Pan-cancer classification by regularized multi-task learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Accurate pan-cancer classification is vital for personalized cancer diagnosis and treatment.
• Gene expression patterns offer a promising avenue for distinguishing between diverse cancer types.
• Existing machine learning models face challenges in reliably classifying multiple cancer types simultaneously.

Purpose of the Study:

• To develop and evaluate PC-RMTL, a novel regularized multi-task learning (RMTL) model for pan-cancer classification.
• To compare the performance of PC-RMTL against established machine learning algorithms using RNASeq data.
• To identify key gene expression patterns that discriminate between various cancer types.

Main Methods:

• Utilized RNASeq data from The Cancer Genome Atlas (TCGA) for 21 cancer types and adjacent normal samples.
• Developed a pan-cancer classification model, PC-RMTL, based on regularized multi-task learning (RMTL).
• Compared PC-RMTL performance against Support Vector Machine (SVM) with linear and radial basis function kernels, Random Forest (RF), k-Nearest Neighbours (kNN), and Decision Trees (DT).

Main Results:

• PC-RMTL achieved 96.07% accuracy and 95.80% MCC score on an independent test set.
• PC-RMTL outperformed SVM-Lin, SVM-RBF, RF, kNN, and DT in pan-cancer classification accuracy.
• SVM-Lin was a competitive alternative only when using complete feature sets for training.

Conclusions:

• PC-RMTL represents a significant advancement in pan-cancer classification, surpassing existing methods in reliability.
• The model's high accuracy demonstrates its potential for improving cancer diagnosis and guiding treatment strategies.
• Functional enrichment analysis of discriminating genes provides insights into cancer-specific molecular mechanisms.