Cross-Species Aging Knowledge Integration into Agentic AI Platform Uncovers Conserved Mechanisms

Cross-Species Aging Knowledge Integration into Agentic AI Platform Uncovers Conserved Mechanisms | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Cross-Species Aging Knowledge Integration into Agentic AI Platform Uncovers Conserved Mechanisms Gaurav Ahuja, Arushi Sharma, Ankit Singh, Shekhar Kedia, Abhinav Sharma, and 16 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9060414/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Aging research has been advanced largely through the use of model organisms, where short lifespans and genetic tractability enable the systematic discovery of molecular pathways influencing longevity and age-related decline. However, knowledge about aging remains fragmented across species-specific repositories and domain-focused databases, limiting our ability to identify evolutionarily conserved mechanisms and translate findings to human biology. To address this gap, we developed EvoAge, a unified, multi-species knowledge graph that integrates aging-specific and general biomedical resources into a systems-level framework. EvoAge harmonizes 48 public datasets into a graph comprising 1.04 billion triples across six key species. A human-centric orthology framework reconciles more than 80,000 gene entries, expanding accessible organism-level aging knowledge by up to 1,700-fold compared with existing resources. To operationalize the graph for biological reasoning, we optimized knowledge graph embedding models and deployed a large language model (LLM)-assisted agentic interface that supports natural-language querying, link prediction, and hypothesis testing. In internal benchmarking using recent pre-print aging literature, EvoAge significantly outperformed state-of-the-art LLMs in distinguishing biologically plausible from implausible hypotheses. Importantly, EvoAge recommended a previously unrecognized Alzheimer’s disease (AD) mechanism involving nanoscale redistribution of BACE1 within synaptic compartments. We experimentally validated this EvoAge-supported prediction using patient-derived iPSCs carrying a familial PSEN1 mutation, demonstrating disease-associated remodeling of β-secretase, defined by altered localization, nanoscale clustering, and compartment-specific enrichment. We further confirmed the predicted evolutionary conservation of this BACE1–pathology relationship in additional AD systems, including transgenic mice and postmortem human brain tissue. Biological sciences/Computational biology and bioinformatics/Computational platforms and environments Biological sciences/Computational biology and bioinformatics/Computational models Biological sciences/Computational biology and bioinformatics/Data integration Biological sciences/Computational biology and bioinformatics/Data mining Knowledge Graphs Evolution Graph Modeling Aging Synapse Alzheimer Figures Figure 1 Figure 2 Figure 3 Figure 4 INTRODUCTION The quest to understand the biological mechanisms of aging represents one of the most complex challenges in biomedical science 1 . Its complexity arises from the interplay of countless molecular and cellular processes that unfold over decades, varying across tissues, and are influenced by both genetics and environment 2 . Given the ethical and practical constraints of longitudinal human aging research, the field has been profoundly shaped by studies in model organisms, from the short-lived nematode Caenorhabditis elegans to the laboratory mouse, where genetic tractability and shorter lifespans enable high-throughput interrogation of longevity and age-related decline 3 . A cornerstone insight from these studies is that core metabolic and protective pathways governing aging, such as insulin signaling, mTOR, and mitochondrial function, are evolutionarily conserved 1,4–6 . However, this foundational insight also presents a major bottleneck: the resulting deluge of data is profoundly fragmented. Critical insights are often confined to species-specific silos and specialized databases, thereby demarcating them from the broader context of human systems biology 5,7–12 . This fragmentation obstructs a unified, systems-level understanding of how mechanisms identified in models translate to human aging and pathology. Knowledge graphs (KGs) have become pivotal in systems biology for integrating disparate data into structured networks, enabling the discovery of non-obvious biological relationships. Special-purpose KGs (e.g., Hetionet, PrimeKG, CKG) 13–15 have proven effective for generating mechanistically grounded hypotheses in drug repurposing, while aging-specific resources (e.g., Aging Atlas 10 , DrugAge 16 , GenDR 17 , AgeXtend 9 ) provide critical expert-curated data on longevity. However, these specialized KGs are often constrained by a narrow scope, excelling in specific data types but lacking the multi-scale, cross-species context necessary for holistic discovery. This limitation prevents the seamless translation of findings, such as linking a longevity gene in yeast to its conserved pathway in mice and its phenotypic consequence in the human system. Bridging this gap requires a multi-species KG that deeply integrates specialized aging knowledge with general biology. A further challenge is utility; the scale of billion-triple KGs makes them inaccessible to many biologists. The synergy between KGs and large language models (LLMs) offers a transformative solution. While LLMs offer intuitive natural language interaction, they risk hallucination 18,19 . In contrast, KGs offer a verifiable, structured knowledge base. Their integration creates an agentic framework where the LLM serves as a natural language interface, querying the KG and generating evidence-based explanations, thereby making deep knowledge accessible for hypothesis generation and testing 20,21 . Here, we introduce EvoAge, an agentic AI platform built on a unified, multi-species knowledge graph for translational aging research. EvoAge integrates 48 aging-focused and general biomedical resources into a harmonized network of 1.04 billion triples across six species, centered on an orthology-driven framework that enables biologically meaningful, cross-species reasoning. An LLM-assisted interface further transforms this large-scale graph into an accessible system supporting natural-language exploration, link prediction, and formal hypothesis testing. To evaluate its reasoning capability, we benchmarked EvoAge against recent aging-related preprints and observed markedly superior discrimination between plausible and implausible hypotheses compared with leading general-purpose LLMs. We then applied EvoAge to uncover a previously unrecognized mechanism in Alzheimer’s disease (AD) involving nanoscale redistribution of Beta-secretase (BACE1) at synapses. This prediction was experimentally validated using familial AD patient-derived iPSC neurons and the isogenic controls, and further confirmed in transgenic mouse models and postmortem human brain tissue. Collectively, EvoAge provides a scalable framework integrating evolutionary biology, structured biomedical knowledge, and AI reasoning to accelerate mechanistic discovery in aging and age-related disease. In the case of BACE1, the aging context was essentia, EvoAge revealed that its redistribution consistently aligned with conserved aging-linked pathways such as synaptic decline and vesicle trafficking across species, distinguishing a true disease-associated shift from normal age-related variation and enabling its nomination as a testable mechanism. RESULTS A Systems-Scale Knowledge Graph Integrating Aging Biology Across Model Species To construct a foundational resource for cross-species aging research, we first established the rationale for a multi-organism framework. Historical publication trends revealed exponential growth in aging studies centered on six core species, including humans (Supplementary Figure 1a, Supplementary Table 1) . Comparative survival analysis further underscored the analytical advantage of lifespan diversity for identifying conserved aging mechanisms (Supplementary Figure 1b) . Guided by these observations, we developed a computational pipeline to assemble the EvoAge Knowledge Graph (KG) (Figure 1a) , integrating data from 48 public sources, including 12 aging-specific and 36 general biomedical databases (Figure 1b, Supplementary Table 2) . To assess the structure of the integrated graph, we analyzed source-level contributions across biological relationship types. A comparative heatmap revealed a dense, heterogeneous relationship matrix in EvoAge, where individual sources specialize in distinct interaction categories, contrasting the sparse architecture of the baseline Aging KG (Supplementary Figure 2a, b) . Species-level provenance analysis confirmed that each organism is anchored by its dedicated model organism databases. This examination demonstrated that knowledge for each organism is dominated by its dedicated Model Organism Database (MOD), with Saccharomyces Genome Database (SGD) 22 and YeastNet (v3) 23 serving as the primary source for Saccharomyces cerevisiae (Supplementary Figure 3c), FlyBase 24 for Drosophila melanogaster (Supplementary Figure 3b) , WormBase 25 for Caenorhabditis elegans (Supplementary Figure 3d) , ZFIN 26 for Danio rerio (Supplementary Figure 3e) , and MGI (Mouse Genome Informatics) for Mus musculus 27 (Supplementary Figure 3a) . Collectively, these patterns demonstrate that EvoAge functions as a “mosaic of experts,” combining broad general-purpose knowledge with deep organism-specific annotations. Data harmonization was achieved through a structured schema encompassing 14 biological entity types and a unified identifier-mapping workflow (Supplementary Figure 1c) . Duplicate records were merged while maintaining full provenance, resulting in a final graph containing >1.04 billion triples (Supplementary Table 3) . To enable true cross-species reasoning, we implemented an orthology-based unification layer that mapped over 80,000 gene entries to human identifiers (Figure 1c) , with mapping success reflecting phylogenetic distance. For instance, we successfully mapped 18,999 genes from the closely related mouse to human counterparts, but only 6,707 genes from the more distant worm (Supplementary Figure 1d) . We next sought to quantify the transformative impact of data integration. The initial Aging KG contained only 616,928 triples, was overwhelmingly human-centric (601,099 triples, 97.5%), and chemically skewed with minimal systems-level context (Supplementary Figure 1e, f) . Through systematic integration, we observed dramatic, species-specific enrichment in the expanded EvoAge KG (Figures 1d and e) ,with particularly notable expansions for model organisms, where Drosophila melanogaster gene nodes increased from 170 to 13,747, and Danio rerio grew from zero to 15,182 gene nodes. To characterize the final composition, we analyzed node type distributions across the integrated resource (Figure 1e). This revealed EvoAge as a literature-first resource dominated by 26.7 million publication nodes, with a strong genomic foundation including 1.5 million mutation nodes and ~1.01 million chemical nodes (Supplementary Figure 1g) , while maintaining distinct species-specific node type distributions. Finally, we performed a comprehensive connectivity analysis to understand how integration affected biological relationships. Connectivity analysis revealed a non-linear, species-specific enrichment, with counts increasing by 22,033-fold for S. cerevisiae (420 to 9.25M), 19,332-fold for D. melanogaster (528 to 10.2M), 5,847-fold for C. elegans (2.4K to 14.2M), and 2,747-fold for M. musculus (12.5K to 34.2M), contributing to 1700 fold expansion of knowledge. Even the human-centric data expanded by 207-fold (601K to 124.3M), while D. rerio grew from zero to 28.7 million triples (Figure 1f-g) . Importantly, EvoAge exhibits a scale-free topology, high modularity, low betweenness centrality, and markedly increased clustering properties, all of which are characteristic of real biological networks and are favorable for computational inference (Figure 1h-k) . Node-level hub analysis further confirmed a paradigm shift: from conceptual hallmarks dominating the baseline KG to structural and functional anchors, including tissue architecture, translational machinery, and GPCR signaling, in the final EvoAge KG (Supplementary Figures 4a and 4b; Supplementary Tables 5 and 6) . Together, these analyses establish EvoAge as a deeply layered, evolutionarily informed, and computationally optimized resource positioned to support mechanistic discovery at scale. Systematic Optimization of Knowledge Graph Embeddings Reveals Architectural Dependencies With the EvoAge KG established, we next sought to convert the resource into a predictive inference engine through knowledge graph embedding (KGE) modeling. We first characterized the relational space of both graphs. The EvoAge KG contained 66 unique relation types spanning 14 biological entity classes, with PMID, chemical-gene, and gene-gene associations representing the most abundant edges (Supplementary Figure 5a; Supplementary Table 4) . In contrast, the baseline Aging KG contained only 27 relation types, predominantly limited to gene-disease relationships (Supplementary Figure 5b) . A co-occurrence heatmap confirmed that nearly all permissible node-type pairs in EvoAge participated in at least one relationship, establishing the necessity of relation-aware learning (Supplementary Figure 5c) . To ensure unbiased evaluation, both graphs were partitioned into 80% training, 10% validation, and 10% testing sets, with stratification across all relation types (27 for the Aging KG and 66 for EvoAge) confirmed in each subset (Supplementary Figure 5d-f) . We next benchmarked six state-of-the-art KGE architectures: Translating Embeddings (TransE) 28 , Rotation-based Knowledge Graph Embedding (RotatE) 29 , Simple Embedding (SimplE) 30 , Diagonal Matrix (DisMult) 31,32 , RElational SCALar (RESCAL) 33 , and Complex Embeddings (ComplEx) 34 , (Figure 2a) using a standardized embedding size of 64 dimensions and a fixed training budget (Supplementary Figure 5g, h) . Performance was evaluated using Mean Rank (MR), Mean Reciprocal Rank (MRR), and hit@k metrics (Supplementary Figure 5i; Supplementary Table 7) . Model behavior was strongly scale-dependent: the tensor-based RESCAL model performed best on the smaller Aging KG, achieving the lowest MR and highest hit@10 and MRR values (Figures 2b and c; Supplementary Figure 5j) , with rapid and stable convergence (Figure 2d, left) . Conversely, for the billion-triple EvoAge KG, the geometric RotatE model consistently outperformed all alternatives (Figures 2b and c; Supplementary Figure 5j) , demonstrating smooth loss reduction over 4 million training steps (Figure 2d, right) . We then performed an ablation study to determine optimal embedding dimensionality (64-512 dimensions) for each selected model (Figure 2e) . For the Aging-RESCAL model, link prediction performance remained uniformly high across all dimensions (Figure 2f, h) . However, edge-type prediction improved substantially at higher dimensionality, peaking at 512 dimensions with a hit@1 score of 0.869 (Figure 2g) , with no evidence of overfitting (Supplementary Figure 5k) . Based on these findings, we selected the 512-dimensional RESCAL model for the Aging KG to maximize accuracy on the most difficult inference tasks. Notably, the EvoAge-RotatE model showed a different capacity profile: 64-dimensional embeddings yielded poor link prediction performance (Figure 2f, i) , whereas performance improved at 128 and 256 dimensions. The 512-dimensional model exhibited significant degradation in edge-type prediction (Figure 2g) , consistent with the phenomenon of overfitting (Supplementary Figure 5l) . After careful comparison, we selected the 128-dimensional RotatE model as the optimal compromise, delivering superior performance on edge type prediction with nearly identical link prediction performance to the 256-dim model (Figure 2f, g, i) . Finally, to evaluate cross-species reasoning, a central goal of EvoAge, we constructed a held-out benchmark containing 3.14 million triples, evenly balanced between within-species and cross-species relationships (Figures 2j and k) . The final 128-dimensional EvoAge-RotatE model achieved uniformly high hit@k scores across all six species and, critically, performed comparably on the cross-species subset for both link prediction (Figure 2l) and edge-type classification (Figure 2m) . These results demonstrate that the EvoAge embeddings capture biologically coherent structure, enabling inference beyond memorized facts and supporting accurate prediction across evolutionary boundaries, fulfilling the design objective of EvoAge as an engine for cross-species discovery. A Conversational AI Interface Leverages the EvoAge KG for Discovery and Validation After constructing the EvoAge knowledge graph and optimizing its embedding models, we next developed an intelligent and accessible interface to operationalize the resource at scale. To achieve this, we developed the EvoAge AI Platform, designed to bridge natural language reasoning with large-scale computational biology. The platform supports three primary capabilities (Figure 3a) : retrieving existing knowledge from the Neo4j database through a natural language-based Search Database interface, performing Link Prediction using KGE inference to identify missing yet plausible biological relationships (for example, “Predict potential disease associations for the gene FOXO3.”), and executing a Test Hypothesis mode, in which a complete triple (for example, “does aging associate with cellular senescence?”) is assigned a quantitative plausibility score and an interpretable verdict. These functionalities are powered by a three-tier system architecture (Figure 3b,c; Supplementary Figure 6a) in which a Streamlit-based frontend (Figure 3g) interacts with a FastAPI backend orchestrating the Neo4j graph database 35 , the internal DGL-KE inference engine 36 , and external LLM services, including Google Gemini 2.5 Flash-Lite. Reasoning within the platform is governed by the EvoAge-Agent, which is implemented using the Kani framework 37 and GPT-4o-mini (OpenAI). The query-processing workflow (Figure 3c) begins with entity extraction, where the model identifies biological entities (for example, FOXO3) and invokes the Search Biological Entity tool (Supplementary Figure 6d) . The agent then infers user intent and selects the appropriate operational mode. Link prediction follows a defined 12-step execution sequence (Figure 3d) in which the system validates the query, loads trained embedding artifacts, generates relation-specific scores, and returns ranked results. Hypothesis testing follows a more complex logic path (Figure 3e) , combining a raw triple score from the DGL-KE model (Supplementary Figure 6e) with evidence retrieval from Neo4j, followed by interpretation from Gemini. The final response is synthesised into natural language (Figure 6f) , and the full workflow is optimised for interactivity, with most inference calls completing in under 30 seconds. To ensure scientific rigor in hypothesis evaluation, we developed a statistical calibration framework that converts raw embedding scores into interpretable thresholds. For each of the 66 relation types in EvoAge, we computed optimal cutoffs using Youden’s J statistic applied to score distributions generated from 1,000 true and 1,000 synthetic negative triples (Supplementary Figure 6b; Supplementary Table 9) . Thresholds varied substantially across relation categories, with frequently observed cellular component associations requiring lower (more negative) values, whereas sparse relations, such as chemical-mutation links, required substantially higher cutoffs (Supplementary Figure 6c) . Once integrated, this calibration step enabled reliable discrimination between plausible and implausible assertions. Finally, the platform was deployed as an interactive chat-based assistant, enabling users to navigate a billion-triple knowledge space through natural language. A representative example query for FOXO3 yields a ranked set of disease associations with numerical prediction scores (Figures 3g and Supplementary Figure 6f) , illustrating the platform’s ability to translate deep graph complexity into intuitive and actionable scientific insights. EvoAge Enables Accurate Hypothesis Generation and Experimental Validation To quantitatively assess the discovery capability of EvoAge, we developed a benchmarking framework that compared its hypothesis-testing performance with that of state-of-the-art general-purpose LLMs (Figure 4a) . We first collected 1,282 aging-related preprints from bioRxiv (January-September 2025) and manually refined this to 980 studies focused on molecular and cellular aging biology (Supplementary Table 10) . From their abstracts, we used Google Gemini to generate a balanced test set comprising 2,850 positive and 2,850 negative hypotheses, ensuring that all models were evaluated on content not present in their training data. We then challenged four systems: three general-purpose LLMs (Gemma, Llama3.1, and Qwen2.5) 38–40 and the EvoAge chatbot, to evaluate these hypotheses. All responses were automatically scored for biological plausibility by BioMistral-7B 41 , a model trained on millions of PubMed abstracts and used as an expert grading system (Supplementary Table 11-12) . EvoAge demonstrated a marked performance advantage, achieving significantly higher clarity scores than all other systems (p < 0.0001, pairwise Wilcoxon test with BH adjustment) (Figure 4b) . While the general-purpose LLMs demonstrated moderate accuracy on positive hypotheses (Figure 4c) , they often failed to reject implausible statements, as indicated by high scores on negative hypotheses (Figure 4d) . In contrast, EvoAge produced a distinctly skewed score distribution with consistently low values for false hypotheses, indicating strong discriminatory power derived from its structured knowledge base. Next, to investigate whether EvoAge’s knowledge-driven hypothesis evaluation framework translates into experimentally verifiable biology, we examined an EvoAge-supported mechanism involving the nanoscale redistribution of the β-secretase enzyme BACE1, which is involved in the amyloidogenic processing of Amyloid Precursor Protein (APP) in Alzheimer’s disease (AD) (Supplementary Figure 8a) . EvoAge assigned this hypothesis a high plausibility score and did not classify it as species-restricted, suggesting that if present, the mechanism should manifest across phylogenetically divergent biological systems (Supplementary figure 8b) . To experimentally validate this hypothesis, we quantified BACE1 nanoscale organisation using STED nanoscopy in multiple models of AD: human iPSC-derived glutamatergic neurons carrying a familial AD (FAD) mutation and their isogenic controls, postmortem human brain tissue, and wild-type versus AD transgenic mouse brain (Figure 4f) . We co-mapped BACE1 (magenta) relative to the Perisynapse/Postsynapse markers (green) in these conditions and observed significant clustering and enrichment in the predicted zones, shown here for iPSCs (Figure 4g-j) . Shapiro-Wilk testing on species (mouse and human), conditions (Healthy vs AD), and compartmental measurements (postsynaptic vs perisynaptic) demonstrated significant deviation from normality (p < 10 -14 to 10 -63 ) (Supplementary Table 13) . Gaussian Mixture Modeling confirmed that two-component distributions consistently provided a superior fit compared to single Gaussian models, with significant improvements in the Bayesian Information Criterion Supplementary Table 14) . This suggests that BACE1 exists not as a uniform entity at the nanoscale, but rather as at least two discrete biological subtypes that differ in composition, spatial architecture, or functional state. Quantitative compartmental analysis revealed a conserved directional shift in BACE1 intensity from postsynaptic density (PSD) toward perisynaptic endocytic regions in disease. In iPSC-derived human neurons, perisynaptic nanodomain intensity modestly exceeded postsynaptic levels in controls (67.45 vs. 61.90; fold-change +0.0897), but the difference increased