A robust and lightweight deep attention multiple instance learning algorithm for predicting genetic alterations.

Area of Science:

• Digital Pathology
• Computational Biology
• Genomics

Background:

• Self-attention algorithms offer interpretability in digital pathology but are computationally intensive.
• Predicting gene mutations from histopathology images is crucial for cancer research and treatment.

Purpose of the Study:

• To develop a computationally efficient and accurate model for predicting gene mutations in digital pathology.
• To introduce the Attention-based Multiple Instance Mutation Learning (AMIML) model.

Main Methods:

• Developed a novel, lightweight Attention-based Multiple Instance Mutation Learning (AMIML) model.
• Compared AMIML against standard self-attention, deep learning, and traditional machine learning models.
• Utilized TCGA data for 24 genes across UCEC, BRCA, GBM, and KIRC cancer cohorts.

Main Results:

• AMIML significantly reduced model parameters by approximately 70% compared to standard self-attention.
• AMIML demonstrated superior robustness and outperformed baseline models for most tested genes.
• AMIML's attention mechanism effectively identified predictive image patches.

Conclusions:

• The AMIML model provides an efficient and interpretable approach for gene mutation prediction in digital pathology.
• AMIML offers a robust alternative to existing methods, outperforming them in accuracy and efficiency.
• The model's ability to pinpoint predictive regions enhances its utility in pathological analysis.