DICE-SynBa: Addressing Synergy Shortcut Learning via Disentangled and Compressed Embeddings

Drug combinations are central to modern cancer therapies and drug synergy prediction models have demonstrated strong performance on benchmark datasets, yet recent evidence suggests that many approaches may rely on shortcut learning. To address this challenge, we propose DICE-SynBa, a framework built upon DeepSynBa that incorporates feature compression, permutation-invariant representations, and disentangled parameter prediction. We evaluate DICE-SynBa in leave-cell-line-out and leave-drug-out scenarios that assess generalization to unseen entities. Across both settings, DICE-SynBa outperforms existing methods, demonstrating improved generalization and reduced shortcut learning.