markedly in FAD-mutant neurons (76.66 vs. 59.60; fold-change +0.2863; MW p = 4.6×10 -14 ). Postmortem human cortex exhibited a substantially stronger redistribution, with perisynaptic levels showing nearly double the relative separation (fold-change +0.8331 in AD vs. +0.3981 in control; MW p = 1.22×10 -53 ; t-test p = 1.68×10 -67 ), and effect size increased to a measurable medium magnitude (Cliff Δ = 0.445). Mouse tissue retained the expected physiological pattern of PSD enrichment (101.38 to 32.02; fold-change -0.6842) but underwent a significant shift toward perisynaptic enrichment in Tg (114.68 to 74.89; fold-change -0.3469; MW p = 1.77×10 -33 ) (Figure 4k) (Supplementary Figure 7a) (Supplementary Table 15) . This pattern reflects a species-scaled trajectory, nanoscale mislocalization of BACE1 in human iPSC neurons, consolidation in human AD brain tissue, and full directional polarity shift in mouse models. Further, nanodomain area and length measurements revealed a biphasic structural remodeling. In human iPSC neurons, perisynaptic nanodomains were slightly larger than postsynaptic ones in both control and mutant neurons (+0.043 and +0.012 fold-change, respectively), although effect sizes were negligible, indicating subtle early remodeling. In human and mouse brain tissue, a striking collapse of the perisynaptic nanodomain area emerged. In transgenic mouse (Tg), perisynaptic area declined from 0.00651 to 0.00179 (fold-change -0.779; p < 1×10 -20 ), mirrored closely in human AD cortex, suggesting compaction into dense nanoclusters (Figure 4l, Supplementary Figure 7c) . Length measurements further demonstrated interesting observations across species and different models of AD. iPSC-derived neurons showed negligible structural differences (fold-change +0.0056 to +0.0149, non-significant) (Figure 4m, Supplementary Figure 7b) . Human samples demonstrated elongation of perisynaptic nanodomains in AD (fold-change +0.442; t-test p = 1.95×10 -6 ), consistent with early architectural expansion prior to collapse. In contrast, mouse AD showed the strongest compaction phenotype (fold-change -0.388; MW p = 4.24×10 -133 ), suggesting terminal pathological confinement (Supplementary Table 15) . These findings demonstrate that AD induces a reproducible and quantifiable nanoscale redistribution of BACE1, from PSD anchoring toward perisynaptic endocytic nanodomains, followed by progressive structural remodeling from diffuse domains into compact stable nanoclusters. This redistribution aligns with known amyloidogenic processing architecture, where endocytic microenvironments of acidic pH enhance β-secretase catalytic efficiency and trafficking activity. Importantly, the consistency of this nanoscale signature across these species and disease stages validates EvoAge’s prediction and reveals a conserved molecular progression underlying synaptic pathology in AD. DISCUSSION Aging is an inherently evolutionary process 42 , yet aging research has remained predominantly species-segregated, concept-fragmented, and methodologically siloed 43 . Although the conservation of aging pathways across phylogeny is well recognized 9,44–47 , translating insights across species remains technically constrained and conceptually fragmented 48 . In this study, we introduce EvoAge, a multispecies, AI-enabled knowledge framework that bridges evolutionary relationships with systems-level biological knowledge. By integrating 48 public resources into a billion-triple, orthology-aligned graph spanning six species, EvoAge demonstrates that aging biology can be computationally represented in a form that is both evolutionarily coherent and mechanistically navigable. The resulting architecture is not merely comprehensive; it reflects the scale-free 49 , modular topology characteristic of real biological systems, reaffirming that biological complexity 50 emerges naturally when previously isolated information streams are unified. A major conceptual contribution of this work is the clarification of the complementary roles of structured knowledge and generative AI in the scientific reasoning process 51 . Our benchmarking demonstrates that while general-purpose LLMs excel at linguistic fluency and memorization, they struggle with biological plausibility discrimination, frequently validating false hypotheses or producing confident hallucinations 52 . EvoAge’s hybrid design addresses this limitation by positioning the knowledge graph as the factual substrate, the embedding model as the probabilistic reasoning layer, and the LLM as the interpretive and interactive interface 53,54 . This division of labor yielded tangible improvements: EvoAge significantly outperformed state-of-the-art general LLMs on a challenging benchmark of 5,700 hypotheses derived from recent preprints, particularly in its ability to reject implausible hypotheses, a fundamental requirement for scientific inference. The ultimate test of a discovery platform, however, is not computational accuracy but biological relevance and experimental tractability. EvoAge met this criterion by prioritizing and supporting a mechanistically grounded hypothesis in AD that β-secretase (BACE1) undergoes nanoscale synaptic redistribution in AD, shifting from postsynaptic to perisynaptic and endocytic compartments where amyloidogenic processing occurs. Guided by EvoAge’s species-agnostic evaluation, we validated this hypothesis using human iPSC-derived neurons, human postmortem brain tissue, and mouse models. Across all systems, STED imaging confirmed increased perisynaptic BACE1 in AD, whereas healthy tissue showed the inverse distribution (Posterior > Perisynaptic), aligning with canonical synaptic organization. Statistical modeling further substantiated this finding. Gaussian mixture models revealed multimodal nanoscale populations rather than uniform distributions, and Bayesian hierarchical inference demonstrated consistent directional shifts toward perisynaptic enrichment across systems, with posterior confidence rising from 0.13-0.59 in raw data to 0.81-0.85 after outlier filtering. Despite small to moderate effect sizes, typical of high-content synaptic datasets and synaptic heterogeneity, the signal was robust, reproducible, and cross-species consistent. These results support a model in which AD induces both biochemical (increased BACE1 intensity) and morphological (altered area and length) nanoscale remodeling, forming compact perisynaptic nanodomains that may enhance catalytic efficiency via molecular crowding and vesicular retention properties, which are mechanistically favorable for increased processing of APP through the amyloidogenic pathway. While powerful, EvoAge is not yet complete. Its content reflects the biases and gaps of source databases, and its current orthology model simplifies many-to-many evolutionary relationships 55 . Furthermore, although EvoAge can evaluate hypotheses with high precision, transitioning from hypothesis validation to intervention prioritization necessitates integrating causal perturbation data 56 , gene regulatory dynamics, and spatial omics 57,58 . Future development will focus on expanding biological granularity, incorporating dynamic updates from new preprints and datasets, refining evolutionary logic to capture paralogy and gene replacement events, and integrating causal inference frameworks capable of predicting interventional outcomes. In the longer term, we envision EvoAge functioning as a federated scientific ecosystem where users contribute private or unpublished data, iteratively refine predictions, and collaboratively generate experimentally actionable insights. In summary, EvoAge represents a methodological shift, moving from fragmented knowledge and static databases toward a continuously evolving, evolution-informed computational framework that supports mechanistic reasoning across species. By unifying evolutionary theory, systems biology, and agentic AI, EvoAge is not simply a repository of aging knowledge. It is a platform that evaluates hypotheses, prioritizes mechanistic plausibility, and guides experimental discovery. As the biology of aging continues to expand in scale and complexity, such integrative frameworks will be essential for uncovering universal principles and actionable intervention targets in aging and age-associated disease. MATERIAL AND METHODS Data Source Curation and Collection Data were programmatically aggregated from 48 publicly available biological databases and knowledge graphs. This included 12 specialized aging resources (AgeAnno 7 , AgeAnnoMO 8 , AgeXtend 9 , Aging Atlas 10 , CellAge 5 , Digital Ageing Atlas 59 , DrugAge 16 , GenDR 17 , GeneAge 60 , HALD 11 , Literature, and MetaboAge 61 and 36 general-purpose biological databases spanning molecular interactions (STRING (v12.0) 62 , STITCH (v5.0) 63 , BioGRID (5.0) 64 , chemicals and drugs (ChEMBL (v33, released 31-May-2023) 65 , BindingDB (downloaded 28-May-2023) 66 , DRKG 67 , model organism databases (FlyBase 24 , WormBase WS287 25 , Worm Interactome Database 68 , ZFIN 26 , MGI 27 , MouseNet 27 , SGD 22 , YeastNet (v3) 23 ), and large-scale knowledge graphs (CKG) 15 , BioGrakn 69 , CROssBAR 70 , Hetionet 13 , MonarchKG 71 , PrimeKG 14 , Harmonizome 72 , TarKG 73 , DTInet 74 , GP-KG 75 , PharmKG 76 , iBKH 77 , TTD 78 . All data were retrieved in their latest available versions as of January 2025 (Supplementary Table 2) . Data Preprocessing and Harmonization A systematic preprocessing pipeline was implemented to handle the inherent heterogeneity of the source data. Each dataset was parsed to extract entities (nodes) and relationships (edges) conforming to a predefined biological entity types (Gene, Protein, Chemical, Disease, Phenotype, etc.). Entity names and relation labels were standardized to consistent naming conventions. To ensure interoperability, a rigorous identifier mapping framework was applied, converting all source identifiers to community-standard reference systems: NCBI 79 Gene for genes, UniProt 80 for proteins, PubChem 81 /DrugBank 82 for chemicals, DOID 83 /MONDO 84 for diseases, UBERON 85 for anatomy, and KEGG 86 /Reactome 87 for pathways. Gene Ontology (GO) 88 is used for Biological Process, Cellular Component, and Molecular Function, and HPO 89 for phenotypes. Duplicate entities were identified via exact and semantic matching and merged into single, harmonized nodes while preserving all original source identifiers for full data provenance. During this process, entities originally classified as chemicals, drugs, metabolites, or phytochemicals were all consolidated under the single 'Chemical' node type. This unification was performed by merging all nodes that share the same canonical SMILES string using openbabel 90 , ensuring a single, unique representation for each distinct molecule. Orthology-Based Cross-Species Integration To enable true cross-species querying, a one-to-one human ortholog mapping strategy was implemented. Genes from S. cerevisiae , C. elegans , D. melanogaster , D. rerio , and M. musculus were mapped to their H. sapiens orthologs using the g:Profiler2 (G:Orth) tool 91 , which queries the Ensembl Compara database. For genes with multiple potential orthologs, only the first-listed ortholog was retained to maintain a clean, unambiguous graph. The original species-specific gene identifiers in the graph were subsequently replaced with their corresponding human ortholog identifiers, genetically unifying the knowledge base. Graph Database Implementation The harmonized data were loaded into a Neo4j graph database (version 5.26.10) 92 using Cypher queries and APOC (version 5.22.0) utilities for batch operations and integrity checks. The final EvoAge Knowledge Graph comprises 30,159,124 nodes and 1,043,615,837 edges (triples) spanning 66 relation types, creating a scalable, queryable resource for multi-species biological exploration. Knowledge Graph Embedding (KGE) Training Vector representations of entities and relations were learned using the Deep Graph Library (DGL, version 1.1.2) 93 , and DGL-KE (version 0.1.0) 36 (Supplementary Table 8) . Six KGE models (TransE, RotatE, SimplE, DisMult, RESCAL, ComplEx) 28–31 , 33,34 were systematically evaluated. The triples were split into training (80%), validation (10%), and test (10%) sets. Models were trained with an embedding dimension of 64, a batch size of 2048, and a margin loss function. The optimal architectures were identified as RESCAL for the smaller Aging KG and RotatE for the full EvoAge KG. An ablation study on embedding size (64, 128, 256, 512) determined the final configurations: a 512-dimensional RESCAL model for the Aging KG and a 128-dimension RotatE model for the EvoAge KG to balance performance and avoid overfitting. Computational Infrastructure All training and evaluation were executed on a high-performance Linux workstation (Ubuntu 22.04 LTS) equipped with two NVIDIA GPUs, an RTX 5000 Ada Generation (32 GB VRAM), and a GeForce RTX 3090 (24 GB VRAM) running CUDA 12.0. This hardware configuration enabled parallelized computation and efficient handling of large-scale graph data, comprising over 30 million nodes and 1.04 billion edges. Type-Constrained Link Prediction and Inference Optimization To ensure that EvoAge generates semantically valid and biologically meaningful predictions, we extended the default DGL-KE inference pipeline to incorporate explicit type constraints. An entity-type mapping was integrated into the system, enabling the inference module to identify the required semantic category for each relation. During prediction, candidate entities are first filtered by their assigned type, and only those matching the expected head or tail type of the queried relation are considered for scoring and ranking. This modification prevents invalid cross-type predictions-for example, ensuring that a gene–disease relation can only return disease entities as plausible targets, and thereby improves both the interpretability and precision of link prediction and hypothesis generation. To enable real-time inference within the EvoAge chatbot, we optimized the DGL-KE based prediction pipeline by introducing an internal caching mechanism for model artifacts. In the default workflow, essential components such as the entity and relation dictionaries are reloaded from disk for every query, whether the request involves hypothesis generation or link prediction. This reload step typically takes 30-50 seconds, and repeating it for every query makes interactive analysis infeasible. To overcome this limitation, we modified the DGL-KE library so that these artifacts are loaded only once and then retained in memory. With this caching mechanism, subsequent inference calls bypass the expensive reload phase entirely, reducing artifact access time to approximately 2-4 seconds. This optimization substantially lowers end-to-end latency and enables near real-time performance for iterative hypothesis testing and link prediction within EvoAge. Hypothesis Testing and Youden's Thresholding A data-driven approach established objective thresholds for the "Test Hypothesis" function. For each of the 66 relation types, the KGE model was evaluated using 1,000 known true triples and 1,000 generated false triples. The Youden's J statistic (J = True Positive Rate - False Positive Rate) was applied to these score distributions to determine the optimal, relation-specific cutoff for accepting a user-proposed triple as plausible. Backend API and Service Architecture The EvoAge platform backend was built as a FastAPI server, providing RESTful endpoints for core functions: /search_biological_entities for faceted full-text search (powered by a Neo4j/Lucene index); /subgraph for 1-hop neighborhood retrieval; /entity_relationships for quantitative adjacency summarization; /check_relationship for paired-node verification; /get_sample_nodes for retrieving sample nodes from the KG; and /get_sample_triples for retrieving sample triples from the KG. Additionally, the platform offers advanced computational endpoints. A Link Prediction endpoint calls the KGE scoring service, while a dedicated Hypothesis testing endpoint orchestrates calls to this same service, the Neo4j graph for fact-checking, and the Gemini API for natural language interpretation. Agentic Framework and Query Processing The EvoAge-agent, built using the Kani framework and GPT-4o-mini 94 , serves as the central reasoning engine. It processes user queries by first performing biological entity extraction and enrichment via the search API. It then detects user intent to route the query to one of three specialized tool paths: direct KG lookup, link prediction, or hypothesis testing. The agent synthesizes the results from these tools into a final, coherent natural language response. Frontend User Interface The graphical user interface was developed using Streamlit (v1.48.0). It features a secure authentication module and a chat-based interface where users can submit natural language queries and view responses that combine explanatory text with structured, ranked predictions, including KGE confidence scores. Computational Benchmarking A rigorous benchmark was designed to evaluate the hypothesis-testing capability. From 1,282 recent BioRxiv preprints (Jan-Sep 2025) on aging, 980 were manually curated. Using Gemini 2.5 Flash Lite, 5,700 balanced positive and negative hypothesis questions were generated. EvoAge and three general-purpose LLMs answered these questions. The specific models evaluated were "google/gemma-2bit" 38 , "llama3.1:8b-instruct-q4_K_M" 95 , and "Qwen/Qwen2.5-7B-Instruct" 40 . All answers were graded on a 1-10 scale for biological feasibility by a "BioMistral/BioMistral-7B" 41 teacher model. All LLMs were run using the vLLM inference engine (v0.10.1.1) 96,97 . Performance was assessed using a composite "clarity score," calculated as (Positive Grade) + (11 - Negative Grade). This metric rewards models that both accept plausible hypotheses (high positive score) and reject implausible ones (low negative score). Ethical Compliance The human brain and transgenic mouse data were sourced as reported previously. 98 and were repurposed in this study for novel analyses. All human stem cell (iPSCs) data presented here were newly generated. All human stem cell work for iPSCs was carried out in accordance with approval from the Institutional Human Ethics Committee and Institutional Biosafety Committee at the Institute for Stem Cell Science and Regenerative Medicine, Bengaluru, India. iPSC Generation and Gene Editing The induced pluripotent stem cells (iPSCs) were generated from skin fibroblasts of a 58-year-old male carrying the Familial Alzheimer’s Disease (FAD) mutation L150P in the Presenilin1 (PSEN1) protein. The patient carried a heterozygous point mutation (c.449C → T) in exon 6 of the PSEN1 gene. The iPSCs were generated by the electroporation of the fibroblasts with episomal plasmids containing hOCT4, hSOX2, hNANOG, hKLF4, hMyc, hLIN28, and shRNA against TP53. The patient-derived iPSCs have been characterized for normal karyotype, expression of pluripotency markers, and differentiation into the three germ layers 99 . The isogenic control line for the same iPSCs was generated using CRISPR/Cas9 gene editing. A 23bp guide RNA was used to revert the point mutation (c.449C → T) to obtain the gene-corrected iPSC line. The gene-corrected iPSCs have been characterized for normal karyotype and expression of pluripotency markers 100 . Stem Cell Culture, Maintenance, and Differentiation The iPSC colonies were maintained in mTeSR1 complete medium supplemented with 1% Penicillin-Streptomycin under the conditions of 37˚C, 5% CO 2 and 21% O 2 . When confluent, they were split in a 1:3 ratio onto pre-coated Matrigel dishes using a dissociation mixture prepared with Collagenase Type IV (1 mg/ml), Trypsin (0.25%), Knockout Serum Replacement (20%), and Calcium Chloride (1 mM). The protocol for neural differentiation was the same as published previously 101 . In brief, iPSCs were expanded in mTeSR until 70-80% confluency was reached. The Neural Basic Media (NBM) for differentiation contained 50% DMEM F-12, 50% Neurobasal, 0.1% PenStrep, Glutamax, N 2 , and B27 without Vitamin A. Neural induction was initiated in the iPSC monolayer via Dual SMAD inhibition. This was achieved by replacing the existing media with Neural Induction Media (NIM), containing the Neural Basic Media (NBM) supplemented with the TGFβ pathway inhibitor SB431542 (10 µM) and the BMP pathway inhibitor LDN193189 (0.1 µM). The cells were subjected to neural induction for 12-15 days by changing the NIM every day. After induction, the monolayer was dissociated using Accutase, and the cells were plated in NIM containing 10 µM ROCK inhibitor overnight on pre-coated poly-L-ornithine/laminin dishes. Poly-L-Ornithine (1:10 dilution) followed by Laminin (5µg/ml) coating was used for the maintenance of neural progenitors and their terminal differentiation. Expansion of neural progenitor cells was carried out in Neural Expansion Media (NEM), which is composed of NBM supplemented with FGF (10 ng/mL) and EGF (10 ng/mL). Neuronal maturation and terminal differentiation were achieved by plating the neural stem cells at a density of 25,000-35,000 cells/cm2 in the Neural Maturation Media (NMM) composed of NBM supplemented with BDNF (20ng/ml), GDNF (10ng/ml), L-Ascorbic Acid (200 µM), and db-Camp (50µM). The neurons were subjected to maturation for a period of 60-70 days by supplementing them with NMM every 4-5 days. The differentiated neurons were characterized by their expression of glutamatergic, axonal, and dendritic markers 102 . The list of materials used are detailed in (Supplementary Table 16) . Immunocytochemistry For iPSCs, the immunocytochemistry was performed as reported previously ( 103 , 98 , 104 ). Briefly, cells were fixed with 4% paraformaldehyde plus 4% sucrose in PBS at 4°C for 10 minutes, followed by quenching with 0.1M glycine in PBS at room temperature and permeabilization with 0.25% Triton X-100 for 5 minutes, and then blocked with 10% Bovine Serum Albumin (BSA) in PBS for 30 minutes at room temperature. This was followed by incubation with the appropriate primary antibody for 1-2 hr. The primary antibodies used were Anti-BACE1 (Biolegend/Covance, #840101) (1:200), Anti-Shank2 (Synaptic Systems, #162204) (1:500), and Anti-Clathrin (Abcam, #ab2731) (1:100). Following washing, cells were then incubated with a suitable secondary antibody for 45 minutes. The secondary antibodies used include Alexa Fluor 594 (Life Technologies, #A11037) (1:200), Abberior Star Red (Abberior, #2-0112-011-8 and #2-0002-011-2) (1:200). Following washing, cells were mounted with Prolong (Molecular Probes, cat. no. MAN0010261) for STED imaging. Immunohistochemistry (Mice) Immunohistochemical data from mouse samples were obtained from experiments previously performed and reported 98 . Coronal cryosections (25 µm) from Tg(APPswe/PS1ΔE9) mice (JAX Stock #004462) and age-matched control littermates were stained and mounted in ProLong with DAPI (Molecular Probes, cat. no. P36962) for confocal and STED imaging. The primary and secondary antibodies used are listed in 98 . Imaging was conducted in the CA1–CA2 stratum radiatum of the hippocampus. Immunohistochemistry (Human) Human immunohistochemical data were similarly sourced from previously performed and reported experiments 98 . Neuroanatomical sampling followed NIA-AA guidelines for the neuropathological assessment of Alzheimer’s disease (reference). Ethical clearance for the collection, storage, and distribution of human brain tissue was obtained from the Human Brain Tissue Repository (Brain Bank) at NIMHANS, Bangalore. Neuropathological classification followed standardized criteria incorporating three parameters: Aβ plaque score, Braak and Braak neurofibrillary tangle (NFT) stage, and CERAD neuritic plaque score to derive an ABC score categorized into four levels: not, low, intermediate, and high. The Blessed Dementia Rating Scale score for the control case was 1.5/17. A summary of staging and pathological features for all samples has been reported previously 98 . Paraffin-embedded human tissue sections were stained with primary and secondary antibodies, followed by mounting with ProLong containing DAPI (Molecular Probes, cat. no. P36962) as reported previously 98 . Stimulated Emission Depletion microscopy (STED) A commercial