Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Abstract

Geroscience offers a transformative paradigm by targeting shared aging hallmarks to enable simultaneous modulation of multiple age-related disorders (ARDs). Yet, current geroprotective interventions often lack mechanistic breadth, as targeting isolated pathways yields limited benefits compared to interventions modulating interconnected regulators of aging biology. To bridge this gap, a systems-level strategy was designed around four key targets, including, Nrf2/Keap1, mTORC1, AMPK, and SIRT1, responsible for regulating oxidative stress, mitochondrial dysfunction, proteostasis, and autophagy. Concurrent regulation of these targets was identified to potentially induce a concerted and sustained geroprotective effect across diverse ARDs. A machine learning-based geroprotector classification model was developed to identify natural compounds capable of executing this integrated strategy. Subsequent drug-likeness screening confirmed favorable pharmacokinetic properties of the predicted compounds, while molecular docking revealed compounds with strong binding affinities with all four geroprotective targets, thereby leading to the identification of a subset of natural compounds with the potential to induce a coordinated geroprotective response. Finally, a graph neural network-based synergy prediction model, trained on known ARD drug combinations, identified five high-confidence pairs composed of four natural compounds, including Baicalein, Pectolinarigenin, Phloretin, and Demethoxycurcumin. These computationally predicted combinations hold the potential to elicit synergistic and comprehensive geroprotective effects across multiple ARDs.

Keywords

geroscience geroprotectors aging machine learning graph neural networks natural compounds

Introduction

The rising incidence rate of age-related disorders (ARDs) including neurological disorders, type 2 diabetes, heart diseases and chronic kidney disease due to global population aging has driven significant interest in geroscience and the development of geroprotective therapeutic strategies. Geroprotective approaches, which aim to delay the onset or progression of ARDs, offer a comprehensive solution by addressing the underlying biology of aging rather than merely managing disease symptoms (Fekete et al., 2024; Kroemer et al., 2025). By targeting the intricate network of aging-related pathways, these strategies enable the restoration and maintenance of cellular homeostasis, making these strategies concurrently applicable in the treatment of all ARDs. However, most of these geroprotective therapeutic strategies tend to address individual aspects of the aging biology, which has highly complex and intricate nature and thus the effectiveness and sustainability of such geroscience-based strategies fail to compete with that of standard therapeutic interventions (Cummings et al., 2023; Forman et al., 2023; Kroemer et al., 2025; Rolland et al., 2023). Considering the highly dynamic and interconnected nature of processes and pathways involved in the intricate network of the biology of aging and its association with several diseases, it is necessary to devise such geroprotective interventions capable enough to address each hub of this interconnected network and thereby induce a comprehensive and sustainable therapeutic effect in multiple ARDs. In order to devise such all-encompassing strategies, extensive literature review was conducted to identify the primary hubs of the aging biology network involved in ARD progression as well as in the associated therapeutics and it was observed that processes including oxidative stress, mitochondrial dysfunction, proteostasis (including proper protein folding and efficient protein degradation) and autophagy, play a crucial role in the onset and progression of ARDs and have been explored for their individual therapeutic potential multiple times (Cheon et al., 2019; Hipp et al., 2019; Kumar et al., 2023; Santos et al., 2024; Somasundaram et al., 2024; Tabibzadeh, 2023; Wu et al., 2024).

Mitochondrial dysfunction, oxidative stress, impaired proteostasis and defective autophagy are highly interconnected hallmarks of aging that collectively contribute to the onset and progression of ARDs. Aging-associated mitochondrial dysfunction increases reactive oxygen species (ROS) production, causing oxidative damage that further exacerbates mitochondrial impairment and disrupts proteostasis (Somasundaram et al., 2024). Simultaneously, declining efficiency of molecular chaperones, the ubiquitin-proteasome system and autophagy promotes accumulation of damaged proteins and organelles, thereby aggravating cellular dysfunction and ARD progression (Kumar et al., 2023; Tabibzadeh, 2023). Since autophagy is responsible for clearance of dysfunctional proteins and mitochondria, its decline further amplifies oxidative stress, mitochondrial dysfunction and proteostatic imbalance through interconnected feedback mechanisms (Aran and Singh, 2023; Lévy et al., 2019; Ornatowski et al., 2020).

Given these cyclical and intricately cross-regulated sub-networks, their deeper mechanistic exploration led to the identification of four key components: Nuclear factor erythroid 2-related factor 2 (Nrf2)/Kelch-like ECH-associated protein 1 (Keap1), mammalian target of rapamycin complex 1 (mTORC1), Sirtuin 1 (SIRT1) and AMP-activated protein kinase (AMPK). Nrf2, a master regulator of antioxidant response, is normally sequestered in the cytoplasm by Keap1, leading to its ubiquitination and proteasomal degradation (Ngo and Duennwald, 2022). Under oxidative stress, Nrf2 dissociates from Keap1, translocates to the nucleus and activates transcription of antioxidant enzymes and proteasomal components, regulating both redox balance and proteostasis (Monsalvo-Maraver et al., 2024). Conversely, mTORC1 hyperactivation suppresses autophagy, causing buildup of dysfunctional mitochondria and protein aggregates (Guillén and Benito, 2018; Querfurth and Lee, 2021); its inhibition, for instance by rapamycin, has extended lifespan in model organisms and remains a key geroprotective target (Konopka et al., 2023; Mannick and Lamming, 2023). In contrast, SIRT1 and AMPK regulate oxidative stress, mitochondrial homeostasis, proteostasis, and autophagy through multiple interconnected pathways and also form a positive feedback loop (Abu-Baih et al., 2024; Marino et al., 2021; Ottens et al., 2021; Ruderman et al., 2010; Sacitharan et al., 2020; Singh et al., 2018; Wang et al., 2022; You and Liang, 2023). Together with Nrf2 activation and mTORC1 inhibition, these regulators collectively coordinate antioxidant defense, autophagy, mitochondrial quality control and proteostasis through interconnected mechanisms involving lysosomal biogenesis, mitophagy, proteasomal regulation and cellular stress-response pathways (Armeli et al., 2024; Białopiotrowicz-Data et al., 2023; Esteras and Abramov, 2022; Garcia and Shaw, 2017; Herzig and Shaw, 2018; Huang et al., 2024; Laplante and Sabatini, 2012; Lee, 2019; Monsalvo-Maraver et al., 2024; Ngo and Duennwald, 2022; Panwar et al., 2023). While each of these regulators can independently modulate aging-associated cellular processes, their mechanistic overlap and coordinated regulation suggest that simultaneous activation of Nrf2, SIRT1, and AMPK alongside inhibition of mTORC1 could potentially induce synergistic restoration of cellular homeostasis across multiple ARDs through integrated modulation of interconnected aging-associated pathways (Fig. 1).

FIG. 1.

Synergistic interplay associated with concurrent Nrf2 activation, mTORC1 inhibition and SIRT1 and AMPK activation. The solid lines represent a direct role in the regulation of a particular process, whereas dashed lines represent indirect regulation. The curved arrows between SIRT1 activation and AMPK activation represent a positive feedback loop. AMPK, AMP-activated protein kinase; SIRT1, sirtuin 1.

This cooperative engagement of geroprotective pathways forms a strong conceptual foundation for combinatorial therapeutic strategies. Targeting these four regulators in unison offers a promising approach to mitigate age-associated cellular dysfunction, thereby restoring cellular homeostasis and ultimately inducing comprehensive therapeutic effects in multiple ARDs. And in line with this therapeutic strategy, the present study introduces a computational framework developed to screen potential geroprotective combinations of natural compounds (phytochemicals) with the ability to simultaneously target all the aforementioned factors including Nrf2, mTORC1, SIRT1, and AMPK. Here, machine learning-based geroprotector classification and molecular docking were conducted to identify natural compounds with geroprotective characteristics and computationally predicted ability to concurrently engage with all the desired targets in order to modulate their intricate network. Subsequently, a graph neural network (GNN)-based synergy prediction model was developed to predict potential combinations among the screened natural compounds.

