Frontiers in Science · Oct 22, 2026
Targeting G protein-coupled receptor networks for healthy aging: grounding systems geroprotection in biophysical and data realities
By Filizola M
In their lead article in Frontiers in Science, Boeringer et al. propose targeting networks of G protein-coupled receptors (GPCRs) as a promising strategy for geroprotection (Boeringer et al. 2026). This approach builds on the remarkable versatility of GPCRs, which sense a wide range of extracellular signals, engage diverse signaling pathways, are broadly expressed across tissues, and play important roles in many disease processes. By framing the GPCRome as a central, druggable regulator of multiple hallmarks of aging, from inflammaging and mitochondrial dysfunction to stem cell depletion, the authors advance an intriguing hypothesis in which systems pharmacology could be used to rebalance several of these processes simultaneously, potentially delaying the onset of chronic disease and extending healthspan. The proposed integration of artificial intelligence (AI), digital twins, and quantum computing to rationally manipulate such complex signaling networks is bold and thought-provoking.At the same time, translating this compelling vision into therapeutic reality presents formidable biophysical, computational, and experimental challenges. The authors explicitly acknowledge several of these limitations, including the sparsity and variable quality of available training data, the context dependence of GPCR signaling, and, importantly, the fact that network-and AI-guided polypharmacology does not yet constitute a validated or generalizable framework for designing multi-receptor interventions. Several further considerations may help define what will be required to move this exciting concept from a compelling systems-level hypothesis toward an experimentally testable and ultimately translatable framework for GPCR network pharmacology.The proposed networks contain many receptors with established disease relevance, but deciding that a receptor belongs in an aging network does not tell us how it should be modulated. A therapeutic network requires knowing whether a receptor should be activated, inhibited, partially activated, allosterically modulated, biased toward one effector, or manipulated in a tissue-restricted manner, and by how much. This distinction matters because GPCR effects are strongly dependent on receptor density, coupling repertoire, accessory proteins, cellular state, ligand concentration, and tissue context.Especially problematic is the situation for biased ligands. Compared with binding data, efficacy data are sparse and functional measurements are not interchangeable across assays.A ligand that appears biased in one cellular background may not retain the same functional preference in another because receptor expression levels, membrane lipid composition, expression reserve, transducer abundance, signal amplification, and regulatory machinery differ. Consequently, pooling efficacy measurements from heterogenous assays may obscure the underlying pharmacology and limit the predictive utility of models trained on such data.The challenge, therefore, is not simply to assemble more data, but to assemble the right data, including their experimental context.The proposed GPCRome is necessarily constrained by measurements reported in the literature. This creates an important circularity: well-studied receptors become highly connected nodes, while poorly characterized receptors, including orphan receptors (i.e., receptors without known endogenous or synthetic ligands) or incompletely deorphanized GPCRs, may be absent or misrepresented despite mediating potentially important biology, leaving major blind spots in any GPCRome network reconstruction. Boeringer et al. also highlight receptors for which human or translational evidence remains limited, with several proposed nodes supported primarily by in vitro cellular assays and animal models. It remains uncertain whether network topologies derived from controlled rodent experiments will reproduce the networks operating in complex, heterogeneous human populations.A useful next step would therefore be a community-wide effort to construct a standardized, machine-readable GPCR functional atlas. But the harder question is what such an atlas should actually measure, as binding affinity alone is insufficient. Harmonized measurements of ligand association and dissociation, receptor expression and activation, G-protein coupling, β-arrestin engagement, pathway-specific efficacy, receptor trafficking and internalization, desensitization, and signaling duration, are urgently needed-ideally across multiple physiologically relevant cellular backgrounds. Rather than reducing heterogeneous experiments to a single activity label, negative results and experimental metadata should be collected as well, and ideally shared in standardized, accessible format. Such a resource could make network inference more robust and, importantly, reveal where uncertainty really lies. While ligands reshape populations of GPCR conformations, different transducers both select receptor states and undergo their own conformational changes. Thus, signaling cannot be reduced to a ligand-induced receptor state. Rather, AI-based structure prediction, molecular dynamics and enhanced sampling, structural biology, and biophysical measurements should be considered complementary approaches for characterizing