Mapping macaque to human cortex with natural scene responses

Kasper Vinken, Saloni Sharma, Margaret S. Livingstone

TL;DR

This paper introduces a novel, data-driven method to map corresponding brain regions across species (macaque and human) by comparing their responses to a large set of natural scenes. This approach bypasses the limitations of traditional methods that rely on narrow hypotheses and hand-picked stimuli, successfully resolving long-standing ambiguities in the alignment of face-processing areas and suggesting a more anterior location for human homologs of macaque brain regions than previously assumed.

What Problem Does This Paper Address?

Neuroscience often relies on studying animal models, particularly macaques, to understand human brain function. However, a major challenge is accurately identifying functionally equivalent brain regions across species, especially in higher-order cortical areas. This is difficult because:

  • Narrow Hypotheses: Traditional methods often assume specific functions (e.g., "face selectivity") and use limited, carefully chosen stimuli to test these assumptions. This can miss the full complexity of neural representations.
  • Measurement Differences: Human studies often use fMRI (measuring blood flow), while animal studies frequently use electrophysiology (measuring electrical activity). Bridging these different measurement types is complex.
  • Anatomical Ambiguity: While broad anatomical landmarks can be helpful, they don't always perfectly align, and cortical areas can evolve and reorganize over time.

The ventral face patch system, a network of brain areas involved in processing faces, is a prime example of this challenge. Researchers have debated the precise correspondence between macaque and human face patches, with different studies proposing conflicting alignments based on limited functional data. This paper aims to provide a more robust and generalizable method for cross-species mapping, particularly in these complex higher-order areas.

Methods

The core innovation of this paper is a data-driven, hypothesis-agnostic approach to cross-species functional alignment. Instead of pre-defining what a brain region should do, they compare the patterns of neural responses to a rich set of stimuli.

  1. Stimulus Set: They used 700 natural scene images, encompassing a wide variety of objects, animals, and humans, presented in context. Crucially, most of these images did not prominently feature faces, allowing for a broader assessment of neural tuning.
  2. Macaque Electrophysiology: They recorded neural activity from multi-unit sites in the macaque inferotemporal cortex (IT), specifically targeting areas known to be involved in face processing (CIT, analogous to ML, and AIT, analogous to AL).
  3. Human fMRI: They utilized existing human fMRI data from the Natural Scenes Dataset (NSD), which also used the same natural scene stimuli.
  4. Response Pattern Similarity: The key step was to compute the similarity between the response patterns of macaque neural units and human fMRI voxels across all 700 images. This was done by calculating Pearson correlations between their response vectors.
  5. Normalization and Mapping: To account for differences in signal reliability between the two modalities, the correlation values were normalized by a "noise ceiling" derived from split-half reliability measures. This generated "similarity maps" showing how well macaque responses aligned with different parts of the human brain.
  6. Hypothesis Testing: They then used these similarity maps to evaluate three competing hypotheses about the alignment of macaque ML and AL with human face areas (OFA, FFA-1, FFA-2, and ATL).

This method is novel in its reliance on the entire response profile to a large, diverse set of natural stimuli, rather than focusing on specific stimulus categories or predefined functional properties.

Datasets

  • Macaque Neural Recordings: Data collected by the authors from five adult male macaques (four rhesus, one pigtailed) using chronic microelectrode arrays implanted in the inferotemporal cortex (CIT and AIT).
  • Human fMRI Data: The Natural Scenes Dataset (NSD) (Allen et al., 2022), a large-scale fMRI dataset from seven human subjects who viewed natural scenes during a recognition task. This dataset provided both whole-brain fMRI responses and localized face-selective regions of interest (ROIs).

The paper uses a shared set of 700 natural scene images for both macaque and human data collection/analysis, ensuring direct comparability.

