Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data".
Marcus: Here's the gist of the paper. DNA methylation (DNAm) serves as one of the most robust molecular biomarkers of biological aging <ref:2607.06583#pg2> To reconstruct these continuous dynamics, we frame lifelong human epigenetic aging as a trajectory inference problem across discrete age snapshots derived from widely available cross-sectional data <ref:2607.06583#pg2> By resolving this RUOT formulation using the DeepRUOT framework, our model fluidly accommodates population-level density shifts like survivorship bias and cellular attrition without requiring rigid biological priors <ref:2607.06583#pg3> Finally, by decoding continuous latent paths back to individual CpG sites, we reconstruct and empirically verify distinct biological aging archetypes <ref:2607.06583#pg2>.
Ines: Pruning redundant reasoning tokens The paper introduces a two-stage computational pipeline to infer continuous human epigenetic aging trajectories from static cross-sectional DNA methylation data <ref:2607.06583#pg2> This pipeline first maps high-dimensional CpG profiles into a chronologically ordered latent manifold using an age-regularized Variational Autoencoder (VAE) <ref:2607.06583#pg2> Subsequently, it models the continuous movement across this latent space via Regularized Unbalanced Optimal Transport (RUOT) to capture deterministic drift, random diffusion, and non-conservative mass changes <ref:2607.06583#pg3>.
Marcus: How the training signal is built The age-regularized VAE architecture is optimized via a composite loss function LVAE which includes a reconstruction loss (Lrecon), a Kullback-Leibler divergence (LKL), and an age penalty term (Lage) <ref:2607.06583#pg2> The Lage term represents the Mean Squared Error of the predicted chronological age derived from the latent vector z <ref:2607.06583#pg2>.
Ines: This joint regularization ensures that the latent space aligns monotonically with biological time, as enforced by the penalty γ <ref:2607.06583#pg2>.
Paper summary: Ines: So we're looking at this paper "Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data" and what they claim is that DNA methylation is a really good marker for how old someone is, but the existing clocks only give you one number.
Marcus: Exactly. Conventional epigenetic clocks treat aging like a static thing, meaning they just output a single score rather than showing you how the whole profile changes over time. This paper frames lifelong human epigenetic aging as this trajectory inference problem across snapshots from cross-sectional data.
Yuki: That’s interesting because it tries to get past that single score and see the continuous dynamics, which is what we need to understand biological processes better. But you're saying they use these static snapshots to reconstruct a continuous path?
Ines: Right. They build this two-stage pipeline where first, an age-regularized Variational Autoencoder compresses those high dimensional CpG profiles into a latent space that is inherently age aware, keeping the original methylation space accessible through a decoder bridge.
Marcus: Then they take that latent space and model the continuous movement across it using what they call regularized optimal transport to capture deterministic drift, random diffusion, and non conservative mass changes. This is important because population aging isn't a simple mass conserving system.
Yuki: You mean like survivorship bias or tissue specific attrition can happen, which violates those traditional mass preserving optimal transport frameworks? That sounds like it tackles a real problem when looking at large cohorts.
Ines: Precisely. They argue that traditional methods structurally can't reconcile younger dense cohorts with older scattered survivors without artificially distorting the underlying paths, and this approach addresses that by modeling the continuous movement across the aging manifold.
Marcus: The training signal for this whole thing comes from a composite loss function called LVAE, which includes a reconstruction loss, a Kullback Leibler divergence, and an age penalty term. This age penalty term is based on the mean squared error of the predicted chronological age derived from that latent vector.
Yuki: So they are forcing the latent space to align with biological time by penalizing any misalignment between what's in there and what time it should be? That seems like a very strong constraint.
Paper summary: Ines: It is, because they use a weight called gamma for that age penalty, which dictates the topological constraints of the manifold, ensuring this latent space aligns monotonically with biological time. This helps preserve chronological progression through the entire process.
Marcus: The core of their dynamic inference comes from DeepRUOT, which models continuous aging dynamics by parameterizing a velocity field, a growth field, and a score function. They optimize this using an energy loss term that uses Monte Carlo sampling and a neural ordinary differential equation solver to optimize the dynamic energy penalty derived from the fisher regularization form.
Yuki: So they are essentially modeling how fast things move and how they grow over time in this latent space, rather than just looking at where they are at one point. That gives us a picture of the actual movement between ages.
