A Hybrid Machine Learning Framework for Mechanistically Interpretable Latent Parameter Inference in a Spatiotemporal CAR-T Therapy Model for Solid Tumours

CAR-T cell therapy remains ineffective in most solid tumours because effector cells infiltrate poorly, undergo exhaustion, and face antigen escape within an immunosuppressive microenvironment. To address this, we developed a hybrid framework that combines a mechanistic spatiotemporal model with machine learning for limited individual-level mechanistic personalisation under data constraints. At its core, we employed a reaction–diffusion–chemotaxis model describing functional and exhausted CAR-T cells, antigen-positive and antigen-negative tumour subpopulations, a chemoattractant, an immunosuppressive factor, and hypoxia. Gradient boosting combined with nested cross-validation was used to recover model-consistent latent-parameter pseudo-labels generated by a limited inverse problem. Within this surrogate-target setting, parameters characterising the tumour microenvironment and CAR-T cell exhaustion were reproduced most robustly, whereas antigen escape and individualised initial conditions were substantially less well constrained. As an auxiliary reference point, we also considered a direct empirical baseline for binary clinical outcomes. This baseline indicated that the observed clinical features contained a more stable signal for disease control than for objective response. A favourable response was associated with high CAR-T cell infiltration and cytotoxic potency, whereas resistance was linked to exhaustion, antigen escape, and a suppressive microenvironment. Overall, the proposed approach should be interpreted as an internally validated, hypothesis-generating proof-of-concept platform for mapping clinical features to mechanistically interpretable surrogate latent targets, rather than as evidence for validated recovery of true patient-specific biological parameters.