Not all individuals age at the same rate. Methods such as the ‘methylation clock’ are invasive, rely on expensive assays of tissue samples and infer the ageing rate by training on chronological age, which is used as a reference for prediction errors. Here, we develop models based on convoluted neural networks through training on non-invasive three-dimensional (3D) facial images of approximately 5,000 Han Chinese individuals that achieve an average difference between chronological or perceived age and predicted age of ±2.8 and 2.9 yr, respectively. We further profile blood transcriptomes from 280 individuals and infer the molecular regulators mediating the impact of lifestyle on the facial-ageing rate through a causal-inference model. These relationships have been deposited and visualized in the Human Blood Gene Expression—3D Facial Image (HuB-Fi) database. Overall, we find that humans age at different rates both in the blood and in the face, but do so coherently and with heterogeneity peaking at middle age. Our study provides an example of how artificial intelligence can be leveraged to determine the perceived age of humans as a marker of biological age, while no longer relying on prediction errors of chronological age, and to estimate the heterogeneity of ageing rates within a population.
This is a preview of subscription content, access via your institution
Open Access articles citing this article.
LensAge index as a deep learning-based biological age for self-monitoring the risks of age-related diseases and mortality
Nature Communications Open Access 06 November 2023
International Journal of Obesity Open Access 25 July 2023
Complex & Intelligent Systems Open Access 19 July 2023
Access Nature and 54 other Nature Portfolio journals
Get Nature+, our best-value online-access subscription
$29.99 / 30 days
cancel any time
Subscribe to this journal
Receive 12 digital issues and online access to articles
$119.00 per year
only $9.92 per issue
Rent or buy this article
Prices vary by article type
Prices may be subject to local taxes which are calculated during checkout
The regulatory genes used in building the causal-inference network are from the KEGG pathway (https://www.genome.jp/kegg/pathway.html), Reactome (https://www.reactome.org/), Animal Transcription Factor Database (http://www.bioguo.org/AnimalTFDB/), human DEPhOsphorylation Database (http://www.depod.bioss.uni-freiburg.de/), IUPHAR/BPS database (http://www.guidetopharmacology.org/) and ImmPort42 (https://www.immport.org/shared/genelists). ImageNet (http://image-net.org/) was used for pretraining.
Results of facial-image, transcriptome and lifestyle associations are searchable at http://www.picb.ac.cn/hanlab/hub-fi/. Three-dimensional images and other metadata sensitive to personal identification cannot be publicized or shared according to our participant consent agreement. Individual sequencing raw data, as they contain genetic information, will be available on request under the condition of approval of the ethics committee of Shanghai Institute of Nutrition and Health, Chinese Academy of Sciences abiding China Human Genetic Resource law. Mapped read counts and FPKM expression values of coding genes from the RNA-seq are deposited to the public repository NODE at https://www.biosino.org/node/project/detail/OEP001041.
Peters, M. J. et al. The transcriptional landscape of age in human peripheral blood. Nat. Commun. 6, 8570 (2015).
Lehallier, B. et al. Undulating changes in human plasma proteome profiles across the lifespan. Nat. Med. 25, 1843–1850 (2019).
Hannum, G. et al. Genome-wide methylation profiles reveal quantitative views of human aging rates. Mol. Cell 49, 359–367 (2013).
Horvath, S. DNA methylation age of human tissues and cell types. Genome Biol. 14, 3156 (2013).
Unschuld, P. U. & Tessenow, H. Huang Di Nei Jing Su Wen (University of California Press, 2011).
Chen, W., Xia, X., Huang, Y., Chen, X. & Han, J.-D. J. Bioimaging for quantitative phenotype analysis. Methods 102, 20–25 (2016).
Chen, W. et al. Three-dimensional human facial morphologies as robust aging markers. Cell Res. 25, 574–587 (2015).
López-Otín, C., Blasco, M. A., Partridge, L., Serrano, M. & Kroemer, G. The Hallmarks of Aging. Cell 153, 1194–1217 (2013).
Gao, X. W., Hui, R. & Tian, Z. Classification of CT brain images based on deep learning networks. Computer Methods Programs Biomed. 138, 49–56 (2017).
Esteva, A. et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature 542, 115–118 (2017).
Millstein, J., Zhang, B., Zhu, J. & Schadt, E. E. Disentangling molecular relationships with a causal inference test. BMC Genet. 10, 23 (2009).
Debrabant, B. et al. DNA methylation age and perceived age in elderly Danish twins. Mechanisms Ageing Dev. 169, 40–44 (2018).
Lu, T. et al. Gene regulation and DNA damage in the ageing human brain. Nature 429, 883–891 (2004).
