DeepCIA: a novel deep-learning model for cancer type identification using class activation map via transcription...

Area of Science:

• Bioinformatics
• Computational Biology
• Oncology

Background:

• Deep learning excels in large-scale data analysis but faces challenges in cancer data due to scale.
• Accurate cancer type identification is crucial for effective treatment strategies.

Purpose of the Study:

• Introduce DeepCIA, a novel deep learning model for cancer type identification.
• Utilize transcription factor expression and class activation maps for diagnostic insights.
• Validate the model's performance against established machine learning algorithms.

Main Methods:

• Selected transcription factor expression profiles from eight cancer types across TCGA (3496 samples) and ICGC (552 samples) databases.
• Developed and validated a 1D-Convolutional Neural Network (1D-CNN) model, comparing it with Support Vector Machine (SVM) and K-Nearest Neighbors (KNN).
• Employed Class Activation Maps for visualization and identification of key molecular drivers.

Main Results:

• The 1D-CNN model (DeepCIA) achieved a 98% average accuracy in external validation, surpassing SVM and KNN by 10-12%.
• Demonstrated high performance across multiple metrics: 98.2% Recall, 98.1% Precision, 98.2% F-score, 99.8% Specificity, 99.8% AUC, and 99.0% Balanced Accuracy.
• Identified the Cys2Hys2 zinc finger group as highly distributed across all cancer types.

Conclusions:

• DeepCIA serves as a powerful decision support system for cancer diagnosis.
• The model effectively classifies unknown primary cancers, highlighting its utility in clinical settings.
• Transcription factor expression analysis via deep learning offers a promising avenue for advancing cancer diagnostics.