Materials and Methods

Dataset preparation, data preprocessing, and machine learning-based classification

A dataset comprising 466 FDA-approved drugs was constructed using the DrugBank database and thorough literature review (Knox et al., 2024). The dataset included drugs approved for treating ARDs, such as neurological disorders, type 2 diabetes, heart diseases and chronic kidney disease, as well as non-ARDs (NARDs), including infectious diseases, cancer and various genetic disorders. The drugs were assigned numerical codes based on their application in different diseases, resulting in a dataset with 466 FDA-approved drugs labeled with numerical values including “0” for NARDs and 1’ for ARDs (Supplementary Table S1). Thereafter, 1024-bit molecular fingerprints were calculated for all the drugs using RDKit and these fingerprints were then used as features to train the classification model (Landrum and Greg, 2006). Fingerprints were standardized by z-score normalization and dimensionality reduction was performed using Principal Component Analysis (PCA), retaining 50 components that captured the majority of dataset variance while minimizing noise and improving computational efficiency. The dataset was split into training and testing sets in an 80:20 ratio and multiple classifiers, including Random Forest, Decision Tree, Support Vector Machine (SVM), Logistic Regression, k-Nearest Neighbors (KNN), Gradient Boosting, and XGBoost, were trained and evaluated (Chen and Guestrin, 2016; Pedregosa et al., 2011). Model performance was assessed through classification metrics and receiver operating characteristic (ROC) analysis, complemented by five-fold stratified cross-validation for robustness evaluation. In addition, external validation was performed using a dataset of 144 geroprotectors retrieved from the Geroprotectors.org database, which contains compounds with experimentally reported lifespan-extending and geroprotective effects across various ARDs (Moskalev et al., 2015). Only those compounds were selected that were not present in the training or test datasets to ensure an unbiased evaluation. Molecular fingerprints were generated for all compounds, followed by the same preprocessing pipeline used during model development, including normalization and PCA. The trained model was then applied to this dataset and its performance was evaluated in terms of prediction accuracy, representing the coverage of known geroprotectors.

Chemical space projection of FDA-approved ARD drugs and predicted geroprotective natural compounds

In order to assess potential distribution shift between the FDA-approved ARD drugs used for training the classification model and natural compounds predicted as geroprotectors and to further validate the robustness of the geroprotector classification model, chemical space projection was performed using Uniform Manifold Approximation and Projection (UMAP) (McInnes and Healy, 2018). Molecular fingerprints were used as representations/features for each compound and were then projected into a two-dimensional plot using UMAP to visualize their relative distribution and structural relationships. This analysis was used to examine the overlap, proximity, and potential extensions between the two groups, thus generating the qualitative assessment of model generalizability across structurally diverse compounds.

Screening of natural compounds for drug-likeness

After screening the natural compounds for their geroprotective roles using the classification model, the natural compounds predicted as geroprotectors were further evaluated for drug-likeness using SWISS ADME (Daina et al., 2017). Compounds were analyzed for their adherence to Lipinski’s Rule of Five and those with any violations were excluded (Lipinski et al., 2001).

Molecular docking of natural compounds with geroprotective targets

The 3D structures of the target proteins were retrieved from the Protein Data Bank (PDB) (Berman et al., 2000), with PDB IDs as follows: Keap1 (PDB ID: 6QME), mTORC1 (PDB ID: 8ERA), AMPK (PDB ID: 5KQ5) and SIRT1 (PDB ID: 4ZZH). Protein preparation was performed in PyMOL (Schrödinger LLC, 2015), involving the removal of water molecules, bound ligands, and unnecessary chains, along with the addition of polar hydrogens.

The 3D conformers of the natural compounds were obtained from PubChem (Kim et al., 2025), and docking was conducted using the AutoDock Vina program (Trott and Olson, 2010) within the PyRx GUI (Dallakyan and Olson, 2015). The grid box parameters for each protein were as follows:

Keap1: Center (x = 11.18, y = 61.74, z = 38.03), Dimensions: 25 × 25 × 25

mTORC1: Center (x = 109.56, y = 220.50, z = 151.21), Dimensions: 25 × 25 × 25

AMPK: Center (x = 53.10, y = −78.25, z = −41.90), Dimensions: 25 × 25 × 25

SIRT1: Center (x = 16.62, y = −48.34, z = 15.40), Dimensions: 25 × 25 × 25

Binding affinities (in kcal/mol) were calculated for the natural compounds across the four proteins, providing critical insights into their potential as geroprotectors. Top candidates were then identified by applying a binding affinity filter. The selected candidates were subsequently subjected to site-specific redocking validation using BIOVIA Discovery Studio (DASSAULT SYSTÈMES, 2025).

Drug pair dataset preparation for GNN-based synergy prediction model

A dataset of 27 drug combinations either approved or under investigation for the treatment of various ARDs was curated (Table 1).

Table 1.

Drug Combinations (Approved/under Investigation) Used in the Treatment of ARDs

Drug_1	Drug_2	Disease(s)	Reference(s)
Carbidopa	Levodopa	Parkinson’s disease	Boshes, 1981
Carbidopa	Entacapone	Parkinson’s disease	Hauser, 2004
Levodopa	Entacapone	Parkinson’s disease	Hauser, 2004
Donepezil	Memantine	Alzheimer’s disease	Yaghmaei et al., 2024
Dapagliflozin	Metformin	Type 2 diabetes	Maksud et al., 2024
Empagliflozin	Metformin	Type 2 diabetes	Goldman, 2018
Linagliptin	Metformin	Type 2 diabetes	Haak, 2015
Sitagliptin	Metformin	Type 2 diabetes	St Onge et al., 2012
Saxagliptin	Metformin	Type 2 diabetes	Minze et al., 2013
Pioglitazone	Metformin	Type 2 diabetes	Derosa and Salvadeo, 2008
Alogliptin	Metformin	Type 2 diabetes	Neumiller and Holland, 2016
Canagliflozin	Metformin	Type 2 diabetes	Davidson and Sloan, 2017
Semaglutide	Metformin	Type 2 diabetes	Ren et al., 2025
Exenatide	Metformin	Type 2 diabetes	Derosa et al., 2013
Sacubitril	Valsartan	Heart diseases	Kaplinsky, 2016
Atorvastatin	Ezetimibe	Heart diseases	Luo et al., 2016
Rosuvastatin	Ezetimibe	Heart diseases	Yang et al., 2017
Sitagliptin	Ertugliflozin	Type 2 diabetes	Miller et al., 2018
Metformin	Ertugliflozin	Type 2 diabetes	Athyros et al., 2021
Dapagliflozin	Saxagliptin	Type 2 diabetes	Williams and Stephens, 2015
Metformin	Rapamycin	Type 2 diabetes, neurodegenerative disorders	Norradee et al., 2023; Reifsnyder et al., 2022
Metformin	Acarbose	Type 2 diabetes	Wang et al., 2013
Rapamycin	Acarbose	Neurodegenerative disorders	Wezeman et al., 2024; Strong et al., 2022
Dasatinib	Quercetin	Diabetic kidney disease	Guo et al., 2024
Fisetin	Quercetin	Type 2 diabetes	Iwasaki et al., 2023
Nicotinamide Mononucleotide	Resveratrol	Heart diseases	Bai et al., 2022
Metformin	Resveratrol	Type 2 diabetes, nonalcoholic fatty-liver disease	Afshari et al., 2023; Frendo-Cumbo et al., 2016

Thereafter, 1024-bit molecular fingerprints of each drug in the combination were calculated using RDKit (Landrum and Greg, 2006), which were used as node features and the known combination information as edges to build a GNN-based synergy prediction model in order to predict potential combinations among the screened natural compounds.

GNN-based geroprotective combinations prediction

The graph was constructed using PyTorch Geometric’s Data object (Fey and Lenssen, 2019), with molecular fingerprint vectors as node features and undirected edge indices (from known drug pair combination knowledge). A three-layer graph convolutional network (GCN) was defined to learn node embeddings of the compounds, and an edge prediction model was built on top, incorporating the GCN and a fully connected layer that concatenated embeddings of compound pairs and applied a sigmoid activation to predict the likelihood of an edge between them (edge probability) as synergy probability (Kipf and Welling, 2016). The model was trained using binary cross-entropy loss, optimized with AdamW, and stabilized by early stopping. To guarantee reproducibility, fixed random seeds were applied across NumPy, PyTorch, and CUDA backends (Pedregosa et al., 2011). Rather than selecting negative samples at random, non-synergistic pairs were generated through a hard-negative strategy in which compound pairs were ranked by the cosine similarity of their molecular fingerprints and negatives were drawn from this similarity-based distribution. This ensured that the negative set was chemically informed and avoided trivial random assignment, compelling the model to learn deeper and meaningful relational features. Model’s robustness was evaluated using five-fold stratified cross-validation, reporting accuracy, precision, recall, F1-score, and ROC–area under curve (AUC) values for each fold.

To benchmark the proposed GNN-based synergy prediction framework against a simpler baseline, a similarity-based approach was implemented using cosine similarity between molecular fingerprints of compounds. Specifically, 1024-bit molecular fingerprints were used to compute pair-wise cosine similarity scores between drug pairs, which were then directly treated as proxy scores for synergy likelihood. A threshold-based classification was applied to derive binary predictions and performance was evaluated using standard ROC–AUC. This baseline serves as a chemically intuitive reference model that relies solely on structural similarity, without incorporating network topology or learned representations, thereby enabling assessment of the added value provided by the GNN-based framework.

