MTTLmsup6 supA

MTTLmsup6 supA predicts base-resolution N6-methyladenosine (m^6A) sites in mRNA by leveraging multi-task transfer learning with an enhanced transformer-based architecture to integrate low-resolution and base-resolution datasets.

Key Features:

• Multi-Task Transfer Learning: Integrates low-resolution and base-resolution m^6A site prediction and leverages low-resolution data from RMBase, including Saccharomyces cerevisiae entries.
• Enhanced Transformer Architecture: Combines convolutional neural network (CNN) layers with a multi-head-attention framework to capture complex RNA sequence patterns.
• One-Hot Sequence Encoding: Encodes RNA sequences using one-hot encoding as standardized model input.
• Probability Assignment: Assigns probabilities to predicted sites to indicate prediction confidence.
• Transfer Learning for Base-Resolution Prediction: Applies transfer learning to extrapolate base-resolution predictions from low-resolution datasets to improve accuracy and generalization.

Scientific Applications:

• Base-resolution m^6A site prediction: Predicts base-resolution m^6A sites to support studies of gene expression regulation, RNA stability, and other m^6A-related processes.
• Validation and cross-species performance: Validated on Saccharomyces cerevisiae m^6A data (AUROC 77.13%) and on Homo sapiens m^1A data (AUROC 92.9%), demonstrating cross-species generalization and performance gains over existing models.

Methodology:

RNA sequences are encoded using one-hot encoding, and a multi-task model combining CNNs and multi-head-attention is trained to predict low-resolution and base-resolution m^6A sites using transfer learning to enhance base-resolution accuracy.