Mechanistically Constrained Latent-State Modeling for Therapeutic Intervention Prioritization in Pediatric Diffuse Midline Glioma

Pediatric diffuse midline glioma and diffuse intrinsic pontine glioma are lethal childhood brain tumors with few effective therapeutic options. We introduce a mechanistically constrained framework that learns a chloride-vulnerability state from single-cell transcriptomics, transfers that state to unpaired spatial transcriptomics, and evaluates candidate interventions through an explicit chloride homeostasis model. SLC12A2 and SLC12A5 are excluded from encoder inputs, encouraging the model to capture broader chloride-associated transcriptional structure. Across 4,058 H3K27M glioma cells, a frozen latent probe reached an AUROC of 0.674 and an average precision of 0.548. An independent chloride/GABA gene-set score remained positively associated with GPI. The calibrated ODE reproduced intracellular chloride states of 52.0 mM at baseline and 14.0 mM under complete NKCC1 block. Validation GPI Spearman reached 0.4876 while normalized ODE inconsistency fell from 0.9591 to 0.7865 at the selected checkpoint. Mechanism-controlled sweeps produced monotonic chloride reductions for bumetanide and torsemide. Together, these results connect molecular state, tissue context, and transporter physiology in one framework for prioritizing chloride-directed interventions in a rare pediatric cancer