Furthermore, the model was subjected to external validation using an independently curated dataset of geroprotective drug combinations with experimentally reported therapeutic synergy in various ARD models. This dataset was compiled through an extensive literature review. Molecular fingerprints of the constituent compounds were used as node features, while the experimentally validated combinations were represented as edges. The trained GNN model was then applied to this dataset by embedding each compound into the drug-drug network through connections to structurally similar drugs, enabling integration of new nodes into the existing graph and subsequent generation of node embeddings for the unseen compounds. This approach allows the GNN model to propagate information across the network through message passing and facilitates the generation of context-aware embeddings, which is consistent with prior graph-based frameworks that integrate similarity-driven neighborhood construction with network-based learning (Peng et al., 2022; Zhao et al., 2025). The expanded graph was then processed through the trained GNN model to generate node embeddings, which were subsequently used to compute synergy probabilities for each combination, enabling assessment of its ability to generalize beyond the training data. Once validated, the trained model was applied to the set of screened natural compounds and all possible pairs among them. Each compound was similarly embedded into the drug–drug network by connecting it to structurally similar drugs and the expanded graph was processed through the trained GNN to generate node embeddings. All possible natural compound pairs were then evaluated by the synergy prediction model, producing synergy probability scores for each pair and a highly stringent probability threshold (≥0.999) was applied to identify the most promising natural compound combinations.

Visualization of predicted natural compound interaction network

The graph of natural compound pairs predicted by the GNN-based synergy prediction model was visualized using NetworkX (Hagberg et al., 2008).

Results and Discussion

The machine learning-based geroprotector classification model was trained on 1024 molecular fingerprints of FDA-approved drugs for treatment of ARDs, as the positive control set and drugs for NARDs, as the negative control set (Supplementary Table S1). Multiple iterations were conducted for different classifiers, including Random Forest, Decision Tree, SVM, Logistic Regression, XGBoost, KNN, and Gradient Boosting, in order to find an ideal classifier (Chen and Guestrin, 2016; Pedregosa et al., 2011). Consequently, XGBoost emerged as the most effective classifier, achieving an ROC–AUC value of 0.91 and robust performance metrics, indicating strong discriminative ability in distinguishing between the two classes (Fig. 2a and Table 2).

FIG. 2.

(a) Receiver operating characteristic (ROC) curves of different machine learning classifiers. The XGBoost classifier achieved the highest area under the curve (AUC = 0.91), whereas other classifiers exhibited comparatively lower performance. The higher ROC–AUC of XGBoost highlights its robust ability to balance sensitivity and specificity for geroprotector classification. (b) Comparative performance of machine learning classifiers evaluated using five-fold stratified cross-validation. The heatmap summarizes mean values for accuracy, precision, recall, and F1-score across folds. Among the tested models, the XGBoost classifier consistently achieved the best overall balance of metrics (highlighted in red), implying its robustness and suitability for geroprotectors classification. (c) UMAP projection of molecular fingerprints illustrating the chemical space distribution of FDA-approved ARD drugs (blue) and 69 predicted geroprotective natural compounds (green). The substantial overlap between the two groups indicates that a majority of the natural compounds reside within the applicability domain of the model. The absence of clearly separated clusters suggests minimal distribution shift and supports the model’s generalizability across structurally diverse compounds. In addition, natural compounds interspersed within the dense clusters of ARD drugs highlight the presence of shared feature space and local structural similarity with clinically validated therapeutics. Moreover, the subset of natural compounds occupying adjacent yet continuous regions that remain proximal to the distribution of FDA-approved ARD drugs indicates the presence of novel scaffolds of potential therapeutic significance.

Table 2.

Classification Report of the XGBoost Classifier

Class/Metric	Precision	Recall	F1-score	Support
0	0.803	0.891	0.845	55
1	0.818	0.692	0.750	39
Accuracy	0.809			94
Macro average	0.811	0.792	0.797	94
Weighted average	0.809	0.809	0.805	94

In order to validate the model’s (XGBoost classifier) performance, five-fold stratified cross-validation was conducted, which resulted in mean accuracy of 0.800, mean precision of 0.803, mean recall of 0.800, and mean F1-score of 0.798 across folds (Fig. 2b). In addition, the model was evaluated on an independent dataset of experimentally validated geroprotectors retrieved from the Geroprotectors.org database (Moskalev et al., 2015). The model achieved an accuracy of 0.861, correctly identifying 124 out of 144 geroprotectors, with an average confidence score of 0.764 (Supplementary Table S2). Notably, well-established geroprotectors were predicted with significantly high confidence, including Resveratrol (confidence score = 0.977), Quercetin (confidence score = 0.853), and Curcumin (confidence score = 0.976) (Supplementary Table S2) (Proshkina et al., 2024; Rivero-Segura et al., 2024). These results indicate that the model is capable of capturing key features associated with geroprotective activity and maintains predictive consistency on an independent dataset as well, thereby supporting its robustness and applicability for identifying potential geroprotectors. Subsequently, the dataset of 1152 natural compounds, retrieved from PubChem using the “Phytochemicals” search query (Supplementary Table S3), with their molecular fingerprints, was fed to the trained and validated geroprotector classification model. The model predicted 69 natural compounds to be potentially therapeutic for ARDs such as neurological disorders, type 2 diabetes, heart diseases, and chronic kidney disease with a confidence score ≥0.99 (Supplementary Table S4).

To assess potential distribution shift, UMAP was applied to the molecular fingerprints of FDA-approved ARD drugs used for training the classification model and 69 predicted geroprotective natural compounds (Supplementary Tables S1 and S4) (McInnes and Healy, 2018). The projection showed significant overlap between the two groups, with no clearly isolated clusters observed for the natural compounds (Fig. 2c). This indicates that the model operates within a relevant applicability domain and is not extrapolating into entirely unseen chemical regimes, thereby supporting its reliability when applied to structurally distinct natural compounds (Fig. 2c). In addition, several natural compounds were found to be interspersed within the dense clusters of ARD drugs, highlighting the presence of shared feature space and local structural similarity with clinically validated therapeutics (Fig. 2c). Moreover, a subset of natural compounds was observed to occupy adjacent yet continuous regions that remain proximal to the distribution of FDA-approved ARD drugs, which indicates the presence of novel scaffolds of potential therapeutic importance (Fig. 2c). Collectively, these findings demonstrate that the model achieves a balance between generalization and specificity. The observed overlap suggests that many predicted geroprotective natural compounds possess ARD drug-like structural and physicochemical properties, while the adjacent extensions indicate the presence of potential scaffolds of therapeutic significance not widely explored in current ARD therapeutics. These natural compounds were further tested for their drug-likeness using SWISS ADME and 55 out of the 69 natural compounds met the criteria with zero violations of Lipinski’s Rule of Five and were selected for further screening (Daina et al., 2017; Lipinski et al., 2001).

In order to find natural compounds with the ability to simultaneously engage all the targets of interest including Nrf2, mTORC1, SIRT1, and AMPK and potentially, induce a comprehensive and synergistic geroprotective effect, molecular docking was conducted directly with the targets or their regulators. In order to activate Nrf2, molecular docking was conducted with its negative regulator Keap1 at the site involved in its interaction with Nrf2 (Heightman et al., 2019), whereas for mTORC1, docking was done at the site involved in known bisteric inhibitor binding (Burnett et al., 2023) and for AMPK and SIRT1, the molecular docking was done at the site involved in known activator binding (Cameron et al., 2016; Dai et al., 2015). The grid box parameters for each target protein are described in Section Materials and Methods. Subsequently, binding affinities of all 55 natural compounds were obtained (Supplementary Table S5). By applying a binding affinity threshold of ≤ –7.5 for Keap1, mTORC1, and AMPK, 12 natural compounds were identified with comparatively stronger binding affinities across all four targets, suggesting their potential to simultaneously modulate the activity of intricately linked geroprotective regulators (Table 3).

Table 3.

