Ranjani Koushik^1,2 (), Alex Abate³, Frederik Kamps⁴, N. Apurva Ratan Murty^1,2; ¹Cognition and Brain Science, Georgia Institute of Technology, ²Center for Computational Cognition, Georgia Institute of Technology, ³Harvard Medical School, ⁴University of Edinburgh

The parahippocampal place area (PPA) is central to the brain’s scene processing system, yet its finer-grained functional organization has remained easier to describe anatomically than functionally. If the PPA contains functionally distinct subregions, those subregions should reflect distinct computations, require distinct computational models, and be independently dissociable with distinct images. Here, we integrated data-driven analyses of neural response structure, subregion computational models, in silico hypothesis testing, and closed-loop experimental validation to define the roles of PPA subregions in scene processing. We first used data-driven clustering to identify PPA voxel groups with similar responses to a large image set (185 images, 20 repetitions per image, N = 4). Across participants, more than 95% of response variance was captured by just two anterior-posterior clusters, consistent with prior work (Baldassano et al., 2013; Nasr et al., 2013; Cukur et al., 2016). To uncover what functionally distinguishes these subregions, we built ANN-based encoding models for each. The models predicted held-out responses well (R > 0.70, P < 0.00001) and identified image pairs expected to strongly drive one subregion while suppressing the other. We tested these model-derived predictions in a new pre-registered fMRI experiment (N=15), revealing a clear double dissociation: the anterior subregion responded stronger to sparse spatial layouts, whereas the posterior preferred images rich in objects and texture. We next asked whether this dissociation reflected a broader organizing principle across the scene-selective network. It did not— the pattern was specific to PPA and did not generalize reliably to other scene-selective regions, including OPA and MPA/RSC (F(2,22) ≥ 14.62, p = .001). Together, these findings reveal reproducible anterior-posterior functional organization within human PPA, with a functional dimensionality of at least two. We further provide localizers for both subregions, computational models for each, and initial insight into their roles in scene recognition.