BioEdge-RGC: Compact Neural Policy Transfer from Model-Predictive Control in a Transcriptomics-Informed Optic Nerve Injury Simulation

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Background/Objectives: Optic nerve injury involves heterogeneous neuronal and microenvironmental responses that may require sequential, state-dependent intervention. BioEdge-RGC was developed as a transcriptomics-informed computational benchmark for transferring a globally planned model-predictive-control policy to compact regional neural models operating with incomplete observations and reduced communication. Methods: Public mouse optic-nerve-crush transcriptomic datasets were summarized into seven retinal ganglion cell modules and eight retinal environment modules and fused into a 15-dimensional temporal reference state for a 16-region, five-step simulator. A candidate-constrained MPC expert was approximated by a central neural teacher and local, neighbour-aware, and coordinated regional students trained using hybrid knowledge distillation or direct MPC supervision. Missing observations, severe simulated injury, INT8 quantization, local parameter perturbations, and an independent GSE229033 transcriptomic plausibility assessment were evaluated. Results: The central teacher retained 99.90% of the MPC objective, and the coordinated distilled student retained 99.69% of teacher performance while reducing modeled communication from 5520 to 1080 bytes per episode. The coordinated-versus-local mean-objective difference was +0.000053 and did not meet the predefined +0.008 threshold; direct MPC supervision performed at least as well as hybrid distillation. The principal objective-based conclusions were preserved across all nine local parameter conditions, whereas the coordinated failure-rate advantage was not parameter-robust. In GSE229033, six of seven modules changed in the predefined favorable direction, but the oriented composite bootstrap interval included zero. Conclusions: Compact regional neural policies reproduced MPC performance with minimal objective loss and substantially lower modeled communication. The external transcriptomic analysis provided limited plausibility support for the RGC state orientation, while neither analysis validated simulator dynamics, controller efficacy, biological treatment effects, or clinical applicability.