Key Insights and Ideas

  • Natural Scenes as a Universal Language: The study demonstrates that rich, naturalistic stimuli can serve as a common ground for comparing brain activity across species and measurement modalities, revealing functional correspondences that might be missed by traditional, narrowly defined approaches.
  • Beyond Face Selectivity: By moving beyond just "face" stimuli, the research shows that the pattern of response to a broad range of visual input is a more powerful tool for mapping brain areas than focusing solely on category-specific responses.
  • Resolving Long-Standing Ambiguities: The data-driven approach successfully resolved conflicting hypotheses about the alignment of macaque ML and AL with human FFA and ATL, providing a clear answer based on empirical evidence.
  • Anterior Shift in Human Homologs: The findings suggest that macaque AL corresponds to human ATL, implying that the human homolog is located more anteriorly than some previous studies suggested, highlighting the importance of considering broader cortical organization.
  • Anatomical and Functional Convergence: The proposed mapping aligns with predictions from full-brain anatomical warping, suggesting that large-scale cortical layout is a strong predictor of functional organization, even when functional data is the primary driver of the mapping.

Key Results

  • Macaque ML aligns with Human FFA: The neural responses from macaque CIT (ML) arrays showed the strongest similarity to human FFA (both FFA-1 and FFA-2).
  • Macaque AL aligns with Human ATL: Neural responses from macaque AIT (AL) arrays showed the strongest similarity to human ATL.
  • Rejection of Scenario 1: The results were inconsistent with the hypothesis that macaque ML aligns with human OFA and AL with FFA.
  • Support for Scenario 3: The findings strongly support the scenario where macaque ML corresponds to human FFA and macaque AL corresponds to human ATL.
  • Quantitative Similarity: The study provides quantitative measures of similarity (Pearson's r / reliability) and statistical tests (t-tests, Z-transformations) to support these findings, showing significant differences in similarity between macaque AIT and CIT arrays with human ATL versus FFA.

Limitations and Future Work

  • Single Array per Monkey: The study recorded from only one array per macaque, meaning inter-individual differences are confounded with between-area differences. While leave-one-subject-out analyses showed no single monkey drove the results, within-subject sampling would be ideal for finer distinctions.
  • ATL Detectability: The anterior temporal lobe face patch (ATL) was only reliably localized in a subset of human participants, meaning the full extent of potential human homologs for anterior macaque regions couldn't be fully assessed.
  • Ecological Validity of Stimuli: While natural scenes were used, the stimulus set was drawn from a human-centric dataset and might not perfectly reflect the ecological niches of macaques. Perceptual meaning could also differ across species.
  • Learned Associations: The approach might be complicated for representations that involve learned semantic or abstract associations, which could differ between species.

Future Work:

  • Expand to Other Brain Regions: Apply this method to other visual and cognitive networks beyond the face system.
  • Denser Sampling: Conduct more extensive recordings within individual macaques to resolve finer distinctions.
  • Incorporate Dynamic Stimuli and Behavior: Use movies, active viewing paradigms, and behavioral tasks to capture more complex representations and disentangle top-down influences.
  • Integrate with Structural Data: Complement functional alignment with cytoarchitecture, connectivity, and developmental lineage to establish phylogenetic homology more rigorously.
  • Leverage Full-Brain Macaque fMRI: Utilize whole-brain macaque fMRI data to support anatomical warping and cross-species comparisons.

Summary updated

Abstract

Neuroscience has long relied on macaque studies to infer human brain function, yet identifying functionally corresponding brain regions across species and measurement modalities remains a fundamental challenge. This is especially true for the higher-order cortex, where functional interpretations are constrained by narrow hypotheses and anatomical landmarks are often nonhomologous. We present a data-driven approach for mapping functional correspondence across species using rich, naturalistic stimuli. By directly comparing macaque electrophysiology with human fMRI responses to 700 natural scenes, we identify fine-grained alignment based on response pattern similarity, without relying on predefined tuning concepts or hand-picked stimuli. As a test case, we examine the ventral face patch system, a well-studied but contested domain in cross-species alignment. Our approach resolves a long-standing ambiguity by supporting a correspondence between macaque ML and human FFA and between AL and more anterior temporal cortex in humans. This result is consistent with full-brain anatomical warping but inconsistent with prior studies limited by narrow functional hypotheses. These findings show that natural image-evoked response patterns provide a robust foundation for cross-species functional alignment, supporting scalable comparisons as large-scale primate recordings become more widespread.
BibTeX
@inproceedings{vinken2025mapping,
  author = {Kasper Vinken and Saloni Sharma and Margaret S. Livingstone},
  title = {{Mapping macaque to human cortex with natural scene responses}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2025},
  doi = {10.1073/pnas.2512619122},
}
Figure