SGDSYNERGY: LEVERAGING MULTIMODAL DATA FOR ACCURATE DRUG COMBINATION PREDICTION
Ziwei Ning · Bioinformatics and Computational Medicine · 2025
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
Bioinformatics and Computational Medicine · Upubscience Publisher · article · 2025
Abstract
The treatment of cancer and other complex diseases require sophisticated therapeutic strategies due to their intricate pathophysiological mechanisms and resistance profiles. While combination therapy has emerged as a cornerstone of precision medicine, many existing computational approaches primarily focus on integrating only simple modalities such as molecular graphs and SMILES sequences, which limits their ability to model the interaction dynamics in an interpretable and biologically meaningful manner. To address these challenges, we propose SGDSynergy that introduces several key innovations in multi-modal fusion.
Specifically, in the drug feature extraction stage, we adopt a CLIP-style contrastive learning mechanism to align molecular representations derived from SMILES sequences and molecular graphs, enabling the generation of semantically enriched and modality-aware drug embeddings. For cell line representation, SGDSynergy incorporates external features derived from DDIs data and heterogeneous graph structures, which significantly enhance the biological context of cancer cell line embeddings.
Moreover, in the synergy evaluation module, we introduce a Bayesian attention mechanism integrated with MC-dropout to probabilistically weigh synergistic features, thereby reducing overfitting and explicitly quantifying prediction uncertainty. These integrated innovations allow SGDSynergy to achieve superior accuracy in predicting drug combination effects, especially in capturing non-linear dynamics and rare synergy patterns, offering a biologically informed and clinically actionable framework for designing personalized combination therapies and overcoming treatment resistance.
Open the resolved locator · 10.61784/bcm3003
Metadata licence: OpenAlex metadata (CC0)