RBP-TSTL

RBP-TSTL predicts RNA-binding proteins at genome scale across diverse species using a two-stage deep transfer learning framework that leverages self-supervised protein sequence embeddings and fine-tuned deep models.

Key Features:

• Two-Stage Transfer Learning Framework: Stage one extracts feature embeddings from a self-supervised pre-trained model to represent protein sequences, and stage two initializes a customized deep learning model with an annotated RBPs dataset then fine-tunes it on target species datasets.
• Self-Supervised Pre-Training: Uses knowledge extracted from self-supervised pre-trained models to generate feature embeddings that enhance protein sequence representation for downstream prediction.
• Customized Deep Learning Model: Initializes model parameters using an annotated dataset of RBPs and adapts the model by fine-tuning on target-species datasets.
• Performance Benchmarking: Benchmarks performance against models using hand-crafted encoding features to quantify improvements from self-supervised pre-training and transfer learning.
• Genome-Scale Predictions: Produces genome-scale RBP predictions and a computational compendium of putative RBP candidates for organisms including Homo sapiens, Arabidopsis thaliana, Escherichia coli, and Salmonella.

Scientific Applications:

• RBP characterization: Enables large-scale identification of putative RBPs to support studies of sequence–structure–function relationships in RNA-binding proteins.
• Cross-species RBP discovery: Facilitates genome-scale prediction of RBPs across multiple species to inform biomedical and biotechnological research.

Methodology:

Apply a two-stage deep transfer learning pipeline: (1) extract protein sequence embeddings from a self-supervised pre-trained model; (2) initialize a customized deep learning model with an annotated RBPs dataset and fine-tune it on target species datasets; performance is benchmarked against models using hand-crafted encoding features.