Newman, A. M. et al. Robust enumeration of cell subsets from tissue expression profiles. Nat. Methods 12, 453–457 (2015).
Jefferson, A. L. et al. Inflammatory biomarkers are associated with total brain volume. Neurology 68, 1032–1038 (2007).
Frąckiewicz, J. et al. Hematological parameters and all-cause mortality: a prospective study of older people. Aging Clin. Exp. Res. 30, 517–526 (2018).
Chatthanawaree, W. Biomarkers of cobalamin (vitamin B12) deficiency and its application. J. Nutr. Health Aging 15, 227–231 (2011).
Conigrave, K. M., Davies, P., Haber, P. & Whitfield, J. B. Traditional markers of excessive alcohol use. Addiction 98, 31–43 (2003).
Liu, Y. et al. Epigenome-wide association data implicate DNA methylation as an intermediary of genetic risk in rheumatoid arthritis. Nat. Biotechnol. 31, 142–147 (2013).
Ahmed, Z. et al. Accelerated lipofuscinosis and ubiquitination in granulin knockout mice suggest a role for progranulin in successful aging. Am. J. Pathol. 177, 311–324 (2010).
He, Z., Ong, C. H. P., Halper, J. & Bateman, A. Progranulin is a mediator of the wound response. Nat. Med. 9, 225–229 (2003).
Elkabets, M. et al. Human tumors instigate granulin-expressing hematopoietic cells that promote malignancy by activating stromal fibroblasts in mice. J. Clin. Invest 121, 784–799 (2011).
Chitramuthu, B. P., Bennett, H. P. J. & Bateman, A. Progranulin: a new avenue towards the understanding and treatment of neurodegenerative disease. Brain 140, 3081–3104 (2017).
Knupp, D. & Miura, P. CircRNA accumulation: a new hallmark of aging? Mechanisms Ageing Dev. 173, 71–79 (2018).
Ruiz, R. et al. Sterol regulatory element-binding protein-1 (SREBP-1) is required to regulate glycogen synthesis and gluconeogenic gene expression in mouse liver. J. Biol. Chem. 289, 5510–5517 (2014).
Oishi, Y. et al. SREBP1 contributes to resolution of pro-inflammatory TLR4 signaling by reprogramming fatty acid metabolism. Cell Metab. 25, 412–427 (2017).
Li, S. et al. Metabolic phenotypes of response to vaccination in humans. Cell 169, 862–877.e817 (2017).
Schoenborn, N. L. et al. Preferred clinician communication about stopping cancer screening among older US adults: results from a national survey. JAMA Oncol. 4, 1126–1128 (2018).
World Medical Association Inc Declaration of Helsinki. Ethical principles for medical research involving human subjects. J. Indian Med. Assoc. 107, 403–405 (2009).
Guo, J., Mei, X. & Tang, K. Automatic landmark annotation and dense correspondence registration for 3D human facial images. BMC Bioinformatics 14, 232 (2013).
King, D. E. Dlib-ml: a Machine Learning Toolkit J. Mach. Learn. Res. 10, 1755–1758 (2009).
Szegedy, C. et al. in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 1–9 (2015).
Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognition. Preprint at https://https://arxiv.org/abs/1409.1556 (2014).
He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. in IEEE Conference on Computer Vision and Pattern Recognition 770–778 (2016).
Kingma, D. & Ba, J. Adam: A method for stochastic optimization. Preprint at https://arxiv.org/abs/1412.6980 (2014).
Harrow, J. et al. GENCODE: the reference human genome annotation for The ENCODE Project. Genome Res. 22, 1760–1774 (2012).
Trapnell, C. et al. Differential gene and transcript expression analysis of RNA-seq experiments with TopHat and Cufflinks. Nat. Protoc. 7, 562–578 (2012).
Zhang, X. O. et al. Diverse alternative back-splicing and alternative splicing landscape of circular RNAs. Genome Res. 26, 1277–1287 (2016).
Wold, S., Sjöström, M. & Eriksson, L. PLS-regression: a basic tool of chemometrics. Chemometrics Intell. Lab. Syst. 58, 109–130 (2001).
Subramanian, A. et al. Gene set enrichment analysis: a knowledge-based approach for interpreting genome-wide expression profiles. Proc. Natl Acad. Sci. USA 102, 15545–15550 (2005).
Dennis, G. et al. DAVID: database for annotation, visualization, and integrated discovery. Genome Biol. 4, P3 (2003).
Bhattacharya, S. et al. ImmPort, toward repurposing of open access immunological assay data for translational and clinical research. Sci. Data 5, 180015 (2018).
Coppé, J.-P., Desprez, P.-Y., Krtolica, A. & Campisi, J. The senescence-associated secretory phenotype: the dark side of tumor suppression. Annu Rev. Pathol. 5, 99–118 (2010).