Ines: Yes, and this optimization happens across pseudo time t from zero to T using a composite loss function called LRUOT, which is what models the continuous movement across the aging manifold we talked about earlier. This gives them those continuous velocity and growth fields that describe population level epigenetic drift.
Marcus: When they look at their results, they verify distinct biological archetypes by analyzing the kinetic properties of individual continuous curves, beta i of t, across the simulated lifespan. They found four primary archetypes: Linear Accumulators, defined by the maximum difference between age eighty and age zero; Linear Decay is defined by the minimum difference between those two points; Exponential and Late Onset Drift is defined by the maximum absolute curvature or second derivative; and Age Invariant Maintenance Sites are those with minimum temporal variance.
Yuki: Those four types sound like distinct kinetic behaviors, which suggests that epigenetic aging isn't just one thing but a composite of different processes happening simultaneously across different parts of the genome. Linear Accumulators show steady gain in methylation, while Linear Decay represents a global loss of methylation.
Ines: That’s what they found when they mapped the continuous paths back to individual CpG sites, generating those continuous aging curves beta i of t for every site. They showed that these paths map back to distinct kinetic behaviors including Linear Accumulators, which exhibit monotonic steady gain in methylation, and Age Invariant Maintenance Sites representing static housekeeping components.
Paper summary: Marcus: For quantitative evaluation they use Wasserstein one distance to measure structural accuracy and Total Mass Variation to quantify localized density adjustment. They found relatively stable Wasserstein one distances across held out experiments, indicating smooth interpolation between the data points.
Yuki: But the Total Mass Variation reveals a distinct shift in aging dynamics, which is minimal during early to mid life transitions but requires a substantial increase to predict the oldest age bin. That's a tangible difference in how they are changing over time as you get older.
Ines: And their growth term shows a prominent localized surge in late life states, which they hypothesize mirrors the biological onset of stochastic epigenetic drift. This is what they think drives those accelerated late life changes we see in some populations.
Marcus: In terms of what this means for us, the paper title "Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data" suggests that instead of just getting a single age score, we can infer the actual path someone’s methylation profile is taking over their lifetime.
Yuki: So if we only listen to this show, it means understanding aging isn't just about where you are now on a scale; it's about the continuous journey you've taken.
Ines: Exactly. It changes how we think about the biology, moving from static measurements to dynamic systems that capture how molecular senescence progresses across time.
Marcus: The authors also point out some limitations with this framework, specifically mentioning that the coarse temporal constraint introduced by multi decade age discretization might smooth over rapid life stage transitions. Plus, they note that using a universal pan tissue strategy suppresses localized tissue specific metabolic kinetics.
Yuki: That’s a fair caution. If the method is looking at the whole human system, it might miss important things happening in specific organs or tissues during certain life stages. We need to see how these trajectories are rewired in pathological settings, like cancer, where DNA methylation instability drives oncogenic transformation and progression.
Ines: Right. The framework bridges macroscopic population phenomena with microscopic molecular kinematics, but we still need to figure out how those individual trajectories behave when things go wrong in disease states. That's the next step for this research direction.
Conclusion: Ines: So we're wrapping up this look at the paper "Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data." Basically, they've taken those static snapshots of DNA methylation data and figured out a way to map them onto a continuous path showing how aging actually happens over time.
Marcus: Yeah, so instead of just giving you one score for how old someone is right now, this model tries to reconstruct the actual journey that profile took from birth through life. It's about inferring the continuous dynamics behind those discrete measurements.
Yuki: From a population genetics standpoint, what this means is we move beyond seeing a static snapshot of aging and start modeling the underlying biological processes driving it across different life stages. It suggests there isn't just one single trajectory for aging in humans across all people and all time.
Ines: Right, and the authors show that by using this DeepRUOT framework, they can identify specific biological archetypes—like linear accumulators or exponential drift—that correspond to these continuous paths. It's about finding the distinct kinetic behaviors happening across different parts of the genome.
Marcus: The main thing I take away is how they handle those tricky population issues, like survivorship bias and things where cells just attrition out over time. They use this method to accommodate those density shifts without needing super rigid biological rules upfront.
