Refining cancer prediction with DNA sequencing and combined machine learning approaches

Area of Science:

• Computational biology
• Genomic medicine
• Machine learning in oncology

Background:

• Accurate early cancer detection is crucial for effective treatment.
• Existing DNA-based prediction models have limitations in accuracy and interpretability.
• Developing robust computational tools is essential for advancing personalized oncology.

Purpose of the Study:

• To develop and validate a high-accuracy DNA-based cancer risk predictor.
• To compare the performance of a blended ensemble model against individual algorithms and state-of-the-art methods.
• To provide an interpretable and effective tool for early cancer prediction.

Main Methods:

• A blended ensemble model combining Logistic Regression and Gaussian Naive Bayes was developed.
• Hyperparameter optimization was performed using grid search.
• The model was trained and validated on a cohort of 390 patients across five cancer types (BRCA1, KIRC, COAD, LUAD, PRAD).

Main Results:

• The blended ensemble achieved 100% accuracy for BRCA1, KIRC, and COAD.
• Accuracies of 98% were obtained for LUAD and PRAD, outperforming recent benchmarks by 1-2%.
• A micro- and macro-average ROC AUC of 0.99 was achieved, demonstrating strong predictive performance.

Conclusions:

• The developed DNA-based cancer risk predictor demonstrates superior accuracy and performance compared to existing methods.
• The blended ensemble approach offers a lightweight, interpretable, and highly effective solution for early cancer prediction.
• This study highlights the potential of machine learning in enhancing genomic medicine and personalized cancer risk assessment.