Effects of antidepressant medications on brain tissue microstructure in persistent depressive disorder across two randomized controlled trials Authors: Ravi Bansal, David J. Hellerstein, Siddhant Sawardekar, Ying Chen, Bradley S. Peterson Published: 2026 ABSTRACT Highlights•PDD was treated with antidepressants in two prospective, double-blind RCTs.•Longitudinal DTI data were collected at baseline and upon trial completion.•DTI-derived measures were used to assess changes in tissue microstructure.•Tissue microstructure normalized in medication- but not placebo-treated patients.•Reduction in symptoms likely reduced the need for compensatory tissue changes. Introduction Persistent depressive disorder(PDD) is characterized by a depressed mood for two or more years(American Psychiatric Association, 2000, Li et al., 2022, Qiu et al., 2020, Sansone and Sansone, 2009) and by symptoms of depression that may or may not meet the criteria for Major Depressive Disorder(MDD) at any time.(American Psychiatric Association, 2000) The pathogenesis of PDD is thought to lie along a continuum with MDD(Solomon et al., 2001) as they generally respond to the same medications.(Griffiths et al., 2000). The neural systems involved in PDD have largely been inferred from prior MRI studies of MDD. Cross-sectional anatomical MRI studies in MDD implicated the limbic system in its pathophysiology, reporting smaller hippocampal volumes(Schmaal et al., 2016) and thinner orbitofrontal, cingulate, insula, and temporal cortices.(Schmaal et al., 2017, Tebartz van Elst et al., 1999) DTI studies of MDD have reported abnormalities in white matter tracts connecting cortical and subcortical regions,(Liao et al., 2013, van Velzen et al., 2020) as well as lower fractional anisotropy(FA) in the deep white matter, including inferior longitudinal fasciculus,(Versace et al., 2010) corona radiata,(Schmaal et al., 2020) anterior cingulate, and dorsolateral prefrontal cortex.(Blood et al., 2010) On the other hand, anatomical studies of PDD have reported enlarged amygdalae(Tebartz van Elst et al., 1999) and smaller volumes of the genu and posterior midbody of the corpus callosum.(Lyoo et al., 2002) Similarly, cross-sectional DTI studies of PDD have reported lower FA in the uncinate and inferior fronto-occipital and cerebrospinal tracts.(Vilgis et al., 2017). Naturalistic, longitudinal cohort studies in MDD have reported that volumes of the amygdala and hippocampus increase and limbic cortices thicken over time in patients who remit or have a milder clinical course, whereas volumes decline and the cortex thins in the same regions in those who do not remit and have a more severe clinical course.(Phillips et al., 2015, Yuksel et al., 2018) Longitudinal studies of MDD have reported that responders compared with nonresponders to selective serotonin-reuptake inhibitors(SSRIs) or selective norepinephrine-reuptake inhibitors(SNRIs) have higher FA values prior to treatment within the forceps minor and superior longitudinal fasciculus(SLF),(Vieira et al., 2021) and remitters compared with nonremitters have a greater cortical thickening during treatment in the superior temporal, precuneus, rostral middle frontal, pars opercularis, inferior parietal, and supramarginal cortices.(Saricicek Aydogan et al., 2019) Task-based fMRI studies in MDD have reported antidepressant medications normalized abnormal neural activity at baseline in the amygdala and ventral anterior cingulate cortex,(Arnone, 2019) whereas one resting-state fMRI study showed that reduced functional connectivity between the subcortical network and the ventral attention network at baseline normalized to control values during SSRI treatment.(Zhang et al., 2023). Brain and symptom severity changes in these naturalistic, longitudinal cohort studies could be attributable either to antidepressant medications or to nonspecific placebo effects. The only way to discern whether medication causes these changes is through the use of randomized controlled trials(RCT). RCTs randomly assign patients to receive either active medication or a placebo, thereby controlling for self-selection and ascertainment biases, as well as placebo effects. Blinding patients and physicians to treatment assignment reduces or eliminates potential rater and participant bias on outcome measures. Combining MRI with an RCT study design thereby supports causal inferences for medication effects on brain outcome measures.(Horga et al., 2014). We acquired MRI data at baseline and study endpoint in two independent, double-blind RCTs of PDD to study the effects of antidepressant medications on the brain.