Yuki: It’s interesting because it bridges that gap between the big picture of population aging and the tiny molecular events at the CpG sites themselves. It gives us a dynamic way to look at how different groups of people might age differently over their lifetimes.
Ines: Exactly, so this work opens up a new way to interpret epigenetic data—not just as a current state, but as part of a historical movie of biological change. It's about understanding the continuous movement that underlies what we measure on the page.
Marcus: And looking at the authors' focus on those four archetypes, it tells us that aging isn't one smooth process; it’s actually a mix of different things happening simultaneously across the body and over time. That complexity is where the real biological insight lies.
The Institute of Cancer Research · University of Cambridge
q-bio.QM, cs.LG
Submitted: 2026-07-04
Updated: 2026-11-11
Comments: 16 pages, 7 figures
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 98/100
The gist: DNA methylation (DNAm) serves as one of the most robust molecular biomarkers of biological aging <ref:2607.06583#pg2> To reconstruct these continuous dynamics…
Key concepts
- DNA Methylation (DNAm)
- This refers to chemical modifications on DNA that often act as a molecular clock for biological age. The study uses these methylation patterns across different ages to track how the epigenome changes continuously over a lifespan.
- DeepRUOT
- This is a dynamic inference framework used to model continuous movement through the latent space derived from the data. It allows researchers to capture complex aging dynamics, including smooth drift and sudden shifts, by parameterizing velocity, growth, and score functions.
- Aging Archetypes
- These are distinct kinetic patterns observed in how different parts of the genome age. The study identifies four main types: Linear Accumulators (steady gain), Linear Decay (global loss), Exponential Drift (accelerated late-life change), and Maintenance Sites (stable components).
- RUOT Formulation
- Regularized Unbalanced Optimal Transport is a mathematical tool used to model the continuous flow or transport of data points over time. It helps capture non-conservative changes in DNA methylation density as individuals age.
Terminology
Summary
DNA methylation (DNAm) serves as one of the most robust molecular biomarkers of biological aging <ref:2607. p. 0 uk
Pietro Liò
University of Cambridge
United Kingdom
pl219@cam. uk
Abstract
DNA methylation (DNAm) serves as one of the most robust molecular biomarkers of biological aging. While conventional epigenetic clocks accurately predict
chronological age from high-dimensional CpG profiles, they treat aging as a static
regression task, meaning they can only output a single score rather than simulating how an entire profile continuously changes over time. To reconstruct these
continuous dynamics, we frame lifelong human epigenetic aging as a trajectory
inference problem across discrete age snapshots derived from widely available
cross-sectional data. Second, we model the
continuous movement across this latent space via Regularized Unbalanced Optimal Transport (RUOT) that unifies deterministic drift, random diffusion, and
non-conservative mass changes. By resolving this RUOT formulation using the
DeepRUOT framework, our model fluidly accommodates population-level density
shifts like survivorship bias and cellular attrition without requiring rigid biological
priors. Evaluated on a large-scale, 80-year pan-tissue dataset, our model demonstrates robust distribution interpolation and uncovers a prominent late-life surge in
the learned growth field that mathematically captures the variance expansion driven
by stochastic epigenetic drift. Finally, by decoding continuous latent paths back
to individual CpG sites, we reconstruct and empirically verify distinct biological
aging archetypes, offering a rigorous, generative paradigm for simulating human
molecular aging.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Pruning redundant reasoning tokens
The paper introduces a two-stage computational pipeline to infer continuous human epigenetic aging trajectories from static cross-sectional DNA methylation data <ref:2607.06583#pg2> This pipeline first maps high-dimensional CpG profiles into a chronologically ordered latent manifold using an age-regularized Variational Autoencoder (VAE) <ref:2607. p. 0 To reconstruct these
continuous dynamics, we frame lifelong human epigenetic aging as a trajectory
inference problem across discrete age snapshots derived from widely available
cross-sectional data. We introduce a two-stage computational pipeline: first, an
