Abstract
Precision oncology plays a pivotal role in contemporary healthcare, aiming to optimize treatments for each patient based on their unique characteristics. This objective has spurred the emergence of various cancer cell line drug response datasets, driven by the need to facilitate pre-clinical studies by exploring the impact of multi-omics data on drug response. Despite the proliferation of machine learning models for Drug Response Prediction (DRP), their validation remains critical to reliably assess their usefulness for drug discovery, precision oncology and their actual ability to generalize over the immense space of cancer cells and chemical compounds. Scientific contribution In this paper we show that the commonly used evaluation strategies for DRP methods can be easily fooled by commonly occurring dataset biases, and they are therefore not able to truly measure the ability of DRP methods to generalize over drugs and cell lines ("specification gaming"). This problem hinders the development of reliable DRP methods and their application to experimental pipelines. Here we propose a new validation protocol composed by three Aggregation Strategies (Global, Fixed-Drug, and Fixed-Cell Line) integrating them with three of the most commonly used train-test evaluation settings, to ensure a truly realistic assessment of the prediction performance. We also scrutinize the challenges associated with using IC50 as a prediction label, showing how its close correlation with the drug concentration ranges worsens the risk of misleading performance assessment, and we indicate an additional reason to replace it with the Area Under the Dose-Response Curve instead.