Adversarial training improves model interpretability in single-cell RNA-seq analysis

Area of Science:

• Computational biology
• Artificial intelligence in medicine
• Machine learning for genomics

Background:

• Reliable predictive computational models are crucial for biology and medicine.
• Model robustness (insensitivity to input variations) and interpretability (explainability of decisions) are key for trust.
• Existing methods often address robustness and interpretability independently, with their interaction being poorly understood.

Purpose of the Study:

• To investigate the impact of adversarial training on the robustness and interpretability of deep learning models.
• To explore the application of these models in predicting cell types from single-cell RNA sequencing data.

Main Methods:

• Adversarial training was applied to a deep learning model for cell type prediction.
• Model interpretability was assessed using standard methods to identify important genes for classification.
• Single-cell RNA sequencing data was used as the example task.

Main Results:

• Adversarial training significantly improved the robustness of the cell type prediction model.
• Surprisingly, adversarial training also enhanced model interpretability, as evidenced by gene importance identification.
• The findings suggest adversarial training is a valuable technique for improving deep learning models.

Conclusions:

• Adversarial training can simultaneously enhance the robustness and interpretability of deep learning models.
• This approach holds potential for broader applications in scientific research, particularly in areas requiring reliable and explainable AI.
• Further evaluation of adversarial training across diverse tasks is recommended.