Source-linked AI summary
A Trust-Network-Based Federated Learning Framework for Multi-Center Aging Clock Prediction
Chunxu Zhang, Bo Li, Wenliang Wang, Yang Liu, Di Jiang, Yuan Huang, Yo-ichi Nabeshima, Akinori Yamamura, Bo Yang, Qiang Yang
TL;DR
TNFL addresses multi-center aging-clock learning when molecular data cannot be centrally shared and centers have limited, heterogeneous, directionally trusted data. It propagates an age-aware mixture-of-experts model through directed trust relations with generative replay, identifying interpretable protein effects and higher-order organization while maintaining effective prediction and robustness. The paper concludes that TNFL captures coherent aging-related molecular subnetworks beyond isolated pairwise associations, although its current framework assumes a pre-specified trust network and chronological-age supervision and analyzes each omics modality separately.
Problem
Multi-center aging-clock modeling requires joint learning from distributed molecular data under limited local samples, constrained directional trust, interpretability demands, and heterogeneous updates.
Method
TNFL progressively propagates an age-aware mixture-of-experts model along directed pairwise trust relations without central aggregation, using generative replay to retain information across center updates.
Results
TNFL achieves effective prediction with limited local data, supports interpretable age-dependent protein patterns, and identifies pairwise interactions that repeatedly organize into higher-order aging-related subnetworks.
Takeaways & Limitations
TNFL provides a unified framework for multi-center aging-clock prediction and for examining molecular relationships across individual, pairwise, and higher interaction orders.
Takeaways & Limitations
The framework assumes a pre-specified trust network and chronological-age supervision, while analyzing each omics modality separately.
Abstract
from arXiv · showhide
Aging clocks quantify biological aging and help characterize individual health status. What protein interactions are important for accurate aging clocks, and are they zeroth-order or higher-order? Addressing these questions requires learning from large molecular datasets distributed across medical centers, where privacy constraints prevent centralized data sharing. Federated learning offers a natural solution but faces four challenges in this setting: limited local sample sizes, sparse and directional inter-center trust, the need to retain discriminative age prediction while supporting interpretation, and model drift and forgetting under heterogeneous cross-center data. We propose TNFL, a trust-network-based federated learning framework that progressively propagates models along directed pairwise trust relations without centralized aggregation. TNFL combines an age-aware mixture-of-experts model with generative replay to preserve previously learned information and reduce forgetting and drift. Experiments across multiple molecular datasets show that TNFL enables effective aging-clock prediction with limited local data, provides interpretable age-dependent prediction patterns, and maintains stable performance across interaction orders. To investigate the biological questions, we analyze TNFL-identified pairwise protein interactions and their higher-order organization through functional and network analyses. The identified interactions repeatedly form coordinated higher-order subnetworks spanning multiple aging-related biological systems, with several proteins recurring across subnetworks. These findings suggest that TNFL captures molecular relationships beyond isolated pairwise associations and reveals coherent higher-order biological organization associated with aging.
1 Results
TNFL identifies age-dependent protein effects, coherent pairwise relationships, and recurring higher-order subnetworks while delivering effective and robust multi-center aging-clock prediction. Its performance remains compatible with diverse backbones and benefits from generative replay under heterogeneous sequential updates.
- Individual Protein Relevance to Aging-Clock Prediction: AgeMoE routing varies systematically with age, with Experts 1 and 2 decreasing and Experts 3 and 4 increasing in routing weight.This routing behavior provides an interpretable view of age-dependent prediction patterns.
- Individual Protein Relevance to Aging-Clock Prediction: The top 100 routing-influential proteins form 20 significantly enriched ontology clusters spanning extracellular-matrix integrity, deregulated nutrient sensing, and neurobiological themes.Several proteins also overlap with GenAge and the SASP Atlas, while most are secreted or ECM-associated.
- Pairwise Protein Interactions: The top-10 protein pairs show endocrine, metabolic, neuroinflammatory, neuroendocrine, cardiac-stress, inflammatory, and extracellular-matrix functional convergence, including both STRING-supported and non-linked pairs.These coherent pairwise relationships motivate testing whether the interactions organize into higher-order subnetworks.