(Hellerstein et al., 2019, Hellerstein et al., 2012) One study was a 10-week double-blind RCT of duloxetine, and the other was a 12-week double-blind RCT of desvenlafaxine.(Hellerstein et al., 2019) Both medications inhibit serotonin and norepinephrine reuptake transporters, though duloxetine potently inhibits both,(Bymaster et al., 2001) whereas desvenlafaxine has a relatively lower affinity for norepinephrine transporters.(Deecher et al., 2006) We previously reported that duloxetine significantly reduced symptom severity,(Hellerstein et al., 2012) whereas desvenlafaxine did not,(Hellerstein et al., 2019) more than placebo. We previously showed that duloxetine normalized baseline cortical hypertrophy,(Bansal et al., 2018b) baseline resting-state hyperconnectivity in the default mode network,(Posner et al., 2013) and baseline elevated metabolite concentrations.(Bansal et al., 2019). These findings together suggested that duloxetine’s effects on brain structure, function, and metabolism may operate through the mechanism of neural plasticity, likely remodeling dendritic arbors and synapses in cortical tissue. Recent studies have shown activity-dependent reorganization of white matter.(Ganguly and Poo, 2013) Neuronal activity both remodels preexisting myelin sheaths(Almeida and Lyons, 2017, Fields, 2015, Gibson et al., 2014, Just et al., 2004, Micu et al., 2016, Orkand et al., 1966, Stevens et al., 2002, Yeung et al., 2014, Young et al., 2013) and induces differentiation of oligodendrocyte precursor cells into new oligodendrocytes.(Almeida and Lyons, 2017, Rivers et al., 2008, Sampaio-Baptista and Johansen-Berg, 2017) This activity-dependent plasticity of myelinated axons alters neural communication and can subsequently affect cognition and behavior.(Fields, 2008) Plasticity-induced changes in gray and white matter(Ganguly and Poo, 2013) will alter the diffusion of free water molecules that can be measured using diffusion tensor imaging(DTI). We therefore paired DTI with RCT study design to assess whether duloxetine or desvenlafaxine changes tissue microstructure. Although our study builds on prior imaging biomarker studies of antidepressant and placebo response, integration of MRI with rigorously controlled RCTs remains relatively uncommon. Based on our prior anatomical, functional, and spectroscopy MRI findings in these studies, we hypothesized that active medications would normalize DTI measures of gray and white matter microstructure within limbic regions of the brain. We also hypothesized that declining symptom severity during the trial would mediate medication effects in normalizing DTI-derived measures of tissue microstructure. Methods and materials Duloxetine and desvenlafaxine were FDA-approved at the time of study. Study protocols were approved by the Institutional Review Board of the New York State Psychiatric Institute (NYSPI), New York, NY. All participants provided written informed consent. All patients agreed to defer the use of other psychotropic medications and psychotherapy during the study. Research psychiatrists evaluated depressing symptoms using the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID-I)(First et al., 1997), and 24-item Hamilton Depression Rating Scale (HDRS-24).(Hamilton, 1960) A statistical power analyses for detecting treatment effects on symptom severity showed that 30 participants in each arm provided an 80% power at a significance level of 0.05.(Hellerstein et al., 2019, Hellerstein et al., 2012) Inclusion criteria: age 20–65 years, 24-item Hamilton Depression Rating Scale (HDRS) score > 12, with chronic depression lasting > 2 years at the time of recruitment. Exclusion criteria: DSM-5 diagnosis of current bipolar disorder, schizophrenia or other psychotic disorder, or major depressive disorder; current or within the past 6-month drug or alcohol abuse disorder; current use of psychoactive medication; serious risk for suicide during the trial; unstable medical conditions; current or planned pregnancy; dementia or other cognitive disorder; current eating disorder; inability to consent to study participation. Patients underwent a washout period of ≥ 7 days for ineffective psychotropic medications or ≥ 28 days for fluoxetine. Potential side effects (sexual dysfunction, weight gain, cognitive impairment) were assessed at each clinical visit using structured scales. Serious adverse event information was collected throughout the trials. Patients were withdrawn from study if clinical condition worsened significantly as assessed by the Clinical Global Impressions (CGI) scale. Duloxetine study Patients were recruited from the Depression Evaluation Service of NYSPI starting July 2006 until November 2011.