Guoyuan Yang

Alignment between human high-level visual representations and those of multimodal large language models (MLLMs) offers a quantitative framework for understanding information processing in human vision. However, similarities between model and brain representations do not directly reveal the semantic content underlying their correspondence, as this content is difficult to isolate in entangled MLLM representations. Here, we introduce SAIL, an interpretable framework that formulates voxel interpretation as a multi-unit sparse semantic correspondence problem between MLLMs and human vision. First, we train unified crosslayer sparse autoencoders (SAEs) on MLLM representations of Natural Scenes Dataset images and captions, and align each brain voxel with multiple SAE units. Second, we quantify semantic similarity between their top-response image sets to map semantic alignment across the cortex. We project multiple semantic labels back onto the cortex to interpret the content of this alignment. Finally, we assess these correspondences using category responses measured in independent fLoc experiments. Compared with SAEs trained independently at each layer, SAIL exhibits lower redundancy in SAE unit response profiles and higher semantic alignment. We find that semantic alignment is organized along the human ventral visual pathway across the models and subjects examined. The projected labels reveal semantic preferences within canonical visual regions, while fLoc experiments show agreement between predicted and measured category response patterns. Exploratory analyses further reveal cortical distributions of action-related semantics consistent with established functional accounts of lateral occipitotemporal and posterior parietal cortex. Overall, SAIL provides an interpretable account of the semantic correspondences between MLLM representations and human visual responses.

Normative models of brain morphometry quantify individual deviations from typical anatomical patterns and hold promise for enhancing clinical decision-making. However, their clinical utility depends critically on demonstrating generalizability across diverse ethnoracial populations. We previously developed sex-specific, race-neutral normative models for cortical thickness, surface area, and subcortical volumes using brain scans from a large international sample of healthy individuals, as part of the CentileBrain Project, a global initiative to provide open-access, neuroimaging reference models. The primary aim of the present study was to empirically evaluate the generalizability and accuracy of these pretrained models across multiple ethnoracial groups. To this end, we tested model performance in independent samples of healthy individuals from Africa, Asia, Europe, and the Americas, with ethnoracial classification defined either by self-identification or genetic ancestry (N = 4,862). We further compared performance against normative models developed exclusively from a single-population Chinese cohort. Across all groups, as well as in the pooled sample, the pretrained CentileBrain models demonstrated consistently high accuracy, with relative mean absolute error values below 10% for subcortical volume and surface area and below 5% for cortical thickness. Model performance was highly concordant across self-identified and ancestry-defined groups. In a separate analysis, the CentileBrain models performed comparably to a population-specific model when applied to an independent ancestry-matched sample. These findings provide empirical support for the generalizability of race-neutral normative models developed on large and diverse samples and underscore their potential utility for individualized neuroimaging assessment across ethnoracially diverse populations.

Cocaine use disorder (CUD) is highly prevalent and characterized by widespread gray matter atrophy across the cerebral cortex. Yet, it remains unclear whether and how connectome-based circuits and biological features shape these structural abnormalities. We mapped cortical atrophy patterns in CUD (discovery cohort: N = 53; replication cohort: N = 74; controls: N = 364) onto the brain's structural connectome, functional connectivity, transcriptomic similarity and receptor similarity architecture. Using a multimodal and multiscale connectivity-based framework, we identified CUD epicenters and evaluated their spatial correspondence with therapeutic brain stimulation targets and individual variations in clinical symptoms. We found that CUD-related regional atrophy is constrained by the white matter (WM) structural connectome. Along these WM pathways, regions that share similar haemodynamic activity and molecular features are more likely to exhibit convergent atrophy profiles. By integrating the structural connectome with multiple connectivity blueprints, we subsequently identified CUD epicenters and revealed that the prefrontal and visual cortices serve as core systems. Furthermore, we linked these epicenters to cortical transcriptomic patterns and receptor architectures, identifying synaptic and neural homeostasis-related gene enrichment and the strongest spatial correspondence with serotonergic (5-HT_1B/5-HT₄) and dopaminergic (D₂) receptors. Finally, we demonstrated that the spatial distribution of these epicenters correlates with cocaine craving-response maps derived from repeated transcranial magnetic stimulation and can track individual variations in clinical behavioural representations, suggesting their potential as targets for therapeutic intervention. Altogether, our findings establish a structurally constrained framework for the spread of pathology underlying cortical atrophy in CUD, where initial perturbations propagate via structural connectome pathways to vulnerable regions shaped by neural activity and molecular landscapes.

The globus pallidus (GP) is a critical basal ganglia nucleus for motor and cognitive control. However, traditional single-modality parcellations provide an incomplete understanding of its internal organization, thereby increasing the risk of imprecision and side effects during therapeutic targeting of its subregions. In this study, we developed a multi-modal approach that combines structural and functional connectivity, termed hybrid pallidum parcellation (HPP), to parcellate the GP into anterior (aGP), middle (mGP), and posterior (pGP) subregions. This approach was rigorously evaluated for reproducibility and homogeneity in healthy individuals and behavioral prediction accuracy in healthy individuals and patients with Parkinson's disease (PD). Connectivity and behavioral prediction analyses were further used to explore the structural and functional differences in the GP subregions. Compared with unimodal (structural or functional connectivity-based) and two public parcellations, HPP demonstrated superior reproducibility and homogeneity. In addition, the HPP-based subregions revealed a posterior-to-anterior structural and functional connectivity gradient in GP, reflecting a transition from primary to higher-order networks. Specifically, the pGP dominated in sensorimotor function prediction and the aGP in emotional processing, while the mGP exhibited a more flexible pattern in cognition and higher-order motor control. Furthermore, the HPP parcellation demonstrated generalizability and superiority in better predicting behavioral measures than other parcellations in both healthy individuals and patients with PD. Overall, HPP provides new insights into the GP's hierarchical functional organization and its subregional roles in PD motor and non-motor symptoms, offering potential to improve patient stratification for PD and related disorders.