Natural Compounds with Most Significant Binding Affinities with All Four Target Proteins

Name	Compound CID	Keap1 binding affinity	mTORC1 binding affinity	AMPK binding affinity	SIRT1 binding affinity
Phloretin	4788	−8.1	−7.9	−7.7	−5.5
Magnolol	72300	−7.6	−8.4	−7.9	−5.8
Daphnoretin	5281406	−9.5	−9.9	−9.1	−6.1
Baicalein	5281605	−9.2	−9.4	−9	−5.7
Pectolinarigenin	5320438	−8.8	−9.6	−8.6	−5.5
Demethoxycurcumin	5469424	−8	−8.8	−7.9	−5.8
Bergamottin	5471349	−7.9	−9.2	−7.6	−6.1
20(S),24(R)-Ocotillol	15886258	−8.4	−8	−7.5	−5.6
Camptothecin	24360	−10	−8.7	−10	−5.7
Rubitecan	472335	−10.7	−9.3	−9.3	−5.7
Calycosin 7-o-beta-D-glucoside	5318267	−9	−9.5	−9.2	−5.5
Dehydrodiisoeugenol	5379033	−8.1	−8.1	−7.2	−5.5

In addition, site-specific redocking validation was performed to assess the structural plausibility of the predicted interactions. For each target protein, the cocrystallized ligand, including J6Q (inhibitor) for Keap1, XYU (inhibitor) for mTORC1, 6VT (allosteric activator) for AMPK, and 4TO (activator) for SIRT1 was redocked into its respective binding site using the same docking parameters in BIOVIA Discovery Studio (Berman et al., 2000; Burnett et al., 2023; Cameron et al., 2016; Dai et al., 2015; DASSAULT SYSTÈMES, 2025; Heightman et al., 2019). Redocking was performed both individually and in the presence of the 12 screened natural compounds, respectively, for each target protein, in order to confirm the colocalization of natural compounds with the cocrystallized ligand within the same binding pocket of the target protein. The resulting binding poses revealed that these compounds consistently occupied the same functional binding regions as the respective co-crystallized ligands, despite structural diversity (Fig. 3). Notably, the natural compounds exhibited spatial proximity and orientation within the desired sites comparable to the standard/reference ligands, indicating specific and structurally consistent interactions (Fig. 3). These findings support the potential of the identified natural compounds to significantly engage with the target proteins and provide a rational basis for further experimental validation to confirm these computationally derived inferences.

FIG. 3.

Docking validation showing co-localization of the 12 screened natural compounds with co-crystallized ligand within the binding pocket of the target protein. (a) SIRT1 with co-crystallized ligand 4TO (activator); (b) SIRT1 with co-crystallized ligand 4TO (activator) and the 12 natural compounds; (c) AMPK with co-crystallized ligand 6VT (allosteric activator); (d) AMPK with co-crystallized ligand 6VT (allosteric activator) and the 12 natural compounds; (e) Keap1 with co-crystallized ligand J6Q (inhibitor); (f) Keap1 with co-crystallized ligand J6Q (inhibitor) and the 12 natural compounds; (g) mTORC1 with co-crystallized ligand XYU (inhibitor); and (h) mTORC1 with co-crystallized ligand XYU (inhibitor) and the 12 natural compounds. The observed co-localization within the experimentally defined binding pockets suggests that the natural compounds exhibit binding modes consistent with known ligand-target interactions, supporting their potential for specific and robust engagement with the target proteins.

Identified as potential geroprotectors by a well-trained and validated classification model, these 12 natural compounds exhibited zero violations of Lipinski’s Rule of Five, indicating favorable drug-like properties. Subsequent molecular docking analysis revealed their potential to regulate the core factors implicated in various ARDs. Taken together, these computational findings position the identified natural compounds as candidates with the potential to simultaneously engage key geroprotective factors and modulate their interconnected network, to ultimately induce a comprehensive and coordinated therapeutic effect across multiple ARD models, subject to further experimental validation.

Owing to the importance of combinatorial strategies, particularly in therapeutic interventions focused on multiple cross-regulatory target engagement, such as the one presented here, the study further explored potential geroprotective combinations among the screened 12 natural compounds using GNN (GCN, in particular) (Kipf and Welling, 2016). A GNN model was trained on 27 drug combinations either approved or under investigation for treatment of various ARDs (Table 1), where 1024-bit molecular fingerprints calculated for each individual drug were provided as node features and known combination information was used as edges, consequently generating 27 such edges. As discussed earlier, the GNN model was trained to predict synergy probabilities and was evaluated periodically for loss convergence using binary cross-entropy loss. The model showed a consistent decrease in loss over the epochs and subsequently the loss flattened close to zero, suggesting convergence and thus indicating effective learning and model stability (Fig. 4a). This consistent minimization of loss suggests that the model successfully captured the underlying structural patterns of compound-compound interactions. The model was further evaluated for its robustness and reliability via five-fold stratified cross-validation, where the model displayed consistent performance across each fold and achieved noteworthy performance metrics, including mean accuracy of 0.8143 ± 0.0806, precision of 0.7761 ± 0.1927, recall of 0.9124 ± 0.0978, F1 score of 0.8135 ± 0.1146 and ROC-AUC of 0.9151 ± 0.0856 (Fig. 4b). In order to further validate the model’s robustness, an independent dataset was curated through a comprehensive literature review, consisting of 15 experimentally validated combinations reported to induce significant therapeutic effects in various ARD models. The GNN-based synergy prediction model was then applied to this independent dataset and it correctly identified 11 out of 15 geroprotective combinations with a synergy probability ≥ 0.872 for each predicted combination (Table 4). This level of recovery indicates that the model is capable of capturing key ARD-specific interaction patterns and generalizing beyond the training dataset, thus supporting its applicability in computationally identifying potential synergistic combinations for various ARDs.

FIG. 4.

(a) Loss versus Epoch curve for the GNN-based synergy prediction model. The plot demonstrates rapid convergence, with the loss ultimately plateauing near zero, indicating effective learning and model stability. (b) Five-fold stratified cross-validation performance of the GNN-based synergy prediction model. The following boxplot summarizes accuracy, precision, recall, F1-score, and ROC–AUC across folds, showing that all metrics consistently lie between 0.8 and 1.0 range, indicating strong model stability and generalizability in predicting potential geroprotective combinations.

Table 4.

External Validation of the GNN-Based Synergy Prediction Model Using an Independent Dataset of Experimentally Validated Geroprotective Combinations across Multiple ARD Models.

Compound A	Compound B	Synergy probability	Reference (s)
Curcumin	Quercetin	0.999996305	Hasan et al., 2022; Matacchione et al., 2022; Chittasupho et al., 2021
Resveratrol	Quercetin	0.999962091	Zhao et al., 2017
Epigallocatechin gallate	Quercetin	0.999531865	Liu et al., 2021; Cialdella-Kam et al., 2017
Silymarin	Chlorogenic acid	0.999105871	Abbas et al., 2020; Al-Rasheed et al., 2014
Resveratrol	Curcumin	0.998350859	Le Boeuf et al., 2026; Zaky et al., 2017
Metformin	Sirolimus	0.996139586	Reifsnyder et al., 2022
Epigallocatechin gallate	Curcumin	0.977092624	Prathumwon et al., 2025; Rahman et al., 2023
Epigallocatechin gallate	Chlorogenic acid	0.937843621	Wei et al., 2023
Berberine	Metformin	0.927581966	Li et al., 2025; Lyu et al., 2022
Resveratrol	Sirolimus	0.895130873	Leontieva et al., 2013
Spermidine	Sirolimus	0.872257531	Hofer et al., 2024
Metformin	Chlorogenic acid	0.108994327	Khalafani et al., 2023; Zamani-Garmsiri et al., 2021
Curcumin	Piperine	0.020285396	Chakraborty et al., 2017
Aspirin	L-Ascorbic acid	0.000481732	Candelario-Jalil et al., 2006
Silymarin	Berberine	1.35713E-05	Fogacci et al., 2019

The table lists compound pairs along with their predicted synergy probabilities and supporting literature references. The model successfully identified 11 out of 15 combinations with high synergy probabilities (≥0.872), demonstrating its ability to capture core ARD-specific interaction patterns and generalize beyond the training dataset

Despite its relatively simple architecture and limited training data, the model was able to identify several experimentally validated combinations with high confidence and showed consistent performance across cross-validation and external validation datasets. This suggests that the learned node embeddings were able to capture meaningful structural and relational patterns within the drug-drug interaction network. By leveraging a graph convolutional architecture, the model encodes relationships between compounds through both feature representation and network structure. While more advanced synergy prediction models utilize larger datasets (Liu et al., 2024; Torkamannia et al., 2023), the current results indicate that even a relatively lightweight GNN-based approach can capture relevant interaction patterns between compounds when guided by appropriate structural representations and network context. To further contextualize model performance, a similarity-based baseline was evaluated using cosine similarity between molecular fingerprints as a proxy for synergy prediction. The baseline approach yielded substantially lower performance (ROC-AUC = 0.555), indicating limited discriminative ability when relying solely on chemical similarity (Supplementary Fig. S1). In contrast, the GNN-based framework achieved a markedly higher ROC-AUC (0.915), suggesting that the proposed approach is able to capture interaction patterns that are not readily reflected by structural similarity alone (Supplementary Fig. S1). Collectively, these findings highlight the potential of GNN-based frameworks as tools for exploring synergistic interactions even in data-constrained settings, supporting their practical utility in early-stage computational screening, particularly in scenarios where large-scale datasets are not readily available. Nevertheless, as a computational framework, this approach requires further validation through experimental studies to further enhance its reliability in early-stage screening, particularly in comparison to more data-intensive models.

