Modeling Antidepressant-Induced Manic Switch in a Unified Computational Framework: Insights from Ketamine, SSRIs, and Neurosteroids

Background: Major depressive disorder involves impaired neural plasticity, yet antidepressants spanning glutamatergic (e.g., ketamine), monoaminergic (e.g., SSRIs), and GABAergic (e.g., neurosteroids) classes differ in onset, durability, and risk of treatment-emergent mania, particularly in bipolar disorder. Computational models offer controlled comparison of these pathways, but few integrate manic liability. Methods: We extended a magnitude-based pruning model (95% sparsity) of depression in feed-forward networks trained on a Gaussian classification task. From identical pruned states, three interventions were simulated: ketamine-like gradient-guided regrowth (50% reinstatement) with consolidation; SSRI-like prolonged low-rate training with tapering noise and escalating excitability gain; neurosteroid-like global tonic inhibition (0.7× damping, tanh activations, reduced gain). Efficacy was assessed via accuracy under clean, noisy, and combined stress conditions; resilience to graded noise and additional pruning; manic risk via biased positive perturbation and activation magnitude. Results were averaged across 10 random seeds. Results: All treatments restored near-ceiling baseline performance. Ketamine-like regrowth yielded superior extreme-stress resilience (76.8% ± 3.6%) and relapse protection (−0.1% ± 0.2% drop), reducing sparsity to 47.5%. Neurosteroid-like modulation matched acute combined-stress recovery (97.7% ± 0.2%) but showed state-dependence (19% off-modulation loss) and weaker extreme buffering (42.9% ± 1.3%). SSRI-like refinement lagged (90.8% ± 3.1% combined, 50.1% ± 3.7% extreme) with highest relapse vulnerability (8.8% ± 4.3%). Manic proxies ranked SSRI-like highest risk (biased accuracy 47.6% ± 12.8%, gain 1.60), ketamine-like moderate (84.2% ± 8.5%, gain 1.25), neurosteroid-like lowest (50.5% ± 7.4%, effective gain ~0.59). Conclusions: Antidepressants operate via distinct plasticity routes—structural rebuilding (durable, moderate risk), reversible stabilization (rapid, low risk), gradual optimization (vulnerable, high risk)—reproducing clinical trade-offs and supporting mechanism-based selection in unipolar and bipolar depression.