Multiscale Generative Modelling of Cellular Senescence and Extracellular Matrix Degeneration: From Molecular Tokenization to Verifiable Tissue-Scale Predictions

Cellular senescence and the degeneration of the extracellular matrix (ECM) are coupled drivers of tissue ageing, yet the causal chain that connects a molecular lesion to a tissue-scale outcome spans at least four orders of organisational magnitude and is rarely modelled as a single object. We present a framework that threads five computational layers into one pipeline: (i) a molecular tokenization layer that represents real ECM and senescence-associated protein sequences as subword tokens and learns a distributional embedding; (ii) a molecular-mechanism layer that models matrix metalloproteinase (MMP)-driven collagen degradation as Michaelis-Menten ki-netics with competitive inhibition by tissue inhibitors of metalloproteinases (TIMPs); (iii) a cellular layer in which paracrine SASP signalling produces a bistable senescence switch with hysteresis; (iv) a tissue layer in which a two-dimensional reaction-diffusion system generates a travelling senescence front that collapses ECM stiffness in its wake; and (v) a generative layer that learns a low-dimensional manifold of coupled multiscale trajectories and samples new virtual tissues. Crucially, we add a verifiability layer: a probabilistic predictor with a deep ensemble and post-hoc recalibration that turns point predictions into certified credible intervals. Every figure is computed from either real sequence data (12 UniProt entries, 7,720 residues) or from the models themselves; no result is hand-drawn. On a held-out set the recalibrated predictor reduces the expected calibration error from 0.161 to 0.065 while retaining a point-prediction RMSE of 0.022. We argue that verifiability-calibrated, uncertainty-aware prediction-is the missing ingredient that would let generative multiscale models of ageing be trusted, tested and falsified.