- Higher-Order Organization of Protein Interactions: Identified interactions repeatedly organize into higher-order subnetworks spanning neural, vascular, endocrine, metabolic, inflammatory, and stress-related systems, with recurrent proteins acting as shared signals across subnetworks.The recurring proteins include GFAP, NEFL, GDF15, GIP, ELN, LTBP2, and CXCL17, and these patterns emerged without prior interaction or pathway knowledge during training.
- Predictive Performance: TNFL lowers MAE versus local training, matches or exceeds FedAvg across datasets, and remains effective across heterogeneous centers and predictive backbones.On UKB-Center, average MAE reductions versus local training are 0.69 years for TNFL and 0.70 years for FedAvg; TNFL further improves over FedAvg by 0.12 years on UKB-Imbalance and 0.34 years on GEO-Methylation.
- Robustness Evaluation: TNFL maintains robust sequential behavior: earlier-client performance is maintained or improved, later updates usually do not degrade prior performance, and overall behavior remains stable across interaction orders and trust-network conditions.Removing generative replay reduces cross-client generalization, especially under stronger distributional heterogeneity.
2 Discussion
TNFL combines effective multi-center aging-clock prediction with biological interpretation across individual proteins, pairwise interactions, and higher-order subnetworks. Its findings suggest coherent aging-related organization, while the framework remains bounded by assumptions about trust, targets, modalities, and deployment settings.
- Biological interpretation: TNFL identifies individual proteins associated with aging-clock prediction and supports their relevance through aging-related biological processes and external annotations.
- Biological interpretation: Identified pairwise protein interactions repeatedly organize into higher-order subnetworks spanning neural, vascular, endocrine, metabolic, inflammatory, and stress-related systems.
- Computational performance: The framework achieves effective aging-clock prediction with limited local data and remains compatible with different predictive backbones.
- Computational performance: Generative replay supports cross-center generalization, while TNFL preserves information from previously visited centers and maintains stable behavior across interaction orders.
- Limitations and future directions: Current collaboration assumes a pre-specified trust network and feasible propagation sequences, motivating adaptive routing under dynamic trust relations.
- Limitations and future directions: Chronological-age supervision may favor strongly age-associated features, while separate modality analyses capture only part of molecular complexity.
3 Methods
TNFL performs aging-clock learning through progressive propagation over directed, pairwise trust relations rather than centralized aggregation. It combines an interpretable AgeMoE predictor with generative replay to retain information across heterogeneous centers.
- TNFL formulates multi-center aging-clock prediction as supervised regression from high-dimensional molecular features to chronological age, with datasets remaining locally held.
- TNFL organizes collaboration through authorized, directed model transmissions and feasible trust-constrained sequences instead of a globally trusted coordinator.
- At each client, filtered pseudo-samples are combined with local data to update the aging clock, while the generative model is updated using local data and propagated onward.
- The AgeMoE predictor exposes age-dependent routing behavior for interpretation while modeling heterogeneous aging patterns through mixture-of-experts representations.
- Generative replay synthesizes information from previously visited centers so the aging clock can retain acquired knowledge during continued adaptation to heterogeneous local distributions.
1. Pseudo-sample generation and prediction-consistency filtering
TNFL generates age-conditioned pseudo-samples representing previously visited clients and filters them by consistency with the received aging clock before replay.
- The received generative model produces pseudo-samples conditioned on target ages and age-bin identifiers to represent information from previously visited clients.
- The received aging clock predicts each pseudo-sample’s age, and samples are retained only when predictions are sufficiently consistent with their conditioning labels.
- Prediction-consistency filtering excludes pseudo-samples with inconsistent age predictions and improves the reliability of generated supervision for subsequent training.
2. Aging clock update
The aging clock is updated using both the current center’s local data and filtered replay samples, preserving information from previously encountered client distributions.
- TNFL combines filtered pseudo-samples with the current center’s data when updating the aging-clock parameters.