(Hellerstein et al., 2012) We approached 350 patients, of whom 65 were eligible and 57 consented to participate. Twenty-eight patients were randomized to receive placebo and 29 to receive duloxetine. In the placebo arm, we collected DTI data from 20 patients at both baseline and trial completion; in the duloxetine arm, we collected DTI data from 21 patients at both time points (Table 1 & Fig.S1). Patients were randomized 1:1 after baseline assessments to receive 30 to 120 mg/day of duloxetine (active medication) or gelatin shell capsules with sugar balls (placebo) for 10 weeks. We have previously published the detailed randomization procedures.(Hellerstein et al., 2019) In the absence of sufficient response (CGI-I score ≥ 3; that is, depression is minimally improved, unchanged, or worsened) and the absence of an adverse event, the dose was increased to 60 mg/day after one week, and then increased by 30 mg/day every week for up to 120 mg/day dose. The mean dose of duloxetine was 89 mg/day (SD 28.3 mg/day), and the mean placebo equivalent was 100.7 mg/day (SD 27.3 mg/day). Symptom severity and psychosocial functioning were assessed at baseline and upon trial completion at week 10. Patients with MRI who were randomized to receive duloxetine did not differ on either age or baseline symptom severity from those randomized to receive placebo (Supplementary Table S1). However, significantly more males were randomized to receive placebo than duloxetine (χ2 = 4.1, df = 1, p = 0.048). Although symptom severity decreased significantly for both duloxetine-treated (p=<0.0001) and placebo-treated patients (p = 0.004), duloxetine-treated patients compared to placebo-treated patients had significantly (p < 0.0001) lower symptom severity.Table 1Collected and Usable DTI Data at Baseline and at Trial Completion for patients in the desvalafaxine and duloxetine studies. The CONSORT diagrams (Fig. S1 & S2) provide detailed flow charts for the number of participants approached, eligible, consented, randomized, drop outs, and DTI collected at baseline and upon trial completion (TC).DesvenlafaxineDuloxetinePlaceboMedicatiionPlaceboMedicatiionDTI collected, Baseline21202021DTI usable, Baseline19202018DTI collected, TC20182021DTI Usable, TC15162021TC = trial completion Table 1Collected and Usable DTI Data at Baseline and at Trial Completion for patients in the desvalafaxine and duloxetine studies. The CONSORT diagrams (Fig. S1 & S2) provide detailed flow charts for the number of participants approached, eligible, consented, randomized, drop outs, and DTI collected at baseline and upon trial completion (TC).DesvenlafaxineDuloxetinePlaceboMedicatiionPlaceboMedicatiionDTI collected, Baseline21202021DTI usable, Baseline19202018DTI collected, TC20182021DTI Usable, TC15162021TC = trial completion Collected and Usable DTI Data at Baseline and at Trial Completion for patients in the desvalafaxine and duloxetine studies. The CONSORT diagrams (Fig. S1 & S2) provide detailed flow charts for the number of participants approached, eligible, consented, randomized, drop outs, and DTI collected at baseline and upon trial completion (TC). TC = trial completion Desvenlafaxine study Patients were recruited from local clinics, psychiatric listservs, hospital bulletin boards, newspapers, and Craigslist from February 2012 to December 2016, in the same geographical area as the duloxetine study. Of 350 patients screened, 71 were eligible to participate and were not using any psychoactive medication. Ten patients either declined to participate, were ineligible at the time of randomization, or were lost before randomization. Thirty patients were randomized to receive placebo and 31 to receive desvenlafaxine. In the placebo arm, we collected DTI data from 21 patients at baseline and 20 patients at trial completion; in the duloxetine arm, we collected DTI data from 20 patients at baseline and 18 patients at trial completion (Table 1 & Fig.S2). We have previously published the detailed randomization procedures.(Hellerstein et al., 2019) Two patients randomized to receive placebo had a current diagnosis of obsessive–compulsive disorder, and two randomized to desvenlafaxine had a current diagnosis of generalized anxiety disorder. Patients received initial doses of 50 mg capsules of either placebo or desvenlafaxine. The dose was increased to 100 mg after week 4 if patients tolerated the 50 mg dose and were still depressed (defined as CGI-I ≥ 3).