MiRNA-Drug Resistance Association Prediction Through the Attentive Multimodal Graph Convolutional Network

Yanqing Niu, Congzhi Song, Yuchong Gong, Wen ZhangFrontiers in Pharmacology 2022 · DOI

In the high-stakes world of drug discovery, "resistance" is the ultimate spoiler. A drug might be designed perfectly to hit a target, but if the body’s internal regulatory machinery—specifically microRNAs (miRNAs)—interferes, the treatment fails. Predicting these associations is vital for personalized medicine, yet doing so in a lab is slow and expensive.

A recent paper published in Frontiers in Pharmacology introduces AMMGC (Attentive Multimodal Graph Convolutional Network), a deep learning framework designed to predict miRNA-drug resistance associations with high accuracy by looking at the problem from multiple biological "angles" simultaneously.

TL;DR

The authors developed AMMGC, a deep learning model that integrates diverse drug and miRNA features through four parallel graph convolution sub-networks. By using an attention mechanism to weigh these different data types, the model achieves state-of-the-art performance (AUC of 0.9467), outperforming previous methods in identifying which miRNAs are likely to drive drug resistance.

What Problem Does This Paper Address?

Drug resistance is a complex phenomenon often driven by miRNAs—small, non-coding RNA molecules that regulate gene expression. If an miRNA down-regulates a gene necessary for a drug to work, the patient becomes resistant.

The challenge for researchers is that the relationship between a drug and an miRNA isn't just a simple A-to-B link; it involves chemical structures, genetic expression profiles, and functional similarities. Previous computational models often relied on limited data types (like just drug fingerprints). The authors identified a need for a "multimodal" approach that could synthesize different types of biological and chemical information into a single predictive framework.

Methods

The core of the paper is the AMMGC architecture, which stands out due to two main features:

  1. Multimodal Graph Convolutional Networks (GCNs): Instead of one large network, the authors built four specialized sub-networks. Each sub-network processes a unique combination of features:
  2. Drug Features: PubChem substructure fingerprints (chemical structure) and SMILES label encoding (sequence-based representation).
  3. miRNA Features: Expression profiles (how much miRNA is present in different tissues) and Gene Ontology (GO) functional similarity.
  4. Attention Mechanism: Not all data is equally useful for every prediction. AMMGC employs an attention neural network that learns to assign "weights" to the outputs of the four sub-networks. This allows the model to prioritize the most relevant features for a specific drug-miRNA pair before making a final prediction via an inner product of the learned representations.

Datasets

The study utilized the ncDR database, a comprehensive resource for non-coding RNA-drug resistance associations. The specific dataset included:

  • 3,338 known associations between 754 miRNAs and 106 drugs.
  • Side information included 920-dimensional drug fingerprints and 2,587-dimensional miRNA GO similarity scores.

Key Insights and Ideas

  • Feature Synergy: The "Aha!" moment of the paper is that chemical fingerprints and SMILES encodings are complementary. While fingerprints describe the presence of chemical groups, SMILES encodings capture the sequence and arrangement. Combining them provides a much richer "latent representation" of the drug.
  • The Power of Attention: The ablation studies showed that the attention mechanism isn't just a bell or whistle—it significantly improves performance. The model learned that Modals 1 and 3 (which utilize miRNA expression profiles) generally carried more weight than those using GO similarity, suggesting expression data is a stronger predictor of resistance.
  • Biological Relevance: By removing biological features (the "w/o BF" test), the AUPR dropped by over 4%. This confirms that drug resistance prediction cannot rely on chemical informatics alone; the biological context of the miRNA is essential.

Key Results

AMMGC was evaluated using 5-fold cross-validation and compared against several baselines (like Collaborative Filtering and Label Propagation) and the previous state-of-the-art model, GCMDR.

  • Top-Tier Accuracy: AMMGC achieved an AUC of 0.9467 and an AUPR of 0.2399.
  • Superior Recall: It outperformed GCMDR by nearly 5% in Recall (REC), which is critical in drug discovery because it means the model is better at finding actual resistance links that shouldn't be missed.
  • Case Study Validation: To prove real-world utility, the authors used AMMGC to predict new associations. It successfully identified links like Oxaliplatin and hsa-mir-146a, which were later confirmed by independent PubMed literature (PMID: 26396533).

Limitations and Future Work

While AMMGC is powerful, the authors note a few areas for growth:

  • Data Sparsity: The model relies on "side information" (like GO similarity) which isn't available for every single miRNA. Currently, they use average values to fill these gaps, which could be refined.
  • Expanding the Graph: The current model focuses on a bipartite graph (Drug-miRNA). Future work could integrate a "tripartite" graph that includes miRNA-gene-drug interactions to better model the biological pathway of resistance.
  • Over-smoothing: The authors found that adding too many GCN layers (more than 2) actually decreased performance due to "over-smoothing," a common issue where node representations become too similar. Finding ways to deepen the network without losing feature distinctiveness remains a challenge.

Summary updated

Abstract

MiRNAs can regulate genes encoding specific proteins which are related to the efficacy of drugs, and predicting miRNA-drug resistance associations is of great importance. In this work, we propose an attentive multimodal graph convolution network method (AMMGC) to predict miRNA-drug resistance associations. AMMGC learns the latent representations of drugs and miRNAs from four graph convolution sub-networks with distinctive combinations of features. Then, an attention neural network is employed to obtain attentive representations of drugs and miRNAs, and miRNA-drug resistance associations are predicted by the inner product of learned attentive representations. The computational experiments show that AMMGC outperforms other state-of-the-art methods and baseline methods, achieving the AUPR score of 0.2399 and the AUC score of 0.9467. The analysis demonstrates that leveraging multiple features of drugs and miRNAs can make a contribution to the miRNA-drug resistance association prediction. The usefulness of AMMGC is further validated by case studies.
BibTeX
@inproceedings{niu2022mirna,
  author = {Yanqing Niu and Congzhi Song and Yuchong Gong and Wen Zhang},
  title = {{MiRNA-Drug Resistance Association Prediction Through the Attentive Multimodal Graph Convolutional Network}},
  booktitle = {Frontiers in Pharmacology},
  year = {2022},
  doi = {10.3389/fphar.2021.799108},
}
Figure