Following successful training and validation, the GNN-based synergy prediction model was applied to the dataset of 12 natural compounds and all 66 possible combinations of compound pairs were generated for synergy probability prediction. The model successfully predicted the synergy probabilities between all the natural compound pairs (Supplementary Table S6). To identify high-confidence predictions, a stringent screening threshold of synergy probability ≥ 0.999 was applied, which yielded five natural compound pairs involving four distinct natural compounds, including Baicalein, Pectolinarigenin, Phloretin and Demethoxycurcumin (Table 5).

Table 5.

Top Natural Compound Pairs with Synergy Probability ≥0.999

Compound A	Compound B	Synergy probability
Baicalein	Pectolinarigenin	0.999990821
Phloretin	Baicalein	0.999910355
Phloretin	Pectolinarigenin	0.999909639
Baicalein	Demethoxycurcumin	0.999816716
Pectolinarigenin	Demethoxycurcumin	0.999809444

The graph of all 66 natural compound pairs formed using synergy probabilities predicted by the GNN-based synergy prediction model was visualized using NetworkX, where synergy probabilities were used as edge weights (Fig. 5).

FIG. 5.

Graph visualization of predicted natural compound combinations. Nodes shown in green represent compound pairs with a synergy probability ≥0.999, connected by solid edges, while the remaining nodes are shown in gray and are connected by dashed edges, indicating lower predicted synergy.

Conclusion

This study presents a systems biology-informed computational framework for the identification of natural compound combinations with the potential to modulate multiple interconnected geroprotective regulators across diverse ARDs. By integrating machine learning-based compound classification, molecular docking for multi-target interaction assessment and GNN-based synergy prediction, the proposed framework provides a systematic and scalable approach for early computational screening of geroprotective candidates and their potential combinations. The present study identified 12 natural compounds as promising candidates with computationally predicted geroprotective capability, favorable drug-like properties and the ability to interact with multiple targets of therapeutic significance, including Nrf2, mTORC1, AMPK, and SIRT1. While these findings suggest possible multi-target engagement, the proposed effects on cellular processes such as oxidative stress regulation, autophagy, proteostasis, and mitochondrial function remain computationally inferred and require experimental validation. To further investigate combinatorial potential, a GNN-based synergy prediction model was employed to identify potential geroprotective combinations among the screened natural compounds. Despite a relatively small dataset and simple architecture, the model demonstrated consistent performance across cross-validation and external validation and was able to recover several high-confidence combinations during validation, suggesting that the learned embeddings capture ARD-specific interaction patterns within the network, likely driven by the GCN-based encoding of relationships between compounds through both feature representation and network structure.

Nevertheless, a set of high-confidence natural compound combinations was identified using this framework, involving Baicalein, Pectolinarigenin, Phloretin, and Demethoxycurcumin. These combinations align with the proposed multi-target geroprotective framework and show the potential to elicit comprehensive and coordinated geroprotective effects. Future improvements could be achieved by incorporating larger datasets and in vitro/in vivo validation to test biological efficacy for enhanced robustness and reliability of our approach.

Authors’ Contributions

Y.S.: Conceptualization, data curation, methodology, writing—original draft. A.D.: Supervision, writing—review and editing, validation.

Data Availability Statement

All data supporting the findings of this study are included within the article and its Supplementary Data. Additional related data are available from the corresponding author upon reasonable request.

Supplemental Material

sj-docx-1-omi-10.1177_15578100261464019 — Supplemental material for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Supplemental material, sj-docx-1-omi-10.1177_15578100261464019 for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks by Yuvraj Sharma, and Asmita Das

Supplemental Material

sj-xlsx-2-omi-10.1177_15578100261464019 — Supplemental material for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Supplemental material, sj-xlsx-2-omi-10.1177_15578100261464019 for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks by Yuvraj Sharma, and Asmita Das

Supplemental Material

sj-xlsx-3-omi-10.1177_15578100261464019 — Supplemental material for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Supplemental material, sj-xlsx-3-omi-10.1177_15578100261464019 for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks by Yuvraj Sharma, and Asmita Das

Supplemental Material

sj-xlsx-4-omi-10.1177_15578100261464019 — Supplemental material for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Supplemental material, sj-xlsx-4-omi-10.1177_15578100261464019 for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks by Yuvraj Sharma, and Asmita Das

Supplemental Material

sj-xlsx-5-omi-10.1177_15578100261464019 — Supplemental material for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Supplemental material, sj-xlsx-5-omi-10.1177_15578100261464019 for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks by Yuvraj Sharma, and Asmita Das

Supplemental Material

sj-xlsx-6-omi-10.1177_15578100261464019 — Supplemental material for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Supplemental material, sj-xlsx-6-omi-10.1177_15578100261464019 for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks by Yuvraj Sharma, and Asmita Das

Supplemental Material

sj-xlsx-7-omi-10.1177_15578100261464019 — Supplemental material for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks

Supplemental material, sj-xlsx-7-omi-10.1177_15578100261464019 for Synergistic Geroprotectors Mapping through Systems Machine Learning and Graph Neural Networks by Yuvraj Sharma, and Asmita Das

Footnotes

Acknowledgments

The present study has been carried out in the Department of Biotechnology, Delhi Technological University, following all ethical principles of the university. The authors acknowledge the Junior Research Fellowship (JRF) awarded to Y.S. by Council of Scientific and Industrial Research (CSIR), India.

Author Disclosure Statement

Authors declare no conflict of interest.

Funding Information

The authors declare that no funding or grants were received to support the conduct of this study.

Abbreviations Used

References

Abbas

NAT

, Awad

, Nafea

. Silymarin in combination with chlorogenic acid protects against hepatotoxicity induced by doxorubicin in rats: Possible role of adenosine monophosphate-activated protein kinase pathway. Toxicol Res (Camb) 2020;9(6):771–777; doi: 10.1093/toxres/tfaa080

Abu-Baih

, Abu-Baih

, Abdel-Hafez

SMN

, Fathy

. Activation of SIRT1/Nrf2/HO-1 and Beclin-1/AMPK/MTOR autophagy pathways by eprosartan ameliorates testicular dysfunction induced by testicular torsion in rats. Sci Rep 2024;14(1):12566; doi: 10.1038/s41598-024-62740-6

Afshari

, Noori

, Zarghi

. A novel combination of metformin and resveratrol alleviates hepatic steatosis by activating autophagy through the CAMP/AMPK/SIRT1 signaling pathway. Naunyn Schmiedebergs Arch Pharmacol 2023;396(11):3135–3148; doi: 10.1007/s00210-023-02520-7

Al-Rasheed

, Al-Rasheed

, Faddah

, Mohamed

, Mohammad

, Al-Amin

. Potential impact of silymarin in combination with chlorogenic acid and/or melatonin in combating cardiomyopathy induced by carbon tetrachloride. Saudi J Biol Sci 2014;21(3):265–274; doi: 10.1016/j.sjbs.2013.09.006

Aran

, Singh

. Mitochondrial dysfunction and oxidative stress in Alzheimer’s disease–a step towards mitochondria based therapeutic strategies. Aging Health Res 2023;3(4):100169; doi: 10.1016/j.ahr.2023.100169

Armeli

, Mengoni

, Laskin

, Businaro

. Interplay among oxidative stress, autophagy, and the endocannabinoid system in neurodegenerative diseases: Role of the Nrf2-P62/SQSTM1 pathway and nutraceutical activation. Curr Issues Mol Biol 2024;46(7):6868–6884; doi: 10.3390/cimb46070410

Athyros

, Boutari

, Karagiannis

. Ertugliflozin + metformin as a treatment option for type 2 diabetes. Expert Opin Pharmacother 2021;22(16):2105–2111; doi: 10.1080/14656566.2021.1939676

Bai

L-B

, Yau

L-F

, Tong

T-T

, Chan

W-H

, Zhang

, Jiang

Z-H

. Improvement of tissue-specific distribution and biotransformation potential of nicotinamide mononucleotide in combination with ginsenosides or resveratrol. Pharmacol Res Perspect 2022;10(4):e00986; doi: 10.1002/prp2.986

Berman

, Westbrook

, Feng

, et al. The protein data bank. Nucleic Acids Res 2000;28(1):235–242; doi: 10.1093/nar/28.1.235

10.

Białopiotrowicz-Data

, Noyszewska-Kania

, Jabłońska

, et al. SIRT1 and HSP90α feed-forward circuit safeguards chromosome segregation integrity in diffuse large B cell lymphomas. Cell Death Dis 2023;14(10):667; doi: 10.1038/s41419-023-06186-0

11.

Boshes

. Sinemet and the treatment of parkinsonism. Ann Intern Med 1981;94(3):364–370; doi: 10.7326/0003-4819-94-3-364

12.