age-regularized Variational Autoencoder (VAE) maps high-dimensional CpG profiles onto a chronologically ordered latent manifold while preserving a generative
decoder bridge back to the original methylation space. We introduce a two-stage computational pipeline: first, an
age-regularized Variational Autoencoder (VAE) maps high-dimensional CpG profiles onto a chronologically ordered latent manifold while preserving a generative
decoder bridge back to the original methylation space. Second, we model the
continuous movement across this latent space via Regularized Unbalanced Optimal Transport (RUOT) that unifies deterministic drift, random diffusion, and
non-conservative mass changes.Where the paper says it
Where the paper says it
How the training signal is built
The age-regularized VAE architecture is optimized via a composite loss function LVAE which includes a reconstruction loss (Lrecon), a Kullback-Leibler divergence (LKL), and an age penalty term (Lage) <ref:2607.06583#pg2> The Lage term represents the Mean Squared Error of the predicted chronological age derived from the latent vector z <ref:2607. p. 2 (3) Latent aging dynamics are modeled using regularized optimal transport, inferring the continuous
velocity and growth fields that describe population-level epigenetic drift across the aging manifold. encoder qϕ(z|x), a generative decoder pθ(x|z), and the regressor fψ(z)—is optimized via a composite
loss function:
LVAE = Lrecon + βLKL + γLage (1)
where Lage represents the Mean Squared Error (MSE) of the predicted chronological age derived from
the latent vector z. Following the β-VAE framework [23], the penalty β enforces smoothness and structural
disentanglement within the latent distribution, while the age penalty γ forces this latent space to
align monotonically with biological time. This joint regularization ensures that our discrete crosssectional cohorts form a continuous, age-aware progression suitable for downstream optimal transport
modeling.Where the paper says it
Where the paper says it
Where the gains come from
The core of the dynamic inference is achieved through DeepRUOT, which models continuous aging dynamics by parameterizing a velocity field (vθv), a growth field (gθg), and a score function (sθs) <ref:2607.06583#pg2>. The energy loss term LEnergy utilizes Monte Carlo sampling and a neural ODE solver to optimize the dynamic energy penalty derived from the Fisher regularization form <ref:2607. p. 3 This converts the underlying SDE system into an equivalent deterministic path-mapping system parameterized by three independent neural networks: a velocity field
vθv
(z, t), a growth field gθg
(z, t), and a scaled log-density score network sθs
(z, t) = σ
2
(t)
2
log p(z, t). The entire system is trained end-to-end by minimizing a composite loss function over pseudo-time
t ∈ [0, T] across K discrete age cohorts:
LRUOT = LEnergy + λrLRecons (3)
Energy Loss Utilizing Monte Carlo sampling and a neural ODE solver along the characteristic
paths dz
dt = vθv
(z, t), the dynamic energy penalty derived from the Fisher regularization form is
optimized via:
LEnergy = Ez0∼p0
Z T
0
1
2
∥vθv
∥
2
2 +
1
2
∥∇zsθs
∥
2
2 −
σ
2
(t)
2
+ sθs
gθg
−
(σ
2
(t))′
σ
2(t)
sθs + αΨ(gθg
)
wθg
(t) dt
(4)
where wθg
(t) = exp R t
0
gθg
(z(s), s) ds
acts as a dynamic particle weight tracker. The entire system is trained end-to-end by minimizing a composite loss function over pseudo-time
t ∈ [0, T] across K discrete age cohorts:
LRUOT = LEnergy + λrLRecons (3)
Energy Loss Utilizing Monte Carlo sampling and a neural ODE solver along the characteristic
paths dz
dt = vθv
(z, t), the dynamic energy penalty derived from the Fisher regularization form is
optimized via:
LEnergy = Ez0∼p0
Z T
0
1
2
∥vθv
∥
2
2 +
1
2
∥∇zsθs
∥
2
2 −
σ
2
(t)
2
+ sθs
gθg
−
(σ
2
(t))′
σ
2(t)
sθs + αΨ(gθg
)
wθg
(t) dt
(4)
where wθg
(t) = exp R t
0
gθg
(z(s), s) ds
acts as a dynamic particle weight tracker. Reconstruction Loss To align the continuous trajectories with unnormalized empirical distributions
at subsequent age checkframes, the reconstruction loss unifies local mass matching and normalized
transport:
LRecons = λmLMass + λdLOT (5)
Here, LMass =
PK−1
k=1 Mk computes a local mass matching error using a localized cardinality
mapping hk between observed cohort data points Ak and their nearest simulated counterparts
Aˆ
k.Where the paper says it
Where the paper says it
Empirical Verification of Aging Archetypes
The model successfully reconstructs and verifies distinct biological archetypes by analyzing the kinetic properties of individual continuous curves βi(t) across the simulated lifespan <ref:2607.06583#pg2> The four primary aging archetypes identified are Linear Accumulators, defined by max(βi(80) − βi(0)), Linear Decay, defined by min(βi(80) − βi(0)), Exponential & Late-Onset Drift, defined by the maximum absolute curvature (second derivative), and Age-Invariant Maintenance Sites, defined as the minimum temporal variance <ref:2607. p. 6 • Exponential & Late-Onset Drift: Defined by the maximum absolute curvature (second
