2026-10-27

In the developing brain, neural circuits are elaborated by the activity-dependent rewiring of synaptic connections. Our earlier self-organization model, which was built on the assumption that axon terminals compete with each other to form synaptic contacts to single dendritic spines, successfully reproduced spatiotemporal receptive fields and orientation-direction joint maps (Tanaka et al., 2020). It has been reported that most synaptic contacts are en passant types rather than axonal terminals (Turegano-Lopez et al., 2024). This finding contradicts our assumption that each axon terminal selects an appropriate dendritic spine. Recent microscopic observations have shown that dendritic spines are more motile than presynaptic boutons (Bonhoeffer and Yuste, 2002; Deng and Dunaevsky, 2005). In addition, the current view of synaptic rewiring in early development is that filopodia spontaneously generated along dendrites transform into spines by establishing contacts with neighboring axonal boutons to form synapses whereas spines of inappropriate synapses are eliminated in an activity-dependent manner (Deng and Dunaevsky, 2005). In this theoretical study, we mainly focus on the processes of spine generation, maturation, and elimination to rebuild a mathematical model of synaptic rewiring, incorporating recent experimental findings on actin polymerization and depolymerization (Ridley, 2006) under the activation of Rho-GTPases (Mackay and Hall, 1998) mediated by the binding of variant forms of brain-derived neurotrophic factor (BDNF) (Binder and Scharfman, 2004; Lesmann and Brigadski, 2009; Kuczewski et al., 2009) to neurotrophin receptors such as tropomyosin kinase receptor B (TrkB) (Klein et al., 1989) and the low-affinity-nerve growth factor receptor (p75NTR) (Yamashita et al., 1999). TrkB receptors are dephosphorylated by chondroitin sulphate proteoglycans (CSPGs) (de Vivo et al., 2013) mediated by protein-tyrosine phosphatase sigma (PTPsigma (Kurihara and Yamashita, 2012) in an age-dependent manner. The model describes the generation or elimination of dendritic spines of visual cortical neurons under actin polymerization and depolymerization to establish appropriate synaptic contacts on afferent fibers of cells in the lateral geniculate nucleus. The simulations based on the new model reproduced not only simple-cell-like receptive fields (DeAngelis et al., 1999) and orientation-direction maps (Shmuel and Grinvald, 1996) but also a characteristic profile of the total number of synapses, which shows an early increase and a delayed decrease during the critical period (Li et al., 2006; White and Fitzpatrick, 2007). It is also demonstrated that whereas orientation selectivity is elevated almost monotonically throughout the critical period, the formation of direction selectivity is delayed behind the formation of orientation selectivity, which is likely promoted by the pruning of inappropriate synapses (Grigonis et al., 1988; Bourgeois et al., 1993; Elston et al., 2009; Faust et al., 2021).

Semantic knowledge encompasses facts, concepts, and their relationships. Recent work has proposed that some forms of semantic knowledge could be encoded in low-dimensional neural representations, such as the abstract, disentangled representational geometries observed in many cognitive brain areas. Experimental work has shown that hippocampal (HPC) lesions impair the formation of flexible and generalizable semantic knowledge. However, HPC representations are often high-dimensional and sparse. How can these representations contribute to the formation of low-dimensional abstract representation across these cognitive cortical areas? Here, we reconcile these observations using multiregional neural network models. Modeling HPC as a sparse, decorrelating autoencoder, we show that its high-dimensional representations promote the robust and efficient formation of low-dimensional abstract representations in downstream regions such as prefrontal cortex. This effect is most pronounced when the geometry of sensory inputs is misaligned with the geometry of desired outputs, requiring a more complex input-output transformation. Mathematical analysis shows that hippocampal expansion recoding suppresses misleading low-dimensional input structure and reshapes the loss landscape of downstream networks, allowing abstract semantic geometry to emerge reliably. The resulting abstract representations provide an inductive bias for sample-efficient learning of novel semantic-to-action mappings. Finally, we show that consolidating hippocampus-dependent representations into a hippocampus-independent pathway improves out-of-distribution generalization based on input structure. Together, these results suggest a circuit mechanism by which high-dimensional hippocampal representations can support, rather than oppose, the emergence of low-dimensional abstract representations.