A Deep Learning Framework for Robust and Accurate Prediction of ncRNA-Protein Interactions Using Evolutionary Information

Hai-Cheng Yi, Zhu‐Hong You, De-Shuang Huang, Li Xiao, Tonghai Jiang, Liping LiMolecular 2018 · DOI

TL;DR

This paper introduces RPI-SAN, a novel deep learning framework that accurately predicts interactions between non-coding RNAs (ncRNAs) and proteins. By leveraging evolutionary information and a stacked auto-encoder architecture, RPI-SAN significantly outperforms existing methods, offering a more efficient and reliable computational tool for understanding crucial biological processes.

What Problem Does This Paper Address?

Non-coding RNAs (ncRNAs) are increasingly recognized for their vital roles in gene regulation and various biological processes, often by interacting with proteins. Identifying these ncRNA-protein interactions (ncRPIs) is fundamental to understanding their functions. However, experimental methods for discovering these interactions are typically expensive and time-consuming. While computational approaches exist, they often rely on hand-crafted features or struggle to capture the complex patterns within sequence data, leading to limitations in accuracy and robustness. This paper aims to bridge this gap by developing a more powerful and accurate computational method for predicting ncRPIs.

Methods

The core of the RPI-SAN method lies in its innovative combination of deep learning and ensemble techniques.

  • Feature Extraction:
    • RNA Sequences: ncRNA sequences are transformed into a k-mers sparse matrix (specifically, 4-mers) to capture nucleotide composition and order. Singular Value Decomposition (SVD) is then applied to extract a 256-feature vector.
  • Protein Sequences: Evolutionary information is captured using Position-Specific Scoring Matrices (PSSMs). Pseudo-Zernike Moments (PZMs) are then used to extract robust features from these PSSMs, resulting in a 240-feature vector.
  • Deep Learning Architecture (Stacked Auto-Encoder Network - SAN):
    • A stacked auto-encoder network is employed to automatically learn high-level, hierarchical features from the extracted RNA and protein sequence features. This deep learning approach avoids the need for manual feature engineering.
  • Prediction Model:
    • The learned high-level features from the SAN are fed into a Random Forest (RF) classifier to predict ncRPIs.
  • Stacked Assembling:
    • To further enhance robustness and accuracy, RPI-SAN utilizes a stacked assembling strategy. This involves integrating the outputs of multiple individual predictors (including variations of the SAN and other methods) into a final prediction layer, often a logistic regression model. This ensemble approach leverages the strengths of different predictors.

Datasets

The study utilized four benchmark datasets for training and evaluation:

  • RPI488: A dataset of ncRNA-protein interactions based on structure complexes.
  • RPI1807: A dataset of RNA-protein interactions.
  • RPI2241: Another dataset of RNA-protein interactions.
  • NPInter v2.0: A dataset of ncRNA-protein interactions derived from non-structure-based sources, serving as an independent test set.

The paper notes that RPI488 is smaller due to the limited availability of ncRNA-protein complexes in the Protein Data Bank.

Key Insights and Ideas

  • Deep Learning for Feature Learning: The study highlights the power of deep learning (specifically, stacked auto-encoders) in automatically learning complex, high-level features from raw sequence data, bypassing the limitations of manual feature engineering.
  • Evolutionary Information is Crucial: Incorporating evolutionary information via PSSMs and PZMs for protein sequences significantly improves the ability to capture relevant biological context for interaction prediction.
  • Ensemble Methods Boost Performance: Stacked assembling, by integrating multiple predictors, demonstrably improves prediction accuracy and robustness, especially on datasets with lower inter-predictor correlation.
  • Sequence-Based Prediction is Viable: The success of RPI-SAN underscores that sequence information alone, when processed effectively with advanced computational techniques, can be highly predictive of ncRPIs.

Key Results

  • Superior Performance: RPI-SAN achieved state-of-the-art performance across multiple benchmark datasets. Notably, it reported accuracies of 90.77% on RPI2241, 89.7% on RPI488, 96.1% on RPI1807, and 99.33% on the independent NPInter v2.0 dataset.
  • Outperforming Existing Methods: RPI-SAN consistently outperformed established prediction tools like RPI-Pred, IPMiner, RPISeq-RF, and lncPro in terms of accuracy, sensitivity, specificity, precision, and MCC. For instance, on the RPI2241 dataset, RPI-SAN improved accuracy by nearly 7% and specificity by over 16% compared to the best previous methods.
  • Effective Ensemble: The stacked assembling strategy in RPI-SAN provided a significant performance boost, achieving an AUC of 0.962 on the RPI2241 dataset, outperforming individual classifiers and simpler ensemble methods like averaging.
  • Case Study Validation: A case study on Homo sapiens interactions identified 15 potential ncRPIs with high probability, which were further verified by Gene Ontology, demonstrating the model's practical utility in discovering novel interactions.

Limitations and Future Work

  • Handling of Long Sequences: The paper mentions that some existing tools have limitations with very long sequences (e.g., RNAsubopt processing only the first 4,095 nucleotides). While RPI-SAN's k-mer approach is more general, the computational cost for extremely long sequences might still be a consideration.
  • Dataset Size and Correlation: The authors note that RPI-SAN shows a better effect on large-scale datasets than small datasets and plan to continue studying this aspect. They also observed that stacked assembling is more impactful on datasets with lower inter-predictor correlation.
  • Future Directions: The authors anticipate RPI-SAN can serve as a valuable computational tool for biomedical research, guiding the discovery of potential ncRPIs. Further research into its performance on larger and more diverse datasets is implied.

Summary updated

Abstract

The interactions between non-coding RNAs (ncRNAs) and proteins play an important role in many biological processes, and their biological functions are primarily achieved by binding with a variety of proteins. High-throughput biological techniques are used to identify protein molecules bound with specific ncRNA, but they are usually expensive and time consuming. Deep learning provides a powerful solution to computationally predict RNA-protein interactions. In this work, we propose the RPI-SAN model by using the deep-learning stacked auto-encoder network to mine the hidden high-level features from RNA and protein sequences and feed them into a random forest (RF) model to predict ncRNA binding proteins. Stacked assembling is further used to improve the accuracy of the proposed method. Four benchmark datasets, including RPI2241, RPI488, RPI1807, and NPInter v2.0, were employed for the unbiased evaluation of five established prediction tools: RPI-Pred, IPMiner, RPISeq-RF, lncPro, and RPI-SAN. The experimental results show that our RPI-SAN model achieves much better performance than other methods, with accuracies of 90.77%, 89.7%, 96.1%, and 99.33%, respectively. It is anticipated that RPI-SAN can be used as an effective computational tool for future biomedical researches and can accurately predict the potential ncRNA-protein interacted pairs, which provides reliable guidance for biological research.
BibTeX
@inproceedings{yi2018deep,
  author = {Hai-Cheng Yi and Zhu‐Hong You and De-Shuang Huang and Li Xiao and Tonghai Jiang and Liping Li},
  title = {{A Deep Learning Framework for Robust and Accurate Prediction of ncRNA-Protein Interactions Using Evolutionary Information}},
  booktitle = {Molecular},
  year = {2018},
  doi = {10.1016/j.omtn.2018.03.001},
}
Figure