Inhibition-Biased Pruning and Cognitive Reserve Amplification in Bipolar Disorder: A Computational Framework with Insights into Glutamatergic Therapeutics

Background: Bipolar disorder (BD) and major depressive disorder (MDD) share genetic overlaps in synaptic and immune pathways, yet differ markedly in clinical course—persistent deficits in MDD versus episodic instability in BD. Emerging evidence implicates excessive synaptic pruning as a core mechanism, with recent genetic analyses suggesting inhibition-biased pruning and cognitive reserve amplification drive BD-specific manic vulnerability. Computational models are needed to mechanistically test these hypotheses and evaluate plasticity-enhancing treatments, such as ketamine, which carry differential risks across disorders. Methods: We developed a gated recurrent unit (GRU)-based neural network simulation to model pruning phenotypes. MDD was induced via severe (95%), unbiased magnitude pruning; BD variants used moderate sparsity (75–85%) with inhibition bias (1.3–2.0). Cognitive reserve was modulated as a hidden-state scalar, and stress as additive noise. Plasticity treatment was simulated through iterative gradient-guided regrowth (0.4 fraction) and fine-tuning (up to 10 cycles). Outcomes included accuracy, stress resilience, excitation/inhibition (E/I) ratios, and instability metrics (variance under sustained drive). Relapse vulnerability was tested via additional pruning (depressive) or reserve amplification (manic). Results: Post-pruning, MDD networks showed profound accuracy collapse under stress (27.1% resilience) with balanced E/I (1.02) and low manic variance. BD phenotypes retained baseline accuracy and superior stress resilience (77.9–86.7%) but exhibited reduced E/I (0.09–0.32) and elevated manic variance. Chronic treatment restored full function across phenotypes, with BD models showing progressive E/I normalization (e.g., 0.22 to 0.77 in BD-classic). Acute treatment prevented depressive relapse but transiently increased manic variance; chronic regimens provided dual protection. Conclusions: This recurrent simulation supports a unified pruning framework where inhibition-biased moderate pruning yields "pruned-but-potent" circuits prone to manic escalation under reserve load, distinguishing BD from MDD's capacity-loss phenotype. Chronic, iterative plasticity enhancement emerges as superior for BD stability, offering mechanistic rationale for cautious use of rapid-acting glutamatergic agents and emphasizing maintenance therapy.