DeepRelCDR: A Deep Relation-Aware Molecular Graph Attention Framework for Cancer Drug Response Prediction
Abstract Predicting cancer cell-line response to therapeutic compounds is central to precision oncology, yet graph neural network models typically represent each drug only through covalent-bond connectivity, leaving ring-system and functional-group structure implicit; the relation set used here is fixed by chemical domain knowledge and evaluated on a single pharmacoge-nomic resource rather than learned or validated across cohorts. We introduce DeepRelCDR, a deep relation-aware molecular graph attention framework that represents each drug as three explicit relations, covalent bonds, ring co-membership, and RDKit-derived functional-group co-membership, each aggregated independently in a depthwise-separable fashion by a transform-free attention head on raw atom features and fused by a single shared pointwise layer; the resulting drug embedding is combined with a cell-line gene-expression embedding to predict a standard drug-sensitivity readout, LN IC50. Trained on 235 drugs and 199,903 drug–cell-line pairs from GDSC2, DeepRelCDR reaches a test R 2 of 0.762 (95% CI [0.756, 0.769]), outperforming a single-relation attention baseline by 6.0 points, a Ridge-regression baseline by 13.6 points, and a GraphDRP-style reimplementation by 6.7 points of R 2. Gradient-based gene attribution recovers Schlafen-11, an independently established chemosensitivity biomarker gene, as the top camptothecin-sensitivity gene across three cancer types, and a personalized drug-ranking analysis reaches an 80.3% top-1 hit rate versus 53.5% for a non-personalized baseline.
Authors
- Masoud Daneshtalab (ORCID: https://orcid.org/0000-0001-6289-1521)
- Maytham N. Meqdad (ORCID: https://orcid.org/0000-0001-8590-6722)
- Majid Sepahvand (ORCID: https://orcid.org/0000-0002-4451-2054)
- Sarvin Chitsaz
Institutions
- Arak University (IR)
- Al-Mustaqbal University (IQ)
- Mälardalen University (SE)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-12
- DOI
- https://doi.org/10.21203/rs.3.rs-10988328/v1
- Primary Topic
- Computational Drug Discovery Methods
- Type
- preprint