A review and comparative study of cancer detection using machine learning: SBERT and SimCSE application

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Machine learning and deep learning are widely used for cancer detection using various data types.
• Previous methods often involve complex preprocessing and feature extraction from medical data.
• Focus on lung, breast, prostate, and colorectal cancers, the most prevalent worldwide.

Approach:

• This study reviews existing machine learning methods for cancer detection.
• A novel approach utilizes sentence transformers (SBERT and SimCSE) to represent raw DNA sequences.
• These DNA representations are then classified using machine learning algorithms like XGBoost, Random Forest, LightGBM, and CNNs.

Key Points:

• Sentence transformers (SBERT, SimCSE) are applied to DNA sequences for cancer detection for the first time.
• XGBoost classifier achieved the highest accuracy, reaching 75% with SimCSE embeddings.
• SimCSE embeddings provided a marginal performance improvement over SBERT embeddings.

Conclusions:

• Sentence transformers offer a new avenue for representing DNA sequences in cancer detection.
• Machine learning models, particularly XGBoost, can effectively classify cancer using these DNA representations.
• This approach simplifies feature engineering by directly using DNA sequences.