Burnett

, Yang

, Aggen

, et al. Discovery of RMC-5552, a selective Bi-steric inhibitor of MTORC1, for the treatment of MTORC1-activated tumors. J Med Chem 2023;66(1):149–169; doi: 10.1021/acs.jmedchem.2c01658

13.

Cameron

, Kung

, Kalgutkar

, et al. Discovery and preclinical characterization of 6-chloro-5-[4-(1-hydroxycyclobutyl)phenyl]-1H-indole-3-carboxylic acid (PF-06409577), a direct activator of Adenosine Monophosphate-Activated Protein Kinase (AMPK), for the potential treatment of diabetic NEPH. J Med Chem 2016;59(17):8068–8081; doi: 10.1021/acs.jmedchem.6b00866

14.

Candelario-Jalil

, Akundi

, Bhatia

, et al. Ascorbic acid enhances the inhibitory effect of aspirin on neuronal cyclooxygenase-2-mediated prostaglandin E2 production. J Neuroimmunol 2006;174(1–2):39–51; doi: 10.1016/j.jneuroim.2006.01.003

15.

Chakraborty

, Bhattacharjee

, Kamath

. Cardioprotective effect of curcumin and piperine combination against cyclophosphamide-induced cardiotoxicity. Indian J Pharmacol 2017;49(1):65–70; doi: 10.4103/0253-7613.201015

16.

Chen

, Guestrin

., and XGBoost: A scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ′16. Association for Computing Machinery: New York, NY; 2016, pp. 785–794; doi:10.1145/2939672.2939785

17.

Cheon

, Kim

, Rubinsztein

, Lee

. Autophagy, cellular aging and age-related human diseases. Exp Neurobiol 2019;28(6):643–657; doi: 10.5607/en.2019.28.6.643

18.

Chittasupho

, Manthaisong

, Okonogi

, Tadtong

, Samee

. Effects of quercetin and curcumin combination on antibacterial, antioxidant, in vitro wound healing and migration of human dermal fibroblast cells. Int J Mol Sci 2021;23(1):142; doi: 10.3390/ijms23010142

19.

Cialdella-Kam

, Ghosh

, Meaney

, Knab

, Shanely

, Nieman

. Quercetin and green tea extract supplementation downregulates genes related to tissue inflammatory responses to a 12-week high fat-diet in mice. Nutrients 2017;9(7):773; doi: 10.3390/nu9070773

20.

Cummings

, Leisgang Osse

, Kinney

. Geroscience and Alzheimer’s disease drug development. J Prev Alzheimers Dis 2023;10(4):620–632; doi: 10.14283/jpad.2023.103

21.

Dai

, Case

, Riera

, et al. Crystallographic structure of a small molecule SIRT1 activator-enzyme complex. Nat Commun 2015;6(1):7645; doi: 10.1038/ncomms8645

22.

Daina

, Michielin

, Zoete

. SwissADME: A free web tool to evaluate pharmacokinetics, drug-likeness and medicinal chemistry friendliness of small molecules. Sci Rep 2017;7(1):42717; doi: 10.1038/srep42717

23.

Dallakyan

, Olson

. Small-molecule library screening by docking with PyRx BT. In: Chemical Biology: Methods and Protocols. ( Hempel

, Williams

, and Hong

, eds.) Springer New York: New York, NY; 2015, pp. 243–250; doi:10.1007/978-1-4939-2269-7_19

24.

DASSAULT SYSTÈMES. 2025. BIOVIA Discovery Studio. DASSAULT SYSTÈMES.

25.

Davidson

, Sloan

. Fixed-dose combination of canagliflozin and metformin for the treatment of type 2 diabetes: An overview. Adv Ther 2017;34(1):41–59; doi: 10.1007/s12325-016-0434-2

26.

Derosa

, Salvadeo

SAT

. Pioglitazone and metformin fixed-dose combination in type 2 diabetes mellitus: An evidence-based review of its place in therapy. Core Evid 2008;2(3):189–198.

27.

Derosa

, Cicero

AFG

, Franzetti

, et al. Effects of exenatide and metformin in combination on some adipocytokine levels: A comparison with metformin monotherapy. Can J Physiol Pharmacol 2013;91(9):724–732; doi: 10.1139/cjpp-2012-0300

28.

Esteras

, Abramov

. Nrf2 as a regulator of mitochondrial function: Energy metabolism and beyond. Free Radic Biol Med 2022;189:136–153; doi: 10.1016/j.freeradbiomed.2022.07.013

29.

Fekete

, Major

, Feher

, Fazekas-Pongor

, Lehoczki

. Geroscience and pathology: A new frontier in understanding age-related diseases. Pathol Oncol Res 2024;30:1611623. Available from: https://www.por-journal.com/journals/pathology-and-oncology-research/articles/10.3389/pore.2024.1611623

30.

Fey

, Lenssen

. Fast graph representation learning with PyTorch geometric. ArXiv 2019. Available from: https://api.semanticscholar.org/CorpusID:70349949

31.

Fogacci

, Grassi

, Rizzo

, Cicero

AFG

. Metabolic effect of berberine-silymarin association: A meta-analysis of randomized, double-blind, placebo-controlled clinical trials. Phytother Res 2019;33(4):862–870; doi: 10.1002/ptr.6282

32.

Forman

, Kuchel

, Newman

, et al. Impact of geroscience on therapeutic strategies for older adults with cardiovascular disease: JACC Scientific Statement. J Am Coll Cardiol 2023;82(7):631–647; doi: 10.1016/j.jacc.2023.05.038

33.

Frendo-Cumbo

, MacPherson

REK

, Wright

. Beneficial effects of combined resveratrol and metformin therapy in treating diet-induced insulin resistance. Physiol Rep 2016;4(15):e12877; doi: 10.14814/phy2.12877

34.

Garcia

, Shaw

. AMPK: Mechanisms of cellular energy sensing and restoration of metabolic balance. Mol Cell 2017;66(6):789–800; doi: 10.1016/j.molcel.2017.05.032

35.

Goldman

. Combination of empagliflozin and metformin therapy: A consideration of its place in type 2 diabetes therapy. Clin Med Insights Endocrinol Diabetes 2018;11:1179551418786258; doi: 10.1177/1179551418786258

36.

Guillén

, Benito

. MTORC1 overactivation as a key aging factor in the progression to type 2 diabetes mellitus. Front Endocrinol (Lausanne) 2018;9:621. Available from: https://www.frontiersin.org/journals/endocrinology/articles/10.3389/fendo.2018.00621

37.

Guo

, Wen

, Wang

, et al. Senolytic combination of dasatinib and quercetin attenuates renal damage in diabetic kidney disease. Phytomedicine 2024;130:155705; doi: 10.1016/j.phymed.2024.155705

38.

Haak

. Combination of linagliptin and metformin for the treatment of patients with type 2 diabetes. Clin Med Insights Endocrinol Diabetes 2015;8:1–6; doi: 10.4137/CMED.S10360

39.

Hagberg

, Schult

, Swart

. and Exploring network structure, dynamics, and function using NetworkX. In: Proceedings of the 7th Python in Science Conference. 2008; pp. 11–15; doi:10.25080/tcwv9851

40.

Hasan

, Tatarskiy

, Kalinina

. Synthetic pathways and the therapeutic potential of quercetin and curcumin. Int J Mol Sci 2022;23(22):14413; doi: 10.3390/ijms232214413

41.

Hauser

. Levodopa/Carbidopa/Entacapone (Stalevo). Neurology 2004;62(1 Suppl 1):S64–S71; doi: 10.1212/WNL.62.1_suppl_1.S64

42.

Heightman

, Callahan

, Chiarparin

, et al. Structure–activity and structure–conformation relationships of ARYL propionic acid inhibitors of the Kelch-like ECH-Associated Protein 1/Nuclear Factor Erythroid 2-related factor 2 (KEAP1/NRF2) protein–protein interaction. J Med Chem 2019;62(9):4683–4702; doi: 10.1021/acs.jmedchem.9b00279

43.

Herzig

, Shaw

. AMPK: Guardian of metabolism and mitochondrial homeostasis. Nat Rev Mol Cell Biol 2018;19(2):121–135; doi: 10.1038/nrm.2017.95

44.

Hipp

, Kasturi

, Hartl

. The proteostasis network and its decline in ageing. Nat Rev Mol Cell Biol 2019;20(7):421–435; doi: 10.1038/s41580-019-0101-y

45.

Hofer

, Daskalaki

, Abdellatif

, et al. A surge in endogenous spermidine is essential for rapamycin-induced autophagy and longevity. Autophagy 2024;20(12):2824–2826; doi: 10.1080/15548627.2024.2396793

46.

Huang

, Ren

, Yuan

, et al. Targeting the AMPK/Nrf2 pathway: A novel therapeutic approach for acute lung injury. J Inflamm Res 2024;17(null):4683–4700; doi: 10.2147/JIR.S467882

47.

