Gene Mutation Classification through Text Evidence Facilitating Cancer Tumour Detection

Area of Science:

• Computational Biology
• Bioinformatics
• Medical Informatics

Background:

• Cancer tumors harbor numerous genetic mutations, complicating the identification of driver mutations crucial for tumor growth.
• Manual classification of genetic mutations based on clinical evidence is time-consuming and subjective, hindering efficient cancer research and treatment.
• Existing methods lack automated approaches for classifying genetic mutations into distinct clinical evidence categories.

Purpose of the Study:

• To propose and evaluate a multiclass classifier using Natural Language Processing (NLP) techniques for automated classification of genetic mutations based on clinical text descriptions.
• To compare the performance of various text transformation and machine learning models in distinguishing driver mutations from neutral genetic mutations.
• To leverage clinical evidence to improve the accuracy and efficiency of cancer mutation analysis.

Main Methods:

• Utilized a dataset from Memorial Sloan Kettering Cancer Center (MSKCC) containing genetic mutation descriptions.
• Applied text transformation models: CountVectorizer, TfidfVectorizer, and Word2Vec.
• Implemented machine learning classifiers: Logistic Regression (LR), Random Forest (RF), XGBoost (XGB), and a Recurrent Neural Network (RNN).
• Evaluated classifier performance using accuracy scores derived from confusion matrices.

Main Results:

• The Recurrent Neural Network (RNN) model demonstrated superior performance compared to other evaluated classifiers.
• The RNN model achieved the highest accuracy of 70% in classifying genetic mutations based on clinical evidence.
• Empirical results indicate the potential of deep learning approaches for automated mutation classification.

Conclusions:

• The proposed NLP-based multiclass classifier, particularly the RNN model, offers a promising automated solution for distinguishing driver from neutral genetic mutations.
• This approach can significantly reduce the manual effort and time required by pathologists, accelerating cancer research.
• Further development and validation of deep learning models can enhance the precision of cancer mutation analysis and support clinical decision-making.