Interpretable Transfer Learning for Cancer Drug Resistance: Candidate Target Identification

Area of Science:

• Computational biology and bioinformatics
• Genomics and transcriptomics
• Cancer research and precision oncology

Background:

• Tumor drug resistance is highly variable across cancer types due to diverse molecular mechanisms.
• Predicting and understanding drug resistance is crucial for effective cancer treatment and personalized medicine.

Purpose of the Study:

• To develop a pan-cancer transfer learning framework for predicting tumor drug resistance.
• To identify novel, tumor-specific drug resistance biomarkers with clinical significance.

Main Methods:

• A residual variational autoencoder (Res VAE) backbone was integrated into a transfer learning framework.
• Five models, including three novel large language model (LLM)-integrated VAEs (VAE_LL, VAE_LS, VAE_LD) and two ensemble methods (RF, XGB), were trained on the Genomics of Drug Sensitivity in Cancer (GDSC) dataset.
• The top-performing models were validated on five TCGA cohorts (1,836 patients) for 180 drug-cancer prediction tasks.

Main Results:

• The VAE_LD model demonstrated superior performance, achieving a mean AUC of 0.81 and F1 score of 0.92 on the GDSC benchmark.
• VAE_LD maintained strong predictive power in clinical validation across multiple cancer types.
• Interpretation analyses identified clinically significant biomarkers, such as TFF1 in lung adenocarcinoma and OPALIN, LTF, IL2RA, SLC17A7 in glioblastoma, associated with resistance to specific drugs (Gefitinib, Temozolomide).

Conclusions:

• The VAE_LD model provides a high-performing and interpretable approach for predicting pan-cancer drug resistance.
• The framework facilitates the discovery of actionable biomarkers for targeted therapies.
• This study offers a robust computational tool for advancing precision oncology applications.