[Paper Review] A Noise-Filtering Approach for Cancer Drug Sensitivity Prediction

Turki Turki, Zhi Wei|arXiv (Cornell University)|Dec 2, 2016

Computational Drug Discovery Methods12 references3 citations

TL;DR

This paper proposes a noise-filtering approach for cancer drug sensitivity prediction that improves machine learning performance by identifying and removing noisy cell lines using numerical linear algebra and information retrieval techniques. By filtering out low-quality samples before model training, the method achieves significantly higher AUC scores—up to 0.878—compared to baseline methods, demonstrating superior stability and predictive accuracy on clinical trial data for breast cancer and multiple myeloma.

ABSTRACT

Accurately predicting drug responses to cancer is an important problem hindering oncologists' efforts to find the most effective drugs to treat cancer, which is a core goal in precision medicine. The scientific community has focused on improving this prediction based on genomic, epigenomic, and proteomic datasets measured in human cancer cell lines. Real-world cancer cell lines contain noise, which degrades the performance of machine learning algorithms. This problem is rarely addressed in the existing approaches. In this paper, we present a noise-filtering approach that integrates techniques from numerical linear algebra and information retrieval targeted at filtering out noisy cancer cell lines. By filtering out noisy cancer cell lines, we can train machine learning algorithms on better quality cancer cell lines. We evaluate the performance of our approach and compare it with an existing approach using the Area Under the ROC Curve (AUC) on clinical trial data. The experimental results show that our proposed approach is stable and also yields the highest AUC at a statistically significant level.

Motivation & Objective

• To address the degradation of machine learning performance in cancer drug sensitivity prediction caused by noisy or low-quality cancer cell lines.
• To improve the accuracy and stability of predictive models by filtering out noisy samples before training.
• To integrate numerical linear algebra and information retrieval techniques to identify high-quality cell lines from genomic and drug response data.
• To evaluate the proposed method against existing approaches using real clinical trial data for breast cancer and multiple myeloma.

Proposed method

• Construct a distance matrix D using the Manhattan distance between all pairs of cell line feature vectors in the training set.
• Apply spectral decomposition to D to obtain eigenvectors and eigenvalues, using the smallest eigenvector to project the data into a noise-filtered feature space.
• Transform the original training input into a new representation using the dominant eigenvectors of the distance matrix to emphasize low-noise features.
• Use cosine similarity between original and transformed feature vectors to rank and retrieve the highest-quality cell lines based on minimum angular deviation.
• Train machine learning models—specifically support vector regression and ridge regression—on the filtered, high-quality subset of cell lines.
• Apply the trained models to predict drug sensitivity on test sets, evaluating performance using AUC on clinical trial outcomes.

Experimental results

Research questions

• RQ1 Can filtering out noisy cancer cell lines improve the predictive performance of machine learning models in drug sensitivity prediction?
• RQ2 How does the integration of numerical linear algebra and information retrieval techniques enhance the identification of high-quality cell lines?
• RQ3 Does the proposed noise-filtering approach yield more stable and accurate predictions than existing baseline methods on real clinical data?
• RQ4 To what extent does the removal of low-quality samples improve AUC in predicting sensitivity to docetaxel in breast cancer and bortezomib in multiple myeloma?

Key findings

• The proposed method achieved an AUC of 0.878 for docetaxel sensitivity prediction in breast cancer, significantly outperforming the baseline B+SVR+S (AUC 0.857) and B+RR (AUC 0.821).
• For bortezomib sensitivity in multiple myeloma, the PA+SVR+S model achieved an AUC of 0.658, surpassing the baseline B+SVR+S (AUC 0.602) and B+RR (AUC 0.608).
• The mean AUC (MAUC) across runs was 0.871 for PA+SVR+S in breast cancer, compared to 0.858 for the baseline B+SVR+S, indicating consistent superiority.
• The PA+SVR+S model showed highly significant p-values (p = 117 × 10⁻⁵) in distinguishing sensitive vs. resistant patients in breast cancer, confirming statistical reliability.
• The method reduced the training set size by filtering out low-quality cell lines, with q+PA values ranging from 459 to 478, while maintaining or improving predictive performance.
• The t-test p-values for predicted sensitivity differences between responders and non-responders were highly significant (p < 0.0001) for all proposed models, confirming robustness.