REVIVE-Flow

Epigenetic clocks read biological age off a blood methylome accurately, but they are diagnostic: they report how old a sample looks without indicating what would make it look younger. REVIVE-Flow (REjuVenatIon Via Epigenetic Flow), posted as a preprint in September 2025 by Biostate AI and Stanford University, closes that gap by turning epigenetic aging from a readout into a controllable process.

The model treats aging as a smooth flow through methylation space. Blood methylation profiles are projected into a linear latent space by principal component analysis, and a neural ordinary differential equation is trained there with a flow-matching objective calibrated to chronological age, so the learned vector field carries units of latent change per year. Because the field is continuous and the embedding exactly invertible, it can be integrated forward to age a methylome or backward to trace a rejuvenation trajectory following the dynamics the data exhibit rather than an arbitrary direction in feature space.

The reusable artifact is that pretrained flow; a separate per-sample convex optimization decodes a backward trajectory into the smallest set of CpG edits that moves a methylome toward it. On EPIC-Italy (GSE51032), a cohort of more than 800 individuals withheld from preprocessing, PCA fitting, training, and hyperparameter selection alike, the dynamics were applied without refitting.

#Key Features

• Prescriptive, not just predictive: Where a clock outputs an age estimate, REVIVE-Flow outputs a list of CpG sites with the direction and magnitude of change at each one.
• Age-calibrated flow matching: The training target is the velocity between two real samples divided by their age gap, reducing the objective to a stable regression whose output is latent drift per year.
• Dose-controllable sparsity: A single L1 penalty tunes how many CpGs get edited against how much age reversal results, tracing a Pareto frontier from a handful of edits to thousands.
• Strict external hold-out: One entire cohort was excluded from every fitting step, so reported performance reflects transfer of fixed dynamics to an unseen study, not in-sample fit.
• Artifact controls: A shuffled-age negative control abolishes the effect, and reference-based deconvolution confirms edits do not merely rebalance leukocyte proportions.

#Technical Details

The vector field is a Transformer encoder that treats each of 1,024 principal components as a token, conditioned on age via sinusoidal Fourier features and on sex via a one-hot vector, both injected through FiLM modulation; that latent dimension captures 67% of variance. Training pairs same-sex samples separated by at least five years and applies a random directional flip so the field learns both aging and rejuvenation. The compendium comprises 6,402 blood samples from 32 cohorts filtered out of MethAgingDB for high-density Illumina arrays covering at least 400,000 CpGs; beta values are logit-transformed to M-values and standardized per study to suppress batch effects. Trajectories are integrated with a fixed-step RK4 solver over 300 steps, and the intervention is solved by ADMM against fidelity to that trajectory, an L1 sparsity cost, and an invariance penalty weighted by inverse training variance to protect low-variance loci.

Efficacy is scored by a Ridge regressor trained on the training-set M-values, acting as an independent age judge. On EPIC-Italy with no sparsity penalty, realized age reduction tracks commanded reduction with a slope of 0.396 (R² = 0.985); across leave-one-study-out folds the slope is 0.291 ± 0.148 with R² of 0.910 ± 0.116. Raising the penalty to 0.02 collapses the slope to 0.052.

#Applications

The model is aimed at aging researchers and epigenetic-editing programs that need to prioritize candidate loci before committing to wet-lab work. Given a methylome and a target age, it returns a budgeted edit list rather than a diffuse genome-wide shift, matching the constraints of targeted methylation and demethylation tools. The selected sites are enriched in CpG islands and shores and in gene bodies and first exons, and depleted at promoter-proximal regions, favoring loci with dynamic range across aging over constitutively unmethylated promoters. Their associated genes are enriched for cell–cell adhesion and calcium ion binding, consistent with age-linked declines in barrier integrity.

#Impact

REVIVE-Flow reframes epigenetic aging as a control problem and shows that a fixed, pretrained flow transfers to an entirely unseen cohort — a complement to methylation predictors such as MethylAI and MethylProphet, which model where methylation sits rather than how it moves. The results carry real limits. The work is a preprint that has not been peer reviewed, and the headline figure of roughly 0.4 years reversed per commanded year is an in-silico measurement from the model's own Ridge judge, not evidence that the prescribed edits rejuvenate cells. The system covers blood only and conditions on age and sex alone. The training code and the full compendium are public — the latter under CC-BY-4.0 on Zenodo — but the repository carries no license file and ships no pretrained checkpoint, so users reproduce the model by training it on the released data.