BioEdge-Rx: From Regenerative Control States to Biologically Compilable Therapeutic Policies for Optic Nerve Repair

Background/Objectives: Distinct responses to optic-nerve injury motivate multidimensional, state-dependent control rather than a universal regeneration score. We tested whether public perturbational transcriptomes could define interpretable states and support computational policy compression. Methods: Five public mouse datasets were analyzed using six prespecified biological modules. Zero-preserving transformed treatment–control effects defined nine perturbation anchors. A 16-region engineering simulator was used to retrain a gradient-boosting student and select an explicit decision tree in disjoint simulated episode cohorts. Results: Longitudinal treatment-response correlations were 0.177–0.406, while same-time PTEN-derived interventions remained non-identical (pairwise Spearman rho 0.670–0.787). A newly trained student attained 85.67% action-label accuracy and macro-F1 0.598. A depth-3 tree (eight leaves, six features) retained 98.59% and 98.62% of the recalibrated MPC objective in 120 independent Standard and 120 OOD episodes, with failure-rate differences of +3.33 and +0.83 percentage points. The 95% bootstrap interval for the Standard failure increase exceeded the five-percentage-point selection limit. Greedy one-step control performed similarly to or slightly better than the approximate MPC. Additional post-review OOD simulations revealed sensitivity to episode sampling and alternative engineering dynamics, limiting the generalizability of the initial benchmark estimates. Conclusions: BioEdge-Rx demonstrates computational compression of a state-dependent policy within an assumed simulator. It does not establish causal treatment transitions, biological efficacy, sensor-level realizability, or superiority of multistep planning.