BiT age: A transcriptome-based aging clock near the theoretical limit of accuracy

Abstract

Aging clocks dissociate biological from chronological age. The estimation of biological age is important for identifying gerontogenes and assessing environmental, nutritional, or therapeutic impacts on the aging process. Recently, methylation markers were shown to allow estimation of biological age based on age-dependent somatic epigenetic alterations. However, DNA methylation is absent in some species such as Caenorhabditis elegans and it remains unclear whether and how the epigenetic clocks affect gene expression. Aging clocks based on transcriptomes have suffered from considerable variation in the data and relatively low accuracy. Here, we devised an approach that uses temporal scaling and binarization of C. elegans transcriptomes to define a gene set that predicts biological age with an accuracy that is close to the theoretical limit. Our model accurately predicts the longevity effects of diverse strains, treatments, and conditions. The involved genes support a role of specific transcription factors as well as innate immunity and neuronal signaling in the regulation of the aging process. We show that this binarized transcriptomic aging (BiT age) clock can also be applied to human age prediction with high accuracy. The BiT age clock could therefore find wide application in genetic, nutritional, environmental, and therapeutic interventions in the aging process.

Keywords: Caenorhabditis elegans; RNA sequencing; aging; aging clock; biological aging; biomarkers; transcriptome.

FIGURE 1

Data overview. Overview of the processed published data utilized in the training of the model. Pie charts show the distribution of different genotypes (blue), treatments (brown), and RNAis (green). The convoluted pie chart on the right shows the overlap of the three classes. The partition “Sterile” contains multiple different genotypes that cannot give rise to progeny and daf‐2, as well as eat‐2, might contain additional mutations. For a more detailed view, see Table S1

FIGURE 2

Biological age prediction. (a) Results of the biological age prediction computed by cross‐validation. The x‐axis shows the rescaled biological age in days starting from adulthood additionally corrected by the second rescaling approach. The y‐axis shows the predicted age computed by the elastic net regression after the second rescaling approach on binarized gene expression data. Every blue dot displays one RNA‐seq sample. The regression line with the 95% confidence interval is shown in blue, and the dotted line shows the perfect linear correlation. The distribution of the data is shown on the side of the plot. r ² = coefficient of determination, Pearson = Pearson correlation, Spearman = Spearman correlation, MAE = mean absolute error in days, MAD = median absolute deviation in days, RMSE = root‐mean‐square‐error in days. (b) Prediction of the model on eight independent datasets consisting of 94 samples at different time points. The x‐axis shows the rescaled biological age in days starting from adulthood additionally corrected by the second rescaling approach. The y‐axis shows the predicted age computed by the elastic net regression after the second rescaling approach on binarized gene expression data. For more details on the data, see Table S1. (c) The y‐axis shows the r ² of a given prediction. The box plot displays 1,000 random models with 576 genes. The prediction by our final model with an r ² of 0.96 is shown as a blue dot and indicated by the arrow. The dotted line shows the theoretical limit of prediction given by the limit of accuracy in the chronological age annotation as well as variance in the lifespan data used for rescaling

FIGURE 3

Biological age prediction of short‐ and long‐lived mutants. The box plots show the predicted biological age in days on the y‐axis. Assuming the properties of a uniform temporal rescaling, a lower predicted age will equal a longer lifespan. The corresponding whole dataset was set aside for the training of the final model for the corresponding plot. Blue dots display single RNA‐seq samples. (a) The lifespan‐extending daf‐2(e1370) strain is predicted to be biologically younger than WT samples of the same chronological age (4.5 days). Note that the WT strain in this publication had a longer lifespan (19.4 days) than the standard 15.5 days and is thereby also predicted to be biologically younger than its chronological age. Data from GSE36041. (b) Dietary restriction (DR) and the long‐lived double mutant daf‐2(e1370); rsks‐1(ok1255) are predicted to be significantly younger than WT samples of the same chronological age (4 days). Data from GSE119485. (c) The lifespan‐shortening mir‐71(n4115) mutation significantly increased the predicted biological age compared to samples of the same chronological age (5 days). Data from GSE72232. (d) The gain‐of‐function mutant skn‐1(lax188) significantly increased the biological age, while an additional mutation in the epigenetic regulator wdr‐5 rescues the biological age back to WT levels (2 days). Data from GSE123531. (e) The double mutant tut‐1(tm1297); elpc‐1(tm2149) significantly increases the biological age (chronological age of 1 day). Data from GSE67387. *p < 0.05, **p ≤ 0.01, ***p ≤ 0.001, independent two‐sided t tests were used for comparisons in (a), (c), and (e). One‐way ANOVA with a post hoc Tukey test was used in (b) and (d). Table S3 contains more detailed statistics