derivative) across the trajectory:
max
t
d
2βi
dt2
≈ max
k
|βi(tk+1) − 2βi(tk) + βi(tk−1)| (6)
This class captures markers that remain stable during early and mid-life but undergo sudden,
accelerated divergence. These sites are key molecular indicators of the mathematical
inflection point identified in our growth-rate analysis, serving as a sensitive proxy for the
onset of late-life biological entropy and the non-linear accrual of methylomic variability
[31, 11]. • Application of Unbalanced Trajectory Inference: We deploy the DeepRUOT framework
across this aging manifold, demonstrating that its unbalanced formulation naturally captures the localized mass adjustments required to simulate late-life entropic expansion and
survivorship bias across the discrete age cohorts. • Empirical Verification of Aging Archetypes: We validate the pipeline on a large-scale pantissue dataset, decoding continuous latent paths to successfully reconstruct and empirically
verify distinct biological archetypes: linear clocks, stable maintenance sites, and accelerated
late-onset drift.Where the paper says it
Where the paper says it
Quantitative Evaluation of Dynamics
Quantitative evaluation uses Wasserstein-1 (W1) distance to measure structural accuracy and Total Mass Variation (TMV) to quantify localized density adjustment <ref:2607.06583#pg2> The model yields relatively stable W1 distances across held-out experiments, indicating smooth interpolation <ref:2607. p. 4 Biologically, TMV acts as a mathematical proxy for population-level divergence,
tracking how much the variance of a cohort expands or disperses due to biological noise. As shown in Table 1, the model yields relatively stable W1 distances across all hold-out experiments,
indicating that the learned vector fields can smoothly interpolate cross-sectional epigenetic distributions across unobserved temporal gaps without catastrophic divergence. As shown in Table 1, the model yields relatively stable W1 distances across all hold-out experiments,
indicating that the learned vector fields can smoothly interpolate cross-sectional epigenetic distributions across unobserved temporal gaps without catastrophic divergence. Crucially, while W1 remains
stable, the TMV reveals a distinct shift in aging dynamics, as it is minimal during early-to-mid-life
transitions (0. This indicates a period of stable, predictable biological progression where individuals age along
tightly bundled paths without extreme deviation. However, as trajectories progress into late-life states,
the model exhibits a prominent, localized surge in the growth term.Where the paper says it
Where the paper says it
Where the paper says it
Biological Characterization of Epigenetic Archetypes
The reconstructed continuous paths are mapped back to individual CpG sites to generate continuous aging curves βi(t) for every site <ref:2607.06583#pg2>. These curves are categorized into four distinct archetypes, including Linear Accumulators which exhibit monotonic, steady gain in methylation <ref:2607. p. 6 We define four
primary aging archetypes (Figure 3):
• Linear Accumulators (Hypermethylation): Defined by the maximum positive shift over
the lifespan, max(βi(80) − βi(0)). These sites exhibit a monotonic, steady gain in methylation. • Linear Decay (Hypomethylation): Defined by the maximum negative shift over the
lifespan, min(βi(80) − βi(0)). These markers display a gradual, global loss of methylation. • Exponential & Late-Onset Drift: Defined by the maximum absolute curvature (second
derivative) across the trajectory:
max
t
d
2βi
dt2
≈ max
k
|βi(tk+1) − 2βi(tk) + βi(tk−1)| (6)
This class captures markers that remain stable during early and mid-life but undergo sudden,
accelerated divergence. These sites are key molecular indicators of the mathematical
inflection point identified in our growth-rate analysis, serving as a sensitive proxy for the
onset of late-life biological entropy and the non-linear accrual of methylomic variability
[31, 11]. They represent the static, robust "housekeeping" component of the
epigenome, which is strictly preserved to maintain essential cellular functions and heavily
protected against the stochastic noise that otherwise drives epigenetic drift in unprotected
regions. The divergence in these archetypes highlights that epigenetic aging is not a uniform process, but a
composite of distinct kinetic behaviors.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Limitations and Future Directions
