Pathway-specific short-term synaptic dynamics and lateral inhibition shape frequency-dependent input integration and population-level pattern separation in the dentate gyrus

The dentate gyrus (DG) decorrelates overlapping entorhinal inputs into distinct granule-cell representations, a computation central to pattern separation and to reducing interference in episodic memory. The three excitatory pathways to granule cells -- the lateral perforant path (distal dendrites), the medial perforant path (middle dendrites), and proximal inputs -- carry distinct short-term synaptic dynamics, but how these combine with lateral inhibition to set the frequency dependence of DG integration and pattern separation remains unclear. Here we integrate mechanism, function, and robustness into a single computational modeling study spanning three complementary model tiers, using Tsodyks-Markram short-term synaptic parameters grounded in prior slice electrophysiology. In a biophysically detailed 37-compartment granule-cell model (Tier 1), the distal (lateral) pathway facilitates at low frequency, the middle (medial) pathway depresses, and the proximal pathway is mixed, producing frequency- and pathway-dependent integration; three-pathway summation is mildly sublinear, and a direct granule-cell-to-granule-cell lateral inhibition -- a shunting connection emulating disynaptic feedforward inhibition without an explicit interneuron -- further attenuates it. In a reduced leaky-integrate-and-fire network with the same dynamics (Tier 2), pattern separation is frequency-dependent, rising to a gamma-band maximum at 40 Hz that is reproducible across independently wired networks, whereas the basket-cell-inhibition magnitude varies with the random connectivity. In a three-layer population network (Tier 3), pattern separation is robust: although single granule-cell spike counts are highly sensitive to input noise, the population-level separation code is nearly noise-invariant (a roughly 60-fold dissociation), and separation is governed by the magnitude of local lateral inhibition rather than its targeting. Two claims that hold at the single-cell scale -- a microsecond spike-timing-precision requirement and an advantage of finely targeted inhibition -- do not survive at the network scale. Pathway-specific synaptic dynamics and lateral inhibition thus shape frequency-dependent integration and noise-robust population pattern separation in the DG. Author SummaryThe dentate gyrus (DG) performs pattern separation: it takes overlapping cortical inputs and makes their DG representations more distinct, a computation thought to reduce memory interference. How does the DG do this? We approach the question across three scales in a single computational study. First (mechanism), we show in a biophysically detailed granule-cell model that the three anatomical input pathways carry different short-term synaptic dynamics -- the distal (lateral perforant path) input facilitates, the middle (medial perforant path) input depresses, and the proximal input is mixed -- so that the cells response depends on input frequency and pathway. Second (function), in a reduced network model we show that these dynamics, combined with lateral inhibition, tune pattern separation in a frequency-selective way. Third (robustness), we find that although a single granule cells spike count is highly sensitive to input noise, the population-level separation code is nearly noise-invariant -- a "noise paradox" in which population coding rescues what is fragile at the single-cell level. We also show that two claims that appear at the single-cell scale -- a microsecond spike-timing-precision requirement, and an advantage of finely targeted inhibition -- do not survive at the network scale: what matters is the amount of local inhibition, not how it is distributed. The synaptic parameters are grounded in prior slice recordings; the model integrates the mechanism.