Multi-kernel feature extraction with dynamic fusion and downsampled residual feature embedding for predicting rice RNA N6-methyladenine sites

Mengya Liu, Zhan-Li Sun, Zhigang Zeng, Kin-Man LamBriefings in Bioinformatics 2025 · DOI

TL;DR

This paper introduces MFDm6ARice, a novel deep learning framework for accurately predicting N6-methyladenine (m6A) sites in rice RNA. By employing multi-kernel feature extraction, dynamic global-local fusion, and downsampled residual embedding, MFDm6ARice overcomes limitations of existing methods, achieving superior performance and demonstrating scalability to other plant species like maize.

What Problem Does This Paper Address?

RNA N6-methyladenosine (m6A) is a crucial epigenetic modification influencing plant growth, development, and stress responses. Accurate identification of m6A sites is vital for advancing precision rice breeding and understanding molecular mechanisms. However, existing computational methods for predicting m6A sites in rice face challenges:

  • Variable Sequence Lengths: RNA sequences have varying lengths, requiring methods to handle them. Current approaches often use padding to a fixed maximum length, which can lead to sparse and redundant features, reducing prediction accuracy.
  • Limited Feature Extraction: Many methods focus on local sequence information, potentially missing broader contextual patterns.
  • Computational Inefficiency: High-dimensional features resulting from padding can lead to computational burdens.
  • Early Stage of Rice-Specific Methods: While methods exist for other species, dedicated and advanced computational tools for rice m6A prediction are still developing.

This paper aims to address these issues by proposing an end-to-end deep learning framework that can effectively extract meaningful features from variable-length RNA sequences, fuse them intelligently, and maintain computational efficiency.

Methods

The proposed framework, MFDm6ARice, is an end-to-end deep learning model composed of several key modules:

  • Input Representation: RNA sequences are first converted into numerical representations using label encoding. To handle variable lengths, sequences are padded to a maximum length of 800 nucleotides. This ensures all sequence information is retained, though it can introduce sparsity.
  • Multi-Kernel Feature Fusion (MKFF) Module:
    • Multi-Kernel Feature Extraction: To combat sparsity and extract richer information, this module employs multiple convolutional kernels of different sizes (1, 3, and 5) to capture features at various receptive fields. This creates a set of multi-kernel features.
  • Global-Local Dynamic Fusion (GLDF): Instead of simple concatenation, GLDF intelligently fuses these multi-kernel features. It combines global channel dynamic fusion (aggregating information across channels) and local spatial dynamic fusion (focusing on spatial relationships) to create a more comprehensive and context-aware feature representation.
  • Downsampling Residual Feature Embedding (DRFE) Module: To address high dimensionality and computational inefficiency, this module uses a series of downsampling residual blocks. These blocks progressively reduce feature map resolution while retaining essential information through residual connections, leading to efficient feature compression and improved performance.
  • Contrastive Learning Loss (Lcl): An additional contrastive learning loss is incorporated to optimize the feature space, aiming to pull similar samples closer and push dissimilar ones further apart, enhancing feature discriminability.
  • Output Module: A fully connected layer with a sigmoid activation function predicts the probability of an m6A site. The model is trained using binary cross-entropy loss.

Datasets

The study utilizes the following datasets:

  • Benchmark Dataset: Derived from the first rice m6A dataset proposed by Wang et al. [29]. Positive samples are from m6A-seq peak sequences of Japonica Nipponbare seedling leaves (20-800 nt), and negative samples are of equal length from adjacent regions without m6A sites. CD-HIT was used for homology reduction. 80% of these data form the benchmark dataset for training and validation.
  • Same-Species Independent Test Set: The remaining 20% of the rice data from the benchmark dataset.
  • Cross-Species Independent Test Set: Human (H. sapiens) m6A data from DeepM6ASeq [26]. This set uses precise methylation sites with 101 nt-long sequences.
  • Maize m6A Dataset: Data from SMEP [29] was used to evaluate the model's scalability. This dataset contains variable-length peak sequences.

Key Insights and Ideas

  • Multi-kernel approach is crucial: Using multiple kernel sizes (1, 3, 5) significantly outperforms using a single kernel, indicating that capturing features at different scales is vital for m6A prediction.
  • Dynamic fusion beats simple concatenation: The proposed GLDF module, which intelligently fuses global and local spatial information, significantly outperforms simple concatenation or addition of features. This highlights the importance of context-aware feature integration.
  • Downsampling with residuals is effective: The DRFE module effectively reduces dimensionality and computational cost without sacrificing performance, demonstrating the power of residual connections in maintaining information flow during downsampling.
  • End-to-end learning is beneficial: The MFDm6ARice framework, being end-to-end, allows for joint optimization of all components, leading to better overall performance.
  • Scalability beyond rice: The model's successful application to maize data suggests its potential for broader use in plant epigenomics.

Key Results

  • Superior Performance: MFDm6ARice achieved state-of-the-art performance on the rice benchmark dataset, outperforming methods like SMEP and PEA-m6A in 5-fold cross-validation (5-CV). It achieved an ACC of 0.8321, MCC of 0.6225, AUC of 0.9038, and AUPR of 0.8201, showing an 11.81% improvement over the second-best method.
  • Robustness and Generalization: The model demonstrated strong performance on both same-species and cross-species independent test sets, indicating good generalization capabilities.
  • Maize Application: When applied to a maize m6A dataset, MFDm6ARice remained competitive, showcasing its scalability.
  • Ablation Studies: Removing any of the key components (MKFF, DRFE, or contrastive learning) led to a performance decrease, confirming the effectiveness and necessity of each module. The MKFF module was found to contribute more significantly than DRFE.
  • Visualization: t-SNE visualizations showed that raw padded sequences had poor discriminative power, while features processed by GLDF and DRFE exhibited improved clustering and separability between m6A and non-m6A sites.

Limitations and Future Work

  • Label Encoding Limitations: The paper acknowledges that label encoding with maximum length padding might overlook nucleotide chemical properties and specific sequence dependencies. Future work will explore alternative encoding methods (e.g., biochemical properties, embedding-based representations) to improve feature representation and interpretability.
  • Computational Demands: While DRFE improves efficiency, the model can still be computationally demanding, potentially limiting deployment in resource-constrained environments. Future work will focus on reducing storage and memory consumption through techniques like model pruning, quantization, and lightweight deployment frameworks.
  • Dataset Specificity: The current model is trained on rice data and tested on maize. Future research aims to develop a broader m6A prediction model that accommodates multiple crops or species.
  • Data Bias: The study notes potential biases in the dataset, such as under-representation of certain sequence types or environmental factors. Future work will address this by incorporating more diverse datasets or integrating multi-omics data.
  • Motif Discovery: The model successfully identified known and novel motifs related to stress response and growth in rice, suggesting a link between m6A modification and these biological processes. Further experimental validation of these motifs is warranted.

Summary updated

BibTeX
@article{liu2025multi,
  author = {Mengya Liu and Zhan-Li Sun and Zhigang Zeng and Kin-Man Lam},
  title = {{Multi-kernel feature extraction with dynamic fusion and downsampled residual feature embedding for predicting rice RNA N6-methyladenine sites}},
  journal = {Briefings in Bioinformatics},
  year = {2025},
  doi = {10.1093/bib/bbae647},
}
Figure