STED inverted microscope (Abberior Expert Line 775 nm, Abberior Instruments GmbH, Göttingen, Germany) was used to obtain confocal and super-resolved images of the same region with a sampling of 15 nm. The microscope was equipped with two pulsed excitation lasers at 561 nm and 640 nm, as well as a pulsed depletion laser at 775 nm. The laser powers were adjusted to 70%, 50% and 40% of their respective total power for 561 nm, 640 nm, and 775 nm, respectively, as described previously 98,103 . Semiautomated detection of dendritic compartments and functional zones of an excitatory synapse The active dendritic area of the protein of interest and synapses were distinguished from the rest of the dendrite using a protocol as described previously 98,104,105 . Briefly, the intensity of the epifluorescence/confocal images of markers for different functional zones of the synapse (post/peri) was thresholded to generate the mask of the puncta. A spine morphometry analysis was then performed, and masks were filtered using various morphological filters, such as length, breadth, and area, through the IMA plugin running inside the MetaMorph software (Molecular Devices) 106 . A similar analysis was performed on super-resolution images to detect functional zones of an excitatory synapse (post- and peri-synaptic) that correspond to PSD/EZ functional zones. STED Nanodomain Analysis β-secretase (BACE1) nanodomains were identified in STED images using Palm-Tracer, as described previously 98,103,105,107 . Nanodomains were analyzed using two-dimensional Gaussian fitting, from which morphological and biophysical parameters, such as length (2.3σlong), area, and nanodomain intensity, were computed for each experimental group. Gaussian fitting was performed on every cluster identified as a nanodomain. Nanodomain metric distribution analysis Statistical analyses and visualization were performed using Python, employing the following libraries and versions: pandas (v2.3.2), numpy (v2.2.6), matplotlib (v3.10.3), statsmodels (v0.14.5), and scikit-learn (v1.15.3). The normality of BACE1 intensity data across all experimental groups was first rigorously assessed using the Shapiro–Wilk test (α = 0.05) via SciPy to determine whether the distributions conformed to Gaussian assumptions. Data failing the test (p-value =< 0.05) was categorized as Non-Gaussian. To investigate underlying heterogeneity, nanodomain morphological and biophysical parameters (Length, Area, Intensity) were subjected to Gaussian Mixture Model (GMM) analysis. We compared 1- and 2-component GMMs using the Bayesian Information Criterion (BIC), concluding that the presence of two distinct subpopulations was indicated if the 2-component model yielded a significantly lower BIC. The Mann-Whitney U test (also known as the Wilcoxon rank-sum test) was selected as the non-parametric method of choice for statistical comparisons, given that initial normality tests indicated that the BACE1 nanodomain metrics were non-Gaussian. This robust test was applied to two critical comparisons: 1) Synaptic Compartment Comparison, evaluating the differences in BACE1 properties between the Postsynaptic and Perisynaptic regions within a single genotype; and 2) Genotype Comparison, assessing the effect of the disease (AD/Mutant) versus Control (GC/Wild-Type) on BACE1 nanodomain metrics within a specific synaptic region, across all species models. Normalization was performed by dividing all BACE1 nanodomain metric values (Intensity, Area, and Length) within a specific species and measurement by the median Postsynaptic Control value for that same species/measurement. This step ensured that the absolute scale differences between models (iPSC, Mouse, Human) were removed, making the relative remodeling effect comparable across species. Following normalization, the Mann-Whitney U test was selected as the non-parametric method for statistical comparisons. This robust test was applied to the normalized data to compare the AD genotype against the Healthy within both the Postsynaptic (PSD) and Perisynaptic (EZ) compartments across all three models. The goal was to statistically confirm that the observed nanoscale shift is a fundamental, conserved feature of Alzheimer's pathogenesis. Statistical testing was performed in R using ggpubr (v0.6.2), gghalves (v0.1.4) for violin geometries, ggpubr (v0.6.2) for statistical annotations. Statistical Analysis All statistical analyses were performed using R (v4.2.3), with significance set at p < 0.05 (* p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001). For the computational benchmark, graded score distributions were determined to be non-parametric and were compared using the Wilcoxon Rank-Sum (Mann-Whitney U) test with Benjamini-Hochberg FDR correction, using the ggpubr R package (v0.6.2). This was confirmed with a one-way ANOVA and Tukey's HSD post-hoc test. For experimental datasets, statistical comparisons between genotypes (Healthy vs. AD) and regions (postsynaptic vs. perisynaptic) were performed using non-parametric Mann–Whitney U tests (wilcox.test), along with Cliff’s delta effect sizes (effsize package). Outliers were removed using an IQR-based rule, and significance annotations were generated using ggpubr and custom R functions. All visualizations were generated using tidyverse (v2.0.0), gghalves (v0.1.4), ggpubr (v0.6.2), ggprism (v1.0.7), gridExtra (v2.3), broom(v1.0.10). Declarations Data and Code Availability The complete source code for the EvoAge platform is publicly available on GitHub (https://github.com/the-ahuja-lab/EvoAge). The EvoAge chatbot web server is publicly accessible at https://evoage.ahujalab.iiitd.edu.in/. The full dataset generated and used in this study, including the final knowledge graph, is archived on Zenodo at https://doi.org/10.5281/zenodo.17711173. For enhanced reproducibility, a pre-configured Docker container is also available on Docker Hub at https://hub.docker.com/r/ahujalab/evoage-project. Declaration of Interests The authors declare no competing interests. Acknowledgments The authors thank the IT-HelpDesk team at IIIT-Delhi for their assistance with computational resources. We thank all the members of the Ahuja lab for their intellectual contributions at various stages of this project. The Ahuja lab is supported by the Ramalingaswami Re-entry Fellowship (BT/HRD/35/02/2006), and a research grant (BT/PR52020/AI/133/180/2024) by the Department of Biotechnology, Ministry of Science & Technology, Government of India, and an intramural Start-up grant from Indraprastha Institute of Information Technology-Delhi. DN acknowledges the senior research grant from DBT Wellcome Trust India Alliance (IA/S/23/2/507005), support from Centre for Brain Research through Funding for Aging Brain Research & Innovation Collaboration, and Core Research grant from Anusandhan National Research Foundation (CRG/2022/002726) and Intramural support from IISc. Author Contributions The study was conceived and supervised by G.A. A.S. built the entire foundation of EvoAge, and A.S and An.S. designed the architecture of EvoAge Agent. A.K.S worked on frontend, P.S, and A.K, worked on data curation. A.S and A.G. worked on model training. S.K performed the mouse and human experiments and microscopy, under the supervision of D.N. Other experiments were carried out by S.R., and R.M. K.F. provided pateient derived stem cells lines. Statistical analysis, Evoage workflow finalization and testing was performed by V.G, S.S, S.C, S.K, Si.S, S.D, S.A, and R.S. Statistical analysis was guided by D.S. Figures were illustrated by G.A., and A.S., and built by A.S. Manuscript was written by A.S., and G.A. All authors read and approved the manuscript. References Borrás, C. The challenge of unlocking the biological secrets of aging. Front. Aging 2 , 676573 (2021). López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, G. Hallmarks of aging: An expanding universe. Cell 186 , 243–278 (2023). Sarygina, E., Kliuchnikova, A., Tarbeeva, S., Ilgisonis, E. & Ponomarenko, E. Model organisms in aging research: Evolution of database annotation and ortholog discovery. Genes (Basel) 16 , 8 (2024). Guo, J. et al. Aging and aging-related diseases: from molecular mechanisms to interventions and treatments. Signal Transduct Target Ther 7 , 391 (2022). Avelar, R. A. et al. A multidimensional systems biology analysis of cellular senescence in aging and disease. Genome Biol. 21 , 91 (2020). Bitto, A., Wang, A. M., Bennett, C. F. & Kaeberlein, M. Biochemical Genetic Pathways that Modulate Aging in Multiple Species. Cold Spring Harb. Perspect. Med. 5 , (2015). Huang, K. et al. AgeAnno: a knowledgebase of single-cell annotation of aging in human. Nucleic Acids Res. 51 , D805–D815 (2023). Huang, K. et al. AgeAnnoMO: a knowledgebase of multi-omics annotation for animal aging. Nucleic Acids Res. 52 , D822–D834 (2024). Arora, S. et al. Discovering geroprotectors through the explainable artificial intelligence-based platform AgeXtend. Nat. Aging 5 , 144–161 (2025). Aging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. Nucleic Acids Res. 49 , D825–D830 (2021). Wu, Z. et al. HALD, a human aging and longevity knowledge graph for precision gerontology and geroscience analyses. Sci. Data 10 , 851 (2023). Sturm, G. et al. A multi-omics longitudinal aging dataset in primary human fibroblasts with mitochondrial perturbations. Sci. Data 9 , 751 (2022). Himmelstein, D. S. et al. Systematic integration of biomedical knowledge prioritizes drugs for repurposing. Elife 6 , (2017). Chandak, P., Huang, K. & Zitnik, M. Building a knowledge graph to enable precision medicine. Sci. Data 10 , 67 (2023). Santos, A. et al. A knowledge graph to interpret clinical proteomics data. Nat. Biotechnol. 40 , 692–702 (2022). Barardo, D. et al. The DrugAge database of aging-related drugs. Aging Cell 16 , 594–597 (2017). Wuttke, D. et al. Dissecting the gene network of dietary restriction to identify evolutionarily conserved pathways and new functional genes. PLoS Genet. 8 , e1002834 (2012). Anh-Hoang, D., Tran, V. & Nguyen, L.-M. Survey and analysis of hallucinations in large language models: attribution to prompting strategies or model behavior. Front. Artif. Intell. 8 , 1622292 (2025). Bélisle-Pipon, J.-C. Why we need to be careful with LLMs in medicine. Front. Med. (Lausanne) 11 , 1495582 (2024). Feng, Y. et al. Knowledge graph-based thought: a knowledge graph-enhanced LLM framework for pan-cancer question answering. Gigascience 14 , (2025). Gao, Y. et al. Leveraging medical knowledge graphs into large language models for diagnosis prediction: Design and application study. JMIR AI 4 , e58670 (2025). Cherry, J. M. et al. SGD: Saccharomyces Genome Database. Nucleic Acids Res. 26 , 73–79 (1998). Kim, H. et al. YeastNet v3: a public database of data-specific and integrated functional gene networks for Saccharomyces cerevisiae. Nucleic Acids Res. 42 , D731–6 (2014). Öztürk-Çolak, A. et al. FlyBase: updates to the Drosophila genes and genomes database. Genetics 227 , (2024). Sternberg, P. W. et al. WormBase 2024: status and transitioning to Alliance infrastructure. Genetics 227 , (2024). Bradford, Y. et al. ZFIN: enhancements and updates to the Zebrafish Model Organism Database. Nucleic Acids Res. 39 , D822–9 (2011). Baldarelli, R. M. et al. Mouse Genome Informatics: an integrated knowledgebase system for the laboratory mouse. Genetics 227 , (2024). Bordes, A., Usunier, N., García-Durán, A., Weston, J. & Yakhnenko, O. Translating embeddings for modeling multi-relational data. Neural Inf Process Syst 2787–2795 (2013). Sun, Z., Deng, Z.-H., Nie, J.-Y. & Tang, J. RotatE: Knowledge graph embedding by relational rotation in complex space. arXiv [cs.LG] (2019). Kazemi, S. M. & Poole, D. SimplE embedding for link prediction in knowledge graphs. arXiv [stat.ML] (2018). Yang, B., Yih, W.-T., He, X., Gao, J. & Deng, L. Embedding entities and relations for learning and inference in knowledge bases. arXiv [cs.CL] (2014). Tran, H. N. & Takasu, A. Analyzing knowledge graph embedding methods from a multi-embedding interaction perspective. arXiv [cs.LG] (2019). Nickel, M., Tresp, V. & Kriegel, H. A Three-Way model for collective learning on multi-relational data. ICML 809–816 (2011). Trouillon, T., Welbl, J., Riedel, S., Gaussier, É. & Bouchard, G. Complex embeddings for simple link prediction. ICML abs/1606.06357 , 2071–2080 (2016). Stothers, J. A. M. & Nguyen, A. Can Neo4j replace PostgreSQL in healthcare? AMIA Summits Transl. Sci. Proc. 2020 , 646–653 (2020). Zheng, D. et al. DGL-KE: Training Knowledge Graph Embeddings at Scale. in Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (ACM, New York, NY, USA, 2020). doi:10.1145/3397271.3401172. Zhu, A., Dugan, L., Hwang, A. & Callison-Burch, C. Kani: A lightweight and highly hackable framework for building language model applications. in Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023) (eds. Tan, L., Milajevs, D., Chauhan, G., Gwinnup, J. & Rippeth, E.) 65–77 (Empirical Methods in Natural Language Processing, Stroudsburg, PA, USA, 2023). Gemma Team et al. Gemma: Open models based on Gemini research and technology. arXiv [cs.CL] (2024). Grattafiori, A. et al. The Llama 3 herd of models. arXiv [cs.AI] (2024). Qwen et al. Qwen2.5 Technical Report. arXiv [cs.CL] (2024). Labrak, Y. et al. BioMistral: A collection of open-source pretrained large Language Models for medical domains. Annu Meet Assoc Comput Linguistics 5848–5864 (2024). Kirkwood, T. B. Evolution of ageing. Nature 270 , 301–304 (1977). Mitchell, S. J., Scheibye-Knudsen, M., Longo, D. L. & de Cabo, R. Animal models of aging research: implications for human aging and age-related diseases. Annu. Rev. Anim. Biosci. 3 , 283–303 (2015). Kenyon, C. J. The genetics of ageing. Nature 464 , 504–512 (2010). López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, G. The hallmarks of aging. Cell 153 , 1194–1217 (2013). Gautam, V. et al. scCamAge: A context-aware prediction engine for cellular age, aging-associated bioactivities, and morphometrics. Cell Rep. 44 , 115270 (2025). Duari, S., Gautam, V. & Ahuja, G. Protocol for cellular age prediction in yeast and human single cells using transfer learning. STAR Protoc. 6 , 104023 (2025). Campisi, J. et al. From discoveries in ageing research to therapeutics for healthy ageing. Nature 571 , 183–192 (2019). Barabasi, A. L. & Albert, R. Emergence of scaling in random networks. Science 286 , 509–512 (1999). Ideker, T. & Krogan, N. J. Differential network biology. Molecular Systems Biology 8 , (2012). Rajabi, E. & Etminani, K. Knowledge-graph-based explainable AI: A systematic review. J. Inf. Sci. 50 , 1019–1029 (2024). Zhang, H., Deng, H., Ou, J. & Feng, C. Mitigating spatial hallucination in large language models for path planning via prompt engineering. Sci. Rep. 15 , 8881 (2025). Anonymous. Grounding LLM reasoning with knowledge graphs. Preprint at https://openreview.net/pdf/f2d669a35707223183429f92ff75668b839cee34.pdf (2024). Amayuelas, A., Sain, J., Kaur, S. & Smiley, C. Grounding LLM reasoning with knowledge graphs. arXiv [cs.CL] (2025). Gabaldón, T. & Koonin, E. V. Functional and evolutionary implications of gene orthology. Nat. Rev. Genet. 14 , 360–366 (2013). Gasperini, M. et al. A genome-wide framework for mapping gene regulation via cellular genetic screens. Cell 176 , 377–390.e19 (2019). Marbach, D. et al. Wisdom of crowds for robust gene network inference. Nat. Methods 9 , 796–804 (2012). Moffitt, J. R. et al. Molecular, spatial, and functional single-cell profiling of the hypothalamic preoptic region. Science 362 , eaau5324 (2018). Craig, T. et al. The Digital Ageing Atlas: integrating the diversity of age-related changes into a unified resource. Nucleic Acids Res. 43 , D873–8 (2015). de Magalhães, J. P. et al. Human Ageing Genomic Resources: updates on key databases in ageing research. Nucleic Acids Res. 52 , D900–D908 (2024). Bucaciuc Mracica, T. et al. MetaboAge DB: a repository of known ageing-related changes in the human metabolome. Biogerontology 21 , 763–771 (2020). Szklarczyk, D. et al. The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest. Nucleic Acids Res. 51 , D638–D646 (2023). Szklarczyk, D. et al. STITCH 5: augmenting protein-chemical interaction networks with tissue and affinity data. Nucleic Acids Res. 44 , D380–4 (2016). Oughtred, R. et al. The BioGRID database: A comprehensive biomedical resource of curated protein, genetic, and chemical interactions. Protein Sci. 30 , 187–200 (2021). Zdrazil, B. et al. The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. Nucleic Acids Res. 52 , D1180–D1192 (2024). Liu, T., Lin, Y., Wen, X., Jorissen, R. N. & Gilson, M. K. BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities. Nucleic Acids Res. 35 , D198–201 (2007). He, S. & Zhao, X. Drug repositioning model based on knowledge graph embedding. Sci. Rep. 15 , 10298 (2025). Simonis, N. et al. Empirically controlled mapping of the Caenorhabditis elegans protein-protein interactome network. Nat. Methods 6 , 47–54 (2009). Messina, A. et al. BioGrakn: A knowledge graph-based semantic database for biomedical sciences. in Advances in Intelligent Systems and Computing 299–309 (Springer International Publishing, Cham, 2018). Doğan, T. et al. CROssBAR: comprehensive resource of biomedical relations with knowledge graph representations. Nucleic Acids Res. 49 , e96 (2021). Putman, T. E. et al. The Monarch Initiative in 2024: an analytic platform integrating phenotypes, genes and diseases across species. Nucleic Acids Res. 52 , D938–D949 (2024). Diamant, I., Clarke, D. J. B., Evangelista, J. E., Lingam, N. & Ma’ayan, A. Harmonizome 3.0: integrated knowledge about genes and proteins from diverse multi-omics resources. Nucleic Acids Res. 53 , D1016–D1028 (2025). Zhou, C. et al. TarKG: a comprehensive biomedical knowledge graph for target discovery. Bioinformatics 40 , (2024). Yao, Q., Chen, Z., Cao, Y. & Hu, H. Enhancing drug-target interaction prediction with graph representation learning and knowledge-based regularization. Front. Bioinform. 5 , 1649337 (2025). Gao, Z., Ding, P. & Xu, R. KG-Predict: A knowledge graph computational framework for drug repurposing. J. Biomed. Inform. 132 , 104133 (2022). Zheng, S. et al. PharmKG: a dedicated knowledge graph benchmark for bomedical data mining. Brief. Bioinform. 22 , (2021). Su, C. et al. Biomedical discovery through the integrative biomedical knowledge hub (iBKH). iScience 26 , 106460 (2023). Zhou, Y. et al. TTD: Therapeutic Target Database describing target druggability information. Nucleic Acids Res. 52 , D1465–D1477 (2024). NCBI Resource Coordinators. Database resources of the National Center for Biotechnology Information. Nucleic Acids Res. 46 , D8–D13 (2018). UniProt Consortium. UniProt: The universal protein knowledgebase in 2023. Nucleic Acids Res. 51 , D523–D531 (2023). Kim, S. et al. PubChem in 2021: new data content and improved web interfaces. Nucleic Acids Res. 49 , D1388–D1395 (2021). Knox, C. et al. DrugBank 6.0: The DrugBank knowledgebase for 2024. Nucleic Acids Res. 52 , D1265–D1275 (2024). Schriml, L. M. et al. Human Disease Ontology 2018 update: classification, content and workflow expansion. Nucleic Acids Res. 47 , D955–D962 (2019). Vasilevsky, N. A. et al. Mondo: Integrating disease terminology across communities. Genetics (2025) doi:10.1093/genetics/iyaf215. Mungall, C. J., Torniai, C., Gkoutos, G. V., Lewis, S. E. & Haendel, M. A. Uberon, an integrative multi-species anatomy ontology. Genome Biol. 13 , R5 (2012). Kanehisa, M., Furumichi, M., Sato, Y., Ishiguro-Watanabe, M. & Tanabe, M. KEGG: integrating viruses and cellular organisms. Nucleic Acids Res. 49 , D545–D551 (2021). Gillespie, M. et al. The reactome pathway knowledgebase 2022. Nucleic Acids Res. 50 , D687–D692 (2022). Binns, D. et al. QuickGO: a web-based tool for Gene Ontology searching. Bioinformatics 25 , 3045–3046 (2009). Köhler, S. et al. Expansion of the Human Phenotype Ontology (HPO) knowledge base and resources. Nucleic Acids Res. 47 , D1018–D1027 (2019). O’Boyle, N. M. et al. Open Babel: An open chemical toolbox. J. Cheminform. 3 , 33 (2011). Kolberg, L., Raudvere, U., Kuzmin, I., Vilo, J. & Peterson, H. gprofiler2 -- an R package for gene list functional enrichment analysis and namespace conversion toolset g:Profiler. F1000Res. 9 , 709 (2020). Webber, J. A programmatic introduction to Neo4j. in Proceedings of the 3rd annual conference on Systems, programming, and applications: software for humanity (ACM, New York, NY, USA, 2012). doi:10.1145/2384716.2384777. Wang, M. et al. Deep Graph Library: A graph-centric, highly-performant package for graph neural networks. arXiv [cs.LG] (2019). OpenAI. https://openai.com. Touvron, H. et al. LLaMA: Open and efficient foundation language models. arXiv [cs.CL] (2023). Kwon, W. et al. Efficient memory management for large language model serving with PagedAttention. arXiv [cs.LG] (2023). Wolf, T. et al. HuggingFace’s transformers: State-of-the-art natural language processing. arXiv [cs.CL] (2019). Kedia, S. et al. Alteration in synaptic nanoscale organization dictates amyloidogenic processing in Alzheimer’s disease. iScience 24 , 101924 (2021). Tubsuwan, A. et al. Generation of induced pluripotent stem cells (iPSCs) from an Alzheimer’s disease patient carrying a L150P mutation in PSEN-1. Stem Cell Res. 16 , 110–112 (2016). Poon, A. et al. Generation of a gene-corrected isogenic control hiPSC line derived from a familial Alzheimer’s disease patient carrying a L150P mutation in presenilin 1. Stem Cell Res. 17 , 466–469 (2016). Ramakrishna, S. et al. Distinct calcium sources regulate temporal profiles of NMDAR and mGluR-mediated protein synthesis. Life Sci. Alliance 7 , e202402594 (2024). Ochalek, A. et al. Neurons derived from sporadic Alzheimer’s disease iPSCs reveal elevated TAU hyperphosphorylation, increased amyloid levels, and GSK3B activation. Alzheimers. Res. Ther. 9 , 90 (2017). Kedia, S. et al. Real-time nanoscale organization of amyloid precursor protein. Nanoscale 12 , 8200–8215 (2020). Kedia, S., Ramanan, N. & Nair, D. Quantifying molecular aggregation by super resolution microscopy within an excitatory synapse from mouse hippocampal neurons. STAR Protoc. 2 , 100470 (2021). Kedia, S. et al. Nanoscale organization of Nicastrin, the substrate receptor of the γ-secretase complex, as independent molecular domains. Mol. Brain 14 , 158 (2021). MetaMorph Microscopy Automation and Image Analysis Software. https://www.moleculardevices.com/products/cellular-imaging-systems/high-content-analysis/metamorph-microscopy. Nair, D. et al. Super-resolution imaging reveals that AMPA receptors inside synapses are dynamically organized in nanodomains regulated by PSD95. J. Neurosci. 