sample, the ensemble of submodels generates predictions, and a confidence score is calculated based on the percentage of submodels predicting the positive class.

Datasets

The method was evaluated on two publicly available cancer datasets:

  • Acute Myeloid Leukemia (AML) Dataset: This dataset was originally used for the DREAM 9 AML Challenge. It includes 40 clinical covariates and proteomic data (231 total or phosphorylated proteins) from 291 newly diagnosed AML patients. Outcomes were recorded as complete response or resistance to induction therapy. The authors used 191 samples for training and 100 for blind testing.
  • Prostate Cancer Dataset: This dataset was downloaded from NCBI GEO (accession number: GSE10645). It consists of gene expression data from 401 prostate cancer patients who underwent prostatectomy. The outcome was metastatic recurrence within 5 years.

Key Insights and Ideas

  • Evolutionary Conservation as a Biological Filter: The core idea is that nature's "test" of evolutionary selection on genes provides a powerful filter for identifying functionally important elements, complementing purely statistical approaches.
  • Reducing Overfitting and Spurious Markers: By prioritizing conserved genes, the method effectively reduces the number of features needed for accurate prediction, thereby mitigating the risk of overfitting common in high-dimensional "omics" data.
  • Improved Interpretability: The identified biomarkers are more likely to be biologically meaningful, leading to more interpretable models and potentially uncovering novel drug targets.
  • Synergy of Evolutionary and Statistical Information: The composite weighting schema demonstrates that combining evolutionary conservation with statistical association yields superior results compared to using either alone.
  • Robustness to Data Noise: The AML dataset analysis revealed that statistical associations alone can be noisy, highlighting the benefit of evolutionary information in stabilizing predictions.

Key Results

  • AML Prediction: The evolution-informed model achieved the highest performance among 31 teams in the DREAM 9 AML Challenge, with a balanced accuracy of 77.9% and AUROC of 0.796. This was a significant improvement over models using only statistical weights.
  • Prostate Cancer Metastasis Prediction: The evolution-informed model showed a significant improvement in accuracy (4% higher balanced accuracy and AUROC) when using a small number of features (10-20) compared to models relying solely on statistical associations.
  • Model Complexity Reduction: The evolution-informed approach led to models with fewer features (e.g., an average of 30 features for AML with high accuracy, compared to 43 for a model without weights but lower accuracy), indicating a more parsimonious and potentially more generalizable solution.
  • Biomarker Significance: Identified biomarkers (e.g., PIK3CA and GSK3 in AML) have known implications in tumor progression and are potential drug targets, validating the biological relevance of the selected features.

Limitations and Future Work

  • Clinical Covariate Weighting: The paper acknowledges that assigning statistical weights to clinical covariates is a current limitation, as it doesn't fully capture rich domain knowledge. Future work could involve deriving scores from meta-analyses or other prior knowledge aggregation methods.
  • Integration of Diverse Data Types: The authors suggest that the framework can be extended to integrate various "omics" data (genomic, transcriptomic, proteomic) by applying evolutionary conservation at different molecular levels.
  • Subclonal Evolution: The paper touches upon the challenge of identifying driver subclones and proposes that evolutionary signatures could be integrated at both species and individual levels to address this.
  • Broader Application: The method's success in both a hematological cancer (AML) and a solid tumor (prostate cancer) suggests its broad applicability across different cancer types.

Summary updated

Abstract

Despite wide applications of high-throughput biotechnologies in cancer research, many biomarkers discovered by exploring large-scale omics data do not provide satisfactory performance when used to predict cancer treatment outcomes. This problem is partly due to the overlooking of functional implications of molecular markers. Here, we present a novel computational method that uses evolutionary conservation as prior knowledge to discover bona fide biomarkers. Evolutionary selection at the molecular level is nature's test on functional consequences of genetic elements. By prioritizing genes that show significant statistical association and high functional impact, our new method reduces the chances of including spurious markers in the predictive model. When applied to predicting therapeutic responses for patients with acute myeloid leukemia and to predicting metastasis for patients with prostate cancers, the new method gave rise to evolution-informed models that enjoyed low complexity and high accuracy. The identified genetic markers also have significant implications in tumor progression and embrace potential drug targets. Because evolutionary conservation can be estimated as a gene-specific, position-specific, or allele-specific parameter on the nucleotide level and on the protein level, this new method can be extended to apply to miscellaneous "omics" data to accelerate biomarker discoveries.
BibTeX
@inproceedings{liu2016evolutioninformed,
  author = {Li Liu and Yung Chang and Tao Yang and David P. Noren and Byron L. Long and Steven M. Kornblau and Amina A. Qutub and Jieping Ye},
  title = {{Evolution‐informed modeling improves outcome prediction for cancers}},
  booktitle = {Evolutionary},
  year = {2016},
  doi = {10.1111/eva.12417},
}
Figure