- The combined dataset supports learning from the current center while retaining information acquired from previously visited client distributions.
3. Generative model update
TNFL jointly propagates an aging clock and generative model through trust-constrained sequential updates. The protocol defines message-specific security obligations and outputs the final aging-clock model after the sequence.
- Generative model update: The generative model is updated with the current center’s data and then passed onward together with the aging clock.
- Generative model update: Joint sequential updates incorporate new local distributions into the propagated generator while filtered replay preserves information from previously encountered centers.
- Generative model update: Algorithm 1 specifies pseudo-sample generation, replay filtering, aging-clock optimization, generative-model optimization, and final model release.
- Sequential training protocol: The propagated parameters are initialized from the preceding authorized message, and the final aging-clock model is output after the last center update.
- Trust and security: The trust relation assigns sender authenticity and receiver confidentiality obligations for each authorized message, differing from participant-level behavioral models.
- Trust and security: Along a valid trust sequence, downstream centers do not use received messages to infer private information from upstream centers.
4 Declarations •
The work received support from PolyU and affiliated research institutes through named startup, scholar, AI, and federated-learning funding programs.
- Funding came from the PolyU Start-up Fund, Presidential Young Scholar Scheme, RIAIoT, and Research Institute for Federated Learning.
Appendix A Details About Biological Analysis of Key Molecular Features
The biological-analysis appendix defines model-based feature importance through age-dependent expert routing and controlled perturbations, then details pathway composition for interpretation.
- An expert whose routing weight increases with age provides the aging-related reference for identifying important molecular features.
- Feature influence is measured by independently varying each protein or CpG site across ten empirical quantile levels while holding other features fixed.
- The analysis quantifies influence as the change in expert routing weight between each feature’s highest and lowest quantile levels.
- All 20 significantly enriched ontology clusters are decomposed into constituent proteins for direct inspection of their functional modules.
A.3 Model-based Identification of Synergistic Protein Pairs
The model identifies synergistic protein pairs through exhaustive pair construction and four-state perturbation analysis, then extends the search to higher-order protein groups under complementary settings.
- Pairwise interactions: Candidate pairs are formed by exhaustively combining important proteins and evaluating their joint effects on model-estimated biological age.
- Pairwise interactions: Pairwise synergy compares outputs when both proteins, either protein, or neither protein is retained at its original values.
- Pairwise interactions: Positive synergy denotes joint effects exceeding the sum of individual effects, while negative scores indicate redundancy or antagonism; the top 10 pairs receive main-text analysis.
- Higher-order interactions: Higher-order candidate sets are built from important proteins to test coordinated collective effects beyond pairwise interactions.
- Higher-order interactions: Higher-order synergy is evaluated from either biological-age outputs or latent expert-routing outputs under four complementary analytical settings.
- Higher-order interactions: A beam-search–style procedure expands strong pairwise seeds one protein at a time while retaining the most promising group configurations.
A.5 Biological Analysis on Methylation Data
The methylation analysis tests whether TNFL-derived molecular patterns extend across omics modalities by characterizing influential CpGs, their genomic context, and enriched biological functions.
- The analysis selects the top 300 CpG sites by perturbation-induced changes in expert routing weights for downstream characterization.
- The selected CpGs map predominantly to gene-associated regions and CpG islands, corresponding to approximately 245 genes.
- Enriched methylation terms include cell adhesion, developmental processes, cell-cycle regulation, MAPK signaling, and immune-related functions.
- The mapped genes show relevance to immune-related hematological traits and enrichment for FOXO3 transcription-factor targets implicated in longevity and aging regulation.
- Functionally relevant processes appear in both proteomic and methylation modalities, indicating meaningful molecular signals across omics data types.
Appendix B Analysis of the Generative Component
The analysis tests whether generative replay improves AgeMoE under TNFL, including datasets with heterogeneous client distributions. Across datasets, the generative component improves global predictive performance, especially under stronger cross-client differences.
