From Deep LSTM-Based Gene Expression Modeling to microRNA Disease Association; hsa-miR-133b identified as a Potentially Functionally Conserved Tumor Suppressor Across Cancers

Document Type : Original Article

Authors

1 Computing and Bioinformatics Program, Faculty of Science, Port Said University, Port Said, Egypt

2 Department of Mathematics and Computer Science, Faculty of Science, Port Said University, 42521-Port Said, Egypt

3 Biotechnology program, Zoology Dept., Faculty of Science, Port Said University

Abstract

Background: Cancer remains a global health burden, and early, accurate diagnosis is vital. Most deep learning models focus on binary classification, with limited work on multiclass tasks. The roles of miRNAs in cancer have also been studied. Objective: This study proposed a hybrid deep learning model based on Deep Long Short-Term Memory (D-LSTM) for binary and multi-class cancer classification. The important genes identified by the model were then used to predict target miRNAs. Methods: The D-LSTM model was trained on GEO dataset GSE203024, covering 14 cancers, colon polyps, and normal samples. Gene selection was per- formed using ANOVA-F and CFS, and SMOTE was used to handle class imbalance. Performance was evaluated using the accuracy, precision, recall, and F1-score. miRNAs were identified using miRNet, and regulatory networks were visualized using Cytoscape. Results: The model achieved 98.38% accuracy (binary) and 99.88–100% (multi-class). hsa-miR-133b has been linked to 14 cancers and colon polyps, targeting CTNND1, CCNB1, and SUZ12. Conclusion: While demonstrating strong diagnostic potential, in silico findings require further biological validation. The comprehensive evaluation of hyperparameters, activation functions, and performance metrics of the model provides a flexible framework for cancer detection. Future studies should focus on in vivo/in vitro validation of hsa-miR-133b’s clinical relevance.

Keywords

Main Subjects