Iwasaki

, Abarca

, Aguayo-Mazzucato

. Regulation of cellular senescence in type 2 diabetes mellitus: From mechanisms to clinical applications. Diabetes Metab J 2023;47(4):441–453; doi: 10.4093/dmj.2022.0416

48.

Kaplinsky

. Sacubitril/valsartan in heart failure: Latest evidence and place in therapy. Ther Adv Chronic Dis 2016;7(6):278–290; doi: 10.1177/2040622316665350

49.

Khalafani

, Zamani-Garmsiri

, Panahi

, Meshkani

. Metformin-chlorogenic acid combination reduces skeletal muscle inflammation in C57BL/6 mice on high-fat diets. Mol Biol Rep 2023;50(3):2581–2589; doi: 10.1007/s11033-022-08030-8

50.

Kim

, Chen

, Cheng

, et al. PubChem 2025 Update. Nucleic Acids Res 2025;53(D1):D1516–D1525; doi: 10.1093/nar/gkae1059

51.

Kipf

, Welling

. Semi-supervised classification with graph convolutional networks. ArXiv 2016. Available from: https://api.semanticscholar.org/CorpusID:3144218

52.

Knox

, Wilson

, Klinger

, et al. DrugBank 6.0: The DrugBank Knowledgebase for 2024. Nucleic Acids Res 2024;52(D1):D1265–D1275; doi: 10.1093/nar/gkad976

53.

Konopka

, Lamming

, Grasso

, et al.; EVERLAST Investigators. Blazing a trail for the clinical use of rapamycin as a GeroprotecTOR. Geroscience 2023;45(5):2769–2783; doi: 10.1007/s11357-023-00935-x

54.

Kroemer

, Maier

, Cuervo

, et al. From geroscience to precision geromedicine: Understanding and managing aging. Cell 2025;188(8):2043–2062; doi: 10.1016/j.cell.2025.03.011

55.

Kumar

, Choudhary

, Kinger

, et al. Proteostasis defects: Medicinal challenges of imperfect aging & neurodegeneration. Transl Med Aging 2023;7:87–97; doi: 10.1016/j.tma.2023.09.001

56.

Landrum and Greg. 2006. RDKit: Open-Source Cheminformatics. Available from: http://www.rdkit.org

57.

Laplante

, Sabatini

. MTOR signaling in growth control and disease. Cell 2012;149(2):274–293; doi: 10.1016/j.cell.2012.03.017

58.

Le Boeuf

, Hefny

, Karuturi

, Rao

PPN

. Combination of resveratrol and curcumin with anti-amyloid monoclonal antibodies aducanumab and lecanemab leads to greater inhibition of amyloid-beta aggregation. ACS Chem Neurosci 2026;17(5):987–997; doi: 10.1021/acschemneuro.5c00760

59.

Lee

. Mechanisms and disease implications of sirtuin-mediated autophagic regulation. Exp Mol Med 2019;51(9):1–11; doi: 10.1038/s12276-019-0302-7

60.

Leontieva

, Paszkiewicz

, Demidenko

, Blagosklonny

. Resveratrol potentiates rapamycin to prevent hyperinsulinemia and obesity in male mice on high fat diet. Cell Death Dis 2013;4(1):e472–e472; doi: 10.1038/cddis.2012.202

61.

Lévy

, El Banna

, Baïlle

, et al. Causative links between protein aggregation and oxidative stress: A review. Int J Mol Sci 2019;20(16):3896; doi: 10.3390/ijms20163896

62.

, Chen

Q-W

, Gong

X-L

, et al. Metformin and berberine synergistically improve NAFLD via the AMPK–SREBP1–FASN signaling pathway. Sci Rep 2025;15(1):29400; doi: 10.1038/s41598-025-15495-7

63.

Lipinski

, Lombardo

, Dominy

, Feeney

. Experimental and computational approaches to estimate solubility and permeability in drug discovery and development Settings1PII of Original Article: S0169-409X(96)00423-1. The Article Was Originally Published in Advanced Drug Delivery Reviews 23 (1997) 3. Adv Drug Deliv Rev 2001;46(1–3):3–26; doi: 10.1016/S0169-409X(96)00423-1

64.

Liu

, Wang

, Li

, et al. The synergistic protection of EGCG and Quercetin against Streptozotocin (STZ)-induced NIT-1 pancreatic β cell damage via upregulation of BCL-2 expression by MiR-16-5p. J Nutr Biochem 2021;96:108748; doi: 10.1016/j.jnutbio.2021.108748

65.

Liu

, Srivastava

, Ramanujam

, Brylinski

. SynerGNet: A graph neural network model to predict anticancer drug synergy. Biomolecules 2024;14(3):253; doi: 10.3390/biom14030253

66.

Luo

, Wang

, Zhu

, Du

, Wang

, Ding

. Impact of atorvastatin combined with ezetimibe for the treatment of carotid atherosclerosis in patients with coronary heart disease. Acta Cardiol Sin 2016;32(5):578–585; doi: 10.6515/acs20151013h

67.

Lyu

, Li

, Yuan

, et al. Effects of combination treatment with metformin and berberine on hypoglycemic activity and gut microbiota modulation in Db/Db Mice. Phytomedicine 2022;101:154099; doi: 10.1016/j.phymed.2022.154099

68.

Maksud

, Bera

, Naim

, Alam

. Dapagliflozin: A new hope for the therapeutic treatment of type 2 diabetes mellitus. Eur J Med Chem Rep 2024;11:100167; doi: 10.1016/j.ejmcr.2024.100167

69.

Mannick

, Lamming

. Targeting the biology of aging with MTOR inhibitors. Nat Aging 2023;3(6):642–660; doi: 10.1038/s43587-023-00416-y

70.

Marino

, Hausenloy

, Andreadou

, Horman

, Bertrand

, Beauloye

. AMP-activated protein kinase: A remarkable contributor to preserve a healthy heart against ROS injury. Free Radic Biol Med 2021;166:238–254; doi: 10.1016/j.freeradbiomed.2021.02.047

71.

Matacchione

, Valli

, Silvestrini

, et al. Curcumin, polydatin and quercetin synergistic activity protects from high-glucose-induced inflammation and oxidative stress. Antioxidants (Basel) 2022;11(6):1037; doi: 10.3390/antiox11061037

72.

McInnes

, Healy

. UMAP: Uniform Manifold Approximation and Projection for dimension reduction. ArXiv 2018; doi: 10.48550/arXiv.1802.03426

73.

Miller

, Krumins

, Zhou

, et al. Ertugliflozin and sitagliptin co-initiation in patients with type 2 diabetes: The VERTIS SITA randomized study. Diabetes Ther 2018;9(1):253–268; doi: 10.1007/s13300-017-0358-0

74.

Minze

, Klein

, Terrell

. Saxagliptin and metformin in fixed combination for the treatment of type 2 diabetes in adults. Clin Med Insights Endocrinol Diabetes 2013;6:15–24; doi: 10.4137/CMED.S8510

75.

Monsalvo-Maraver

, Ovalle-Noguez

, Nava-Osorio

, et al.; Students from Programa Delfín 2022. Interactions between the ubiquitin–proteasome system, NRF2, and the cannabinoidome as protective strategies to combat neurodegeneration: Review on experimental evidence. Neurotox Res 2024;42(2):18; doi: 10.1007/s12640-024-00694-3

76.

Moskalev

, Chernyagina

, de Magalhães

, et al. Geroprotectors.org: A new, structured and curated database of current therapeutic interventions in aging and age-related disease. Aging (Albany NY) 2015;7(9):616–628; doi: 10.18632/aging.100799

77.

Neumiller

, Holland

. Alogliptin + metformin combination for the treatment of type 2 diabetes mellitus. Expert Rev Endocrinol Metab 2016;11(1):21–31; doi: 10.1586/17446651.2016.1110484

78.

Ngo

, Duennwald

. Nrf2 and oxidative stress: A general overview of mechanisms and implications in human disease. Antioxidants (Basel) 2022;11(12):2345; doi: 10.3390/antiox11122345

79.

Norradee

, Khwanraj

, Balit

, Dharmasaroja

. Evaluation of the combination of metformin and rapamycin in an MPP+-Treated SH-SY5Y model of Parkinson’s disease. Adv Pharmacol Pharm Sci 2023;2023(1):3830861; doi: 10.1155/2023/3830861

80.

Ornatowski

, Lu

, Yegambaram

, et al. Complex interplay between autophagy and oxidative stress in the development of pulmonary disease. Redox Biol 2020;36:101679; doi: 10.1016/j.redox.2020.101679

81.

Ottens

, Franz

, Hoppe

. Build-UPS and break-downs: Metabolism impacts on proteostasis and aging. Cell Death Differ 2021;28(2):505–521; doi: 10.1038/s41418-020-00682-y

82.