33 , 13204–13224 (2013). Additional Declarations There is NO Competing Interest. Supplementary Files SuppTable1.xlsx Supplementary Table 1 SuppTable3.xlsx Supplementary Table 3 SuppTable2.xlsx Supplementary Table 2 SuppTable15.xlsx Supplementary Table 15 SuppTable11.xlsx Supplementary Table 11 SuppTable9.xlsx Supplementary Table 9 SupplementaryFigure7.pdf Supplementary Figure 7 SuppTable12.xlsx Supplementary Table 12 SuppTable5.xlsx Supplementary Table 5 SupplementaryFigure3.pdf Supplementary Figure 3 SuppTable8.xlsx Supplementary Table 8 SupplementaryFigure8.pdf Supplementary Figure 8 SupplementaryFigure1.pdf Supplementary Figure 1 SuppTable13.xlsx Supplementary Table 13 SuppTable7.xlsx Supplementary Table 7 SuppTable10.xlsx Supplementary Table 10 SuppTable14.xlsx Supplementary Table 14 SupplementaryFigure6.pdf Supplementary Figure 6 SuppTable4.xlsx Supplementary Table 4 SuppTable6.xlsx Supplementary Table 6 SupplementaryFigure4.pdf Supplementary Figure 4 SupplementaryFigure5.pdf Supplementary Figure 5 SupplementaryFigure2.pdf Supplementary Figure 2 SUPPLEMENTARYLEGENDS.docx Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9060414","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":605109814,"identity":"2b8e8af2-0374-4a9e-b4a6-e1c5940b3f40","order_by":0,"name":"Gaurav Ahuja","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA/klEQVRIiWNgGAWjYDADAwaGBIYPBhI8IBoIJIjTwjgDqIWHFC0MzEArGKBacAN59zNmH362McibSyQ8/GxTYCFjz57A+OEHg0UeLi2GZ3KMZ/a2MRjunJGQLJ0DchjPA2bJHgaJYpxaGnKMGXjOMDBuuJGQANEikcAgDfRLYgMuLf1vjBn/nGGwB2pJ/m0B0cL8G58WeYkcY2aeCoZEoJY0aQaIFja8thhIPCtmlqmQSN5w5kGaZQ9Iy5mHbSAGblv6kzczvjGwsd1wPCf5xo8/dfbs7cmHb/yoqMNtywEwBYo4ngSoGGMDJJpw2YIwi/0AbmWjYBSMglEwogEAlK5MNSS+ivEAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-2837-9361","institution":"Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi)","correspondingAuthor":true,"prefix":"","firstName":"Gaurav","middleName":"","lastName":"Ahuja","suffix":""},{"id":605109815,"identity":"f1d40968-4394-4a5f-aa85-a49bc0d12002","order_by":1,"name":"Arushi Sharma","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology","correspondingAuthor":false,"prefix":"","firstName":"Arushi","middleName":"","lastName":"Sharma","suffix":""},{"id":605109816,"identity":"4e4a4dc9-f7ca-48a5-a814-f8cc8fe669a7","order_by":2,"name":"Ankit Singh","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi)","correspondingAuthor":false,"prefix":"","firstName":"Ankit","middleName":"","lastName":"Singh","suffix":""},{"id":605109817,"identity":"48d6c1f4-4112-458a-87d9-62fbfa624eb1","order_by":3,"name":"Shekhar Kedia","email":"","orcid":"https://orcid.org/0000-0002-9322-979X","institution":"University of Cambridge","correspondingAuthor":false,"prefix":"","firstName":"Shekhar","middleName":"","lastName":"Kedia","suffix":""},{"id":605109818,"identity":"bd75f03b-1619-4ccc-af00-69e4bb621bb0","order_by":4,"name":"Abhinav Sharma","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi)","correspondingAuthor":false,"prefix":"","firstName":"Abhinav","middleName":"","lastName":"Sharma","suffix":""},{"id":605109819,"identity":"0bef5879-5118-4608-a767-13651a7ed51c","order_by":5,"name":"Vishakha Gautam","email":"","orcid":"","institution":"Department of Computational Biology, Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi), Okhla, Phase III, New Delhi, 110020, India.","correspondingAuthor":false,"prefix":"","firstName":"Vishakha","middleName":"","lastName":"Gautam","suffix":""},{"id":605109820,"identity":"32a310ea-ff44-43ee-9ba8-707435b17a15","order_by":6,"name":"Pranjal Sharma","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi)","correspondingAuthor":false,"prefix":"","firstName":"Pranjal","middleName":"","lastName":"Sharma","suffix":""},{"id":605109821,"identity":"60e462dd-178a-4e68-96e9-817f7bb6401c","order_by":7,"name":"Aniket Khandelwal","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi)","correspondingAuthor":false,"prefix":"","firstName":"Aniket","middleName":"","lastName":"Khandelwal","suffix":""},{"id":605109822,"identity":"e5df9537-c11e-430c-9455-9fe4934728e2","order_by":8,"name":"Sarayu Ramakrishna","email":"","orcid":"","institution":"Indian Institute of Science","correspondingAuthor":false,"prefix":"","firstName":"Sarayu","middleName":"","lastName":"Ramakrishna","suffix":""},{"id":605109823,"identity":"e85053ff-d103-41b9-a151-5c26dbd2d928","order_by":9,"name":"Ravi Muddashetty","email":"","orcid":"","institution":"Indian Institute of Science","correspondingAuthor":false,"prefix":"","firstName":"Ravi","middleName":"","lastName":"Muddashetty","suffix":""},{"id":605109824,"identity":"6bd5fa89-3ca7-4acc-b858-46cf794707ae","order_by":10,"name":"Kristine Freude","email":"","orcid":"https://orcid.org/0000-0001-9480-2386","institution":"University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Kristine","middleName":"","lastName":"Freude","suffix":""},{"id":605109825,"identity":"f3e8f390-318f-4c7b-af9e-03a77b3f1839","order_by":11,"name":"Saveena Solanki","email":"","orcid":"","institution":"Department of Computational Biology, Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi), Okhla, Phase III, New Delhi, 110020, India.","correspondingAuthor":false,"prefix":"","firstName":"Saveena","middleName":"","lastName":"Solanki","suffix":""},{"id":605109826,"identity":"26033d7e-15b7-4ed8-a735-127b80165879","order_by":12,"name":"Sonam Chauhan","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology","correspondingAuthor":false,"prefix":"","firstName":"Sonam","middleName":"","lastName":"Chauhan","suffix":""},{"id":605109827,"identity":"7fa98b4c-23e7-4ce1-a376-f63bfe881043","order_by":13,"name":"Suvendu Kumar","email":"","orcid":"","institution":"Department of Computational Biology, Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi), Okhla, Phase III, New Delhi, 110020, India.","correspondingAuthor":false,"prefix":"","firstName":"Suvendu","middleName":"","lastName":"Kumar","suffix":""},{"id":605109828,"identity":"4d87ee0c-0738-4459-98c8-a448a1cd4ef2","order_by":14,"name":"Shiva Satija","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology","correspondingAuthor":false,"prefix":"","firstName":"Shiva","middleName":"","lastName":"Satija","suffix":""},{"id":605109829,"identity":"be2f0bd5-abb0-47be-83e0-0613f691c99a","order_by":15,"name":"Subhadeep Duari","email":"","orcid":"","institution":"Department of Computational Biology, Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi), Okhla, Phase III, New Delhi, 110020, India.","correspondingAuthor":false,"prefix":"","firstName":"Subhadeep","middleName":"","lastName":"Duari","suffix":""},{"id":605109830,"identity":"8978f398-9a7c-45c5-8eb8-d3a2c4fa89b5","order_by":16,"name":"Sakshi Arora","email":"","orcid":"","institution":"Department of Computational Biology, Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi), Okhla, Phase III, New Delhi, 110020, India.","correspondingAuthor":false,"prefix":"","firstName":"Sakshi","middleName":"","lastName":"Arora","suffix":""},{"id":605109831,"identity":"6fed09bc-ed53-40b7-8f31-51d867ea9ccd","order_by":17,"name":"Advik Gupta","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology-Delhi (IIIT-Delhi)","correspondingAuthor":false,"prefix":"","firstName":"Advik","middleName":"","lastName":"Gupta","suffix":""},{"id":605109832,"identity":"1fde3e1e-bd06-4701-9fbe-89676c0f5806","order_by":18,"name":"Raidhani Shome","email":"","orcid":"","institution":"Indraprastha Institute of Information Technology","correspondingAuthor":false,"prefix":"","firstName":"Raidhani","middleName":"","lastName":"Shome","suffix":""},{"id":605109833,"identity":"ba09555e-697e-4686-ab89-85d9ba009c4d","order_by":19,"name":"Debarka Sengupta","email":"","orcid":"https://orcid.org/0000-0002-6353-5411","institution":"Indraprastha Institute of Information Technology Delhi","correspondingAuthor":false,"prefix":"","firstName":"Debarka","middleName":"","lastName":"Sengupta","suffix":""},{"id":605109834,"identity":"ff535301-fec1-4b1a-ae33-9c804ce5ca47","order_by":20,"name":"Deepak nair","email":"","orcid":"https://orcid.org/0000-0002-2103-1653","institution":"Indian Institute of Science","correspondingAuthor":false,"prefix":"","firstName":"Deepak","middleName":"","lastName":"nair","suffix":""}],"badges":[],"createdAt":"2026-03-07 19:05:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9060414/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9060414/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":104781931,"identity":"dc1981c1-4a8e-4b58-85bb-be9a147e3e8c","added_by":"auto","created_at":"2026-03-17 07:56:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":152467,"visible":true,"origin":"","legend":"\u003cp\u003eEvoAge-KG Construction Reveals a Billion-Triple Multi-Species Network with Complex Topology\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a) \u003c/strong\u003eConceptual framework of the EvoAge-KG, showing the integration of data from six species and the 14 core biological entity types. \u003cstrong\u003e(b) \u003c/strong\u003eSunburst chart of the 48 data sources integrated into the EvoAge Knowledge Graph (KG). Data sources are categorized by 'Aging' (12 sources) and 'Other' (36 sources). The outer ring details the contribution in millions of triples ('M'), resulting in a total of 1.04 billion triples. The literature datasource includes structured literature data linking hallmarks of aging with specific biological processes. The Species Connection data source refers to triples where allentity nodes belonging to a species are directly connected to that species node, ensuring species-level integrity within the KG. \u003cstrong\u003e(c)\u003c/strong\u003e Alluvial plot of the one-to-one human ortholog mapping, showing gene flow from source ('Aging' vs. 'EvoAge'), through species, to final human ortholog status ('ortho' or 'non-ortho'). \u003cstrong\u003e(d) \u003c/strong\u003eSpecies-specific node counts by entity type (log10 scale). Stacks compare nodes from 'Aging' sources with additional nodes from 'Other' sources. \u003cstrong\u003e(e)\u003c/strong\u003e Total node count per species (log10 scale). Stacks compare 'Aging' and 'EvoAge' contributions. \u003cstrong\u003e(f)\u003c/strong\u003eSunburst plot comparing edge (triple) counts by type and species for the 'Aging' KG and the full 'EvoAge' KG where 'M' = millions. \u003cstrong\u003e(g)\u003c/strong\u003e Total triple count per species (log10 scale). Stacks compare 'Aging' and 'EvoAge' contributions. \u003cstrong\u003e(h-k)\u003c/strong\u003e Topological analysis comparing the 'Aging' and 'EvoAge' KGs, \u003cstrong\u003e(h)\u003c/strong\u003e Scale free node degree distribution (log-log plot). \u003cstrong\u003e(i) \u003c/strong\u003eNode distribution across the top 15 communities. \u003cstrong\u003e(j)\u003c/strong\u003e Betweenness centrality density (log-log plot), measures the influence of a node by quantifying how often it lies on the shortest path between any other two nodes. \u003cstrong\u003e(k)\u003c/strong\u003e Triangles per node density (log-log plot, shows), shows local clustering, counting the number of complete triads (three connected nodes) each node participates in.\u003c/p\u003e","description":"","filename":"Binder11.png","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/4ef690af1ab2b19fe3026526.png"},{"id":104781837,"identity":"1c736a9a-9622-480c-a1b4-13285ceb94d5","added_by":"auto","created_at":"2026-03-17 07:56:25","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":147645,"visible":true,"origin":"","legend":"\u003cp\u003eOptimized Knowledge Graph Embeddings Enable Accurate Cross-Species Prediction\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e Workflow for KGE model selection. \u003cstrong\u003e(b)\u003c/strong\u003e Heatmap of link prediction performance (hit@1, @3, @10) for six KGE models on the 'Aging' and 'EvoAge' KGs (validation and test sets). Starred models (RESCAL, RotatE) were selected. \u003cstrong\u003e(c) \u003c/strong\u003eBar chart comparing Mean Reciprocal Rank (MRR) for the six models on the 'Aging' (pink) and 'EvoAge' (blue) KGs. \u003cstrong\u003e(d)\u003c/strong\u003e Training convergence (average loss) for all models on the 'Aging' KG (left) and 'EvoAge' KG (right). \u003cstrong\u003e(e)\u003c/strong\u003eSchematic of the embedding size (64-512) optimization workflow, benchmarked on link and edge type prediction. \u003cstrong\u003e(f)\u003c/strong\u003e Heatmap of link prediction performance (hit@K) vs. embedding size for the selected models: 'Aging'-RESCAL (left) and 'EvoAge'-RotatE (right). \u003cstrong\u003e(g)\u003c/strong\u003e Heatmap of edge type prediction performance (hit@K) vs. embedding size for 'Aging'-RESCAL (top) and 'EvoAge'-RotatE (bottom). \u003cstrong\u003e(h)\u003c/strong\u003e Mean Rank (MR) and Mean Reciprocal Rank (MRR) vs. embedding size for 'Aging'-RESCAL. \u003cstrong\u003e(i)\u003c/strong\u003e Mean Rank (MR) and MRR vs. embedding size for 'EvoAge'-RotatE. \u003cstrong\u003e(j) \u003c/strong\u003eSchematic of the Knowledge Graph (KG) data splitting strategy for model training and evaluation. The 1.04 billion triples were partitioned into Training and Validation sets. The reserved Heldout Testing set (3.14M triples) was explicitly designed for rigorous benchmarking and is composed of two types of triples: a species-specific test set (1% of triples sampled within each species' data) and a cross-species test set dedicated to evolutionary inference (triples connecting a non human entity to a human entity), sized equal to the human data subset. \u003cstrong\u003e(k)\u003c/strong\u003e Donut chart showing the composition of the 3.14M triple 'heldout' test set, broken down by 'Within-Species' and 'cross-species' triples. \u003cstrong\u003e(l)\u003c/strong\u003e Bar chart of link prediction performance (hit@1, @3, @10) on the 'heldout' test set, shown for each species and the 'cross-species' subset. \u003cstrong\u003e(m)\u003c/strong\u003e Line graph of edge type prediction performance (hit@1, @3, @10) on the 'heldout' test set, shown for each species and the 'cross-species' subset.\u003c/p\u003e","description":"","filename":"Binder13.png","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/4aeb21a48d546b3d07a04937.png"},{"id":104782128,"identity":"73808a87-47a7-489c-847d-dcea2d8f0686","added_by":"auto","created_at":"2026-03-17 07:56:52","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":218485,"visible":true,"origin":"","legend":"\u003cp\u003eIntegrated AI Platform Enables Accessible Biological Discovery\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e Schematic of the platform's three core functionalities: (1) 'Search database' for fact retrieval, (2) 'Link prediction' for inferring missing entities, and (3) 'Test hypothesis' for scoring user-proposed triples. \u003cstrong\u003e(b)\u003c/strong\u003e High-level system architecture, showing the Streamlit GUI, FastAPI server, and backend modules, including Data Stores (Neo4j, Redis), External APIs (Gemini), and the DGL-KE inference engine. \u003cstrong\u003e(c)\u003c/strong\u003e Query processing workflow. The EvoAge-agent (Kani + GPT-4o-mini) performs entity extraction and intent detection, then routes the query to the appropriate tool (KG Query, Link Prediction, or Hypothesis Testing) to generate a final response. \u003cstrong\u003e(d)\u003c/strong\u003e Sequence diagram of the 'Link Prediction' workflow, detailing the process from API request to DGL-KE scoring. \u003cstrong\u003e(e)\u003c/strong\u003e Sequence diagram of the 'Hypothesis Testing' workflow, showing the orchestration of DGL-KE scoring, Neo4j validation, and Gemini-based interpretation. \u003cstrong\u003e(f) \u003c/strong\u003e\u0026nbsp;Sample of the EvoAge chat interface, showing a user's 'Hypothesis Testing' query and the platform's comprehensive response. \u003cstrong\u003e(g)\u003c/strong\u003e Screenshot of the EvoAge platform's graphical user interface (GUI), showing the user authentication and welcome screen.\u003c/p\u003e","description":"","filename":"Binder15.png","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/633a576d8c68507f1288ecfe.png"},{"id":104782482,"identity":"84632ca9-c777-435a-a449-9b5248bab26e","added_by":"auto","created_at":"2026-03-17 07:57:23","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":465071,"visible":true,"origin":"","legend":"\u003cp\u003eEvoAge Platform Benchmarking Demonstrates Superior Performance and Experimental Validation\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e Schematic of the computational benchmarking pipeline. 5,700 novel hypothesis questions (2,850 positive, 2,850 negative) were generated from 980 recent BioRxiv preprints. EvoAge and three general-purpose LLMs (Gemma, Llama, Qwen) were challenged to answer, with responses graded for biological feasibility by a BioMistral-7B \"teacher\" model. \u003cstrong\u003e(b)\u003c/strong\u003e Violin plot comparing the \"Clarity Score\" (a composite metric for discriminative power) of EvoAge (red) against general-purpose LLMs. ****p \u0026lt; 0.0001 (pairwise Wilcoxon test with BH adjustment). \u003cstrong\u003e(c)\u003c/strong\u003e Violin plot of raw scores (1-10) for the 2,850 positive hypothesis questions. \u003cstrong\u003e(d)\u003c/strong\u003e Violin plot of raw scores (1-10) for the 2,850 negative hypothesis questions. A lower score indicates correct rejection of a false hypothesis \u003cstrong\u003e(e) \u003c/strong\u003eresults of statistics of clarity score and ANOVA test\u003cstrong\u003e. (f) \u003c/strong\u003eSchematic illustration of the multi-model experimental pipeline used for validating EvoAge-generated hypotheses. \u003cstrong\u003e(g)\u003c/strong\u003e Representative STED nanoscopy image showing BACE1 (green) relative to the Postsynaptic marker (magenta) in GC (control) iPSC neurons.\u003cstrong\u003e (h)\u003c/strong\u003e Representative STED nanoscopy image showing BACE1 (green) relative to the Postsynaptic marker (magenta) in Mutant iPSC neurons. \u003cstrong\u003e(i)\u003c/strong\u003eRepresentative STED nanoscopy image showing BACE1 (green) relative to the Perisynaptic/Endocytic marker (magenta) in GC (control) iPSC neurons. \u003cstrong\u003e(j)\u003c/strong\u003eRepresentative STED nanoscopy image showing BACE1 (green) relative to the Perisynaptic/Endocytic marker (magenta) in Mutant iPSC neurons; \u003cstrong\u003e(k)\u003c/strong\u003e The ridge plots visually demonstrated the statistical distributions of BACE1 nanodomain (Intensity (a.u)) across all three species models. \u003cstrong\u003e(l)\u003c/strong\u003e The ridge plots visually demonstrated the statistical distributions of BACE1 nanodomain (area (sq. μm)). \u003cstrong\u003e(m) \u003c/strong\u003eThe ridge plots show the statistical distributions of BACE1 nanodomain (length (μm)).\u003c/p\u003e","description":"","filename":"Binder17.png","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/1af77fb2d3a9a1884074fcbe.png"},{"id":104835970,"identity":"f80f8ff8-4fe7-48b6-b137-7839b897cd84","added_by":"auto","created_at":"2026-03-17 17:50:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3490880,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/3d66d5be-a769-4fa9-a6f3-4c8b83d2e3f7.pdf"},{"id":104583218,"identity":"0a299b29-27a6-4be4-b2c4-376654b2b424","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":11975,"visible":true,"origin":"","legend":"Supplementary Table 1","description":"","filename":"SuppTable1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/10de33a1ccbb6056b8cc0204.xlsx"},{"id":104782020,"identity":"846a9a53-a5b5-461f-b7f5-9ff174e7800d","added_by":"auto","created_at":"2026-03-17 07:56:42","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":9706,"visible":true,"origin":"","legend":"Supplementary Table 3","description":"","filename":"SuppTable3.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/896e36a910c24291c0e4666e.xlsx"},{"id":104583216,"identity":"c2d39f98-cd05-4d27-b753-c4546509fcd9","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":10903,"visible":true,"origin":"","legend":"Supplementary Table 2","description":"","filename":"SuppTable2.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/8a12c7c33f01c783569ea584.xlsx"},{"id":104835245,"identity":"741c9193-9912-44e5-a7b8-a5517a300f88","added_by":"auto","created_at":"2026-03-17 17:42:32","extension":"xlsx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":10853,"visible":true,"origin":"","legend":"Supplementary Table 15","description":"","filename":"SuppTable15.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/a1cdd1b42aafcf6362abbeb9.xlsx"},{"id":104583229,"identity":"16f63108-fea3-42a5-955a-a8149f415c4e","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":340641,"visible":true,"origin":"","legend":"Supplementary Table 11","description":"","filename":"SuppTable11.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/cb88e6d7d3d70efc0f764ab9.xlsx"},{"id":104781368,"identity":"82700e13-62d1-4fcd-8890-c339e134f941","added_by":"auto","created_at":"2026-03-17 07:55:33","extension":"xlsx","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":14260,"visible":true,"origin":"","legend":"Supplementary Table 9","description":"","filename":"SuppTable9.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/c7fe28bb122badff8bb18b85.xlsx"},{"id":104583223,"identity":"4f73c819-e59a-40b2-a0a5-e418658b71aa","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"pdf","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":541709,"visible":true,"origin":"","legend":"Supplementary Figure 7","description":"","filename":"SupplementaryFigure7.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/bf3d586bd123d2ff090b0e13.pdf"},{"id":104781612,"identity":"9a28d72d-4c08-4d42-b94e-fa6f28a15898","added_by":"auto","created_at":"2026-03-17 07:56:00","extension":"xlsx","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":11126,"visible":true,"origin":"","legend":"Supplementary Table 12","description":"","filename":"SuppTable12.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/a1fccb1f2020aa764f1b6d48.xlsx"},{"id":104583220,"identity":"36330063-ccbd-4fa0-be59-5b07c64e37d1","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":14801,"visible":true,"origin":"","legend":"Supplementary Table 5","description":"","filename":"SuppTable5.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/8cb61fa2766832fe0e52c2b3.xlsx"},{"id":104781900,"identity":"8b5e6037-329c-4e7a-a84d-e568d01521c5","added_by":"auto","created_at":"2026-03-17 07:56:32","extension":"pdf","order_by":10,"title":"","display":"","copyAsset":false,"role":"supplement","size":425066,"visible":true,"origin":"","legend":"Supplementary Figure 3","description":"","filename":"SupplementaryFigure3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/9ec0bf26284be8a71efb5b2d.pdf"},{"id":104583221,"identity":"224c971d-ea78-45e3-99ad-07053598320a","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":11,"title":"","display":"","copyAsset":false,"role":"supplement","size":13680,"visible":true,"origin":"","legend":"Supplementary Table 8","description":"","filename":"SuppTable8.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/df648fbf7175efc72c6dc7b0.xlsx"},{"id":104781679,"identity":"9dac856d-1b9e-4a04-9715-2e6af1670a59","added_by":"auto","created_at":"2026-03-17 07:56:09","extension":"pdf","order_by":12,"title":"","display":"","copyAsset":false,"role":"supplement","size":217925,"visible":true,"origin":"","legend":"Supplementary Figure 8","description":"","filename":"SupplementaryFigure8.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/796ddfeb27560f9bbc54c091.pdf"},{"id":104583237,"identity":"fb4d2da9-bd65-4ac4-ac90-8901b70ad633","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"pdf","order_by":13,"title":"","display":"","copyAsset":false,"role":"supplement","size":669245,"visible":true,"origin":"","legend":"Supplementary Figure 1","description":"","filename":"SupplementaryFigure1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/713062567234d4713932b3df.pdf"},{"id":104781609,"identity":"06b0b0b7-c1e1-43f5-8530-4d2ac45e23f0","added_by":"auto","created_at":"2026-03-17 07:56:00","extension":"xlsx","order_by":14,"title":"","display":"","copyAsset":false,"role":"supplement","size":11201,"visible":true,"origin":"","legend":"Supplementary Table 13","description":"","filename":"SuppTable13.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/6216ab36dee91a05721f5d85.xlsx"},{"id":104781118,"identity":"327e2dc3-450a-4a19-b50e-a10e907155f7","added_by":"auto","created_at":"2026-03-17 07:54:50","extension":"xlsx","order_by":15,"title":"","display":"","copyAsset":false,"role":"supplement","size":12944,"visible":true,"origin":"","legend":"Supplementary Table 7","description":"","filename":"SuppTable7.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/5bf0884635c1a8b9bc8c04b9.xlsx"},{"id":104583235,"identity":"1ef09a73-aa97-4084-b971-221266079fa9","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":16,"title":"","display":"","copyAsset":false,"role":"supplement","size":885065,"visible":true,"origin":"","legend":"Supplementary Table 10","description":"","filename":"SuppTable10.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/9a2fee73ac35eaa78770d217.xlsx"},{"id":104583234,"identity":"709d997a-7057-462a-8c01-7e8e9d16fd13","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":17,"title":"","display":"","copyAsset":false,"role":"supplement","size":14854,"visible":true,"origin":"","legend":"Supplementary Table 14","description":"","filename":"SuppTable14.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/f805bcd79cf741a2f57d3242.xlsx"},{"id":104583238,"identity":"831b1684-b83e-4fe6-b9be-2f77c19b21ad","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"pdf","order_by":18,"title":"","display":"","copyAsset":false,"role":"supplement","size":292468,"visible":true,"origin":"","legend":"Supplementary Figure 6","description":"","filename":"SupplementaryFigure6.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/42365f62369b59d874a02bea.pdf"},{"id":104782022,"identity":"75865c4b-d3c9-4337-8cd9-9ed80c559959","added_by":"auto","created_at":"2026-03-17 07:56:42","extension":"xlsx","order_by":19,"title":"","display":"","copyAsset":false,"role":"supplement","size":13159,"visible":true,"origin":"","legend":"Supplementary Table 4","description":"","filename":"SuppTable4.