TL;DR
This paper introduces HeteSim-based Small Molecule-miRNA Association Prediction (HSSMMA), a novel computational model that leverages a heterogeneous network and a path-based measurement method to predict potential interactions between small molecules (SMs) and microRNAs (miRNAs). The model demonstrates high accuracy in predicting these associations, offering a valuable tool for drug discovery and disease therapy.
What Problem Does This Paper Address?
Targeting microRNAs (miRNAs) with small molecules (SMs) is an emerging therapeutic strategy for various complex human diseases. However, experimentally identifying these potential miRNA-SM associations is a time-consuming and costly process. There's a critical need for efficient and accurate computational methods to predict these interactions, thereby accelerating pharmaceutical engineering and disease treatment. Existing computational approaches often rely on functional similarity of differentially expressed genes or network-based methods, but this paper aims to improve prediction by incorporating a more nuanced understanding of relationships within a heterogeneous network.
Methods
The core of this work is the HeteSim-based Small Molecule-miRNA Association Prediction (HSSMMA) model. It constructs a heterogeneous network that integrates three key components:
- Known miRNA-SM associations: This forms the foundation of known interactions.
- Integrated miRNA similarity: This captures relationships between different miRNAs based on factors like gene functional consistency and disease phenotype similarity.
- Integrated SM similarity: Similarly, this captures relationships between different SMs based on side effect similarity, chemical structure similarity, and disease phenotype similarity.
HSSMMA then employs the HeteSim algorithm, a path-based measurement method, to calculate the relatedness between SMs and miRNAs within this heterogeneous network. HeteSim considers various "search paths" (sequences of node types) connecting an SM to a miRNA, capturing semantic information embedded in these paths. The model considers paths with lengths up to four, integrating scores from different path types to generate a final association probability.
Datasets
The study utilized the following datasets:
- SM2miR database: This was the primary source for experimentally verified miRNA-SM associations.
- Dataset 1: This dataset included 541 miRNAs, 831 SMs, and 664 known miRNA-SM associations. Notably, not all miRNAs and SMs in this dataset participated in the known associations, allowing for the evaluation of predictions for less-represented entities.
- Dataset 2: This dataset included 39 SMs and 286 miRNAs, with all participating entities involved in the known 664 associations. This dataset was used to assess the model's performance when all entities have known associations.
The paper also references other databases for similarity calculations, including SIDe Effect Resource (SIDER), DrugBank, Therapeutic Targets Database (TTD), TargetScan, Human MicroRNA Disease Database (HMDD), miR2Disease, and PhenomiR.
Key Insights and Ideas
- The power of heterogeneous networks: By integrating diverse data sources (known interactions, SM similarity, miRNA similarity) into a single network, the model can capture complex relationships that might be missed by single-source approaches.
- Path semantics matter: The HeteSim algorithm's ability to analyze "paths" through the network, rather than just direct connections, allows it to infer more nuanced and semantically rich associations. This is akin to understanding how two entities are related through a series of intermediate steps and their types.
- Predicting for "new" entities: The study successfully tested the model's ability to predict associations for SMs that had no prior known miRNA partners, demonstrating its utility beyond known interactions.
- Validation is key: The rigorous cross-validation and case study approach, including confirmation against published literature, lends significant credibility to the model's predictions.
Key Results
The HSSMMA model demonstrated strong predictive performance across various evaluation metrics:
- Global LOOCV: Achieved AUCs of 0.9913 (Dataset 1) and 0.7401 (Dataset 2).
- miRNA-fixed local LOOCV: Achieved AUCs of 0.9902 (Dataset 1) and 0.8466 (Dataset 2).
- SM-fixed local LOOCV: Achieved AUCs of 0.7989 (Dataset 1) and 0.6149 (Dataset 2).
- 5-fold cross-validation: Achieved average AUCs of 0.9910 ± 0.0004 (Dataset 1) and 0.7451 ± 0.0054 (Dataset 2).
In case studies, the model successfully predicted associations that were later confirmed by published literature:
- 2 out of the top 10 predicted associations were confirmed.
- 13 out of the top 50 predicted associations were confirmed.
For new SMs without known miRNA associations, the model predicted associations for 5-fluorouracil (5-FU), 17ẞ-estradiol (E2), and 5-Aza-2'-deoxycytidine (5-Aza-CdR) with a high proportion of confirmed predictions (e.g., 27 of the top 50 for 5-FU).
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Dataset size: The number of experimentally confirmed miRNA-SM associations (664) is considered insufficient. More verified associations would likely improve prediction accuracy.
- Integration of HeteSim scores: The current method for combining HeteSim scores from different paths is relatively simple. Future work could explore more advanced machine-learning techniques for this integration.
- Prediction accuracy: While significantly improved over previous models, the prediction accuracy is still not perfect, motivating the development of even more effective computational models.
- Future research: The authors suggest that future research should focus on developing more advanced computational models that leverage reliable biological datasets to further enhance prediction capabilities.
Summary updated