Local measures of neurotransmitters provide crucial insights into neurobiological changes underlying altered functional connectivity in psychiatric disorders. However, noninvasive neuroimaging techniques such as magnetic resonance spectroscopy (MRS) may cover anatomically and functionally distinct areas, such as p32 and p24 of the pregenual anterior cingulate cortex (pgACC). Here, we aimed to overcome this low spatial specificity of MRS by predicting local glutamate and GABA based on functional characteristics and neuroanatomy in a sample of 88 human participants (35 females), using complementary machine learning approaches. Functional connectivity profiles of pgACC area p32 predicted pgACC glutamate better than chance (R² = 0.324) and explained more variance compared with area p24 using both elastic net and partial least-squares regression. In contrast, GABA could not be robustly predicted. To summarize, machine learning helps exploit the high resolution of fMRI to improve the interpretation of local neurometabolism. Our augmented multimodal imaging analysis can deliver novel insights into neurobiology by using complementary information.

SIGNIFICANCE STATEMENT Magnetic resonance spectroscopy (MRS) measures local glutamate and GABA noninvasively. However, conventional MRS requires large voxels compared with fMRI, because of its inherently low signal-to-noise ratio. Consequently, a single MRS voxel may cover areas with distinct cytoarchitecture. In the largest multimodal 7 tesla machine learning study to date, we overcome this limitation by capitalizing on the spatial resolution of fMRI to predict local neurotransmitters in the PFC. Critically, we found that prefrontal glutamate could be robustly and exclusively predicted from the functional connectivity fingerprint of one of two anatomically and functionally defined areas that form the pregenual anterior cingulate cortex. Our approach provides greater spatial specificity on neurotransmitter levels, potentially improving the understanding of altered functional connectivity in mental disorders.

神经递质的局部测量为精神疾病中功能连接改变背后的神经生物学变化提供了重要的见解。然而，磁共振波谱 (MRS) 等无创神经影像技术可能覆盖解剖学和功能上不同的区域，例如前扣带皮层 (pgACC) 的p32和p24 。在这里，我们的目标是通过使用互补的机器学习方法，根据 88 名人类参与者（35 名女性）样本的功能特征和神经解剖学来预测局部谷氨酸和 GABA，从而克服 MRS 的低空间特异性。 pgACC 区域p32的功能连接概况比机会更好地预测 pgACC 谷氨酸 ( R ² = 0.324)，并使用弹性网和偏最小二乘回归解释与区域p24相比更多的方差。相比之下，GABA 无法稳健预测。总而言之，机器学习有助于利用功能磁共振成像的高分辨率来改善对局部神经代谢的解释。我们的增强多模态成像分析可以通过使用补充信息提供对神经生物学的新颖见解。

意义声明磁共振波谱 (MRS) 可无创地测量局部谷氨酸和 GABA。然而，与 fMRI 相比，传统 MRS 需要较大的体素，因为其固有的信噪比较低。因此，单个 MRS 体素可能覆盖具有不同细胞结构的区域。在迄今为止最大的多模态 7 特斯拉机器学习研究中，我们通过利用 fMRI 的空间分辨率来预测 PFC 中的局部神经递质，从而克服了这一限制。至关重要的是，我们发现前额叶谷氨酸可以根据形成前扣带皮层的两个解剖学和功能定义区域之一的功能连接指纹来可靠且唯一地预测。我们的方法提供了神经递质水平更大的空间特异性，有可能提高对精神障碍中功能连接改变的理解。