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/742c8d3d18ba3005db9f1e99.xlsx"},{"id":104583240,"identity":"c416104e-a630-4f7e-a0f9-933e17b79886","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"xlsx","order_by":20,"title":"","display":"","copyAsset":false,"role":"supplement","size":14662,"visible":true,"origin":"","legend":"Supplementary Table 6","description":"","filename":"SuppTable6.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/830695a89ea2d0126d3b4b62.xlsx"},{"id":104583242,"identity":"8b3c5f2e-3f21-4589-9274-4137ad1b4648","added_by":"auto","created_at":"2026-03-13 15:18:15","extension":"pdf","order_by":21,"title":"","display":"","copyAsset":false,"role":"supplement","size":169756,"visible":true,"origin":"","legend":"Supplementary Figure 4","description":"","filename":"SupplementaryFigure4.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/655bf26b5fe7b6a656d36242.pdf"},{"id":104583243,"identity":"71231890-0636-49da-8b61-c32da06dc840","added_by":"auto","created_at":"2026-03-13 15:18:15","extension":"pdf","order_by":22,"title":"","display":"","copyAsset":false,"role":"supplement","size":422120,"visible":true,"origin":"","legend":"Supplementary Figure 5","description":"","filename":"SupplementaryFigure5.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/c86c2c4b747355cca082c6de.pdf"},{"id":104781673,"identity":"450aacf3-bc70-469a-bdc5-296cd4311be2","added_by":"auto","created_at":"2026-03-17 07:56:08","extension":"pdf","order_by":23,"title":"","display":"","copyAsset":false,"role":"supplement","size":1691385,"visible":true,"origin":"","legend":"Supplementary Figure 2","description":"","filename":"SupplementaryFigure2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/493e0f557a2a3edc423bc4a5.pdf"},{"id":104583239,"identity":"66f1a67f-83ab-421e-a902-cfb067bd6ae3","added_by":"auto","created_at":"2026-03-13 15:18:14","extension":"docx","order_by":24,"title":"","display":"","copyAsset":false,"role":"supplement","size":18824,"visible":true,"origin":"","legend":"","description":"","filename":"SUPPLEMENTARYLEGENDS.docx","url":"https://assets-eu.researchsquare.com/files/rs-9060414/v1/e036d3128f8cd2da6f8e69e4.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Cross-Species Aging Knowledge Integration into Agentic AI Platform Uncovers Conserved Mechanisms","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eThe quest to understand the biological mechanisms of aging represents one of the most complex challenges in biomedical science \u003csup\u003e1\u003c/sup\u003e. Its complexity arises from the interplay of countless molecular and cellular processes that unfold over decades, varying across tissues, and are influenced by both genetics and environment \u003csup\u003e2\u003c/sup\u003e. Given the ethical and practical constraints of longitudinal human aging research, the field has been profoundly shaped by studies in model organisms, from the short-lived nematode \u003cem\u003eCaenorhabditis elegans\u003c/em\u003e to the laboratory mouse, where genetic tractability and shorter lifespans enable high-throughput interrogation of longevity and age-related decline \u003csup\u003e3\u003c/sup\u003e. A cornerstone insight from these studies is that core metabolic and protective pathways governing aging, such as insulin signaling, mTOR, and mitochondrial function, are evolutionarily conserved \u003csup\u003e1,4\u0026ndash;6\u003c/sup\u003e. However, this foundational insight also presents a major bottleneck: the resulting deluge of data is profoundly fragmented. Critical insights are often confined to species-specific silos and specialized databases, thereby demarcating them from the broader context of human systems biology \u003csup\u003e5,7\u0026ndash;12\u003c/sup\u003e. This fragmentation obstructs a unified, systems-level understanding of how mechanisms identified in models translate to human aging and pathology.\u003c/p\u003e\n\n\u003cp\u003eKnowledge graphs (KGs) have become pivotal in systems biology for integrating disparate data into structured networks, enabling the discovery of non-obvious biological relationships. Special-purpose KGs (e.g., Hetionet, PrimeKG, CKG) \u003csup\u003e13\u0026ndash;15\u003c/sup\u003e have proven effective for generating mechanistically grounded hypotheses in drug repurposing, while aging-specific resources (e.g., Aging Atlas \u003csup\u003e10\u003c/sup\u003e, DrugAge \u003csup\u003e16\u003c/sup\u003e, GenDR \u003csup\u003e17\u003c/sup\u003e, AgeXtend \u003csup\u003e9\u003c/sup\u003e) provide critical expert-curated data on longevity. However, these specialized KGs are often constrained by a narrow scope, excelling in specific data types but lacking the multi-scale, cross-species context necessary for holistic discovery. This limitation prevents the seamless translation of findings, such as linking a longevity gene in yeast to its conserved pathway in mice and its phenotypic consequence in the human system. Bridging this gap requires a multi-species KG that deeply integrates specialized aging knowledge with general biology. A further challenge is utility; the scale of billion-triple KGs makes them inaccessible to many biologists. The synergy between KGs and large language models (LLMs) offers a transformative solution. While LLMs offer intuitive natural language interaction, they risk hallucination \u003csup\u003e18,19\u003c/sup\u003e. In contrast, KGs offer a verifiable, structured knowledge base. Their integration creates an agentic framework where the LLM serves as a natural language interface, querying the KG and generating evidence-based explanations, thereby making deep knowledge accessible for hypothesis generation and testing \u003csup\u003e20,21\u003c/sup\u003e. \u003c/p\u003e\n\n\u003cp\u003eHere, we introduce EvoAge, an agentic AI platform built on a unified, multi-species knowledge graph for translational aging research. EvoAge integrates 48 aging-focused and general biomedical resources into a harmonized network of 1.04 billion triples across six species, centered on an orthology-driven framework that enables biologically meaningful, cross-species reasoning. An LLM-assisted interface further transforms this large-scale graph into an accessible system supporting natural-language exploration, link prediction, and formal hypothesis testing. To evaluate its reasoning capability, we benchmarked EvoAge against recent aging-related preprints and observed markedly superior discrimination between plausible and implausible hypotheses compared with leading general-purpose LLMs. We then applied EvoAge to uncover a previously unrecognized mechanism in Alzheimer\u0026rsquo;s disease (AD) involving nanoscale redistribution of Beta-secretase (BACE1) at synapses. This prediction was experimentally validated using familial AD patient-derived iPSC neurons and the isogenic controls, and further confirmed in transgenic mouse models and postmortem human brain tissue. Collectively, EvoAge provides a scalable framework integrating evolutionary biology, structured biomedical knowledge, and AI reasoning to accelerate mechanistic discovery in aging and age-related disease. In the case of BACE1, the aging context was essentia, EvoAge revealed that its redistribution consistently aligned with conserved aging-linked pathways such as synaptic decline and vesicle trafficking across species, distinguishing a true disease-associated shift from normal age-related variation and enabling its nomination as a testable mechanism.\u003c/p\u003e\n\n"},{"header":"RESULTS","content":"\u003ch2\u003eA Systems-Scale Knowledge Graph Integrating Aging Biology Across Model Species\u003c/h2\u003e\n\u003cp\u003eTo construct a foundational resource for cross-species aging research, we first established the rationale for a multi-organism framework. Historical publication trends revealed exponential growth in aging studies centered on six core species, including humans \u003cstrong\u003e(Supplementary Figure 1a, Supplementary Table 1)\u003c/strong\u003e. Comparative survival analysis further underscored the analytical advantage of lifespan diversity for identifying conserved aging mechanisms \u003cstrong\u003e(Supplementary Figure 1b)\u003c/strong\u003e. Guided by these observations, we developed a computational pipeline to assemble the EvoAge Knowledge Graph (KG) \u003cstrong\u003e(Figure 1a)\u003c/strong\u003e, integrating data from 48 public sources, including 12 aging-specific and 36 general biomedical databases \u003cstrong\u003e(Figure 1b, Supplementary Table 2)\u003c/strong\u003e. To assess the structure of the integrated graph, we analyzed source-level contributions across biological relationship types. A comparative heatmap revealed a dense, heterogeneous relationship matrix in EvoAge, where individual sources specialize in distinct interaction categories, contrasting the sparse architecture of the baseline Aging KG \u003cstrong\u003e(Supplementary Figure 2a, b)\u003c/strong\u003e. Species-level provenance analysis confirmed that each organism is anchored by its dedicated model organism databases. This examination demonstrated that knowledge for each organism is dominated by its dedicated Model Organism Database (MOD), with \u003cem\u003eSaccharomyces\u003c/em\u003e Genome Database (SGD) \u003csup\u003e22\u003c/sup\u003e and YeastNet (v3) \u003csup\u003e23\u003c/sup\u003e serving as the primary source for \u003cem\u003eSaccharomyces cerevisiae\u003c/em\u003e \u003cstrong\u003e(Supplementary Figure 3c), \u003c/strong\u003eFlyBase \u003csup\u003e24\u003c/sup\u003e for \u003cem\u003eDrosophila melanogaster\u003c/em\u003e \u003cstrong\u003e(Supplementary Figure 3b)\u003c/strong\u003e, WormBase \u003csup\u003e25\u003c/sup\u003e for \u003cem\u003eCaenorhabditis elegans\u003c/em\u003e \u003cstrong\u003e(Supplementary Figure 3d)\u003c/strong\u003e, ZFIN \u003csup\u003e26\u003c/sup\u003e for \u003cem\u003eDanio rerio \u003c/em\u003e\u003cstrong\u003e(Supplementary Figure 3e)\u003c/strong\u003e\u003cem\u003e, \u003c/em\u003eand MGI (Mouse Genome Informatics) for \u003cem\u003eMus musculus\u003c/em\u003e \u003csup\u003e27\u003c/sup\u003e \u003cstrong\u003e(Supplementary Figure 3a)\u003c/strong\u003e. Collectively, these patterns demonstrate that EvoAge functions as a \u0026ldquo;mosaic of experts,\u0026rdquo; combining broad general-purpose knowledge with deep organism-specific annotations.\u003c/p\u003e\n\n\u003cp\u003eData harmonization was achieved through a structured schema encompassing 14 biological entity types and a unified identifier-mapping workflow \u003cstrong\u003e(Supplementary Figure 1c)\u003c/strong\u003e. Duplicate records were merged while maintaining full provenance, resulting in a final graph containing \u0026gt;1.04 billion triples \u003cstrong\u003e(Supplementary Table 3)\u003c/strong\u003e. To enable true cross-species reasoning, we implemented an orthology-based unification layer that mapped over 80,000 gene entries to human identifiers \u003cstrong\u003e(Figure 1c)\u003c/strong\u003e, with mapping success reflecting phylogenetic distance. For instance, we successfully mapped 18,999 genes from the closely related mouse to human counterparts, but only 6,707 genes from the more distant worm \u003cstrong\u003e(Supplementary Figure 1d)\u003c/strong\u003e. We next sought to quantify the transformative impact of data integration. The initial Aging KG contained only 616,928 triples, was overwhelmingly human-centric (601,099 triples, 97.5%), and chemically skewed with minimal systems-level context \u003cstrong\u003e(Supplementary Figure 1e, f)\u003c/strong\u003e. Through systematic integration, we observed dramatic, species-specific enrichment in the expanded EvoAge KG \u003cstrong\u003e(Figures 1d and e)\u003c/strong\u003e,with particularly notable expansions for model organisms, where\u003cem\u003e Drosophila melanogaster\u003c/em\u003e gene nodes increased from 170 to 13,747, and \u003cem\u003eDanio rerio\u003c/em\u003e grew from zero to 15,182 gene nodes. To characterize the final composition, we analyzed node type distributions across the integrated resource \u003cstrong\u003e(Figure 1e).\u003c/strong\u003e This revealed EvoAge as a literature-first resource dominated by 26.7 million publication nodes, with a strong genomic foundation including 1.5 million mutation nodes and ~1.01 million chemical nodes \u003cstrong\u003e(Supplementary Figure 1g)\u003c/strong\u003e, while maintaining distinct species-specific node type distributions.\u003c/p\u003e\n\u003cp\u003eFinally, we performed a comprehensive connectivity analysis to understand how integration affected biological relationships. Connectivity analysis revealed a non-linear, species-specific enrichment, with counts increasing by 22,033-fold for \u003cem\u003eS. cerevisiae\u003c/em\u003e (420 to 9.25M), 19,332-fold for \u003cem\u003eD. melanogaster\u003c/em\u003e (528 to 10.2M), 5,847-fold for \u003cem\u003eC. elegans\u003c/em\u003e (2.4K to 14.2M), and 2,747-fold for \u003cem\u003eM. musculus\u003c/em\u003e (12.5K to 34.2M), contributing to 1700 fold expansion of knowledge. Even the human-centric data expanded by 207-fold (601K to 124.3M), while \u003cem\u003eD. rerio\u003c/em\u003e grew from zero to 28.7 million triples \u003cstrong\u003e(Figure 1f-g)\u003c/strong\u003e. Importantly, EvoAge exhibits a scale-free topology, high modularity, low betweenness centrality, and markedly increased clustering properties, all of which are characteristic of real biological networks and are favorable for computational inference \u003cstrong\u003e(Figure 1h-k)\u003c/strong\u003e. Node-level hub analysis further confirmed a paradigm shift: from conceptual hallmarks dominating the baseline KG to structural and functional anchors, including tissue architecture, translational machinery, and GPCR signaling, in the final EvoAge KG\u003cstrong\u003e (Supplementary Figures 4a and 4b; Supplementary Tables 5 and 6)\u003c/strong\u003e. Together, these analyses establish EvoAge as a deeply layered, evolutionarily informed, and computationally optimized resource positioned to support mechanistic discovery at scale.\u003c/p\u003e\n\u003ch2\u003eSystematic Optimization of Knowledge Graph Embeddings Reveals Architectural Dependencies\u003c/h2\u003e\n\u003cp\u003eWith the EvoAge KG established, we next sought to convert the resource into a predictive inference engine through knowledge graph embedding (KGE) modeling. We first characterized the relational space of both graphs. The EvoAge KG contained 66 unique relation types spanning 14 biological entity classes, with PMID, chemical-gene, and gene-gene associations representing the most abundant edges \u003cstrong\u003e(Supplementary Figure 5a; Supplementary Table 4)\u003c/strong\u003e. In contrast, the baseline Aging KG contained only 27 relation types, predominantly limited to gene-disease relationships \u003cstrong\u003e(Supplementary Figure 5b)\u003c/strong\u003e. A co-occurrence heatmap confirmed that nearly all permissible node-type pairs in EvoAge participated in at least one relationship, establishing the necessity of relation-aware learning \u003cstrong\u003e(Supplementary Figure 5c)\u003c/strong\u003e. To ensure unbiased evaluation, both graphs were partitioned into 80% training, 10% validation, and 10% testing sets, with stratification across all relation types (27 for the Aging KG and 66 for EvoAge) confirmed in each subset \u003cstrong\u003e(Supplementary Figure 5d-f)\u003c/strong\u003e.\u003c/p\u003e\n\n\u003cp\u003eWe next benchmarked six state-of-the-art KGE architectures: Translating Embeddings (TransE) \u003csup\u003e28\u003c/sup\u003e, Rotation-based Knowledge Graph Embedding (RotatE) \u003csup\u003e29\u003c/sup\u003e, Simple Embedding (SimplE) \u003csup\u003e30\u003c/sup\u003e, Diagonal Matrix (DisMult) \u003csup\u003e31,32\u003c/sup\u003e, RElational SCALar (RESCAL) \u003csup\u003e33\u003c/sup\u003e, and Complex Embeddings (ComplEx) \u003csup\u003e34\u003c/sup\u003e, \u003cstrong\u003e(Figure 2a) \u003c/strong\u003eusing a standardized embedding size of 64 dimensions and a fixed training budget \u003cstrong\u003e(Supplementary Figure 5g, h)\u003c/strong\u003e. Performance was evaluated using Mean Rank (MR), Mean Reciprocal Rank (MRR), and hit@k metrics \u003cstrong\u003e(Supplementary Figure 5i; Supplementary Table 7)\u003c/strong\u003e. Model behavior was strongly scale-dependent: the tensor-based RESCAL model performed best on the smaller Aging KG, achieving the lowest MR and highest hit@10 and MRR values \u003cstrong\u003e(Figures 2b and c; Supplementary Figure 5j)\u003c/strong\u003e, with rapid and stable convergence \u003cstrong\u003e(Figure 2d, left)\u003c/strong\u003e. Conversely, for the billion-triple EvoAge KG, the geometric RotatE model consistently outperformed all alternatives \u003cstrong\u003e(Figures 2b and c; Supplementary Figure 5j)\u003c/strong\u003e, demonstrating smooth loss reduction over 4 million training steps \u003cstrong\u003e(Figure 2d, right)\u003c/strong\u003e. We then performed an ablation study to determine optimal embedding dimensionality (64-512 dimensions) for each selected model \u003cstrong\u003e(Figure 2e)\u003c/strong\u003e. For the Aging-RESCAL model, link prediction performance remained uniformly high across all dimensions \u003cstrong\u003e(Figure 2f, h)\u003c/strong\u003e. However, edge-type prediction improved substantially at higher dimensionality, peaking at 512 dimensions with a hit@1 score of 0.869 \u003cstrong\u003e(Figure 2g)\u003c/strong\u003e, with no evidence of overfitting \u003cstrong\u003e(Supplementary Figure 5k)\u003c/strong\u003e. Based on these findings, we selected the 512-dimensional RESCAL model for the Aging KG to maximize accuracy on the most difficult inference tasks. Notably, the EvoAge-RotatE model showed a different capacity profile: 64-dimensional embeddings yielded poor link prediction performance \u003cstrong\u003e(Figure 2f, i)\u003c/strong\u003e, whereas performance improved at 128 and 256 dimensions. The 512-dimensional model exhibited significant degradation in edge-type prediction \u003cstrong\u003e(Figure 2g)\u003c/strong\u003e, consistent with the phenomenon of overfitting \u003cstrong\u003e(Supplementary Figure 5l)\u003c/strong\u003e. After careful comparison, we selected the 128-dimensional RotatE model as the optimal compromise, delivering superior performance on edge type prediction with nearly identical link prediction performance to the 256-dim model \u003cstrong\u003e(Figure 2f, g, i)\u003c/strong\u003e.\u003c/p\u003e\n\n\u003cp\u003eFinally, to evaluate cross-species reasoning, a central goal of EvoAge, we constructed a held-out benchmark containing 3.14 million triples, evenly balanced between within-species and cross-species relationships \u003cstrong\u003e(Figures 2j and k)\u003c/strong\u003e. The final 128-dimensional EvoAge-RotatE model achieved uniformly high hit@k scores across all six species and, critically, performed comparably on the cross-species subset for both link prediction \u003cstrong\u003e(Figure 2l)\u003c/strong\u003e and edge-type classification \u003cstrong\u003e(Figure 2m)\u003c/strong\u003e. These results demonstrate that the EvoAge embeddings capture biologically coherent structure, enabling inference beyond memorized facts and supporting accurate prediction across evolutionary boundaries, fulfilling the design objective of EvoAge as an engine for cross-species discovery.\u003c/p\u003e\n\u003ch2\u003eA Conversational AI Interface Leverages the EvoAge KG for Discovery and Validation\u003c/h2\u003e\n\u003cp\u003eAfter constructing the EvoAge knowledge graph and optimizing its embedding models, we next developed an intelligent and accessible interface to operationalize the resource at scale. To achieve this, we developed the EvoAge AI Platform, designed to bridge natural language reasoning with large-scale computational biology. The platform supports three primary capabilities \u003cstrong\u003e(Figure 3a)\u003c/strong\u003e: retrieving existing knowledge from the Neo4j database through a natural language-based Search Database interface, performing Link Prediction using KGE inference to identify missing yet plausible biological relationships (for example, \u0026ldquo;Predict potential disease associations for the gene FOXO3.\u0026rdquo;), and executing a Test Hypothesis mode, in which a complete triple (for example, \u0026ldquo;does aging associate with cellular senescence?\u0026rdquo;) is assigned a quantitative plausibility score and an interpretable verdict. These functionalities are powered by a three-tier system architecture \u003cstrong\u003e(Figure 3b,c; Supplementary Figure 6a)\u003c/strong\u003e in which a Streamlit-based frontend \u003cstrong\u003e(Figure 3g)\u003c/strong\u003e interacts with a FastAPI backend orchestrating the Neo4j graph database \u003csup\u003e35\u003c/sup\u003e, the internal DGL-KE inference engine \u003csup\u003e36\u003c/sup\u003e, and external LLM services, including Google Gemini 2.5 Flash-Lite.\u003c/p\u003e\n\n\u003cp\u003eReasoning within the platform is governed by the EvoAge-Agent, which is implemented using the Kani framework \u003csup\u003e37\u003c/sup\u003e and GPT-4o-mini (OpenAI). The query-processing workflow \u003cstrong\u003e(Figure 3c)\u003c/strong\u003e begins with entity extraction, where the model identifies biological entities (for example, FOXO3) and invokes the Search Biological Entity tool \u003cstrong\u003e(Supplementary Figure 6d)\u003c/strong\u003e. The agent then infers user intent and selects the appropriate operational mode. Link prediction follows a defined 12-step execution sequence \u003cstrong\u003e(Figure 3d)\u003c/strong\u003e in which the system validates the query, loads trained embedding artifacts, generates relation-specific scores, and returns ranked results. Hypothesis testing follows a more complex logic path \u003cstrong\u003e(Figure 3e)\u003c/strong\u003e, combining a raw triple score from the DGL-KE model \u003cstrong\u003e(Supplementary Figure 6e)\u003c/strong\u003e with evidence retrieval from Neo4j, followed by interpretation from Gemini. The final response is synthesised into natural language \u003cstrong\u003e(Figure 6f)\u003c/strong\u003e, and the full workflow is optimised for interactivity, with most inference calls completing in under 30 seconds.\u003c/p\u003e\n\n\u003cp\u003eTo ensure scientific rigor in hypothesis evaluation, we developed a statistical calibration framework that converts raw embedding scores into interpretable thresholds. For each of the 66 relation types in EvoAge, we computed optimal cutoffs using Youden\u0026rsquo;s J statistic applied to score distributions generated from 1,000 true and 1,000 synthetic negative triples \u003cstrong\u003e(Supplementary Figure 6b; Supplementary Table 9)\u003c/strong\u003e. Thresholds varied substantially across relation categories, with frequently observed cellular component associations requiring lower (more negative) values, whereas sparse relations, such as chemical-mutation links, required substantially higher cutoffs \u003cstrong\u003e(Supplementary Figure 6c)\u003c/strong\u003e. Once integrated, this calibration step enabled reliable discrimination between plausible and implausible assertions. Finally, the platform was deployed as an interactive chat-based assistant, enabling users to navigate a billion-triple knowledge space through natural language. A representative example query for FOXO3 yields a ranked set of disease associations with numerical prediction scores \u003cstrong\u003e(Figures 3g and Supplementary Figure 6f)\u003c/strong\u003e, illustrating the platform\u0026rsquo;s ability to translate deep graph complexity into intuitive and actionable scientific insights.