# Fast-adapting graph neural network with prior knowledge for drug response prediction across preclinical and clinical data

## TL;DR

A new model called metaDRP improves drug response predictions using few samples and adapts well to new data, helping bridge the gap between lab and clinical settings.

## Contribution

Introduces metaDRP, a few-shot learning model with prior knowledge integration to address data scarcity and out-of-distribution issues in drug response prediction.

## Key findings

- metaDRP performs comparably to state-of-the-art models in drug response prediction with limited sample sizes.
- The model effectively mitigates out-of-distribution issues between in vitro and in vivo settings.
- metaDRP provides interpretable insights into drug mechanisms of action, such as for epothilone B and pemetrexed.

## Abstract

Efficient drug response prediction is crucial for reducing drug development costs and time, but current computational models struggle with limited experimental data and out-of-distribution issues between in vitro and in vivo settings. To address this, we introduced drug response prediction meta-learner (metaDRP), a novel few-shot learning model designed to enhance predictive accuracy with limited sample sizes across diverse drug-tissue tasks. metaDRP achieves performance comparable to state-of-the-art models in both genomics of drug sensitivity in cancer (GDSC) drug screening and in vivo datasets, while effectively mitigating out-of-distribution problems, making it reliable for translating findings from controlled environments to clinical applications. Additionally, metaDRP’s inherent interpretability offers reliable insights into drug mechanisms of action, such as elucidating the pathways and molecular targets of drugs like epothilone B and pemetrexed. This work provides a promising approach to overcoming data scarcity and out-of-distribution challenges in drug response prediction, while promoting the integration of few-shot learning in this field.

Image 1

•MAML-based bilevel optimization framework to boost drug response prediction model’s generalization in low-sample settings.•Prior knowledge-driven model with graph neural networks and sparse linear layers for biomolecular network topological information.•Attention and bio-annotated weights endow model with interpretability to find key drug-related pathways and target genes.

MAML-based bilevel optimization framework to boost drug response prediction model’s generalization in low-sample settings.

Prior knowledge-driven model with graph neural networks and sparse linear layers for biomolecular network topological information.

Attention and bio-annotated weights endow model with interpretability to find key drug-related pathways and target genes.