Predicting combinative drug pairs towards realistic screening via integrating heterogeneous features

Jian-Yu Shi, Jia-Xin Li, Ke Gao, Peng Lei, Siu-Ming YiuBMC Bioinformatics 2017 · DOI

TL;DR

This paper tackles the challenge of predicting effective drug combinations for complex diseases by developing a novel approach that integrates diverse drug features. Their method not only improves prediction accuracy but also uniquely enables the screening of potential combinations involving new drugs that haven't been previously studied together.

What Problem Does This Paper Address?

Treating complex diseases often requires combining multiple drugs, as single-drug therapies can be insufficient. However, experimentally screening all possible drug combinations is prohibitively expensive and time-consuming. Computational approaches aim to accelerate this process, but existing methods face two main hurdles:

  1. Feature Integration: They often struggle to effectively combine various types of information (heterogeneous features) about drugs, leading to inefficient training and overfitting.
  2. Realistic Screening: Most current methods focus on predicting combinations among drugs that already have known pairings. This overlooks the crucial need to identify potential combinations for "new" drugs that have no prior combination data, which is essential for drug discovery pipelines.

Methods

The authors propose a "drug-driven" approach that focuses on drug attributes rather than disease specifics. Their method involves:

  • Feature Extraction: They extract four types of drug features from diverse sources:
  • Drug-Drug Interactions (DDI): Pharmaceutical incompatibilities between drugs, characterized using Singular Value Decomposition (SVD).
  • Anatomical Therapeutic Chemical (ATC) Classification: Semantic similarity based on drug classification codes, calculated using the Tanimoto coefficient.
  • Drug-Target Interactions (DTI): Information about which proteins a drug targets, represented as binary feature vectors.
  • Side Effects (SE): Occurrence of beneficial or adverse side effects, also represented as binary vectors.
  • Dimensionality Reduction: To combat the computational issues of high-dimensional features, they apply PCA to DTI and SE features, and select a subset of SVs for DDI features.
  • Two-Step Fusion Scheme: Instead of simple concatenation, they propose a novel fusion strategy:
  • Individual Classifiers: Each of the DDI, DTI, and SE features is used to train a separate logistic regression classifier.
  • Score Integration: The confidence scores from these classifiers are then combined with the ATC similarity score. The final confidence score is an average of these four values. This "average rule" fusion is shown to be effective.
  • Realistic Cross-Validation: They design three specific cross-validation strategies (S1, S2, S3) to mimic realistic screening scenarios:
  • S1: Predicting new combinations among known drugs.
  • S2: Predicting combinations between known and new drugs.
  • S3: Predicting combinations exclusively among new drugs.

Datasets

The study utilizes a benchmark dataset collected from the Drug Combination Database (DCDB) and the FDA orange book.

  • Initial Dataset: 245 drugs with 239 approved drug pairs.
  • Final Dataset for Analysis: After filtering for drugs with all four feature types (DDI, ATC, DTI, SE), the dataset comprised 159 drugs and 1904 drug pairs (132 positive, 1772 negative).

Key Insights and Ideas

  • SE Features are Highly Informative: Side effect data proved to be the most predictive individual feature for drug combinations.
  • Fusion Outperforms Individual Features: Combining multiple heterogeneous features significantly boosts prediction performance compared to using any single feature type.
  • DDI and ATC Provide Complementary Information: DDI data suggests drugs in a combination are often not directly interacting but are "close" in the interaction network, while ATC similarity indicates a tendency for drugs to act on the same anatomical part.
  • DTI is Less Predictive Individually: While DTI data is considered, it showed the weakest performance as a standalone feature, suggesting that simply sharing targets isn't a strong indicator of combinability.
  • The "Average Rule" is Robust: A simple average fusion scheme, despite being an equal weighting method, achieved near-optimal performance and is practical for integrating diverse features.
  • Novelty in "New Drug" Prediction: The most significant contribution is the development of a framework and validation strategies (S2 and S3) that can predict combinations involving drugs with no prior combination history, a critical gap in existing methods.

Key Results

The proposed fusion scheme achieved superior performance across all three realistic screening scenarios:

  • S1 (Known Drugs): AUC of 0.954, AUPR of 0.821.
  • S2 (Known + New Drugs): AUC of 0.909, AUPR of 0.635.
  • S3 (New Drugs): AUC of 0.809, AUPR of 0.592.

These results demonstrate the effectiveness of their fusion method and its ability to generalize to scenarios involving novel drug combinations.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that while their approach is promising, it focuses on pairwise drug combinations.
  • The DTI feature, while included, showed weaker individual predictive power.
  • Future Work:
    • The authors suggest extending their technique to predict combinations of more than two drugs.
  • Further exploration into optimizing feature weighting beyond the simple average rule could potentially yield even better results.

Summary updated

BibTeX
@article{shi2017predicting,
  author = {Jian-Yu Shi and Jia-Xin Li and Ke Gao and Peng Lei and Siu-Ming Yiu},
  title = {{Predicting combinative drug pairs towards realistic screening via integrating heterogeneous features}},
  journal = {BMC Bioinformatics},
  year = {2017},
  doi = {10.1186/s12859-017-1818-2},
}
Figure