TULIP: An RNA-seq-based Primary Tumor Type Prediction Tool Using Convolutional Neural Networks.

Area of Science:

• Oncology
• Bioinformatics
• Computational Biology

Background:

• Cancer is a leading global cause of death, making accurate primary tumor type prediction essential for understanding genetic factors influencing tumor progression.
• Previous research has utilized machine learning and deep learning for tumor classification using gene expression data.

Purpose of the Study:

• To develop and evaluate deep learning models for classifying primary tumor types using RNA-seq count data.
• To assess the performance of models using all Ensembl genes versus protein-coding genes only.
• To create a user-friendly tool for tumor classification.

Main Methods:

• Four 1-dimensional Convolutional Neural Network (1D-CNN) models were developed to classify 17 or 32 primary tumor types.
• Models were trained and validated on RNA-seq count data (FPKM-UQ) from The Cancer Genome Atlas (TCGA).
• Input data included either all Ensembl genes or protein-coding genes, without gene filtering to avoid bias.

Main Results:

• All 1D-CNN models achieved high overall accuracy, ranging from 94.7% to 97.6% on the test dataset.
• Models utilizing only protein-coding genes demonstrated superior accuracy compared to those using all Ensembl genes.
• Most primary tumor types were classified with an accuracy exceeding 80% across all models.

Conclusions:

• The developed models were packaged into a Python tool named TULIP (TUmor CLassIfication Predictor) for quality control and characterization of unknown cancer samples.
• Further model optimization is recommended to enhance accuracy for specific primary tumor types.