\u003c/p\u003e\n\u003ch2\u003eEvoAge Enables Accurate Hypothesis Generation and Experimental Validation\u003c/h2\u003e\n\u003cp\u003eTo quantitatively assess the discovery capability of EvoAge, we developed a benchmarking framework that compared its hypothesis-testing performance with that of state-of-the-art general-purpose LLMs \u003cstrong\u003e(Figure 4a)\u003c/strong\u003e. We first collected 1,282 aging-related preprints from bioRxiv (January-September 2025) and manually refined this to 980 studies focused on molecular and cellular aging biology \u003cstrong\u003e(Supplementary Table 10)\u003c/strong\u003e. From their abstracts, we used Google Gemini to generate a balanced test set comprising 2,850 positive and 2,850 negative hypotheses, ensuring that all models were evaluated on content not present in their training data. We then challenged four systems: three general-purpose LLMs (Gemma, Llama3.1, and Qwen2.5) \u003csup\u003e38\u0026ndash;40\u003c/sup\u003e and the EvoAge chatbot, to evaluate these hypotheses. All responses were automatically scored for biological plausibility by BioMistral-7B \u003csup\u003e41\u003c/sup\u003e, a model trained on millions of PubMed abstracts and used as an expert grading system \u003cstrong\u003e(Supplementary Table 11-12)\u003c/strong\u003e. EvoAge demonstrated a marked performance advantage, achieving significantly higher clarity scores than all other systems (p \u0026lt; 0.0001, pairwise Wilcoxon test with BH adjustment) \u003cstrong\u003e(Figure 4b)\u003c/strong\u003e. While the general-purpose LLMs demonstrated moderate accuracy on positive hypotheses \u003cstrong\u003e(Figure 4c)\u003c/strong\u003e, they often failed to reject implausible statements, as indicated by high scores on negative hypotheses \u003cstrong\u003e(Figure 4d)\u003c/strong\u003e. In contrast, EvoAge produced a distinctly skewed score distribution with consistently low values for false hypotheses, indicating strong discriminatory power derived from its structured knowledge base. \u003c/p\u003e\n\n\u003cp\u003eNext, to investigate whether EvoAge\u0026rsquo;s knowledge-driven hypothesis evaluation framework translates into experimentally verifiable biology, we examined an EvoAge-supported mechanism involving the nanoscale redistribution of the \u0026beta;-secretase enzyme BACE1, which is involved in the amyloidogenic processing of Amyloid Precursor Protein (APP) in Alzheimer\u0026rsquo;s disease (AD) \u003cstrong\u003e(Supplementary Figure 8a)\u003c/strong\u003e. EvoAge assigned this hypothesis a high plausibility score and did not classify it as species-restricted, suggesting that if present, the mechanism should manifest across phylogenetically divergent biological systems \u003cstrong\u003e(Supplementary figure 8b)\u003c/strong\u003e. To experimentally validate this hypothesis, we quantified BACE1 nanoscale organisation using STED nanoscopy in multiple models of AD: human iPSC-derived glutamatergic neurons carrying a familial AD (FAD) mutation and their isogenic controls, postmortem human brain tissue, and wild-type versus AD transgenic mouse brain \u003cstrong\u003e(Figure 4f)\u003c/strong\u003e. We co-mapped BACE1 (magenta) relative to the Perisynapse/Postsynapse markers (green) in these conditions and observed significant clustering and enrichment in the predicted zones, shown here for iPSCs \u003cstrong\u003e(Figure 4g-j)\u003c/strong\u003e. Shapiro-Wilk testing on species (mouse and human), conditions (Healthy vs AD), and compartmental measurements (postsynaptic vs perisynaptic) demonstrated significant deviation from normality (p \u0026lt; 10\u003csup\u003e-14\u003c/sup\u003e to 10\u003csup\u003e-63\u003c/sup\u003e) \u003cstrong\u003e(Supplementary Table 13)\u003c/strong\u003e. Gaussian Mixture Modeling confirmed that two-component distributions consistently provided a superior fit compared to single Gaussian models, with significant improvements in the Bayesian Information Criterion \u003cstrong\u003eSupplementary Table 14)\u003c/strong\u003e. This suggests that BACE1 exists not as a uniform entity at the nanoscale, but rather as at least two discrete biological subtypes that differ in composition, spatial architecture, or functional state. \u003c/p\u003e\n\n\u003cp\u003eQuantitative compartmental analysis revealed a conserved directional shift in BACE1 intensity from postsynaptic density (PSD) toward perisynaptic endocytic regions in disease. In iPSC-derived human neurons, perisynaptic nanodomain intensity modestly exceeded postsynaptic levels in controls (67.45 vs. 61.90; fold-change +0.0897), but the difference increased markedly in FAD-mutant neurons (76.66 vs. 59.60; fold-change +0.2863; MW p = 4.6\u0026times;10\u003csup\u003e-14\u003c/sup\u003e). Postmortem human cortex exhibited a substantially stronger redistribution, with perisynaptic levels showing nearly double the relative separation (fold-change +0.8331 in AD vs. +0.3981 in control; MW p = 1.22\u0026times;10\u003csup\u003e-53\u003c/sup\u003e; t-test p = 1.68\u0026times;10\u003csup\u003e-67\u003c/sup\u003e), and effect size increased to a measurable medium magnitude (Cliff \u0026Delta; = 0.445). Mouse tissue retained the expected physiological pattern of PSD enrichment (101.38 to 32.02; fold-change -0.6842) but underwent a significant shift toward perisynaptic enrichment in Tg (114.68 to 74.89; fold-change -0.3469; MW p = 1.77\u0026times;10\u003csup\u003e-33\u003c/sup\u003e) \u003cstrong\u003e(Figure 4k)\u003c/strong\u003e \u003cstrong\u003e(Supplementary Figure 7a) (Supplementary Table 15)\u003c/strong\u003e. This pattern reflects a species-scaled trajectory, nanoscale mislocalization of BACE1 in human iPSC neurons, consolidation in human AD brain tissue, and full directional polarity shift in mouse models.\u003c/p\u003e\n\n\u003cp\u003eFurther, nanodomain area and length measurements revealed a biphasic structural remodeling. In human iPSC neurons, perisynaptic nanodomains were slightly larger than postsynaptic ones in both control and mutant neurons (+0.043 and +0.012 fold-change, respectively), although effect sizes were negligible, indicating subtle early remodeling. In human and mouse brain tissue, a striking collapse of the perisynaptic nanodomain area emerged. In transgenic mouse (Tg), perisynaptic area declined from 0.00651 to 0.00179 (fold-change -0.779; p \u0026lt; 1\u0026times;10\u003csup\u003e-20\u003c/sup\u003e), mirrored closely in human AD cortex, suggesting compaction into dense nanoclusters \u003cstrong\u003e(Figure 4l, Supplementary Figure 7c)\u003c/strong\u003e. Length measurements further demonstrated interesting observations across species and different models of AD. iPSC-derived neurons showed negligible structural differences (fold-change +0.0056 to +0.0149, non-significant) \u003cstrong\u003e(Figure 4m, Supplementary Figure 7b)\u003c/strong\u003e. Human samples demonstrated elongation of perisynaptic nanodomains in AD (fold-change +0.442; t-test p = 1.95\u0026times;10\u003csup\u003e-6\u003c/sup\u003e), consistent with early architectural expansion prior to collapse. In contrast, mouse AD showed the strongest compaction phenotype (fold-change -0.388; MW p = 4.24\u0026times;10\u003csup\u003e-133\u003c/sup\u003e), suggesting terminal pathological confinement \u003cstrong\u003e(Supplementary Table 15)\u003c/strong\u003e. These findings demonstrate that AD induces a reproducible and quantifiable nanoscale redistribution of BACE1, from PSD anchoring toward perisynaptic endocytic nanodomains, followed by progressive structural remodeling from diffuse domains into compact stable nanoclusters. This redistribution aligns with known amyloidogenic processing architecture, where endocytic microenvironments of acidic pH enhance \u0026beta;-secretase catalytic efficiency and trafficking activity. Importantly, the consistency of this nanoscale signature across these species and disease stages validates EvoAge\u0026rsquo;s prediction and reveals a conserved molecular progression underlying synaptic pathology in AD. \u003c/p\u003e\n"},{"header":"DISCUSSION","content":"\u003cp\u003eAging is an inherently evolutionary process \u003csup\u003e42\u003c/sup\u003e, yet aging research has remained predominantly species-segregated, concept-fragmented, and methodologically siloed \u003csup\u003e43\u003c/sup\u003e. Although the conservation of aging pathways across phylogeny is well recognized \u003csup\u003e9,44\u0026ndash;47\u003c/sup\u003e, translating insights across species remains technically constrained and conceptually fragmented \u003csup\u003e48\u003c/sup\u003e. In this study, we introduce EvoAge, a multispecies, AI-enabled knowledge framework that bridges evolutionary relationships with systems-level biological knowledge. By integrating 48 public resources into a billion-triple, orthology-aligned graph spanning six species, EvoAge demonstrates that aging biology can be computationally represented in a form that is both evolutionarily coherent and mechanistically navigable. The resulting architecture is not merely comprehensive; it reflects the scale-free \u003csup\u003e49\u003c/sup\u003e, modular topology characteristic of real biological systems, reaffirming that biological complexity \u003csup\u003e50\u003c/sup\u003e emerges naturally when previously isolated information streams are unified.\u003c/p\u003e\n\n\u003cp\u003eA major conceptual contribution of this work is the clarification of the complementary roles of structured knowledge and generative AI in the scientific reasoning process \u003csup\u003e51\u003c/sup\u003e. Our benchmarking demonstrates that while general-purpose LLMs excel at linguistic fluency and memorization, they struggle with biological plausibility discrimination, frequently validating false hypotheses or producing confident hallucinations \u003csup\u003e52\u003c/sup\u003e. EvoAge\u0026rsquo;s hybrid design addresses this limitation by positioning the knowledge graph as the factual substrate, the embedding model as the probabilistic reasoning layer, and the LLM as the interpretive and interactive interface \u003csup\u003e53,54\u003c/sup\u003e. This division of labor yielded tangible improvements: EvoAge significantly outperformed state-of-the-art general LLMs on a challenging benchmark of 5,700 hypotheses derived from recent preprints, particularly in its ability to reject implausible hypotheses, a fundamental requirement for scientific inference. The ultimate test of a discovery platform, however, is not computational accuracy but biological relevance and experimental tractability. EvoAge met this criterion by prioritizing and supporting a mechanistically grounded hypothesis in AD that \u0026beta;-secretase (BACE1) undergoes nanoscale synaptic redistribution in AD, shifting from postsynaptic to perisynaptic and endocytic compartments where amyloidogenic processing occurs. Guided by EvoAge\u0026rsquo;s species-agnostic evaluation, we validated this hypothesis using human iPSC-derived neurons, human postmortem brain tissue, and mouse models. Across all systems, STED imaging confirmed increased perisynaptic BACE1 in AD, whereas healthy tissue showed the inverse distribution (Posterior \u0026gt; Perisynaptic), aligning with canonical synaptic organization. Statistical modeling further substantiated this finding. Gaussian mixture models revealed multimodal nanoscale populations rather than uniform distributions, and Bayesian hierarchical inference demonstrated consistent directional shifts toward perisynaptic enrichment across systems, with posterior confidence rising from 0.13-0.59 in raw data to 0.81-0.85 after outlier filtering. Despite small to moderate effect sizes, typical of high-content synaptic datasets and synaptic heterogeneity, the signal was robust, reproducible, and cross-species consistent. These results support a model in which AD induces both biochemical (increased BACE1 intensity) and morphological (altered area and length) nanoscale remodeling, forming compact perisynaptic nanodomains that may enhance catalytic efficiency via molecular crowding and vesicular retention properties, which are mechanistically favorable for increased processing of APP through the amyloidogenic pathway. \u003c/p\u003e\n\n\u003cp\u003eWhile powerful, EvoAge is not yet complete. Its content reflects the biases and gaps of source databases, and its current orthology model simplifies many-to-many evolutionary relationships \u003csup\u003e55\u003c/sup\u003e. Furthermore, although EvoAge can evaluate hypotheses with high precision, transitioning from hypothesis validation to intervention prioritization necessitates integrating causal perturbation data \u003csup\u003e56\u003c/sup\u003e, gene regulatory dynamics, and spatial omics \u003csup\u003e57,58\u003c/sup\u003e. Future development will focus on expanding biological granularity, incorporating dynamic updates from new preprints and datasets, refining evolutionary logic to capture paralogy and gene replacement events, and integrating causal inference frameworks capable of predicting interventional outcomes. In the longer term, we envision EvoAge functioning as a federated scientific ecosystem where users contribute private or unpublished data, iteratively refine predictions, and collaboratively generate experimentally actionable insights. In summary, EvoAge represents a methodological shift, moving from fragmented knowledge and static databases toward a continuously evolving, evolution-informed computational framework that supports mechanistic reasoning across species. By unifying evolutionary theory, systems biology, and agentic AI, EvoAge is not simply a repository of aging knowledge. It is a platform that evaluates hypotheses, prioritizes mechanistic plausibility, and guides experimental discovery. As the biology of aging continues to expand in scale and complexity, such integrative frameworks will be essential for uncovering universal principles and actionable intervention targets in aging and age-associated disease.\u003c/p\u003e"},{"header":"MATERIAL AND METHODS","content":"\u003ch2\u003eData Source Curation and Collection\u003c/h2\u003e\n\u003cp\u003eData were programmatically aggregated from 48 publicly available biological databases and knowledge graphs. This included 12 specialized aging resources (AgeAnno \u003csup\u003e7\u003c/sup\u003e, AgeAnnoMO \u003csup\u003e8\u003c/sup\u003e, AgeXtend \u003csup\u003e9\u003c/sup\u003e, Aging Atlas \u003csup\u003e10\u003c/sup\u003e, CellAge \u003csup\u003e5\u003c/sup\u003e, Digital Ageing Atlas \u003csup\u003e59\u003c/sup\u003e, DrugAge \u003csup\u003e16\u003c/sup\u003e, GenDR \u003csup\u003e17\u003c/sup\u003e, GeneAge \u003csup\u003e60\u003c/sup\u003e, HALD \u003csup\u003e11\u003c/sup\u003e, Literature, and MetaboAge \u003csup\u003e61\u003c/sup\u003e and 36 general-purpose biological databases spanning molecular interactions (STRING (v12.0) \u003csup\u003e62\u003c/sup\u003e, STITCH (v5.0) \u003csup\u003e63\u003c/sup\u003e, BioGRID (5.0) \u003csup\u003e64\u003c/sup\u003e, chemicals and drugs (ChEMBL (v33, released 31-May-2023) \u003csup\u003e65\u003c/sup\u003e, BindingDB (downloaded 28-May-2023) \u003csup\u003e66\u003c/sup\u003e, DRKG \u003csup\u003e67\u003c/sup\u003e, model organism databases (FlyBase \u003csup\u003e24\u003c/sup\u003e, WormBase WS287 \u003csup\u003e25\u003c/sup\u003e, Worm Interactome Database \u003csup\u003e68\u003c/sup\u003e, ZFIN \u003csup\u003e26\u003c/sup\u003e, MGI \u003csup\u003e27\u003c/sup\u003e, MouseNet \u003csup\u003e27\u003c/sup\u003e, SGD \u003csup\u003e22\u003c/sup\u003e, YeastNet (v3) \u003csup\u003e23\u003c/sup\u003e), and large-scale knowledge graphs (CKG) \u003csup\u003e15\u003c/sup\u003e, BioGrakn \u003csup\u003e69\u003c/sup\u003e, CROssBAR \u003csup\u003e70\u003c/sup\u003e, Hetionet \u003csup\u003e13\u003c/sup\u003e, MonarchKG \u003csup\u003e71\u003c/sup\u003e, PrimeKG \u003csup\u003e14\u003c/sup\u003e, Harmonizome \u003csup\u003e72\u003c/sup\u003e, TarKG \u003csup\u003e73\u003c/sup\u003e, DTInet \u003csup\u003e74\u003c/sup\u003e, GP-KG \u003csup\u003e75\u003c/sup\u003e, PharmKG \u003csup\u003e76\u003c/sup\u003e, iBKH \u003csup\u003e77\u003c/sup\u003e, TTD \u003csup\u003e78\u003c/sup\u003e. All data were retrieved in their latest available versions as of January 2025 \u003cstrong\u003e(Supplementary Table 2)\u003c/strong\u003e.\u003c/p\u003e\n\u003ch2\u003eData Preprocessing and Harmonization\u003c/h2\u003e\n\u003cp\u003eA systematic preprocessing pipeline was implemented to handle the inherent heterogeneity of the source data. Each dataset was parsed to extract entities (nodes) and relationships (edges) conforming to a predefined biological entity types (Gene, Protein, Chemical, Disease, Phenotype, etc.). Entity names and relation labels were standardized to consistent naming conventions. To ensure interoperability, a rigorous identifier mapping framework was applied, converting all source identifiers to community-standard reference systems: NCBI \u003csup\u003e79\u003c/sup\u003e Gene for genes, UniProt \u003csup\u003e80\u003c/sup\u003e for proteins, PubChem \u003csup\u003e81\u003c/sup\u003e/DrugBank \u003csup\u003e82\u003c/sup\u003e for chemicals, DOID \u003csup\u003e83\u003c/sup\u003e/MONDO \u003csup\u003e84\u003c/sup\u003e for diseases, UBERON \u003csup\u003e85\u003c/sup\u003e for anatomy, and KEGG \u003csup\u003e86\u003c/sup\u003e/Reactome \u003csup\u003e87\u003c/sup\u003e for pathways. Gene Ontology (GO) \u003csup\u003e88\u003c/sup\u003e is used for Biological Process, Cellular Component, and Molecular Function, and HPO \u003csup\u003e89\u003c/sup\u003e for phenotypes. Duplicate entities were identified via exact and semantic matching and merged into single, harmonized nodes while preserving all original source identifiers for full data provenance. During this process, entities originally classified as chemicals, drugs, metabolites, or phytochemicals were all consolidated under the single \u0026apos;Chemical\u0026apos; node type. This unification was performed by merging all nodes that share the same canonical SMILES string using openbabel \u003csup\u003e90\u003c/sup\u003e, ensuring a single, unique representation for each distinct molecule.\u003c/p\u003e\n\u003ch2\u003eOrthology-Based Cross-Species Integration\u003c/h2\u003e\n\u003cp\u003eTo enable true cross-species querying, a one-to-one human ortholog mapping strategy was implemented. Genes from \u003cem\u003eS. cerevisiae\u003c/em\u003e, \u003cem\u003eC. elegans\u003c/em\u003e, \u003cem\u003eD. melanogaster\u003c/em\u003e, \u003cem\u003eD. rerio\u003c/em\u003e, and \u003cem\u003eM. musculus\u003c/em\u003e were mapped to their \u003cem\u003eH. sapiens\u003c/em\u003e orthologs using the g:Profiler2 (G:Orth) tool \u003csup\u003e91\u003c/sup\u003e, which queries the Ensembl Compara database. For genes with multiple potential orthologs, only the first-listed ortholog was retained to maintain a clean, unambiguous graph. The original species-specific gene identifiers in the graph were subsequently replaced with their corresponding human ortholog identifiers, genetically unifying the knowledge base.\u003c/p\u003e\n\u003ch2\u003eGraph Database Implementation\u003c/h2\u003e\n\u003cp\u003eThe harmonized data were loaded into a Neo4j graph database (version 5.26.10) \u003csup\u003e92\u003c/sup\u003e using Cypher queries and APOC (version 5.22.0) utilities for batch operations and integrity checks. The final EvoAge Knowledge Graph comprises 30,159,124 nodes and 1,043,615,837 edges (triples) spanning 66 relation types, creating a scalable, queryable resource for multi-species biological exploration.\u003c/p\u003e\n\u003ch2\u003eKnowledge Graph Embedding (KGE) Training\u003c/h2\u003e\n\u003cp\u003eVector representations of entities and relations were learned using the Deep Graph Library (DGL, version 1.1.2) \u003csup\u003e93\u003c/sup\u003e, and DGL-KE (version 0.1.0) \u003csup\u003e36\u003c/sup\u003e\u003cstrong\u003e\u0026nbsp;(Supplementary Table 8)\u003c/strong\u003e. Six KGE models (TransE, RotatE, SimplE, DisMult, RESCAL, ComplEx) \u003csup\u003e28\u0026ndash;31\u003c/sup\u003e\u003csup\u003e,\u003c/sup\u003e\u003csup\u003e33,34\u003c/sup\u003e were systematically evaluated. The triples were split into training (80%), validation (10%), and test (10%) sets. Models were trained with an embedding dimension of 64, a batch size of 2048, and a margin loss function. The optimal architectures were identified as RESCAL for the smaller Aging KG and RotatE for the full EvoAge KG. An ablation study on embedding size (64, 128, 256, 512) determined the final configurations: a 512-dimensional RESCAL model for the Aging KG and a 128-dimension RotatE model for the EvoAge KG to balance performance and avoid overfitting.\u003c/p\u003e\n\u003ch2\u003eComputational Infrastructure\u003c/h2\u003e\n\u003cp\u003eAll training and evaluation were executed on a high-performance Linux workstation (Ubuntu 22.04 LTS) equipped with two NVIDIA GPUs, an RTX 5000 Ada Generation (32 GB VRAM), and a GeForce RTX 3090 (24 GB VRAM) running CUDA 12.0. This hardware configuration enabled parallelized computation and efficient handling of large-scale graph data, comprising over 30 million nodes and 1.04 billion edges.\u003c/p\u003e\n\u003ch2\u003eType-Constrained Link Prediction and Inference Optimization\u003c/h2\u003e\n\u003cp\u003eTo ensure that EvoAge generates semantically valid and biologically meaningful predictions, we extended the default DGL-KE inference pipeline to incorporate explicit type constraints. An entity-type mapping was integrated into the system, enabling the inference module to identify the required semantic category for each relation. During prediction, candidate entities are first filtered by their assigned type, and only those matching the expected head or tail type of the queried relation are considered for scoring and ranking. This modification prevents invalid cross-type predictions-for example, ensuring that a gene\u0026ndash;disease relation can only return disease entities as plausible targets, and thereby improves both the interpretability and precision of link prediction and hypothesis generation.\u003c/p\u003e\n\u003cp\u003eTo enable real-time inference within the EvoAge chatbot, we optimized the DGL-KE based prediction pipeline by introducing an internal caching mechanism for model artifacts. In the default workflow, essential components such as the entity and relation dictionaries are reloaded from disk for every query, whether the request involves hypothesis generation or link prediction. This reload step typically takes 30-50 seconds, and repeating it for every query makes interactive analysis infeasible. To overcome this limitation, we modified the DGL-KE library so that these artifacts are loaded only once and then retained in memory. With this caching mechanism, subsequent inference calls bypass the expensive reload phase entirely, reducing artifact access time to approximately 2-4 seconds. This optimization substantially lowers end-to-end latency and enables near real-time performance for iterative hypothesis testing and link prediction within EvoAge.\u003c/p\u003e\n\u003ch2\u003eHypothesis Testing and Youden\u0026apos;s Thresholding\u003c/h2\u003e\n\u003cp\u003eA data-driven approach established objective thresholds for the \u0026quot;Test Hypothesis\u0026quot; function. For each of the 66 relation types, the KGE model was evaluated using 1,000 known true triples and 1,000 generated false triples. The Youden\u0026apos;s J statistic (J = True Positive Rate - False Positive Rate) was applied to these score distributions to determine the optimal, relation-specific cutoff for accepting a user-proposed triple as plausible.\u003c/p\u003e\n\u003ch2\u003eBackend API and Service Architecture\u003c/h2\u003e\n\u003cp\u003eThe EvoAge platform backend was built as a FastAPI server, providing RESTful endpoints for core functions: /search_biological_entities for faceted full-text search (powered by a Neo4j/Lucene index); /subgraph for 1-hop neighborhood retrieval; /entity_relationships for quantitative adjacency summarization; /check_relationship for paired-node verification; /get_sample_nodes for retrieving sample nodes from the KG; and /get_sample_triples for retrieving sample triples from the KG. Additionally, the platform offers advanced computational endpoints. A Link Prediction endpoint calls the KGE scoring service, while a dedicated Hypothesis testing endpoint orchestrates calls to this same service, the Neo4j graph for fact-checking, and the Gemini API for natural language interpretation.\u003c/p\u003e\n\u003ch2\u003eAgentic Framework and Query Processing\u003c/h2\u003e\n\u003cp\u003eThe EvoAge-agent, built using the Kani framework and GPT-4o-mini \u003csup\u003e94\u003c/sup\u003e, serves as the central reasoning engine. It processes user queries by first performing biological entity extraction and enrichment via the search API. It then detects user intent to route the query to one of three specialized tool paths: direct KG lookup, link prediction, or hypothesis testing. The agent synthesizes the results from these tools into a final, coherent natural language response.