A multi-task deep learning model based on comprehensive feature integration and self-attention mechanism for...

Area of Science:

• Oncology
• Computational Biology
• Biostatistics

Background:

• Immune checkpoint inhibitors (ICIs) are crucial for advanced cancer treatment.
• Predicting ICI efficacy is difficult due to biomarker heterogeneity.
• There is a need for comprehensive biomarker utilization for better treatment prediction.

Purpose of the Study:

• To develop a deep learning model for predicting ICI efficacy.
• To integrate multidimensional biomarkers for enhanced predictive power.
• To address the challenge of predicting treatment response in advanced cancers.

Main Methods:

• Utilized statistical and machine learning techniques.
• Integrated gene expression data from various sources.
• Employed a deep learning model with a self-attention mechanism.
• Performed feature selection across single-cell sequencing, PD-L1, tumor mutational burden (TMB)/mismatch repair (MMR), and somatic copy number alteration (SCNA) levels.

Main Results:

• Identified 96 key feature genes through multidimensional analysis.
• Trained a multi-task deep learning model on a pan-cancer dataset.
• Achieved promising predictive performance in bladder urothelial carcinoma (AUC=0.62-0.66), non-small cell lung cancer (AUC=0.85), and skin cutaneous melanoma (AUC=0.71).

Conclusions:

• The deep learning model shows potential for predicting ICI treatment outcomes.
• Integrating multidimensional features enhances predictive accuracy.
• This approach offers a methodological framework for complex medical data integration.