Predicting potential miRNA-disease associations based on more reliable negative sample selection

Ruiyu Guo, Hailin Chen, Wengang Wang, Guangsheng Wu, Fangliang LvBMC Bioinformatics 2022 · DOI

TL;DR

This paper introduces KR-NSSM, a novel computational method for identifying potential miRNA-disease associations. It addresses the critical challenge of selecting reliable negative samples, which are crucial for supervised machine learning but often unavailable due to lack of experimental validation. KR-NSSM significantly improves prediction accuracy by employing a two-stage semi-supervised approach for more robust negative sample selection.

What Problem Does This Paper Address?

MicroRNAs (miRNAs) are small non-coding RNA molecules that play vital roles in regulating gene expression and are implicated in numerous human diseases. Identifying which miRNAs are associated with specific diseases is crucial for understanding disease mechanisms and developing diagnostic tools. Computational methods, particularly those using supervised machine learning, are increasingly used to predict these associations due to the time and cost of experimental validation.

A fundamental requirement for supervised learning is the availability of both positive (known associations) and negative (known non-associations) samples. While experimentally confirmed miRNA-disease associations (positive samples) are available, experimentally validated non-associations (negative samples) are scarce because they are not typically the focus of research. This scarcity forces researchers to rely on unlabelled data to infer negative samples. Existing methods often randomly select these negative samples or use basic clustering techniques like K-means, which can introduce noise and lead to less reliable predictions. This paper tackles the problem of selecting more accurate and reliable negative samples from unlabelled miRNA-disease pairs to improve the performance of miRNA-disease association prediction models.

Methods

The proposed method, KR-NSSM, employs a two-stage semi-supervised approach for negative sample selection:

  1. SS-Kmeans (Semi-Supervised K-means): This is an improved K-means algorithm that leverages both labelled (experimentally confirmed positive associations) and unlabelled data. It first calculates centroids for the positive sample set and the unlabelled sample set. Then, it uses cosine similarity to classify unlabelled samples into "likely positive" and "likely negative" sets. This step is iterated until stable centroids are achieved.
  2. Rocchio Classification: This algorithm is used for further refinement of the samples identified by SS-Kmeans. It aims to generate more robust prototype vectors representing positive and negative samples.
  3. Rocchio1: Uses the positive samples and the "likely negative" samples from SS-Kmeans to create prototype vectors. Samples in the "likely negative" set are then classified as either "reliable negative" or "reliable positive" based on their similarity to these prototypes.
  4. Rocchio2: Addresses potential errors from Rocchio1 by further dividing the "reliable negative" samples into subsets using K-means. For each subset, new prototype vectors are calculated, and samples are re-classified based on their similarity to these refined prototypes.

The method also constructs feature vectors for miRNA-disease pairs by concatenating miRNA similarity scores and disease similarity scores.

Datasets

The study utilizes the following datasets:

  • Benchmark Dataset: Downloaded from reference [26], containing 5430 experimentally confirmed miRNA-disease associations. These are treated as positive samples.
  • miRNA-miRNA Similarities: Computed from reference [40].
  • Disease-Disease Similarities: Calculated based on semantic values from the MeSH database.
  • Validation Databases: HMDD V3.2 [34] and dbDEMC [35] were used to validate the predicted positive miRNA-disease associations.

No new benchmark datasets were introduced; rather, existing, well-established datasets were used for training and evaluation.

Key Insights and Ideas

  • The critical role of negative sample quality: The paper strongly emphasizes that the accuracy of supervised learning models for miRNA-disease association prediction is highly dependent on the quality of the negative samples used. Noisy or poorly selected negative samples can significantly degrade performance.
  • Leveraging semi-supervised learning for unlabelled data: KR-NSSM effectively bridges the gap of unavailable negative samples by employing semi-supervised techniques that can learn from both labelled and unlabelled data. This is a powerful strategy when ground truth for negative cases is scarce.
  • Two-stage refinement is better than one: The combination of SS-Kmeans for initial screening and Rocchio classification for further refinement proved to be more effective than using either method alone, leading to more reliable negative (and positive) sample sets.
  • Beyond negative samples: Identifying positive associations: The KR-NSSM framework also demonstrated an ability to identify reliable positive miRNA-disease associations, with a significant portion being confirmed by existing databases.

Key Results

  • Improved prediction accuracy: Experiments showed that using negative samples selected by KR-NSSM significantly improved prediction accuracy across six classic classifiers (lightGBM, SVM, Random Forest, LR, XGBoost, MLP) and five existing miRNA-disease association prediction models compared to traditional random selection or K-means-based strategies.
  • Ablation study: An ablation test indicated that integrating both SS-Kmeans and Rocchio classification in KR-NSSM yielded better prediction performance than using either method individually. For instance, using KR-NSSM achieved an AUC of 0.9763, compared to 0.9712 for SS-Kmeans alone and 0.9652 for Rocchio classification alone.
  • Identification of reliable positive associations: KR-NSSM identified 1123 potential positive miRNA-disease associations, of which 469 were confirmed by established databases (HMDD V3.2 and dbDEMC).

Limitations and Future Work

  • Limitations: The paper acknowledges that while KR-NSSM improves negative sample selection, the inherent lack of experimentally validated negative samples means there's always a potential for some noise or misclassification. The performance is also dependent on the quality of the initial similarity scores used for feature construction.
  • Future Work: The authors suggest that KR-NSSM's approach to reliable negative sample selection could be widely applied to other association prediction tasks in bioinformatics, such as drug-target, drug-disease, and lncRNA-disease associations, where similar challenges with negative sample availability exist. Further exploration of different semi-supervised learning algorithms and feature engineering techniques could also enhance performance.

Summary updated

BibTeX
@article{guo2022predicting,
  author = {Ruiyu Guo and Hailin Chen and Wengang Wang and Guangsheng Wu and Fangliang Lv},
  title = {{Predicting potential miRNA-disease associations based on more reliable negative sample selection}},
  journal = {BMC Bioinformatics},
  year = {2022},
  doi = {10.1186/s12859-022-04978-3},
}
Figure