Aging as a dissipative dynamical process

Khodaee, Zandie, Xia & Edelman (MIT; arXiv:2504.13044, npj Aging 2025) model a biological dynamical system (BDS) on a high-dimensional gene-expression state space and propose that aging is dominated by dissipative (non-conservative) dynamics. The framing is explicitly dynamical-systems-theoretic rather than equilibrium-thermodynamic: dissipation here means phase-space contraction / non-recurrence, not (directly) heat export. This is the most concrete computational realization yet of the dissipative-structure intuitions of atoms 5–7 — and the closest published cousin to trajectory-based ("trend") hypotheses of aging.

Invoking the Hopf decomposition, the state space splits into a conservative/recurrent set (Poincaré-recurrent: orbits return arbitrarily close to initial conditions) and a dissipative set (wandering points that escape and do not return). The thesis: aging = growing predominance of the dissipative component, yielding deviation from recurrent states and monotone entropy increase along trajectories. Note the conceptual alignment with the level/trend distinction (atom 7): this is a trajectory claim about the qualitative character of the flow (recurrence loss), not merely a level of some scalar — though here dissipation is dynamical (phase-space), and bridging it to thermodynamic σ remains, as ever, the open task (atom 10).

Lacking closed-form BDS equations, they employ a transformer-based multimodal foundation model over single-cell transcriptomics, treating genes and metadata (including chronological age as a token) as inputs and producing per-token embeddings. The learned embedding geometry serves as a surrogate for the system's state and flow, enabling estimation of age-conditioned dissipation signatures without an explicit vector field — a neural-surrogate approach to dynamical-system identification. This inherits both the power (captures nonlinear, tissue-specific structure) and the opacity (embedding-space quantities are model-dependent) of foundation models.

The CAM is constructed from cell-type- and tissue-resolved embeddings across aging phases. Reported signatures: (1) divergence in gene-embedding space (increasing dispersion — echoing the dysregulation/Mahalanobis picture of atom 14, now in a learned latent space); (2) nonlinear transitions (non-uniform, possibly tipping-point-like dynamics — cf. bifurcations, atom 6); (3) tissue- and cell-type-specific entropy variation. The CAM yields molecular-resolution, tissue-specific aging quantification grounded in a dissipative-dynamics rationale rather than supervised age regression — a meaningfully different construction from the clocks of atom 13.

Strengths: an explicit, computable dynamical-systems formalization of the dissipation intuition, with molecular resolution and signatures consistent with atoms 6/14. Caveats: (1) dissipation here is phase-space contraction / loss of recurrence in a learned embedding, not measured thermodynamic entropy production — the identification with physical σ is asserted by analogy, not derived (atom 10); (2) ergodic/recurrence claims on a neural surrogate inherit the model's inductive biases and training distribution; (3) "age as a token" risks encoding chronological-age correlates rather than causal dynamics (cf. clock critique, atom 13); (4) no derivation yet of the macroscopic hazard law (atom 1) from the CAM. It is, nonetheless, the closest published instantiation of a trajectory/dissipation account — precisely the family in which trend-type hypotheses (the sign of a dissipation measure's change) would be formulated and tested.