We introduce a novel paradigm for non-invasive brain microstructure quantification, where advanced diffusion modeling is merged with cutting-edge diffusion-weighted spectroscopy (DW-MRS) to capture features of cellular morphology that have remained largely ignored by DW-MRI. We elaborate on the finding that metabolite diffusion measured at long diffusion times t_d (up to ~1 second) is fairly stable in the primate brain^1,2, suggesting that metabolites (which, unlike water, are almost exclusively intracellular) are not significantly confined in subcellular regions, but are instead diffusing in long neuronal and astrocytic fibers. However, metabolites should experience long-range structure as t_d keeps increasing. They will explore the ramifications and experience restriction at the extremity of fibers, making their ADC depends on these parameters. To analyze these kinds of data, we propose a model using a compact description of long-range cellular morphology, based on a small set of morphometric statistics, to generate large collections of randomly generated synthetic cells where particles diffusion is simulated. The values of morphometric statistics can then be iterated until calculated ADC matches experimental ADC (Fig.1). After describing the model and investigating its robustness, we apply it on experimental metabolites ADC measured in the monkey brain up to t_d=2 seconds.Assuming the signal attenuation due to displacement perpendicular to fibers is negligible (valid for t_d longer than a few dozen ms), a set of four parameters is used to describe the morphometric statistics (mean and S.D. of segment length L_segment and SD_Lsegment, and the mean and S.D. of embranchments along cell processes N_branch and SD_Nbranch) generating synthetic branched cells (like neurons and astrocytes). N=2000 particles are positioned in each cell and diffusion is simulated by Monte Carlo (with effective intracellular diffusivity D_intra), allowing to compute phase evolution and signal. 80 different cell-graphs are generated to account for cellular heterogeneity. D_intra, L_segment, SD_Lsegment, N_branch and SD_Nbranch are iteratively changed until simulated ADC matches measured ADC (Fig.1). While above parameters have a clear effect on ADC(t_d) (Fig.2A), the number of processes radiating from the cell body doesn’t, and is therefore set to 10±5. The fitting procedure uses a combined Parallel-Tempering and Levenberg-Marquardt approach³, for unsupervised initialization and quick convergence.Code was implemented in Matlab to manage the computation in parallel on GPU device, making it possible to fit experimental ADC in ~3 minutes. Fitting stability relative to noise was assessed by 250 Monte Carlo trials: at each trial, Gaussian noise (15% relative S.D.) was added to a reference ADC(t_d) curve to generate a new dataset, which was analyzed using the fitting pipeline. Stability of the fit was then evaluated by studying the bias and the coefficient of variation (CV) of the estimated parameters. Brain metabolite ADC as a function of t_d was measured in seven healthy macaques, as we have recently described¹. Spectra were acquired at b=0 and 3000 s/mm² using a STEAM sequence (TE=18 ms) with cross-terms cancellation for t_d = 86, 361, 511, 661, 1011 and 2011 ms. Post-processing consisted in scan-to-scan phasing, eddy current correction and subtraction of experimental macromolecule spectrum. Spectra were analyzed with LCModel to estimate the ADC of total N-acetyl aspartate (tNAA=NAA+NAAG), total creatine (tCr), choline compounds (tCho), glutamate (Glu), and myo-inositol (Ins).Accuracy and precision of the fit with respect to noise are reported in Fig.2B and quantified in Table 1. Less than 5% error for all parameters was found, yielding satisfying accuracy and precision despite 15% noise on ADC. Experimental ADC as a function of t_d is reported in Fig.3 for all metabolites. Morphometric statistics evaluated from the fit of experimental data (Fig.3) are reported in Table 2. Generally speaking, cell size and complexity appear realistic. It is striking that the compartments extracted for all five metabolites match the expected cell-specificity: Ins and tCho, which are generally supposed to be mainly astrocytic⁴, are indeed found to be in similar, smaller and simpler synthetic cells, while Glu and tNAA, which are supposed to be neuronal, are found to be in similar, larger and more complex cells. tCr, which is supposed to have no preferential compartmentation, is found in intermediate cells, with larger heterogeneity (mostly SD_Lsegment).The new paradigm introduced here lets us foresee for the first time the possibility to non-invasively extract quantitative information about long-range cellular structure (size, complexity and even heterogeneity) in the brain. The model can of course be refined by adding other morphometric parameters of interest, depending on the context and available data. However, it already gives results that can be compared with real histological data, as explored in another abstract presented at this symposium.