RAMP

Response-Aware Multi-Task Framework for Cancer Drug Sensitivity Prediction

RAMP implements a response-aware multitask prediction (RAMP) framework to predict cancer drug sensitivity from multiomics profiles of individual patients by integrating heterogeneous drug response data and modeling complex feature interactions.

Key Features:

• Response-Aware Multi-Task Learning: Uses Bayesian neural networks within a multitask learning framework to jointly model multiple drug response prediction tasks.
• Soft-Supervised Contrastive Regularization: Applies contrastive regularization to distinguish positive and negative drug response samples and improve predictive robustness.
• Network Embedding with Response-Aware Negative Sampling: Learns representation vectors from heterogeneous networks using cell line–drug response information for response-aware negative sampling.
• Comprehensive Drug Response Feature Selection: Mitigates imbalance in trained response data by systematically incorporating diverse drug response features to reduce bias toward dominant responses.

Scientific Applications:

• Cancer Drug Sensitivity Prediction: Validated on the Genomics of Drug Sensitivity in Cancer dataset with AUC-ROC >89%, AUC-PR >59%, and F1 score >52%, supporting prediction of missing drug responses and personalized cancer therapy selection.

Methodology:

RAMP integrates multiomics patient profiles and heterogeneous cell line–drug response data using Bayesian neural network–based multitask learning. The RAMP framework incorporates soft-supervised contrastive regularization and response-aware negative sampling within network embedding to model complex feature interactions and address response imbalance in public drug sensitivity datasets.