Search Schedule From Shared Calendar. ChikkuTakku is service to search schedule from shared calendar and help share calendar
Title: Do Drug-Response Prediction Models Actually Use Molecular Features? A Systematic Benchmark and Ablation StudyAbstract: Numerous deep learning models have been proposed for drug-response prediction — forecasting gene expression changes following drug treatment. While model architectures have grown increasingly complex, the lack of a unified evaluation framework has made fair cross-model comparison difficult, and the actual contribution of molecular features remains poorly understood. This study introduces a standardised evaluation framework with unified data splits, cross-validation, and metrics to systematically benchmark nine published models. The framework centres on two key components: ablation via zero-replacement and random shuffling of molecular feature inputs to quantify their true contribution, and an MLP baseline using only cell-line features to probe the relationship between architectural complexity and performance. Preliminary results show that a simple three-layer MLP (DEG PCC ≈ 0.637) matches or exceeds most complex models, while several models show little to no performance degradation when molecular features are removed. These findings highlight a widespread underutilisation of molecular information in current approaches and offer a reusable benchmark for guiding future model development in this field.Biography:Jinming Bai is a PhD candidate in the Department of Human Biology, Faculty of Health Sciences, University of Cape Town, having previously completed his MSc at UCT. His research, supervised by Prof Sharon Prince (cancer biology) and co-supervised by Prof Geoff Nitschke (computer science), applies machine learning to predict post-treatment gene expression changes for drug repurposing and targeted cancer therapy.
📍 John Day LT1