(Hellerstein et al., 2019) Symptom severity and psychosocial functioning were assessed at baseline and week 12.(Hellerstein et al., 2019) Patients with MRI who were randomized to desvenlafaxine did not differ from those randomized to placebo on age, sex, or symptom severity at baseline (Supplementary Table S1). Symptom severity decreased significantly for both desvenlafaxine-treated (p < 0.0001) and placebo-treated patients (p = 0.0004). However, upon trial completion, symptom severity in desvenlafaxine-treated patients did not differ from that in placebo-treated patients (p = 0.3). Study differences in patient characteristics Patients in the duloxetine study did not differ from patients in the desvenlafaxine study on age (p = 0.11), sex (p = 0.82), or symptom severity (p = 0.66, HDRS-17), but had significantly (p < 0.001) longer duration of illness (12.8 ± 13.5 years, desvenlafaxine; 22.15 ± 15.11 years, duloxetine). Healthy controls Thirty-five healthy controls (age = 39.73 ± 9.58 years, 22 males) were recruited from the same geographical region as the patient populations. Healthy controls did not differ from patients either in the duloxetine study (age p = 0.09, sex p = 0.36) or in the desvenlafaxine study (age = 0.95, sex p = 0.15). Exclusion criteria IQ < 80, a current DSM-IV Axis I disorder, a lifetime history of Attention-Deficit/Hyperactivity Disorder, tic disorder, psychotic disorder, developmental delay, head trauma with loss of consciousness, or epilepsy. MRI data acquisition MRI scanner MRI data for the duloxetine study were collected on a 3 T GE Signa MRI scanner equipped with an 8-channel head coil at NYSPI, whereas data for desvenlafaxine patients were acquired on a 3 T GE Signa MRI scanner at either NYSPI for 17 patients (10 males, age = 39.45 ± 11.87 years) or Cornell University for 25 patients (7 males, age = 31.84 ± 11.32 years). Desvenlafaxine patients scanned at Cornell were significantly (p = 0.04) younger than desvenlafaxine patients scanned at NYSPI. Desvenlafaxine patients at the two scan sites however did not differ in sex (X2 = 2.81, df = 1, p = 0.09), symptom severity (p = 0.41), or treatment assignment (X2 = 0.0, df = 1, p = 1.0). Detailed pulse sequence parameters and processing of MRI data are in the Supplementary Materials and Methods. Patients with MRI scans In the duloxetine study, we acquired MRI data in 53 patients at baseline and 42 upon trial completion. Useable DTI data were available in 38 patients at both time-points (Fig.S1). In the desvenlafaxine study, we acquired MRI data in 41 patients at baseline and 39 at the end of the trial. Useable DTI data were available in 39 patients at baseline and 31 patients upon trial completion (Fig.S2). Patients with useable DTI data at baseline and trial completion were included in the analyses. Statistical analyses Both duloxetine and desvenlafaxine are SNRIs and, therefore, we combined patients from the two studies to increase statistical power for detecting common medication effects on tissue microstructure. We regressed out the effects of nuisance variables age, sex, and scan site. We applied the cluster size-based familywise error rate(FWER) procedure to control for false positives in multiple hypotheses testing (Supplementary Materials and Methods).(Friston et al., 1994, Worsley et al., 1996). Three-Way interaction (Treatment*Time*Study) effects We first tested the 3-way interaction, time-by-treatment-study effect, to assess whether the effects of differing treatments (active medication versus placebo) on tissue microstructure during trials differed by study. If treatment effects do not differ by study, then patients in the two studies can be combined. A combined, larger cohort would provide greater statistical power to test whether medication effects differ from placebo effects on brain microstructure. We evaluated statistical significance of the 3-way interaction effects using a univariate repeated–measures analysis without subject–level random intercepts or random slopes:yij=β0+β1∗Agei+β2∗Sexi+β3·ScanSitei+β4∗Timeij+β5∗Txi+β5∗Sti+β6∗Timeij∗Txi+β7∗Timeij∗Sti+β8∗Txi∗Sti+β9∗Timeij∗Txi∗Sti+∊ijwhere yij is the DTI measure (FA or ADC) for i th participant at j th assessment (baseline or trial completion), Agei is i th participant age at baseline scan, treatment assigned to the i th patient Txi is 0 for placebo or 1 for medication, and Timeij is 0 for baseline or 1 for trial completion, Sti is 0 for the desvenlafaxine study or 1 for the duloxetine