OptiSyn: an interpretable, multi-omics–driven graph convolutional network framework for synergy-oriented drug combination design in disease treatment

TL;DR

This paper introduces OptiSyn, a new AI framework that uses multi-omics data to design drug combinations for diseases like ankylosing spondylitis, combining traditional Chinese medicine with modern bioinformatics.

Contribution

OptiSyn is a novel interpretable graph convolutional network framework that integrates multi-omics data to design synergistic drug combinations aligned with TCM principles.

Findings

Eight AS-associated hub genes were identified using multi-omics and network analysis.

ASD-A, a TCM formula, reduced pro-inflammatory cytokines and modulated immune cell subsets in experiments.

Formula decomposition analysis clarified the roles of major and auxiliary components in controlling hub gene activity.

Abstract

Bioinformatics and large-scale computational modelling have emerged as essential research fields in modern biomedical science, enabling drug discovery and therapeutic optimisation. A unique and potent technical framework for the modernisation and mechanistic clarification of traditional Chinese medicine (TCM) formulations is provided by the integration of multidimensional data using systems biology and artificial intelligence (AI) techniques.

Ankylosing spondylitis–associated key hub genes were identified using multi-omics datasets, differential gene expression analysis, weighted gene co-expression network analysis, single-cell transcriptomic analysis, Mendelian randomization, and network module partitioning. In order to predict the optimal drug combinations and synergistic principal-auxiliary therapeutic roles, an interpretable, multi-layer graph convolutional network model was built using network topology features, molecular docking data, empirical clinical medication knowledge, and compound clustering similarity.

Eight AS-associated hub genes were found using AS as a representative disease model. A possible TCM formula, ASD-A, comprising Myrrha, Drynariae Rhizoma, Lycii Fructus, Epimedii Folium, Achyranthis Bidentatae Radix, Alpiniae Officinarum Rhizoma, Forsythiae Fructus, Astragali Radix, was prioritised by the proposed model. While ablation studies highlighted the crucial role of multi-source information integration in compound formula screening and the creation of customised intervention methods, model performance evaluation showed strong predictive potential. The identified hub genes were found to be tightly linked to immunological responses and T-cell-mediated immune processes, according to functional enrichment analyses. Experiments conducted both in vivo and in vitro confirmed that ASD-A significantly reduced pro-inflammatory cytokines like IL-6 and TNF-α (P < 0.05), modulated the proportions of CD80 and CD86 cell subsets, and regulated the expression of important genes like KRAS, SMAD2, and MAPK14 (P < 0.05). Furthermore, in activated Jurkat cells, ASD-A significantly decreased IL17 fluorescence while increasing Foxp3 fluorescence (P < 0.05), indicating a rebalancing of the IL17/Foxp3 axis. The roles of major and auxiliary components in controlling hub gene activity were further clarified by formula decomposition analysis of ASD-A.

The suggested AI-driven formula design approach provides new insights into combinatorial therapy approaches for AS while also conforming to the TCM theory of Jun-Chen-Zuo-Shi principle. The significant potential of combining contemporary bioinformatics and AI techniques with traditional medicine is demonstrated by this study, which could facilitate efficient and mechanistically informed disease therapy.

The online version contains supplementary material available at 10.1186/s13020-026-01385-1.

Figures