TUGDA

TUGDA improves generalization of cancer drug-response prediction models by combining multi-task learning, domain adaptation, and uncertainty quantification to translate in vitro findings to in vivo and patient datasets.

Key Features:

• Multi-Task Learning (MTL) Integration: Employs multi-task learning to leverage shared information across related prediction tasks and mitigate limited-sample effects in in vitro drug-response models.
• Domain Adaptation (DA): Incorporates domain adaptation strategies to extend in vitro predictive models to in vivo and patient-derived contexts.
• Uncertainty Quantification: Quantifies predictor uncertainty and weights predictors in shared feature representations to reduce the influence of noisy tasks and domains.
• Reduction of Negative Transfer (NT): Reduces negative transfer, achieving a 63% reduction for drugs with limited data and a 94% overall reduction compared to state-of-the-art methods.
• Unsupervised Learning Capability: Operates with unsupervised training and shows improved predictions for six of twelve drugs in patient-derived xenografts (PDX) and seven of twenty-two drugs in The Cancer Genome Atlas (TCGA) patient datasets.

Scientific Applications:

• Oncology / Precision Medicine: Improves cancer treatment stratification by enhancing reliability of drug-response predictions across cell lines, PDX, and TCGA patient cohorts.
• Integrative Omics and Drug-Response Analysis: Supports integration of high-dimensional omics data and diverse drug-response datasets to inform personalized therapeutic strategies.

Methodology:

Implements a unified framework combining multi-task learning and domain adaptation with uncertainty quantification that prioritizes low-uncertainty predictors; the model can be trained in an unsupervised fashion.