receptor-transducer conformational ensembles (Wootten et al., 2013;Hilger et al., 2018).Kinetics adds another critical dimension. Ligand residence time, receptor activation and deactivation rates, trafficking and recycling, and the duration of downstream signaling may influence in vivo efficacy and therapeutic window beyond what is captured by equilibrium affinity or endpoint efficacy (Copeland et al., 2006). Incorporating these temporal features into network models could improve both the selection of network-level interventions and the prediction of physiological outcomes.Sparse data should not necessarily halt progress, but they should change the modeling strategy. Transfer learning offers a particularly attractive route: models can learn shared representations from data-rich GPCRs and then adapt to understudied receptors, provided that domain differences and uncertainty are explicitly modeled [e.g., (Provasi and Filizola, 2025)]. The goal should not be to manufacture confidence from sparse data, but to quantify uncertainty and use it to prioritize the next experiment. AI can then become an activelearning engine linking prediction to targeted measurements rather than a one-way generator of hypotheses (Nguyen et al., 2024).A related issue is chemical space. Much of the historical GPCR ligand repertoire was optimized for individual targets, not for deliberately engineered polypharmacology. The chemical matter already available to the field may therefore be a poor representation of the multitarget molecules needed for network control. While generative modeling architectures have successfully enabled the exploration of new regions of chemical space, the objective should be defined in terms of multi-receptor and multi-pathway pharmacology, including selectivity, efficacy, kinetics, tissue exposure, and safety, rather than simply potency at one receptor. This is where structure-based methods, functional profiling, generative chemistry, and experimental validation need to be tightly coupled.Perhaps the most important question is whether the proposed network architecture will have translational relevance. Boeringer et al. (Boeringer et al. 2026) present mechanistic hypotheses that draw largely on cellular studies, animal models, and selected examples of clinical pharmacology. However, networks inferred from preclinical systems may look quite different in patients, where age, genotype, comorbidities, medication history, tissue composition, and disease stage can all reshape signaling. The authors' vision of digital twins and patient-specific GPCR interventions is compelling precisely because it points to this missing layer.The next step should be to test these network hypotheses directly against patient-derived data rather than assume that preclinical networks represent stable therapeutic maps. Single-cell and spatial transcriptomics, proteomics, functional biomarkers, clinical phenotypes, and longitudinal treatment-response data could help establish which receptor nodes are conserved in humans and which are specific to particular tissues. Ultimately, the goal should not be to build a larger GPCR network, but a better-calibrated one that makes predictions that can be rigorously tested, challenged, and ultimately falsified in patients.A major strength of the Boeringer et al. article is that it challenges the field to think beyond the single receptor. That is an important conceptual shift. The next challenge is not simply to add more receptors, but to establish the experimental and computational standards needed to make network pharmacology genuinely testable. Such a program would: combine systematic, context-rich functional measurements with patient-derived evidence; use approaches such as transfer learning and uncertainty-aware active learning to deal with sparse data; account for receptor-transducer mechanisms and signaling kinetics rather than relying on static structures alone; and integrate physics-based and data-driven approaches to connect molecular mechanisms with cellular and organismal behavior (Hendrickx et al., 2020).The central question, then, is whether we can generate the data, assays, models, and chemical tools needed to perturb these networks in a predictable way. Turning this framework into something clinically useful will require considerably more experimental grounding and validation than the current computational roadmap implies.diseases might eventually become. But moving from theoretical network diagrams to clinically useful geroprotective therapeutics will require computational models firmly grounded in experimental biophysics and human biology. No single computational or experimental approach will be sufficient. What is needed instead is an integrated framework that brings together community-generated harmonized functional data, kinetic profiling, mechanistic modeling, generative chemistry, and patient-derived evidence. The real measure Filizola, M. Realities of GPCR Network Control 9 of success will not be how comprehensive the resulting GPCR networks appear, but whether they capture signaling well enough to generate predictions that can be experimentally tested and, ultimately, translated to patients.The author declared that generative AI was used in the creation of this manuscript.ChatGPT.Edu (GPT 5.6 Sol) was used to assist with editing. The author remains responsible for checking the factual accuracy of the manuscript and for the final content.