SORFPP: Enhancing rich sequence-driven information to identify SEPs based on fused framework on validation datasets

Area of Science:

• Genomics and Molecular Biology
• Bioinformatics and Computational Biology

Background:

• Short open reading frames (sORFs) within long non-coding RNAs (lncRNAs) encode functional peptides (SEPs) with significant regulatory roles.
• Existing computational methods for SEP prediction lack sufficient feature engineering and effective modeling.
• A high-throughput computational method is essential for accurate and efficient SEP identification.

Purpose of the Study:

• To develop and validate a novel computational method for predicting SEPs.
• To address limitations in current feature extraction and predictive modeling for SEPs.
• To enhance the accuracy and robustness of SEP prediction using integrated approaches.

Main Methods:

• Developed SORFPP, a computational method integrating protein language model ESM-2 and traditional encodings (QSOrder, k-mer).
• Employed CatBoost to handle sparsity in traditional features and a Self-attention model for ESM-2 characterization.
• Utilized an ensemble learning framework combining multiple models, with final predictions from a Logistic Regression model.

Main Results:

• SORFPP demonstrated superior performance compared to state-of-the-art models.
• Achieved significant improvements in Matthew correlation coefficient, ranging from 12.2% to 24.2% across three benchmark datasets.
• Validated the efficacy of integrating ensemble learning, traditional features, and protein language encoding.

Conclusions:

• The proposed SORFPP method effectively predicts SEPs by combining diverse feature information and ensemble learning.
• Integrating ensemble strategies with traditional and protein language encoding methods yields enhanced predictive performance.
• The study provides accessible datasets and code for further research in SEP prediction.