Deepm5C

Deepm5C predicts N^5-methylcytosine (m5C) sites in human RNA using a hybrid deep-learning and conventional machine-learning framework to enable accurate genome-wide identification of RNA m5C modifications.

Key Features:

• Hybrid Framework: Combines deep learning with conventional machine-learning techniques to enhance prediction accuracy.
• Benchmarking Dataset: Built on a newly constructed benchmarking dataset addressing limitations of previous small training datasets.
• Feature-Encoding: Investigates a mixture of three conventional feature-encoding algorithms and a word-embedding-derived feature to capture comprehensive representations.
• Classifier Ensemble: Trains four variants of deep-learning classifiers and four conventional classifiers on encoded features to generate 32 baseline models.
• Stacking Strategy: Integrates outputs from optimal baseline models and trains a one-dimensional (1D) convolutional neural network as the meta-classifier.

Scientific Applications:

• Genome-wide m5C detection: Provides a method for genome-wide identification of RNA N^5-methylcytosine (m5C) sites.
• Functional implication studies: Enables exploration of the functional implications of m5C modifications across cellular processes and disease-related contexts.
• Putative site identification and hypothesis generation: Supports identification of putative m5C sites to facilitate formulation of testable biological hypotheses.

Methodology:

Constructed a comprehensive benchmarking dataset; encoded sequences using three conventional feature-encoding algorithms plus a word-embedding-derived feature; trained four deep-learning classifier variants and four conventional classifiers to produce 32 baseline models; applied a stacking strategy by integrating optimal baseline outputs and training a one-dimensional (1D) convolutional neural network as meta-classifier; validated by cross-validation (MCC 0.697, accuracy 0.855) and independent testing (MCC 0.691, accuracy 0.852).