Simulating Synaptic Pruning and Ketamine-Like Recovery in Depression: Insights from Consolidation Duration and Iterative Regimens on Resilience and Relapse

A computational model of major depressive disorder shows that excessive synaptic pruning during development leaves neural networks brittle, reducing accuracy under noise to about 32% when 95% of connections are removed. A single regrowth cycle that restores half the lost weights, followed by extended consolidation (15–20 epochs of fine-tuning), returns accuracy to roughly 97% and eliminates relapse losses from a second pruning challenge. Repeated smaller regrowth cycles further reduce residual sparsity below 1%, lift extreme-stress accuracy 9–11 points beyond single repair, and provide strong protection against relapse. These simulations suggest that lasting stability depends on both the duration and repetition of plasticity, explaining the clinical advantage of maintenance or multi-dose ketamine protocols.

Study at a glance

Abstract

Background: Major depressive disorder (MDD) is increasingly framed as a failure of neuroplasticity. Excessive synaptic pruning in adolescence can leave circuits fragile; later stress then unmasks the weakness, whereas ketamine and related compounds can quickly rebuild critical synapses. Clinical data show that a single treatment often gives short relief, while repeated courses reduce relapse. To clarify the mechanisms, we expanded an existing pruning-plasticity model to test how the length of post-growth consolidation and the choice between one-off and iterative synaptogenesis shape long-term resilience. Methods: A fully connected feed-forward network with roughly 396 000 weights learned a four-class Gaussian task that included input noise. After training, 95 % of the weights were removed to mimic excessive adolescent pruning. Robustness was gauged under internal activation noise (σ up to 2.5) and further input jitter. Recovery used gradient-guided regrowth that reinstated half of the lost weights—the ones promising the largest loss reduction—representing BDNF/mTOR-driven synaptogenesis. We varied fine-tuning time after regrowth from 0 to 20 epochs to model different consolidation windows. Separate experiments compared single, large regrowth episodes (60 % or full restoration) with chronic schedules of 3–10 smaller cycles (each restoring 40 % of the still-missing weights with brief tuning). A second pruning wave of 40 % served as a relapse challenge. Performance was tracked under combined input and internal noise (σinput = 1.0, σinternal = 0.5). Results: Cutting 95 % of the weights drove accuracy under joint stress down to about 32 %, with a steep failure point once sparsity exceeded 93 %. One regrowth cycle followed by consolidation returned accuracy to roughly 97 % even though nearly half of the synapses were still absent. Extending consolidation raised extreme-stress performance by up to 55 percentage points and all but eliminated relapse losses when fine-tuning lasted 15–20 epochs. Iterative regrowth drove residual sparsity below 1 %, lifted extreme-stress accuracy a further 9–11 points beyond single-cycle repair, and provided strong protection against the second pruning hit—often matching or outdoing full, one-step restoration because connectivity was refined as well as increased. Conclusions: The simulations support a pruning-mediated plasticity deficit model of MDD: over-elimination during development leaves networks brittle, but targeted synaptogenesis can hide the deficit. Lasting stability, however, depends on how long and how often plasticity is engaged. Extended consolidation and repeated growth cycles give superior stress resilience and guard against relapse, offering an explanation for the clinical advantage of maintenance or multi-dose ketamine protocols and a guide for tailoring plasticity-enhancing treatments.