Integrating deep generative model with active learning for predicting immunotherapy responses in gastric cancer

Area of Science:

• Oncology
• Immunotherapy
• Bioinformatics

Background:

• Immune checkpoint therapy (ICT) offers revolutionary gastric cancer (GC) treatment but benefits limited patients.
• Accurate prediction models for immunotherapy response are crucial but hindered by data scarcity.

Purpose of the Study:

• To develop an accurate prediction model for immunotherapy response in GC patients.
• To address data insufficiency using data augmentation and active learning.

Main Methods:

• Utilized generative adversarial networks (GANs) for data augmentation.
• Employed active learning (AL) to select informative samples and mitigate overfitting.
• Evaluated machine learning algorithms, identifying Random Forest (RF) as optimal, and validated on an independent cohort.

Main Results:

• The RF model achieved high accuracy (0.911), recall (0.896), F1-score (0.898), and AUC (0.902) on the real testing set.
• External validation yielded an AUC of 0.838.
• Identified 15 key genes associated with immune processes and two potential therapeutic agents (Docetaxel, Lapatinib).

Conclusions:

• The proposed framework effectively predicts GC immunotherapy response, addressing data limitations.
• Identified key genes and compounds offer insights for combinatorial therapies.
• This approach can improve ICT response prediction accuracy and guide novel therapeutic strategies.