Abstract A008: AI-Predict: Artificial intelligence-mediated drug synergy prediction and validation in cancer models

Abstract Introduction: In precision oncology, monotherapies frequently result in resistance and disease relapse, emphasizing the need for rational drug combinations. Identifying effective combinations remains challenging due to the immense combinatorial space, tumor-specific molecular heterogeneity, and the limited scalability of experimental screening. Drug repurposing offers a promising and cost-effective alternative by leveraging compounds with known safety profiles. In recent years, machine learning approaches have been proposed to predict drug synergy, yet many remain limited by reliance on static datasets, batch effects, and minimal biological validation. Methods: We present a graph neural network (GNN)-based model that predicts drug synergy using molecular structure and gene expression data from cancer cell lines. Drugs are represented as molecular graphs and processed through GATv2 layers to capture structural relationships. Cell lines are encoded via attention-based embeddings of 908 landmark genes, capturing transcriptomic context. The model outputs the probability of synergy for each drug pair–cell line triplet and is trained using binary cross-entropy with L2 regularization. For in vitro validation, human cancer cell lines (ovarian UWB1.289, renal 786-0, and breast BT-549) were cultured under standard conditions and routinely tested for Mycoplasma contamination. Results: Our model was trained on the DrugComb v1.5 dataset (>1.4M triplets), annotated with four synergy scores (Loewe, Bliss, HSA, ZIP) and three consensus-based labels (Majority-2/3/4). We benchmarked performance against classical ML models, DeepSynergy, DeepDDS, and a GNN baseline with standard GAT layers. Across seven datasets, our model consistently outperformed all baselines, with gains of 3–18% in AUPR and 3–6% in AUC. It also maintained strong performance in generalization tasks involving unseen drugs and cell lines. To confirm biological relevance, we tested predicted drug pairs in vitro across three human cancer cell lines: ovarian (UWB1.289), renal (786-0), and breast (BT-549). We selected combinations with high predicted synergy, high antagonism, and low interaction probability. Each pair was tested under different exposure times (72h–8 days) and six-point dose-response curves based on known drug properties. Synergistic pairs showed enhanced cytotoxicity, while antagonistic combinations resulted in reduced efficacy. Neutral predictions aligned with additive effects. Context-dependent differences further highlighted the relevance of transcriptomic integration. Conclusion: Our GNN-based model demonstrates strong predictive performance for drug synergy and generalizes across biological contexts. In vitro experiments validate its translational utility, supporting its application in drug repurposing and personalized combination therapy. Future directions include testing in patient-derived models and leveraging experimental feedback to refine predictions through active learning. Citation Format: Alicia Pliego, Chantal Pauli, Lara Planas-Paz, Michael Krauthammer, Amina Mollaysa, Kyriakos Schwarz, Sarah Kollar, Ahmed Allam. AI-Predict: Artificial intelligence-mediated drug synergy prediction and validation in cancer models [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A008.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.