Explainable AI and debiased networks: The ‘last mile’ for precision medicine delivery

We read Pandey et al.'s article with great interest (Pandey et al., 2025). The authors deeply integrated network science, systems pharmacology and multi omics data, and innovatively proposed the ‘network precision medicine’ framework, providing a one-stop solution for disease classification, biomarker discovery, drug target identification and new use of old drugs, significantly improving research and development efficiency and personalized treatment accuracy. This visionary work not only broadens the boundaries of modern pharmacology but also lays a solid methodological foundation for future artificial intelligence (AI) driven precision medicine, which is admirable. In addition to the limitations pointed out by the authors, we believe there are three issues worth further discussion. Firstly, the complexity and interpretability of computation, as well as the algorithmic black box nature of high-dimensional networks such as integrated genomes, epigenomes and metabolomes, may make it difficult for clinical doctors to understand the predictive basis, hindering clinical adoption (the article only mentions AI interpretability as a ‘future direction’, without discussing its necessity). Secondly, there is the issue of bias in network construction. Existing protein interaction networks (such as STRING; Szklarczyk et al., 2025) tend to focus on more studied proteins (such as kinases and transcription factors), while rare proteins or newly discovered molecules may be systematically overlooked, leading to bias. Then, there is the issue of hierarchical lack of disease heterogeneity. Current network models are mostly based on ‘average patient’ data, but disease subtypes (such as molecular typing of breast cancer; Perou et al., 2000) or microenvironment differences (such as tumour immune infiltration; Gentles et al., 2015) may not be fully modelled. We believe that in the future, by promoting the application of graph neural networks (GNNs; Wu et al., 2021), interpretable AI (such as SAVERUNNER and PrediCT algorithms) in biomedical fields, and innovating algorithms to solve ‘black box’ problems, and by integrating multiple omics data and network models to provide new targets or new strategies for rare diseases (such as genetic metabolic diseases), this work will reshape the understanding framework of diseases and drugs and become a hub connecting basic discovery, clinical translation and industrial applications. In summary, the authors systematically reviewed the latest developments in network medicine and systems pharmacology, demonstrating how to use network methods to identify disease modules, discover biomarkers, guide drug target selection and new use of old drugs, and providing a new paradigm for precision medicine. In the future, with the expansion of multi omics data and the integration of AI, network medicine is expected to achieve real-time customization of individualized diagnosis and treatment plans and accelerate breakthroughs in new drugs and repurposing. Key protein targets and ligands in this article are hyperlinked to corresponding entries in the IUPHAR/BPS Guide to PHARMACOLOGY http://www.guidetopharmacology.org and are permanently archived in the Concise Guide to PHARMACOLOGY 2023/24 (Alexander, Fabbro et al., 2023; Alexander, Kelly et al., 2023). Siyi Liu: Writing—original draft; writing—review and editing; formal analysis. Zekai Yu: Formal analysis; writing—review and editing. The authors declare that, to their knowledge, they have no financial, personal or professional relationships that could be construed as a potential competing interest regarding the work described in this manuscript.