\u003c/p\u003e\n\u003ch2\u003eFrontend User Interface\u003c/h2\u003e\n\u003cp\u003eThe graphical user interface was developed using Streamlit (v1.48.0). It features a secure authentication module and a chat-based interface where users can submit natural language queries and view responses that combine explanatory text with structured, ranked predictions, including KGE confidence scores.\u003c/p\u003e\n\u003ch2\u003eComputational Benchmarking\u003c/h2\u003e\n\u003cp\u003eA rigorous benchmark was designed to evaluate the hypothesis-testing capability. From 1,282 recent BioRxiv preprints (Jan-Sep 2025) on aging, 980 were manually curated. Using Gemini 2.5 Flash Lite, 5,700 balanced positive and negative hypothesis questions were generated. EvoAge and three general-purpose LLMs answered these questions. The specific models evaluated were \u0026quot;google/gemma-2bit\u0026quot; \u003csup\u003e38\u003c/sup\u003e, \u0026quot;llama3.1:8b-instruct-q4_K_M\u0026quot; \u003csup\u003e95\u003c/sup\u003e, and \u0026quot;Qwen/Qwen2.5-7B-Instruct\u0026quot; \u003csup\u003e40\u003c/sup\u003e. All answers were graded on a 1-10 scale for biological feasibility by a \u0026quot;BioMistral/BioMistral-7B\u0026quot; \u003csup\u003e41\u003c/sup\u003e teacher model. All LLMs were run using the vLLM inference engine (v0.10.1.1) \u003csup\u003e96,97\u003c/sup\u003e\u003cstrong\u003e.\u003c/strong\u003e Performance was assessed using a composite \u0026quot;clarity score,\u0026quot; calculated as (Positive Grade) + (11 - Negative Grade). This metric rewards models that both accept plausible hypotheses (high positive score) and reject implausible ones (low negative score).\u003c/p\u003e\n\u003ch2\u003eEthical Compliance\u003c/h2\u003e\n\u003cp\u003eThe human brain and transgenic mouse data were sourced as reported previously. \u003csup\u003e98\u003c/sup\u003e and were repurposed in this study for novel analyses. All human stem cell (iPSCs) data presented here were newly generated. All human stem cell work for iPSCs was carried out in accordance with approval from the Institutional Human Ethics Committee and Institutional Biosafety Committee at the Institute for Stem Cell Science and Regenerative Medicine, Bengaluru, India.\u003c/p\u003e\n\u003ch2\u003eiPSC Generation and Gene Editing\u003c/h2\u003e\n\u003cp\u003eThe induced pluripotent stem cells (iPSCs) were generated from skin fibroblasts of a 58-year-old male carrying the Familial Alzheimer\u0026rsquo;s Disease (FAD) mutation L150P in the Presenilin1 (PSEN1) protein. The patient carried a heterozygous point mutation (c.449C \u0026rarr; T) in exon 6 of the PSEN1 gene. The iPSCs were generated by the electroporation of the fibroblasts with episomal plasmids containing hOCT4, hSOX2, hNANOG, hKLF4, hMyc, hLIN28, and shRNA against TP53. The patient-derived iPSCs have been characterized for normal karyotype, expression of pluripotency markers, and differentiation into the three germ layers \u003csup\u003e99\u003c/sup\u003e. The isogenic control line for the same iPSCs was generated using CRISPR/Cas9 gene editing. A 23bp guide RNA was used to revert the point mutation (c.449C \u0026rarr; T) to obtain the gene-corrected iPSC line. The gene-corrected iPSCs have been characterized for normal karyotype and expression of pluripotency markers \u003csup\u003e100\u003c/sup\u003e.\u003c/p\u003e\n\u003ch2\u003eStem Cell Culture, Maintenance, and Differentiation\u003c/h2\u003e\n\u003cp\u003eThe iPSC colonies were maintained in mTeSR1 complete medium supplemented with 1% Penicillin-Streptomycin under the conditions of 37˚C, 5% CO\u003csub\u003e2\u003c/sub\u003e and 21% O\u003csub\u003e2\u003c/sub\u003e. When confluent, they were split in a 1:3 ratio onto pre-coated Matrigel dishes using a dissociation mixture prepared with Collagenase Type IV (1 mg/ml), Trypsin (0.25%), Knockout Serum Replacement (20%), and Calcium Chloride (1 mM).\u003c/p\u003e\n\u003cp\u003eThe protocol for neural differentiation was the same as published previously \u003csup\u003e101\u003c/sup\u003e. In brief, iPSCs were expanded in mTeSR until 70-80% confluency was reached. The Neural Basic Media (NBM) for differentiation contained 50% DMEM F-12, 50% Neurobasal, 0.1% PenStrep, Glutamax, N\u003csub\u003e2\u003c/sub\u003e, and B27 without Vitamin A. Neural induction was initiated in the iPSC monolayer via Dual SMAD inhibition. This was achieved by replacing the existing media with Neural Induction Media (NIM), containing the Neural Basic Media (NBM) supplemented with the TGF\u0026beta; pathway inhibitor SB431542 (10 \u0026micro;M) and the BMP pathway inhibitor LDN193189 (0.1 \u0026micro;M). The cells were subjected to neural induction for 12-15 days by changing the NIM every day. After induction, the monolayer was dissociated using Accutase, and the cells were plated in NIM containing 10 \u0026micro;M ROCK inhibitor overnight on pre-coated poly-L-ornithine/laminin dishes. Poly-L-Ornithine (1:10 dilution) followed by Laminin (5\u0026micro;g/ml) coating was used for the maintenance of neural progenitors and their terminal differentiation. Expansion of neural progenitor cells was carried out in Neural Expansion Media (NEM), which is composed of NBM supplemented with FGF (10 ng/mL) and EGF (10 ng/mL). Neuronal maturation and terminal differentiation were achieved by plating the neural stem cells at a density of 25,000-35,000 cells/cm2 in the Neural Maturation Media (NMM) composed of NBM supplemented with BDNF (20ng/ml), GDNF (10ng/ml), L-Ascorbic Acid (200 \u0026micro;M), and db-Camp (50\u0026micro;M). The neurons were subjected to maturation for a period of 60-70 days by supplementing them with NMM every 4-5 days. The differentiated neurons were characterized by their expression of glutamatergic, axonal, and dendritic markers \u003csup\u003e102\u003c/sup\u003e. The list of materials used are detailed in \u003cstrong\u003e(Supplementary Table 16)\u003c/strong\u003e.\u003c/p\u003e\n\u003ch2\u003eImmunocytochemistry\u003c/h2\u003e\n\u003cp\u003eFor iPSCs, the immunocytochemistry was performed as reported previously (\u003csup\u003e103\u003c/sup\u003e,\u003csup\u003e98\u003c/sup\u003e, \u003csup\u003e104\u003c/sup\u003e). Briefly, cells were fixed with 4% paraformaldehyde plus 4% sucrose in PBS at 4\u0026deg;C for 10 minutes, followed by quenching with 0.1M glycine in PBS at room temperature and permeabilization with 0.25% Triton X-100 for 5 minutes, and then blocked with 10% Bovine Serum Albumin (BSA) in PBS for 30 minutes at room temperature. This was followed by incubation with the appropriate primary antibody for 1-2 hr. The primary antibodies used were Anti-BACE1 (Biolegend/Covance, #840101) (1:200), Anti-Shank2 (Synaptic Systems, #162204) (1:500), and Anti-Clathrin (Abcam, #ab2731) (1:100). Following washing, cells were then incubated with a suitable secondary antibody for 45 minutes. The secondary antibodies used include Alexa Fluor 594 (Life Technologies, #A11037) (1:200), Abberior Star Red (Abberior, #2-0112-011-8 and #2-0002-011-2) (1:200). Following washing, cells were mounted with Prolong (Molecular Probes, cat. no. MAN0010261) for STED imaging.\u003c/p\u003e\n\u003ch2\u003eImmunohistochemistry (Mice)\u003c/h2\u003e\n\u003cp\u003eImmunohistochemical data from mouse samples were obtained from experiments previously performed and reported \u003csup\u003e98\u003c/sup\u003e. Coronal cryosections (25 \u0026micro;m) from Tg(APPswe/PS1\u0026Delta;E9) mice (JAX Stock #004462) and age-matched control littermates were stained and mounted in ProLong with DAPI (Molecular Probes, cat. no. P36962) for confocal and STED imaging. The primary and secondary antibodies used are listed in \u003csup\u003e98\u003c/sup\u003e. Imaging was conducted in the CA1\u0026ndash;CA2 stratum radiatum of the hippocampus.\u003c/p\u003e\n\u003ch2\u003eImmunohistochemistry (Human)\u003c/h2\u003e\n\u003cp\u003eHuman immunohistochemical data were similarly sourced from previously performed and reported experiments \u003csup\u003e98\u003c/sup\u003e. Neuroanatomical sampling followed NIA-AA guidelines for the neuropathological assessment of Alzheimer\u0026rsquo;s disease (reference). Ethical clearance for the collection, storage, and distribution of human brain tissue was obtained from the Human Brain Tissue Repository (Brain Bank) at NIMHANS, Bangalore. Neuropathological classification followed standardized criteria incorporating three parameters: A\u0026beta; plaque score, Braak and Braak neurofibrillary tangle (NFT) stage, and CERAD neuritic plaque score to derive an ABC score categorized into four levels: not, low, intermediate, and high. The Blessed Dementia Rating Scale score for the control case was 1.5/17. A summary of staging and pathological features for all samples has been reported previously \u003csup\u003e98\u003c/sup\u003e. Paraffin-embedded human tissue sections were stained with primary and secondary antibodies, followed by mounting with ProLong containing DAPI (Molecular Probes, cat. no. P36962) as reported previously \u003csup\u003e98\u003c/sup\u003e.\u003c/p\u003e\n\u003ch2\u003eStimulated Emission Depletion microscopy (STED)\u003c/h2\u003e\n\u003cp\u003eA commercial STED inverted microscope (Abberior Expert Line 775 nm, Abberior Instruments GmbH, G\u0026ouml;ttingen, Germany) was used to obtain confocal and super-resolved images of the same region with a sampling of 15 nm. The microscope was equipped with two pulsed excitation lasers at 561 nm and 640 nm, as well as a pulsed depletion laser at 775 nm. The laser powers were adjusted to 70%, 50% and 40% of their respective total power for 561 nm, 640 nm, and 775 nm, respectively, as described previously \u003csup\u003e98,103\u003c/sup\u003e.\u003c/p\u003e\n\u003ch2\u003eSemiautomated detection of dendritic compartments and functional zones of an excitatory synapse\u003c/h2\u003e\n\u003cp\u003eThe active dendritic area of the protein of interest and synapses were distinguished from the rest of the dendrite using a protocol as described previously \u003csup\u003e98,104,105\u003c/sup\u003e. Briefly, the intensity of the epifluorescence/confocal images of markers for different functional zones of the synapse (post/peri) was thresholded to generate the mask of the puncta. A spine morphometry analysis was then performed, and masks were filtered using various morphological filters, such as length, breadth, and area, through the IMA plugin running inside the MetaMorph software (Molecular Devices) \u003csup\u003e106\u003c/sup\u003e. A similar analysis was performed on super-resolution images to detect functional zones of an excitatory synapse (post- and peri-synaptic) that correspond to PSD/EZ functional zones.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSTED Nanodomain Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u0026beta;-secretase (BACE1) nanodomains were identified in STED images using Palm-Tracer, as described previously \u003csup\u003e98,103,105,107\u003c/sup\u003e. Nanodomains were analyzed using two-dimensional Gaussian fitting, from which morphological and biophysical parameters, such as length (2.3\u0026sigma;long), area, and nanodomain intensity, were computed for each experimental group. Gaussian fitting was performed on every cluster identified as a nanodomain.\u003c/p\u003e\n\u003ch2\u003eNanodomain metric distribution analysis\u003c/h2\u003e\n\u003cp\u003eStatistical analyses and visualization were performed using Python, employing the following libraries and versions: pandas (v2.3.2), numpy (v2.2.6), matplotlib (v3.10.3), statsmodels (v0.14.5), and scikit-learn (v1.15.3). The normality of BACE1 intensity data across all experimental groups was first rigorously assessed using the Shapiro\u0026ndash;Wilk test (\u0026alpha; = 0.05) via SciPy to determine whether the distributions conformed to Gaussian assumptions. Data failing the test (p-value =\u0026lt; 0.05) was categorized as Non-Gaussian. To investigate underlying heterogeneity, nanodomain morphological and biophysical parameters (Length, Area, Intensity) were subjected to Gaussian Mixture Model (GMM) analysis. We compared 1- and 2-component GMMs using the Bayesian Information Criterion (BIC), concluding that the presence of two distinct subpopulations was indicated if the 2-component model yielded a significantly lower BIC.\u003c/p\u003e\n\u003cp\u003eThe Mann-Whitney U test (also known as the Wilcoxon rank-sum test) was selected as the non-parametric method of choice for statistical comparisons, given that initial normality tests indicated that the BACE1 nanodomain metrics were non-Gaussian. This robust test was applied to two critical comparisons: 1) Synaptic Compartment Comparison, evaluating the differences in BACE1 properties between the Postsynaptic and Perisynaptic regions within a single genotype; and 2) Genotype Comparison, assessing the effect of the disease (AD/Mutant) versus Control (GC/Wild-Type) on BACE1 nanodomain metrics within a specific synaptic region, across all species models.\u003c/p\u003e\n\u003cp\u003eNormalization was performed by dividing all BACE1 nanodomain metric values (Intensity, Area, and Length) within a specific species and measurement by the median Postsynaptic Control value for that same species/measurement. This step ensured that the absolute scale differences between models (iPSC, Mouse, Human) were removed, making the relative remodeling effect comparable across species. Following normalization, the Mann-Whitney U test was selected as the non-parametric method for statistical comparisons. This robust test was applied to the normalized data to compare the AD genotype against the Healthy within both the Postsynaptic (PSD) and Perisynaptic (EZ) compartments across all three models. The goal was to statistically confirm that the observed nanoscale shift is a fundamental, conserved feature of Alzheimer\u0026apos;s pathogenesis. Statistical testing was performed in R using ggpubr (v0.6.2), gghalves (v0.1.4) for violin geometries, ggpubr (v0.6.2) for statistical annotations.\u003c/p\u003e\n\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\n\u003cp\u003eAll statistical analyses were performed using R (v4.2.3), with significance set at \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05 (* \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05, ** \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.01, *** \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001, **** \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.0001). For the computational benchmark, graded score distributions were determined to be non-parametric and were compared using the Wilcoxon Rank-Sum (Mann-Whitney U) test with Benjamini-Hochberg FDR correction, using the ggpubr R package (v0.6.2). This was confirmed with a one-way ANOVA and Tukey\u0026apos;s HSD post-hoc test. For experimental datasets, statistical comparisons between genotypes (Healthy vs. AD) and regions (postsynaptic vs. perisynaptic) were performed using non-parametric Mann\u0026ndash;Whitney U tests (wilcox.test), along with Cliff\u0026rsquo;s delta effect sizes (effsize package). Outliers were removed using an IQR-based rule, and significance annotations were generated using ggpubr and custom R functions. All visualizations were generated using tidyverse (v2.0.0), gghalves (v0.1.4), ggpubr (v0.6.2), ggprism (v1.0.7), gridExtra (v2.3), broom(v1.0.10).\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eData and Code Availability\u003c/h2\u003e\n\u003cp\u003eThe complete source code for the EvoAge platform is publicly available on GitHub (https://github.com/the-ahuja-lab/EvoAge). The EvoAge chatbot web server is publicly accessible at https://evoage.ahujalab.iiitd.edu.in/. The full dataset generated and used in this study, including the final knowledge graph, is archived on Zenodo at https://doi.org/10.5281/zenodo.17711173. For enhanced reproducibility, a pre-configured Docker container is also available on Docker Hub at https://hub.docker.com/r/ahujalab/evoage-project.\u003c/p\u003e\n\u003ch2\u003eDeclaration of Interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e\n\u003ch2\u003eAcknowledgments\u003c/h2\u003e\n\u003cp\u003eThe authors thank the IT-HelpDesk team at IIIT-Delhi for their assistance with computational resources. We thank all the members of the Ahuja lab for their intellectual contributions at various stages of this project. The Ahuja lab is supported by the Ramalingaswami Re-entry Fellowship (BT/HRD/35/02/2006), and a research grant (BT/PR52020/AI/133/180/2024) by the Department of Biotechnology, Ministry of Science \u0026amp; Technology, Government of India, and an intramural Start-up grant from Indraprastha Institute of Information Technology-Delhi. DN acknowledges the senior research grant from DBT Wellcome Trust India Alliance (IA/S/23/2/507005), support from Centre for Brain Research through Funding for Aging Brain Research \u0026amp; Innovation Collaboration, and Core Research grant from Anusandhan National Research Foundation (CRG/2022/002726) and Intramural support from IISc.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe study was conceived and supervised by G.A. A.S. built the entire foundation of EvoAge, and A.S and An.S. designed the architecture of EvoAge Agent. A.K.S worked on frontend, P.S, and A.K, worked on data curation. A.S and A.G. worked on model training. S.K performed the mouse and human experiments and microscopy, under the supervision of D.N. Other experiments were carried out by S.R., and R.M. K.F. provided pateient derived stem cells lines. Statistical analysis, Evoage workflow finalization and testing was performed by V.G, S.S, S.C, S.K, Si.S, S.D, S.A, and R.S. Statistical analysis was guided by D.S. Figures were illustrated by G.A., and A.S., and built by A.S. Manuscript was written by A.S., and G.A. All authors read and approved the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eBorr\u0026aacute;s, C. The challenge of unlocking the biological secrets of aging. \u003cem\u003eFront. Aging\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 676573 (2021).\u003c/li\u003e\n\u003cli\u003eL\u0026oacute;pez-Ot\u0026iacute;n, C., Blasco, M. A., Partridge, L., Serrano, M. \u0026amp; Kroemer, G. Hallmarks of aging: An expanding universe. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e186\u003c/strong\u003e, 243\u0026ndash;278 (2023).\u003c/li\u003e\n\u003cli\u003eSarygina, E., Kliuchnikova, A., Tarbeeva, S., Ilgisonis, E. \u0026amp; Ponomarenko, E. Model organisms in aging research: Evolution of database annotation and ortholog discovery. \u003cem\u003eGenes (Basel)\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 8 (2024).\u003c/li\u003e\n\u003cli\u003eGuo, J. \u003cem\u003eet al.\u003c/em\u003e Aging and aging-related diseases: from molecular mechanisms to interventions and treatments. \u003cem\u003eSignal Transduct Target Ther\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 391 (2022).\u003c/li\u003e\n\u003cli\u003eAvelar, R. A. \u003cem\u003eet al.\u003c/em\u003e A multidimensional systems biology analysis of cellular senescence in aging and disease. \u003cem\u003eGenome Biol.\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 91 (2020).\u003c/li\u003e\n\u003cli\u003eBitto, A., Wang, A. M., Bennett, C. F. \u0026amp; Kaeberlein, M. Biochemical Genetic Pathways that Modulate Aging in Multiple Species. \u003cem\u003eCold Spring Harb. Perspect. Med.\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, (2015).\u003c/li\u003e\n\u003cli\u003eHuang, K. \u003cem\u003eet al.\u003c/em\u003e AgeAnno: a knowledgebase of single-cell annotation of aging in human. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, D805\u0026ndash;D815 (2023).\u003c/li\u003e\n\u003cli\u003eHuang, K. \u003cem\u003eet al.\u003c/em\u003e AgeAnnoMO: a knowledgebase of multi-omics annotation for animal aging. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, D822\u0026ndash;D834 (2024).\u003c/li\u003e\n\u003cli\u003eArora, S. \u003cem\u003eet al.\u003c/em\u003e Discovering geroprotectors through the explainable artificial intelligence-based platform AgeXtend. \u003cem\u003eNat. Aging\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 144\u0026ndash;161 (2025).\u003c/li\u003e\n\u003cli\u003eAging Atlas Consortium. Aging Atlas: a multi-omics database for aging biology. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, D825\u0026ndash;D830 (2021).\u003c/li\u003e\n\u003cli\u003eWu, Z. \u003cem\u003eet al.\u003c/em\u003e HALD, a human aging and longevity knowledge graph for precision gerontology and geroscience analyses. \u003cem\u003eSci. Data\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 851 (2023).\u003c/li\u003e\n\u003cli\u003eSturm, G. \u003cem\u003eet al.\u003c/em\u003e A multi-omics longitudinal aging dataset in primary human fibroblasts with mitochondrial perturbations. \u003cem\u003eSci. Data\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 751 (2022).\u003c/li\u003e\n\u003cli\u003eHimmelstein, D. S. \u003cem\u003eet al.\u003c/em\u003e Systematic integration of biomedical knowledge prioritizes drugs for repurposing. \u003cem\u003eElife\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, (2017).\u003c/li\u003e\n\u003cli\u003eChandak, P., Huang, K. \u0026amp; Zitnik, M. Building a knowledge graph to enable precision medicine. \u003cem\u003eSci. Data\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 67 (2023).\u003c/li\u003e\n\u003cli\u003eSantos, A. \u003cem\u003eet al.\u003c/em\u003e A knowledge graph to interpret clinical proteomics data. \u003cem\u003eNat. Biotechnol.\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, 692\u0026ndash;702 (2022).\u003c/li\u003e\n\u003cli\u003eBarardo, D. \u003cem\u003eet al.\u003c/em\u003e The DrugAge database of aging-related drugs. \u003cem\u003eAging Cell\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 594\u0026ndash;597 (2017).\u003c/li\u003e\n\u003cli\u003eWuttke, D. \u003cem\u003eet al.\u003c/em\u003e Dissecting the gene network of dietary restriction to identify evolutionarily conserved pathways and new functional genes. \u003cem\u003ePLoS Genet.\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, e1002834 (2012).\u003c/li\u003e\n\u003cli\u003eAnh-Hoang, D., Tran, V. \u0026amp; Nguyen, L.-M. Survey and analysis of hallucinations in large language models: attribution to prompting strategies or model behavior. \u003cem\u003eFront. Artif. Intell.\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 1622292 (2025).\u003c/li\u003e\n\u003cli\u003eB\u0026eacute;lisle-Pipon, J.-C. Why we need to be careful with LLMs in medicine. \u003cem\u003eFront. Med. (Lausanne)\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1495582 (2024).\u003c/li\u003e\n\u003cli\u003eFeng, Y. \u003cem\u003eet al.\u003c/em\u003e Knowledge graph-based thought: a knowledge graph-enhanced LLM framework for pan-cancer question answering. \u003cem\u003eGigascience\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, (2025).\u003c/li\u003e\n\u003cli\u003eGao, Y. \u003cem\u003eet al.\u003c/em\u003e Leveraging medical knowledge graphs into large language models for diagnosis prediction: Design and application study. \u003cem\u003eJMIR AI\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, e58670 (2025).\u003c/li\u003e\n\u003cli\u003eCherry, J. M. \u003cem\u003eet al.\u003c/em\u003e SGD: Saccharomyces Genome Database. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 73\u0026ndash;79 (1998).\u003c/li\u003e\n\u003cli\u003eKim, H. \u003cem\u003eet al.\u003c/em\u003e YeastNet v3: a public database of data-specific and integrated functional gene networks for Saccharomyces cerevisiae. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, D731\u0026ndash;6 (2014).\u003c/li\u003e\n\u003cli\u003e\u0026Ouml;zt\u0026uuml;rk-\u0026Ccedil;olak, A. \u003cem\u003eet al.\u003c/em\u003e FlyBase: updates to the Drosophila genes and genomes database. \u003cem\u003eGenetics\u003c/em\u003e \u003cstrong\u003e227\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eSternberg, P. W. \u003cem\u003eet al.\u003c/em\u003e WormBase 2024: status and transitioning to Alliance infrastructure. \u003cem\u003eGenetics\u003c/em\u003e \u003cstrong\u003e227\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eBradford, Y. \u003cem\u003eet al.