Millstein, J., Chen, G. K. & Breton, C. V. cit: hypothesis testing software for mediation analysis in genomic applications. Bioinformatics 32, 2364–2365 (2016).
This work was supported by grants from the National Natural Science Foundation of China (91749205), China Ministry of Science and Technology (2016YFE0108700) and Shanghai Municipal Science and Technology Major Project (2017SHZDZX01) to J.-D.J.H.
Authors declare no competing interests.
Peer review information Primary Handling Editor: Pooja Jha.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
a, Deep learning performance in training, validation and testing datasets for age prediction. Average loss (mean average difference (MAD), upper panel) and accuracy (Pearson Correlation Coefficient (PCC), lower panel) were plotted over training epochs. One epoch indicates the network weights updates over the whole training dataset for one time. b, Overlap between FaceCnnAge, FaceCnnPerceivedAge, FacePerceivedAge, and FacePlsAge in younger (left), normal (middle), and older (right) samples. The lower table shows the one-tailed Fisher’s exact test p-values with light red indicate p < 0.05. c, d, Independent validation of FaceCnnAge (c) and FaceCnnPerceivedAge (d) on 332 facial images collected in Beijing, 2012 (left) and 358 facial images collected in Beijing, 2015 (right).
The p value is derived from ANOVA (n=341). All results with p < 0.05 are shown here. Data are presented as mean +/- SD.
a, The standard deviation of randomly guess a number between 20–85 (n=4719, the boxes show 25%, 50% and 75% quantile and whiskers show maximum and minimum value). b, Relationship between age and the standard deviation of four AgeDiffs with bin size as 20 in Jidong cohort. Data are presented as mean +/- SD. c, Heatmap of aging-related facial features, health parameters and RNAs in Beijing (2012) cohort sorted by increasing chronological age (PCC with age, FDR < 0.05). Features were ranked by PCC from low to high (ALB: albumin, A/G: albumin/ globulin, TP: total protein, GGT: glutamyl transpeptidase, ALP: alkline phosphatase, CREA: creatinine, CHO: total cholesterol). d, e, Relationship between age and the standard deviation of four AgeDiffs with bin size as 20 (d) and 10 (e) in Beijing (2012) cohort. Data are presented as mean +/- SD.
a, b, Broken stick regression in Jidong cohort with bin size 100 (a) and 20 (b). c, d, Broken stick regression in Beijing (2012) cohort with bin size 20 (c) and 10 (d).
a–c, Mean absolute difference (MAD) (top panel), Pearson correlation coefficient (PCC) (bottom panel) saturation analysis and correlation against chronological age of transcriptomes PLS age prediction for all the samples (a), female (b) and male (c), respectively. d, Enriched GO biological processes terms of PLS top 10% (upper) or 20% (lower) loading genes. P values are derived from hypergeometric test (Methods). e, Overlap between FacePlsAge, FaceCnnAge, RnaPlsAge, and FaceCnnPerceivedAge in younger (left), normal (middle), and older (right) samples. The lower table shows the one-tailed Fisher’s exact test p-values with light red highlight indicate p < 0.05.
a, Cytokines (left panels) and antigen processing and presentation (right panels) enrichment scores as a function of four AgeDiffs. P values are derived from permutation test (Methods). b, Association of RNA-seq deconvoluted cell type fractions and AgeDiffs (* p<0.1, ** p<0.05, *** p<0.01 derived from two-sided t test, and * Benjamini-Hochberg correction derived FDR < 0.1).
a, The heatmap of expressed lncRNAs (FPKM > 2) significantly related to chronological age (FDR < 0.1). The samples (columns) were sorted by age and lncRNAs were sorted by PCC of expression to age from high to low. b, The heatmap of expressed circRNAs (TPM > 2) significantly related to chronological age (FDR < 0.1). The samples (columns) were sorted by age and circRNAs were sorted by PCC of expression to age from high to low. c, Top three enriched terms for parent genes of age-up (top) and age-down circRNAs. P values are derived from hypergeometric test (Methods). d, The correlation between age and total expression level of circRNAs. P value is derived from two-sided t test.
About this article
Cite this article
Xia, X., Chen, X., Wu, G. et al. Three-dimensional facial-image analysis to predict heterogeneity of the human ageing rate and the impact of lifestyle. Nat Metab 2, 946–957 (2020). https://doi.org/10.1038/s42255-020-00270-x
This article is cited by
LensAge index as a deep learning-based biological age for self-monitoring the risks of age-related diseases and mortality
Nature Communications (2023)
International Journal of Obesity (2023)
Complex & Intelligent Systems (2023)
Science China Life Sciences (2023)