Abstract
Abstract
Cell-type atlases now catalogue hundreds of distinct human cell identities across development, each annotated with mitochondrial abundance, dominant fuel pathway, and lineage of origin, yet no computable model connects the topology of the lineage tree to the metabolic rate the resulting organs display in the adult body. Using a curated 396-node human cell-lineage tree spanning the zygote to terminal somatic identities, we test whether organ-level standard metabolic rate (SMR) is better predicted by developmental time (lineage depth) or by terminal fate identity (anatomical compartment). We find that lineage depth explains essentially none of the variance in a cell’s metabolic tier (r = 0.11, R
2
≈ 1.2%), whereas compartment identity explains roughly three-quarters of it (≈75%), and that the mean mitochondrial volume fraction of an organ’s constituent terminal cell types tracks the organ’s classical literature-derived SMR with r = 0.90 across five canonical reference-man organ groups. These results motivate a five-layer computable framework — lineage as a formal tree grammar, information-theoretic structure of the tree, metabolic switches as runtime operators, compilation to organ-level energetic networks, and execution as organism-level bioenergetics — and a specific falsifiable hypothesis we call the metabolic commitment-horizon model: that a cell’s metabolic tier is fixed at a discrete lineage-commitment event and thereafter held constant, rather than accumulated continuously with differentiation time. In a worked held-out test (5-fold cross-validation, n = 396), a nearest-commitment-ancestor predictor beat a global-mean null (MAE 0.752 vs 0.809, p = 0.015) but was not significantly better than a lineage-depth regression (ΔMAE = −0.042, 95% CI −0.098 to +0.012, p = 0.175), so we report the commitment-horizon model as supported descriptively but not yet discriminated predictively on these data. We then connect this within-species result to the classical, cross-species rate-of-living hypothesis, in which breathing rate scales predictably with body mass (Stahl’s allometric equation, fR = 53.5·M
−0·26
) yet fails to predict lifespan on its own: naked mole rats, little brown bats, humans, and bowhead whales all live several-fold longer than mass alone predicts (longevity quotients of 4.4, 5.1, 4.6, and 2.8 respectively), for documented reasons unrelated to a slower pace of living. We then treat insulin resistance as an acquired perturbation of the clock, and find that insulin responsiveness is orthogonal to mitochondrial density: adipose and skeletal muscle, the two lowest-mitochondrial-density tissues in the organ set, are the principal sites of insulin-stimulated glucose disposal, while brain, the second-highest by per-kilogram rate, is essentially insulin-independent. We survey machine learning and deep learning tools — graph neural networks, single-cell foundation models (Geneformer, scGPT, scVI), and RNA velocity — already available to extend or directly test this framework on primary single-cell data. We report the supporting computations transparently, flag which values in the underlying dataset are measured versus estimated or drawn from the literature, and close with eight testable predictions and their falsification conditions.