Panwar

, Singh

, Bhatt

, et al. Multifaceted role of MTOR (mammalian target of rapamycin) signaling pathway in human health and disease. Signal Transduct Target Ther 2023;8(1):375; doi: 10.1038/s41392-023-01608-z

83.

Pedregosa

, Varoquaux

, Gramfort

, et al. Scikit-learn: Machine learning in python. J. Mach. Learn. Res 2011;12(null):2825–2830.

84.

Peng

, Zhang

, Dou

, Yang

, Zhang

, Yu

. Reinforced neighborhood selection guided multi-relational graph neural networks. ACM Trans. Inf. Syst. 2022;40(4):1–46; doi: 10.1145/3490181

85.

Prathumwon

, Anuchapreeda

, Kiattisin

, et al. Curcumin and EGCG combined formulation in nanostructured lipid carriers for anti-aging applications. Int J Pharm X 2025;9:100323; doi: 10.1016/j.ijpx.2025.100323

86.

Proshkina

, Koval

, Platonova

, et al. Polyphenols as potential geroprotectors. Antioxid Redox Signal 2024;40(7–9):564–593; doi: 10.1089/ars.2023.0247

87.

Querfurth

, Lee

H-K

. Mammalian/Mechanistic Target of Rapamycin (MTOR) complexes in neurodegeneration. Mol Neurodegener 2021;16(1):44; doi: 10.1186/s13024-021-00428-5

88.

Rahman

, Shuvo

, Apu

MMH

, et al. Combination of epigallocatechin 3 gallate and curcumin improves d-galactose and normal-aging associated memory impairment in mice. Sci Rep 2023;13(1):12681; doi: 10.1038/s41598-023-39919-4

89.

Reifsnyder

, Flurkey

, Doty

, Calcutt

, Koza

, Harrison

. Rapamycin/metformin co-treatment normalizes insulin sensitivity and reduces complications of metabolic syndrome in type 2 diabetic mice. Aging Cell 2022;21(9):e13666; doi: 10.1111/acel.13666

90.

Ren

, Pei

, Kong

, Shi

. The effect of semaglutide combined with metformin on liver inflammation and pancreatic beta-cell function in patients with type 2 diabetes and non-alcoholic fatty liver disease. J Diabetes Complications 2025;39(2):108932; doi: 10.1016/j.jdiacomp.2024.108932

91.

Rivero-Segura

, Zepeda-Arzate

, Castillo-Vazquez

, et al. Exploring the geroprotective potential of nutraceuticals. Nutrients 2024;16(17):2835; doi: 10.3390/nu16172835

92.

Rolland

, Sierra

, Ferrucci

, et al.; G.C.T-TF group. Challenges in developing geroscience trials. Nat Commun 2023;14(1):5038; doi: 10.1038/s41467-023-39786-7

93.

Ruderman

, Xu

, Nelson

, et al. AMPK and SIRT1: A long-standing partnership? American Journal of Physiology-Endocrinology and Metabolism 2010;298(4):E751–E60; doi: 10.1152/ajpendo.00745.2009

94.

Sacitharan

, Bou-Gharios

, Edwards

. SIRT1 directly activates autophagy in human chondrocytes. Cell Death Discov 2020;6(1):41; doi: 10.1038/s41420-020-0277-0

95.

Santos

, Simão

, Nóbrega

, Bragança

, Castelo-Branco

, Araújo

IM.

and ALFA Score Consortium. Oxidative stress and aging: Synergies for age related diseases. FEBS Lett 2024;598(17):2074–2091; doi: 10.1002/1873-3468.14995

96.

Schrödinger LLC. The PyMOL Molecular Graphics System, Version 2.0. Schrödinger LLC; 2015.

97.

Singh

, Chhabra

, Ndiaye

, Garcia-Peterson

, Mack

, Ahmad

. The role of sirtuins in antioxidant and redox signaling. Antioxid Redox Signal 2018;28(8):643–661; doi: 10.1089/ars.2017.7290

98.

Somasundaram

, Jain

, Blot-Chabaud

, et al. Mitochondrial dysfunction and its association with age-related disorders. Front Physiol 2024;15:1384966. Available from: https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2024.1384966

99.

St Onge

, Miller

, Clements

. Sitagliptin/metformin (Janumet) as combination therapy in the treatment of type-2 diabetes mellitus. P T 2012;37(12):699–708.

100.

Strong

, Miller

, Cheng

, et al. Lifespan benefits for the combination of rapamycin plus acarbose and for captopril in genetically heterogeneous mice. Aging Cell 2022;21(12):e13724; doi: 10.1111/acel.13724

101.

Tabibzadeh

. Role of autophagy in aging: The good, the bad, and the ugly. Aging Cell 2023;22(1):e13753; doi: 10.1111/acel.13753

102.

Torkamannia

, Omidi

, Ferdousi

. SYNDEEP: A deep learning approach for the prediction of cancer drugs synergy. Sci Rep 2023;13(1):6184; doi: 10.1038/s41598-023-33271-3

103.

Trott

, Olson

. AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. J Comput Chem 2010;31(2):455–461; doi: 10.1002/jcc.21334

104.

Wang

J-S

, Huang

C-N

, Hung

Y-J

, et al.; acarbose/metformin fixed-dose combination study investigators. Acarbose plus metformin fixed-dose combination outperforms acarbose monotherapy for type 2 diabetes. Diabetes Res Clin Pract 2013;102(1):16–24; doi: 10.1016/j.diabres.2013.08.001

105.

Wang

, Li

, Yuan

, Fan

, Cai

. Role of AMPK in autophagy. Front Physiol 2022;13:1015500. Available from: https://www.frontiersin.org/journals/physiology/articles/10.3389/fphys.2022.1015500

106.

Wei

, Su

, Mackenzie

. Chlorogenic acid combined with epigallocatechin-3-gallate mitigates d-galactose-induced gut aging in mice. Food Funct 2023;14(6):2684–2697; doi: 10.1039/D2FO03306B

107.

Wezeman

, Darvas

, Postupna

, et al. A drug cocktail of rapamycin, acarbose, and phenylbutyrate enhances resilience to features of early-stage Alzheimer’s disease in aging mice. bioRxiv 2024; doi: 10.1101/2024.01.26.577437

108.

Williams

, Stephens

. Combination therapy with saxagliptin and dapagliflozin for the treatment of type 2 diabetes. Expert Opin Pharmacother 2015;16(15):2373–2379; doi: 10.1517/14656566.2015.1088830

109.

, Zheng

, Zhou

, et al. Autophagy in aging-related diseases and cancer: Principles, regulatory mechanisms and therapeutic potential. Ageing Res Rev 2024;100:102428; doi: 10.1016/j.arr.2024.102428

110.

Yaghmaei

, Lu

, Ehwerhemuepha

, et al. Combined use of donepezil and memantine increases the probability of five-year survival of alzheimer’s disease patients. Commun Med (Lond) 2024;4(1):99; doi: 10.1038/s43856-024-00527-6

111.

Yang

Y-J

, Lee

S-H

, Kim

, et al. Combination therapy of rosuvastatin and ezetimibe in patients with high cardiovascular risk. Clin Ther 2017;39(1):107–117; doi: 10.1016/j.clinthera.2016.11.014

112.

You

, Liang

. SIRT1 and SIRT6: The role in aging-related diseases. Biochim Biophys Acta Mol Basis Dis 2023;1869(7):166815; doi: 10.1016/j.bbadis.2023.166815

113.

Zaky

, Bassiouny

, Farghaly

, El-Sabaa

. A combination of resveratrol and curcumin is effective against aluminum chloride-induced neuroinflammation in rats. J Alzheimers Dis 2017;60(s1):S221–S235; doi: 10.3233/JAD-161115

114.

Zamani-Garmsiri

, Ghasempour

, Aliabadi

, Hashemnia

SMR

, Emamgholipour

, Meshkani

. Combination of metformin and chlorogenic acid attenuates hepatic steatosis and inflammation in high-fat diet fed mice. IUBMB Life 2021;73(1):252–263; doi: 10.1002/iub.2424

115.

Zhao

, Cen

, Tian

, et al. Combination treatment with quercetin and resveratrol attenuates high fat diet-induced obesity and associated inflammation in rats via the AMPKα1/SIRT1 signaling pathway. Exp Ther Med 2017;14(6):5942–5948; doi: 10.3892/etm.2017.5331

116.

Zhao

, Chen

, Du

, et al. Dual-route embedding-aware graph neural networks for drug repositioning. Brief Bioinform 2025;26(5):bbaf555; doi: 10.1093/bib/bbaf555

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.04 MB

0.01 MB

0.22 MB

0.02 MB

0.01 MB

0.00 MB

0.01 MB