- Generative component: AgeMoE integrates generative replay to preserve information from previously visited clients and augment local training with pseudo-samples.The ablation compares the full model with a variant without the generative module while keeping other training settings unchanged.
- Datasets: The evaluation covers UKB-Center, UKB-Imbalance, GEO-Methylation, and age-partitioned UKB-AgeSplit datasets.UKB-AgeSplit partitions samples into six clients by five-year age intervals to simulate stronger cross-client distributional differences.
- Results: Generative replay improves global predictive performance across the evaluated datasets.Larger gains occur for GEO-Methylation and UKB-AgeSplit, which have stronger cross-client distributional differences.
- Interpretation: Retaining information from earlier clients and adding pseudo-samples helps maintain overall data-distribution coverage and preserve earlier information during heterogeneous training.The reported benefit is particularly pronounced when client data distributions differ substantially.
Appendix C Details on Robustness Metrics
The appendix defines three metrics for evaluating TNFL sensitivity to interaction order: client-level MAE range, cumulative global MAE degradation, and client-level forgetting. These metrics quantify variability, accumulated error, and retained performance across sequential updates.
- Client-level MAE range: Client-level MAE range measures each client’s MAE variability across interaction orders, with smaller values indicating more stable performance.The distributions of per-client ranges are visualized with boxplots for model comparison.
- Cumulative MAE degradation: Cumulative MAE degradation measures how sequential client updates change global-test-set MAE across an interaction order.Lower degradation indicates that sequential updates do not substantially increase error, and values are compared across orders with boxplots.
- Cumulative MAE degradation: The degradation metric aggregates MAE changes over the T update steps in each interaction sequence.T denotes the number of update steps.
- Forgetting rate: Forgetting measures each client’s relative performance loss by comparing MAE after its own update with MAE under the final model.The metric is calculated for clients in the first half of each sequence to emphasize early-stage effects.
- Forgetting rate: Positive forgetting values indicate performance deterioration after subsequent updates, whereas negative values indicate improvement or knowledge retention.Forgetting-rate distributions are summarized across sequences with boxplots.
Appendix D Details About Performance Under Simulated Measurement Variability
The perturbation study tests TNFL under simulated inter- and intra-center measurement variability. TNFL shows only a small decline in predictive performance under the specified perturbations.
- Perturbation design: The study adds client-level shifts and sample-level noise to simulate calibration, protocol, batch, site, and individual measurement variability.The perturbations represent systematic center deviations and independent sample fluctuations.
- Perturbation design: The perturbation magnitudes are set to approximately 35% of global feature standard deviation for client shifts and 12% for sample noise.These settings introduce controlled inter- and intra-center variability.
- Results: 0.07 global mean MAE increase, from 2.33 on original data to 2.40 on perturbed data, indicates minimal performance decline.Performance is evaluated on both individual-client and global test sets.
Appendix E Details About Performance Across Trust-Network Structures
The appendix evaluates TNFL under interaction sequences derived from directed trust networks with controlled sparsity. Predictive performance and robustness remain broadly consistent across valid trust-network sequences and random-permutation sequences.
- Trust-network construction: Trust-network-derived sequences are generated from directed networks with sparsity ranging from 0.1 to 0.5 and variable lengths covering all clients.The sequences respect directed trust relations while testing different connectivity levels.
- Experimental scope: The trust-network experiments use UKB-Center because it has many clients and heterogeneous distributions, while omitting repeated backbone-compatibility experiments.The evaluation focuses on predictive performance and robustness to interaction patterns.
- Predictive performance: TNFL achieves similar predictive performance under random and trust-network-derived sequence generation, remains better than local training, and stays competitive with FedAvg.The comparison includes independent local training, FedAvg, and both TNFL sequence settings.
- Robustness: TNFL maintains low variability, limited cumulative degradation, and no observable increase in forgetting across trust-network-derived interaction sequences.These stability patterns are reported for the same three robustness metrics used in the main study.
- Overall comparison: The observed predictive and robustness trends are broadly consistent between random-permutation and trust-network-derived sequence schemes.This supports stable behavior across different valid interaction sequences from the underlying trust network.