Limitations include the coarse temporal constraint introduced by multi-decade age discretization, which may smooth over rapid life-stage transitions <ref:2607.06583#pg2>. The universal pan-tissue strategy also suppresses localized, tissue-specific metabolic kinetics <ref:2607. p. 7 However, limitations remain: our
multi-decade age discretization introduces a coarse temporal constraint that may smooth over rapid
life-stage transitions (e. , puberty or menopause), while the universal pan-tissue strategy deliberately
suppresses localized, tissue-specific metabolic kinetics. Furthermore, while these analyses offer
novel insights into normal tissue dynamics, it remains to be explored how these trajectories are
rewired in pathological settings, particularly in cancer, where DNA methylation instability frequently
drives pathways linked to oncogenic transformation and progression. These boundaries highlight
compelling avenues for future research. (Bottom-Right) Age-invariant maintenance loci. coupling an age-regularized VAE with the DeepRUOT framework, our pipeline bridges macroscopic
population phenomena with microscopic molecular kinematics.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Continuous Trajectory Inference for Epigenetic Aging: The system can simulate how an entire profile continuously changes over time rather than outputting a single score, allowing for a
generative paradigm for simulating human molecular aging.
-
Modeling Non-Conservative Population Dynamics: By using Regularized Unbalanced Optimal Transport (RUOT), the improved system can accommodate
population-level density shifts like survivorship bias and cellular attrition without requiring rigid biological priors,
which standard mass-preserving optimal transport cannot handle. -
Discovery of Biological Aging Archetypes: The model can reconstruct and empirically verify distinct biological aging archetypes, such as
Linear Accumulators (Hypermethylation),
Linear Decay (Hypomethylation),
and "Exponential & Late-Onset Drift," by analyzing the kinetic properties of individual continuous curves. -
Resolving Stochastic Epigenetic Drift: The system can identify the
prominent late-life surge in the learned growth field that mathematically captures the variance expansion driven by stochastic epigenetic drift,
modeling how older populations become more diverse due to accumulating random methylation errors.
Abstract
DNA methylation (DNAm) serves as one of the most robust molecular biomarkers of biological aging. While conventional epigenetic clocks accurately predict chronological age from high-dimensional CpG profiles, they treat aging as a static regression task, meaning they can only output a single score rather than simulating how an entire profile continuously changes over time. To reconstruct these continuous dynamics, we frame lifelong human epigenetic aging as a trajectory inference problem across discrete age snapshots derived from widely available cross-sectional data. We introduce a two-stage computational pipeline: first, an age-regularized Variational Autoencoder (VAE) maps high-dimensional CpG profiles onto a chronologically ordered latent manifold while preserving a generative decoder bridge back to the original methylation space. Second, we model the continuous movement across this latent space via Regularized Unbalanced Optimal Transport (RUOT) that unifies deterministic drift, random diffusion, and non-conservative mass changes. By resolving this RUOT formulation using the DeepRUOT framework, our model fluidly accommodates population-level density shifts like survivorship bias and cellular attrition without requiring rigid biological priors. Evaluated on a large-scale, 80-year pan-tissue dataset, our model demonstrates robust distribution interpolation and uncovers a prominent late-life surge in the learned growth field that mathematically captures the variance expansion driven by stochastic epigenetic drift. Finally, by decoding continuous latent paths back to individual CpG sites, we reconstruct and empirically verify distinct biological aging archetypes, offering a rigorous, generative paradigm for simulating human molecular aging.
Sources
- Towards a mathematical theory of trajectory inference
- A survey of the Schr\"odinger problem and some of its connections with optimal transport
- Adam: A Method for Stochastic Optimization
Related papers
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
- CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
- Probing and steering biology across Boltz-1s trunk-diffusion boundary
- Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation
- Molecular notation can act as a shortcut in AI benchmark evaluation
- Towards Efficient Auto-Interpretability of Boltz-1 Sparse Autoencoders