\u003c/em\u003e ZFIN: enhancements and updates to the Zebrafish Model Organism Database. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e39\u003c/strong\u003e, D822\u0026ndash;9 (2011).\u003c/li\u003e\n\u003cli\u003eBaldarelli, R. M. \u003cem\u003eet al.\u003c/em\u003e Mouse Genome Informatics: an integrated knowledgebase system for the laboratory mouse. \u003cem\u003eGenetics\u003c/em\u003e \u003cstrong\u003e227\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eBordes, A., Usunier, N., Garc\u0026iacute;a-Dur\u0026aacute;n, A., Weston, J. \u0026amp; Yakhnenko, O. Translating embeddings for modeling multi-relational data. \u003cem\u003eNeural Inf Process Syst\u003c/em\u003e 2787\u0026ndash;2795 (2013).\u003c/li\u003e\n\u003cli\u003eSun, Z., Deng, Z.-H., Nie, J.-Y. \u0026amp; Tang, J. RotatE: Knowledge graph embedding by relational rotation in complex space. \u003cem\u003earXiv [cs.LG]\u003c/em\u003e (2019).\u003c/li\u003e\n\u003cli\u003eKazemi, S. M. \u0026amp; Poole, D. SimplE embedding for link prediction in knowledge graphs. \u003cem\u003earXiv [stat.ML]\u003c/em\u003e (2018).\u003c/li\u003e\n\u003cli\u003eYang, B., Yih, W.-T., He, X., Gao, J. \u0026amp; Deng, L. Embedding entities and relations for learning and inference in knowledge bases. \u003cem\u003earXiv [cs.CL]\u003c/em\u003e (2014).\u003c/li\u003e\n\u003cli\u003eTran, H. N. \u0026amp; Takasu, A. Analyzing knowledge graph embedding methods from a multi-embedding interaction perspective. \u003cem\u003earXiv [cs.LG]\u003c/em\u003e (2019).\u003c/li\u003e\n\u003cli\u003eNickel, M., Tresp, V. \u0026amp; Kriegel, H. A Three-Way model for collective learning on multi-relational data. \u003cem\u003eICML\u003c/em\u003e 809\u0026ndash;816 (2011).\u003c/li\u003e\n\u003cli\u003eTrouillon, T., Welbl, J., Riedel, S., Gaussier, \u0026Eacute;. \u0026amp; Bouchard, G. Complex embeddings for simple link prediction. \u003cem\u003eICML\u003c/em\u003e \u003cstrong\u003eabs/1606.06357\u003c/strong\u003e, 2071\u0026ndash;2080 (2016).\u003c/li\u003e\n\u003cli\u003eStothers, J. A. M. \u0026amp; Nguyen, A. Can Neo4j replace PostgreSQL in healthcare? \u003cem\u003eAMIA Summits Transl. Sci. Proc.\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, 646\u0026ndash;653 (2020).\u003c/li\u003e\n\u003cli\u003eZheng, D. \u003cem\u003eet al.\u003c/em\u003e DGL-KE: Training Knowledge Graph Embeddings at Scale. in \u003cem\u003eProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval\u003c/em\u003e (ACM, New York, NY, USA, 2020). doi:10.1145/3397271.3401172.\u003c/li\u003e\n\u003cli\u003eZhu, A., Dugan, L., Hwang, A. \u0026amp; Callison-Burch, C. Kani: A lightweight and highly hackable framework for building language model applications. in \u003cem\u003eProceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023)\u003c/em\u003e (eds. Tan, L., Milajevs, D., Chauhan, G., Gwinnup, J. \u0026amp; Rippeth, E.) 65\u0026ndash;77 (Empirical Methods in Natural Language Processing, Stroudsburg, PA, USA, 2023).\u003c/li\u003e\n\u003cli\u003eGemma Team \u003cem\u003eet al.\u003c/em\u003e Gemma: Open models based on Gemini research and technology. \u003cem\u003earXiv [cs.CL]\u003c/em\u003e (2024).\u003c/li\u003e\n\u003cli\u003eGrattafiori, A. \u003cem\u003eet al.\u003c/em\u003e The Llama 3 herd of models. \u003cem\u003earXiv [cs.AI]\u003c/em\u003e (2024).\u003c/li\u003e\n\u003cli\u003eQwen \u003cem\u003eet al.\u003c/em\u003e Qwen2.5 Technical Report. \u003cem\u003earXiv [cs.CL]\u003c/em\u003e (2024).\u003c/li\u003e\n\u003cli\u003eLabrak, Y. \u003cem\u003eet al.\u003c/em\u003e BioMistral: A collection of open-source pretrained large Language Models for medical domains. \u003cem\u003eAnnu Meet Assoc Comput Linguistics\u003c/em\u003e 5848\u0026ndash;5864 (2024).\u003c/li\u003e\n\u003cli\u003eKirkwood, T. B. Evolution of ageing. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e270\u003c/strong\u003e, 301\u0026ndash;304 (1977).\u003c/li\u003e\n\u003cli\u003eMitchell, S. J., Scheibye-Knudsen, M., Longo, D. L. \u0026amp; de Cabo, R. Animal models of aging research: implications for human aging and age-related diseases. \u003cem\u003eAnnu. Rev. Anim. Biosci.\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 283\u0026ndash;303 (2015).\u003c/li\u003e\n\u003cli\u003eKenyon, C. J. The genetics of ageing. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e464\u003c/strong\u003e, 504\u0026ndash;512 (2010).\u003c/li\u003e\n\u003cli\u003eL\u0026oacute;pez-Ot\u0026iacute;n, C., Blasco, M. A., Partridge, L., Serrano, M. \u0026amp; Kroemer, G. The hallmarks of aging. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e153\u003c/strong\u003e, 1194\u0026ndash;1217 (2013).\u003c/li\u003e\n\u003cli\u003eGautam, V. \u003cem\u003eet al.\u003c/em\u003e scCamAge: A context-aware prediction engine for cellular age, aging-associated bioactivities, and morphometrics. \u003cem\u003eCell Rep.\u003c/em\u003e \u003cstrong\u003e44\u003c/strong\u003e, 115270 (2025).\u003c/li\u003e\n\u003cli\u003eDuari, S., Gautam, V. \u0026amp; Ahuja, G. Protocol for cellular age prediction in yeast and human single cells using transfer learning. \u003cem\u003eSTAR Protoc.\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 104023 (2025).\u003c/li\u003e\n\u003cli\u003eCampisi, J. \u003cem\u003eet al.\u003c/em\u003e From discoveries in ageing research to therapeutics for healthy ageing. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e571\u003c/strong\u003e, 183\u0026ndash;192 (2019).\u003c/li\u003e\n\u003cli\u003eBarabasi, A. L. \u0026amp; Albert, R. Emergence of scaling in random networks. \u003cem\u003eScience\u003c/em\u003e \u003cstrong\u003e286\u003c/strong\u003e, 509\u0026ndash;512 (1999).\u003c/li\u003e\n\u003cli\u003eIdeker, T. \u0026amp; Krogan, N. J. Differential network biology. \u003cem\u003eMolecular Systems Biology\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, (2012).\u003c/li\u003e\n\u003cli\u003eRajabi, E. \u0026amp; Etminani, K. Knowledge-graph-based explainable AI: A systematic review. \u003cem\u003eJ. Inf. Sci.\u003c/em\u003e \u003cstrong\u003e50\u003c/strong\u003e, 1019\u0026ndash;1029 (2024).\u003c/li\u003e\n\u003cli\u003eZhang, H., Deng, H., Ou, J. \u0026amp; Feng, C. Mitigating spatial hallucination in large language models for path planning via prompt engineering. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 8881 (2025).\u003c/li\u003e\n\u003cli\u003eAnonymous. Grounding LLM reasoning with knowledge graphs. Preprint at https://openreview.net/pdf/f2d669a35707223183429f92ff75668b839cee34.pdf (2024).\u003c/li\u003e\n\u003cli\u003eAmayuelas, A., Sain, J., Kaur, S. \u0026amp; Smiley, C. Grounding LLM reasoning with knowledge graphs. \u003cem\u003earXiv [cs.CL]\u003c/em\u003e (2025).\u003c/li\u003e\n\u003cli\u003eGabald\u0026oacute;n, T. \u0026amp; Koonin, E. V. Functional and evolutionary implications of gene orthology. \u003cem\u003eNat. Rev. Genet.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 360\u0026ndash;366 (2013).\u003c/li\u003e\n\u003cli\u003eGasperini, M. \u003cem\u003eet al.\u003c/em\u003e A genome-wide framework for mapping gene regulation via cellular genetic screens. \u003cem\u003eCell\u003c/em\u003e \u003cstrong\u003e176\u003c/strong\u003e, 377\u0026ndash;390.e19 (2019).\u003c/li\u003e\n\u003cli\u003eMarbach, D. \u003cem\u003eet al.\u003c/em\u003e Wisdom of crowds for robust gene network inference. \u003cem\u003eNat. Methods\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 796\u0026ndash;804 (2012).\u003c/li\u003e\n\u003cli\u003eMoffitt, J. R. \u003cem\u003eet al.\u003c/em\u003e Molecular, spatial, and functional single-cell profiling of the hypothalamic preoptic region. \u003cem\u003eScience\u003c/em\u003e \u003cstrong\u003e362\u003c/strong\u003e, eaau5324 (2018).\u003c/li\u003e\n\u003cli\u003eCraig, T. \u003cem\u003eet al.\u003c/em\u003e The Digital Ageing Atlas: integrating the diversity of age-related changes into a unified resource. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e43\u003c/strong\u003e, D873\u0026ndash;8 (2015).\u003c/li\u003e\n\u003cli\u003ede Magalh\u0026atilde;es, J. P. \u003cem\u003eet al.\u003c/em\u003e Human Ageing Genomic Resources: updates on key databases in ageing research. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, D900\u0026ndash;D908 (2024).\u003c/li\u003e\n\u003cli\u003eBucaciuc Mracica, T. \u003cem\u003eet al.\u003c/em\u003e MetaboAge DB: a repository of known ageing-related changes in the human metabolome. \u003cem\u003eBiogerontology\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 763\u0026ndash;771 (2020).\u003c/li\u003e\n\u003cli\u003eSzklarczyk, D. \u003cem\u003eet al.\u003c/em\u003e The STRING database in 2023: protein-protein association networks and functional enrichment analyses for any sequenced genome of interest. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, D638\u0026ndash;D646 (2023).\u003c/li\u003e\n\u003cli\u003eSzklarczyk, D. \u003cem\u003eet al.\u003c/em\u003e STITCH 5: augmenting protein-chemical interaction networks with tissue and affinity data. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e44\u003c/strong\u003e, D380\u0026ndash;4 (2016).\u003c/li\u003e\n\u003cli\u003eOughtred, R. \u003cem\u003eet al.\u003c/em\u003e The BioGRID database: A comprehensive biomedical resource of curated protein, genetic, and chemical interactions. \u003cem\u003eProtein Sci.\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 187\u0026ndash;200 (2021).\u003c/li\u003e\n\u003cli\u003eZdrazil, B. \u003cem\u003eet al.\u003c/em\u003e The ChEMBL Database in 2023: a drug discovery platform spanning multiple bioactivity data types and time periods. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, D1180\u0026ndash;D1192 (2024).\u003c/li\u003e\n\u003cli\u003eLiu, T., Lin, Y., Wen, X., Jorissen, R. N. \u0026amp; Gilson, M. K. BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, D198\u0026ndash;201 (2007).\u003c/li\u003e\n\u003cli\u003eHe, S. \u0026amp; Zhao, X. Drug repositioning model based on knowledge graph embedding. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 10298 (2025).\u003c/li\u003e\n\u003cli\u003eSimonis, N. \u003cem\u003eet al.\u003c/em\u003e Empirically controlled mapping of the Caenorhabditis elegans protein-protein interactome network. \u003cem\u003eNat. Methods\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 47\u0026ndash;54 (2009).\u003c/li\u003e\n\u003cli\u003eMessina, A. \u003cem\u003eet al.\u003c/em\u003e BioGrakn: A knowledge graph-based semantic database for biomedical sciences. in \u003cem\u003eAdvances in Intelligent Systems and Computing\u003c/em\u003e 299\u0026ndash;309 (Springer International Publishing, Cham, 2018).\u003c/li\u003e\n\u003cli\u003eDoğan, T. \u003cem\u003eet al.\u003c/em\u003e CROssBAR: comprehensive resource of biomedical relations with knowledge graph representations. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, e96 (2021).\u003c/li\u003e\n\u003cli\u003ePutman, T. E. \u003cem\u003eet al.\u003c/em\u003e The Monarch Initiative in 2024: an analytic platform integrating phenotypes, genes and diseases across species. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, D938\u0026ndash;D949 (2024).\u003c/li\u003e\n\u003cli\u003eDiamant, I., Clarke, D. J. B., Evangelista, J. E., Lingam, N. \u0026amp; Ma\u0026rsquo;ayan, A. Harmonizome 3.0: integrated knowledge about genes and proteins from diverse multi-omics resources. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e53\u003c/strong\u003e, D1016\u0026ndash;D1028 (2025).\u003c/li\u003e\n\u003cli\u003eZhou, C. \u003cem\u003eet al.\u003c/em\u003e TarKG: a comprehensive biomedical knowledge graph for target discovery. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eYao, Q., Chen, Z., Cao, Y. \u0026amp; Hu, H. Enhancing drug-target interaction prediction with graph representation learning and knowledge-based regularization. \u003cem\u003eFront. Bioinform.\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 1649337 (2025).\u003c/li\u003e\n\u003cli\u003eGao, Z., Ding, P. \u0026amp; Xu, R. KG-Predict: A knowledge graph computational framework for drug repurposing. \u003cem\u003eJ. Biomed. Inform.\u003c/em\u003e \u003cstrong\u003e132\u003c/strong\u003e, 104133 (2022).\u003c/li\u003e\n\u003cli\u003eZheng, S. \u003cem\u003eet al.\u003c/em\u003e PharmKG: a dedicated knowledge graph benchmark for bomedical data mining. \u003cem\u003eBrief. Bioinform.\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, (2021).\u003c/li\u003e\n\u003cli\u003eSu, C. \u003cem\u003eet al.\u003c/em\u003e Biomedical discovery through the integrative biomedical knowledge hub (iBKH). \u003cem\u003eiScience\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 106460 (2023).\u003c/li\u003e\n\u003cli\u003eZhou, Y. \u003cem\u003eet al.\u003c/em\u003e TTD: Therapeutic Target Database describing target druggability information. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, D1465\u0026ndash;D1477 (2024).\u003c/li\u003e\n\u003cli\u003eNCBI Resource Coordinators. Database resources of the National Center for Biotechnology Information. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e46\u003c/strong\u003e, D8\u0026ndash;D13 (2018).\u003c/li\u003e\n\u003cli\u003eUniProt Consortium. UniProt: The universal protein knowledgebase in 2023. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, D523\u0026ndash;D531 (2023).\u003c/li\u003e\n\u003cli\u003eKim, S. \u003cem\u003eet al.\u003c/em\u003e PubChem in 2021: new data content and improved web interfaces. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, D1388\u0026ndash;D1395 (2021).\u003c/li\u003e\n\u003cli\u003eKnox, C. \u003cem\u003eet al.\u003c/em\u003e DrugBank 6.0: The DrugBank knowledgebase for 2024. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, D1265\u0026ndash;D1275 (2024).\u003c/li\u003e\n\u003cli\u003eSchriml, L. M. \u003cem\u003eet al.\u003c/em\u003e Human Disease Ontology 2018 update: classification, content and workflow expansion. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e47\u003c/strong\u003e, D955\u0026ndash;D962 (2019).\u003c/li\u003e\n\u003cli\u003eVasilevsky, N. A. \u003cem\u003eet al.\u003c/em\u003e Mondo: Integrating disease terminology across communities. \u003cem\u003eGenetics\u003c/em\u003e (2025) doi:10.1093/genetics/iyaf215.\u003c/li\u003e\n\u003cli\u003eMungall, C. J., Torniai, C., Gkoutos, G. V., Lewis, S. E. \u0026amp; Haendel, M. A. Uberon, an integrative multi-species anatomy ontology. \u003cem\u003eGenome Biol.\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, R5 (2012).\u003c/li\u003e\n\u003cli\u003eKanehisa, M., Furumichi, M., Sato, Y., Ishiguro-Watanabe, M. \u0026amp; Tanabe, M. KEGG: integrating viruses and cellular organisms. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e49\u003c/strong\u003e, D545\u0026ndash;D551 (2021).\u003c/li\u003e\n\u003cli\u003eGillespie, M. \u003cem\u003eet al.\u003c/em\u003e The reactome pathway knowledgebase 2022. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e50\u003c/strong\u003e, D687\u0026ndash;D692 (2022).\u003c/li\u003e\n\u003cli\u003eBinns, D. \u003cem\u003eet al.\u003c/em\u003e QuickGO: a web-based tool for Gene Ontology searching. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 3045\u0026ndash;3046 (2009).\u003c/li\u003e\n\u003cli\u003eK\u0026ouml;hler, S. \u003cem\u003eet al.\u003c/em\u003e Expansion of the Human Phenotype Ontology (HPO) knowledge base and resources. \u003cem\u003eNucleic Acids Res.\u003c/em\u003e \u003cstrong\u003e47\u003c/strong\u003e, D1018\u0026ndash;D1027 (2019).\u003c/li\u003e\n\u003cli\u003eO\u0026rsquo;Boyle, N. M. \u003cem\u003eet al.\u003c/em\u003e Open Babel: An open chemical toolbox. \u003cem\u003eJ. Cheminform.\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 33 (2011).\u003c/li\u003e\n\u003cli\u003eKolberg, L., Raudvere, U., Kuzmin, I., Vilo, J. \u0026amp; Peterson, H. gprofiler2 -- an R package for gene list functional enrichment analysis and namespace conversion toolset g:Profiler. \u003cem\u003eF1000Res.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 709 (2020).\u003c/li\u003e\n\u003cli\u003eWebber, J. A programmatic introduction to Neo4j. in \u003cem\u003eProceedings of the 3rd annual conference on Systems, programming, and applications: software for humanity\u003c/em\u003e (ACM, New York, NY, USA, 2012). doi:10.1145/2384716.2384777.\u003c/li\u003e\n\u003cli\u003eWang, M. \u003cem\u003eet al.\u003c/em\u003e Deep Graph Library: A graph-centric, highly-performant package for graph neural networks. \u003cem\u003earXiv [cs.LG]\u003c/em\u003e (2019).\u003c/li\u003e\n\u003cli\u003eOpenAI. https://openai.com.\u003c/li\u003e\n\u003cli\u003eTouvron, H. \u003cem\u003eet al.\u003c/em\u003e LLaMA: Open and efficient foundation language models. \u003cem\u003earXiv [cs.CL]\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003eKwon, W. \u003cem\u003eet al.\u003c/em\u003e Efficient memory management for large language model serving with PagedAttention. \u003cem\u003earXiv [cs.LG]\u003c/em\u003e (2023).\u003c/li\u003e\n\u003cli\u003eWolf, T. \u003cem\u003eet al.\u003c/em\u003e HuggingFace\u0026rsquo;s transformers: State-of-the-art natural language processing. \u003cem\u003earXiv [cs.CL]\u003c/em\u003e (2019).\u003c/li\u003e\n\u003cli\u003eKedia, S. \u003cem\u003eet al.\u003c/em\u003e Alteration in synaptic nanoscale organization dictates amyloidogenic processing in Alzheimer\u0026rsquo;s disease. \u003cem\u003eiScience\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 101924 (2021).\u003c/li\u003e\n\u003cli\u003eTubsuwan, A. \u003cem\u003eet al.\u003c/em\u003e Generation of induced pluripotent stem cells (iPSCs) from an Alzheimer\u0026rsquo;s disease patient carrying a L150P mutation in PSEN-1. \u003cem\u003eStem Cell Res.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 110\u0026ndash;112 (2016).\u003c/li\u003e\n\u003cli\u003ePoon, A. \u003cem\u003eet al.\u003c/em\u003e Generation of a gene-corrected isogenic control hiPSC line derived from a familial Alzheimer\u0026rsquo;s disease patient carrying a L150P mutation in presenilin 1. \u003cem\u003eStem Cell Res.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 466\u0026ndash;469 (2016).\u003c/li\u003e\n\u003cli\u003eRamakrishna, S. \u003cem\u003eet al.\u003c/em\u003e Distinct calcium sources regulate temporal profiles of NMDAR and mGluR-mediated protein synthesis. \u003cem\u003eLife Sci. Alliance\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, e202402594 (2024).\u003c/li\u003e\n\u003cli\u003eOchalek, A. \u003cem\u003eet al.\u003c/em\u003e Neurons derived from sporadic Alzheimer\u0026rsquo;s disease iPSCs reveal elevated TAU hyperphosphorylation, increased amyloid levels, and GSK3B activation. \u003cem\u003eAlzheimers. Res. Ther.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 90 (2017).\u003c/li\u003e\n\u003cli\u003eKedia, S. \u003cem\u003eet al.\u003c/em\u003e Real-time nanoscale organization of amyloid precursor protein. \u003cem\u003eNanoscale\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 8200\u0026ndash;8215 (2020).\u003c/li\u003e\n\u003cli\u003eKedia, S., Ramanan, N. \u0026amp; Nair, D. Quantifying molecular aggregation by super resolution microscopy within an excitatory synapse from mouse hippocampal neurons. \u003cem\u003eSTAR Protoc.\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 100470 (2021).\u003c/li\u003e\n\u003cli\u003eKedia, S. \u003cem\u003eet al.\u003c/em\u003e Nanoscale organization of Nicastrin, the substrate receptor of the \u0026gamma;-secretase complex, as independent molecular domains. \u003cem\u003eMol. Brain\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 158 (2021).\u003c/li\u003e\n\u003cli\u003eMetaMorph Microscopy Automation and Image Analysis Software. https://www.moleculardevices.com/products/cellular-imaging-systems/high-content-analysis/metamorph-microscopy.\u003c/li\u003e\n\u003cli\u003eNair, D. \u003cem\u003eet al.\u003c/em\u003e Super-resolution imaging reveals that AMPA receptors inside synapses are dynamically organized in nanodomains regulated by PSD95. \u003cem\u003eJ. Neurosci.\u003c/em\u003e \u003cstrong\u003e33\u003c/strong\u003e, 13204\u0026ndash;13224 (2013). \u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Knowledge Graphs, Evolution, Graph Modeling, Aging, Synapse, Alzheimer","lastPublishedDoi":"10.21203/rs.3.rs-9060414/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9060414/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Aging research has been advanced largely through the use of model organisms, where short lifespans and genetic tractability enable the systematic discovery of molecular pathways influencing longevity and age-related decline. However, knowledge about aging remains fragmented across species-specific repositories and domain-focused databases, limiting our ability to identify evolutionarily conserved mechanisms and translate findings to human biology. To address this gap, we developed EvoAge, a unified, multi-species knowledge graph that integrates aging-specific and general biomedical resources into a systems-level framework. EvoAge harmonizes 48 public datasets into a graph comprising 1.04 billion triples across six key species. A human-centric orthology framework reconciles more than 80,000 gene entries, expanding accessible organism-level aging knowledge by up to 1,700-fold compared with existing resources. To operationalize the graph for biological reasoning, we optimized knowledge graph embedding models and deployed a large language model (LLM)-assisted agentic interface that supports natural-language querying, link prediction, and hypothesis testing. In internal benchmarking using recent pre-print aging literature, EvoAge significantly outperformed state-of-the-art LLMs in distinguishing biologically plausible from implausible hypotheses. Importantly, EvoAge recommended a previously unrecognized Alzheimer’s disease (AD) mechanism involving nanoscale redistribution of BACE1 within synaptic compartments. We experimentally validated this EvoAge-supported prediction using patient-derived iPSCs carrying a familial PSEN1 mutation, demonstrating disease-associated remodeling of β-secretase, defined by altered localization, nanoscale clustering, and compartment-specific enrichment. We further confirmed the predicted evolutionary conservation of this BACE1–pathology relationship in additional AD systems, including transgenic mice and postmortem human brain tissue.","manuscriptTitle":"Cross-Species Aging Knowledge Integration into Agentic AI Platform Uncovers Conserved Mechanisms","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-13 15:18:09","doi":"10.21203/rs.3.rs-9060414/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-communications","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"NCOMMS","sideBox":"Learn more about [Nature Communications](http://www.nature.com/ncomms/)","snPcode":"","submissionUrl":"https://mts-ncomms.nature.com/","title":"Nature Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Communications","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"e638bb33-adc4-45bd-94ba-9ee4e682bc84","owner":[],"postedDate":"March 13th, 2026","published":true,"recentEditorialEvents":[{"type":"decision","content":"revise","date":"2026-05-06T07:12:40+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"This content is not available.","date":"2026-05-03T05:05:18+00:00","index":5,"fulltext":"This content is not available."},{"type":"editorInvitedReview","content":"This content is not available.","date":"2026-05-02T19:52:53+00:00","index":4,"fulltext":"This content is not available."}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":64397234,"name":"Biological sciences/Computational biology and bioinformatics/Computational platforms and environments"},{"id":64397235,"name":"Biological sciences/Computational biology and bioinformatics/Computational models"},{"id":64397236,"name":"Biological sciences/Computational biology and bioinformatics/Data integration"},{"id":64397237,"name":"Biological sciences/Computational biology and bioinformatics/Data mining"}],"tags":[],"updatedAt":"2026-05-06T07:15:45+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-13 15:18:09","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9060414","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9060414","identity":"rs-9060414","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}