Weak pre-before-post pair biases sort synaptic weights in a conductance-based model of spontaneous network bursting

Abstract

Spontaneous population bursts repeatedly expose synapses to correlated spike activity, but it is unclear whether such events reinforce pre-existing synaptic weight differences when precise firing order carries little information. We asked this in a conductance-based model of a bursting neuronal culture: 48 excitatory and 12 inhibitory stomatogastric-ganglion (STG) model neurons with sparse recurrent connectivity, short-term depression, and pair-based spike-timing-dependent plasticity (STDP), with every prediction and falsifier committed to a version-controlled record before the corresponding run. Across two cohorts of 10 connectivity wirings - realizations of one model, not independent preparations - stronger synapses consistently received greater net potentiation than weaker ones, although the preregistered magnitude criterion was not met. The spike statistic most closely associated with this sorting was not first-spike latency: stronger synapses instead showed a very small but highly consistent excess of pre-before-post spike pairs, 0.2-0.35 percentage points, accumulated over approximately 10^4 pairs per synapse in 60 s (median 8,960 in the out-of-sample cohort; approximately 3 x 10^6 per wiring). In a preregistered intervention, shortening bursts by 41% at approximately matched burst rate increased the sorting measure by about an order of magnitude in all five wirings. A second, physiologically distinct perturbation reproduced the relationship with realized burst duration, but duration could not be separated from per-burst spike count. All interventional evidence is computational. We then preregistered an observational test of the corresponding prediction in 16 archived recordings of developing cortical cultures. The predicted strength-ordered pair excess was not detected, and the predicted negative relationship with burst duration was absent. The former is a bounded null - the recordings exclude effects above +0.32 percentage points but cannot resolve the model's approximately 0.2-point magnitude; the latter was adequately powered under the preregistered criterion. A consistent positive latency-strength relationship, unpredicted by the model, was observed. The study therefore establishes a reproducible sorting phenomenon within a specific computational model and a testable burst-shortening prediction, but does not establish that the mechanism operates in living cortical networks; an intervention on cultured networks with an independently validated measure of synaptic strength is the decisive test.