FIGURE 4

Biological age prediction of a variety of treatments and stressors. The box plots show the predicted biological age in days on the y‐axis. Assuming the properties of a uniform temporal rescaling, a lower predicted age will equal a longer lifespan. The corresponding whole dataset was set aside for the training of the final model for the corresponding plot. Blue dots display single RNA‐seq samples. (a) Heat shock induces a strong increase in the predicted biological age at a chronological age of 3 days in WT. Data from PRJNA523315. (b) Pathogen infection by Pseudomonas aeruginosa at 25°C at a chronological age of day 1 increases significantly the predicted age. Data from GSE122544. (c) Pathogen infection by S. aureus at 25°C at a chronological age of day 1 increases significantly the predicted age. Data from GSE57739. (d) The bacterial strain‐dependent effect of metformin is resembled in the prediction. The box plots show wild‐type worm populations at a chronological age of day 2 with either a standard OP50 E. coli diet or a Metformin‐resistant OP50 (OP50‐MR) strain with or without 50 mM Metformin. A two‐way ANOVA showed a significant treatment effect (p = 0.004). Data from E‐MTAB‐7272. (e) The dosage‐dependent effect of Mianserin is resembled in the prediction. The box plots show wild‐type worm populations at a chronological age of day 10 either treated with water or 50 µM Mianserin on day 3 or day 1. A one‐way ANOVA showed significance (p = 0.0008). Data from GSE63528. (f) The effect of drug combinations at the chronological age of 6 days is resembled in the prediction. A one‐way ANOVA showed significance (p = 0.02). Data from GSE108263. (g) An independent dataset without a reported lifespan sequenced at the chronological age of day 1. Wild‐type worms were treated with either 10 µM or 20 µM of the proteasome inhibitor Bortezomib (BTZ), or RNAi against the proteasomal subunit rpn‐6. Data from GSE124178. (h) An independent dataset without a reported lifespan sequenced at the chronological age of day 3. Data from GSE121920. The predicted median lifespan reduction of 35.7% is similar to the reported lifespan reduction of 33.5% (Pang & Curran, 2014). (i) An independent dataset without a reported lifespan sequenced at the chronological age of day 2. Data from GSE158729. The predicted median lifespan reduction of 63.96% is similar to the reported lifespan reduction of 50%–60.69% (Ratnappan et al., 2014). *p < 0.05, **p ≤ 0.01, ***p ≤ 0.001, independent two‐sided t tests were used for comparisons in (a), (b), (c), (h), and (i). One‐way ANOVA with a post hoc Tukey test was used in (e), (f), and (g). Two‐way ANOVA with a post hoc Tukey test was used in (d). Table S3 contains more detailed statistics

FIGURE 5

Functional analysis of the predictor genes. (a–d) WormExp gene set enrichment analysis for the 576 predictor genes. The x‐axis displays the −log10 of the adjusted p‐value. Only statistically significant (adjusted p < 0.05) enrichments are shown. (a–c) Gene set enrichment analyses for the genes with a coefficient ≤0 for the Development/Dauer/Aging category (a), the TF Targets category (b), and the Tissue category (c). (d) Gene set enrichment analyses for the genes with a coefficient >0 for the Development/Dauer/Aging category. (e) Functional enrichment analysis for the 576 predictor genes by String and geneSCF. The x‐axis displays the −log10 of the FDR. The red line displays an FDR of 0.05. Different enrichment categories are color‐coded

FIGURE 6

Transcriptomic human aging clock. (a) Results of the age prediction computed by cross‐validation on human fibroblast gene expression data. The x‐axis shows the chronological age in years. The y‐axis shows the predicted age computed by an elastic net regression on binarized gene expression data. Every blue dot displays one RNA‐seq sample. The regression line with the 95% confidence interval is shown in blue, and the dotted line shows the perfect linear correlation. The distribution of the data is shown on the side of the plot. r ² = coefficient of determination, Pearson = Pearson correlation, Spearman = Spearman correlation, MAE = mean absolute error in years, MAD = median absolute deviation in years, RMSE = root‐mean‐square‐error in years. Data from GSE113957. (b) Box plots of age predictions of samples from Hutchinson‐Gilford progeria syndrome patients (red) and predictions of age‐matched healthy controls (blue) by the elastic net regression of binarized gene expression data. Progeria samples are predicted to be significantly older than age‐matched healthy controls. Data from GSE113957. **p ≤ 0.01, calculated by an independent two‐sided t test. Table S3 contains more detailed statistics