The cerebellum plays a critical role in large-scale brain functional connectivity structure and synchronizing global waves propagation. Despite extensive studies on cortical spatiotemporal architecture, particularly in zero-lag and time-lag synchrony, leaving the parsimonious cerebellar representational structure largely overlooked. Using complex principal component analysis (CPCA) on resting-state fMRI data from the Human Connectome Project, we delineate three dominant low-dimensional spatiotemporal patterns within the cerebellum, manifesting as unimodal-to-transmodal progression, functional antagonism, and hemispheric asymmetry. These patterns correlate with various empirical functional connectivity topographies, including functional connectome network structure, quasiperiodic pattern, functional connectivity gradients and brain network dynamics hierarchy. Importantly, functional connectivity reconstructed from these patterns preserves the intrinsic structural modularity, proving the robustness of the sparse cerebellar dynamics. Moreover, personalized CPCA-derived features enable accurate sex classification and prediction of cognitive, emotional, alertness and personality traits. Our findings position the cerebellum as an active, trait-sensitive integrator of brain-wide spatiotemporal organization, suggesting propagation-based motifs constitute a foundational mode of individual brain intrinsic function.

Neanderthal endocrania are different in shape, though slightly larger in size than modern humans on average. These shape differences have long been used to suggest Neanderthals differed cognitively from modern humans, e.g., by having inferior linguistic/symbolic ability, poorer executive function, and/or smaller episodic and working memory capacity. However, whether the morphological differences in their brains inferred from their endocrania indicate truly meaningful cognitive/behavioral differences—with real evolutionary implications—is not clear. Recent work using deformation mapping techniques suggested there had been significant brain differences between Neanderthal and the anatomically modern Homo sapiens (am Hs ) that were contemporary with them. However the inferred differences were not put into the context of modern human populational variation in brain anatomy, which is known to be substantial. We estimate MRI brain region volumes in two modern population samples (US and Chinese) using deformation mapping and find similar sized differences to those reported for Neanderthal vs. am Hs , with 9 of 13 brain regions showing larger absolute differences for US vs. Chinese samples as compared to reported Neanderthal vs. am Hs differences. To the extent that brain anatomy differences indicate cognitive differences, this suggests that cognitive differences between Neanderthals and am Hs would have comfortably fit within the range found among modern human populations—which are generally not considered evolutionarily significant. In fact, the endocranial differences between Neanderthal and their contemporaries predict cognitive difference effect sizes of only 0.14 SDs or less. This undermines the suggestion that Neanderthal replacement occurred because of cognitive limitations.

Sleep is an essential biological process that restores the brain. Although numerous advances have been made in uncovering the circuits regulating the sleep-wake cycle, the sleep-related dynamic reconfiguration of brain functional connectome topography remains unknown. Here, we collected simultaneous electroencephalogram and functional magnetic resonance imaging data, along with multiple behavioral assessments, from 130 healthy adults during nocturnal sleep. This rich dataset enabled us to construct a large-scale, sleep-state-dependent suite of system-level brain atlases and to establish their associations with sleep quality and pressure. We found that the action-mode network acts as a highly dynamic system that contracts as sleep deepens, most prominent in subcortical structures. Similar contractions were observed in humans and macaques under anesthesia, reflecting cross-species homologues of network reorganization during low-arousal states. Furthermore, the topological reconfiguration of the action-mode network is a stable biological marker that distinguishes sleep stages, reflects sleep quality, and predicts sleep pressure. Together, our work identifies the action-mode network as a dynamic core that reconfigures across the sleep-wake cycle, potentially serving as a neuro-target for intervening sleep disorders and tracking brain restoration.

Adolescent major depressive disorder (MDD) involves complex and heterogeneous alterations in brain functional organization. A possible explanation is the disruption of individualized brain maturation during adolescence, which typically transitions from lower-order sensory to higher-order association regions. To test this hypothesis, we analyzed resting-state functional MRI data using advanced matrix decomposition techniques to extract organizational gradients, temporal dynamics, and information flow. Clustering sensory–association gradient features in our main sample (N_MDD = 302 and N_Controls = 207; ages 11-17) revealed two subtypes. Subtype 1 showed bottom‑up (sensory‑to‑association) flow, reduced synergy but heightened redundancy in sensory cortices, and a highly modular yet globally inefficient network—patterns consistent with aberrant sensory processing and impaired integration. Subtype 2 instead displayed a top‑down (association‑to‑sensory) flow, reduced synergy in association areas, elevated redundancy in sensory regions, and network efficiency between subtype 1 and healthy controls, suggesting compensatory higher‑order engagement. Subtypes also differed in age-related changes, clinical profiles, and neurobiological processing. These patterns were largely replicated in an independent sample (N_MDD = 73 and N_Controls = 28). Overall, the sensory-association axis distinguishes MDD subtypes, integrating functional heterogeneity, cortical dynamics, developmental trajectories, and genetic influences, offering new insights for pathophysiology and precision psychiatry.