study, and ScanSitei is 1 for scans at NYSPI or 0 for scans at Cornell University. We did not model subject–specific random effects because each participant contributed only two observations: random effects would overparameterize the model leading to instability and inflated standard errors. We tested the null hypothesis H0:β9=0 and the alternate hypothesis Ha:β9≠0 for detecting significant 3-way effects of treatment on DTI measures. Two-Way interaction (Treatment*Time) effects We hypothesized that tissue microstructure in the limbic system would normalize following treatment with medications. In brain regions where time-by-treatment effects did not differ by study (i.e., 3-way interaction effects were absent), we studied whether medication effects differed from placebo effects in the combined cohort of patients. We used a univariate repeated measures analysis to test the two-way interaction effects (the Treatment*Time interaction) on tissue microstructure:yij=β0+β1∗Agei+β2∗Sexi+β3·ScanSitei+β4∗Txi+β5∗Timeij+β6∗Timeij∗Txi+∊ijwhere yij is the DTI measure (FA or ADC) for i th participant at j th assessment (baseline or trial completion), Agei is i th participant age at baseline, Txi is 0 for placebo or 1 for medication, and Timeij is 0 for baseline or 1 for trial completion. We tested the null hypothesis H0:β6=0 and the alternate hypothesis Ha:β6≠0 for detecting the differing effects of active medication relative to placebo on DTI measures. Associations of symptom severity change with change in tissue microstructure We assessed, separately for medication- and for placebo-treated patients, whether the decline in symptom severity was associated with changes in tissue microstructure over the trials using a univariate repeated measures analysis:(Hedeker and Gibbons, 2006)yij=β0+β1∙Agei+β2∙Sexi+β3·ScanSitei+β4∙Sti+β5∙Sx¯i+β6∙Sxij-Sx¯i+εijwhere yij is the DTI measure (FA or ADC) for i th participant at j th assessment (baseline or trial completion), Sx¯i=12Sxi1+Sxi2 is the average of symptom severity across baseline (Sxi1) and endpoint (Sxi2) for the i th patient, β5 is the between-individual correlation of DTI measure with symptom severity, and β6 is the longitudinal, within-individual change in DTI measure accompanying the change in symptom severity. Testing whether changes in symptom severity mediated treatment effects on tissue microstructure We also hypothesized that a decline in symptom severity mediates treatment effects on tissue microstructure. If so, then baseline brain abnormalities are compensatory that help patients cope with depressive symptoms. We used a two-wave, longitudinal mediation analysis with following regression models to test this hypothesis:(Bind et al., 2016, MacKinnon, 2008, Selig and Preacher, 2009)yi2=β0,1+β1,1∙Agei+β2,1∙Sexi+β3,1∙ScanSitei+c1∙Txi+b1∙Sxi1+b2∙Sxi2+g1∙yi1+εi,1Sxi2=β0,2+β1,2∙Agei+β2,2∙Sexi+β3,2∙ScanSitei+a1∙Txi+g2∙Sxi,1+εi,2where yi1 and yi2 are DTI measures (FA or ADC), Sxi1 and Sxi2 are symptom severity at baseline (j=1) and end of the trial (j=2), respectively, and Agei, Sexi, and Txi denote age, sex, and treatment for the i th patient. The temporal, across-time mediated effects are represented by the product a1b1. (Bind et al., 2016, MacKinnon, 2008, Selig and Preacher, 2009) We used the Sobel test(Sobel, 1982) to evaluate the significance of the longitudinal mediation effects a1b1; its standard error is seab=a12∙seb2+(b12∙sea2), where sea, seb are the standard errors of the coefficients a1 and b1, respectively. We computed the p-value under the null hypothesis H0:a1b1=0. Results Three-Way interaction (Treatment*Time*Study) effects Significant Treatment*Time*Study 3-way interaction effects, that is treatment effects (medication versus placebo) over time differed significantly between the two studies, were detected for FA values in deep dorsal white matter tracts (anterior and superior corona radiata, post thalamic radiations, and superior longitudinal fasciculus, and superior frontal gyrus) and for average diffusivity coefficient (ADC) values primarily in frontal cortical gray matter (Fig. 1, Left Panel). The 3-way interaction effects on axial diffusivity (AD) and radial diffusivity (RD) were statistically significant in the same brain regions (Fig.S6). Bar graphs in representative voxels showed that (a) in the duloxetine study, FA values that were higher than normal values in anterior corona radiata (aCR) at baseline normalized for patients treated with duloxetine but not placebo, whereas ADC values in frontal cortices did not change in patients treated with placebo (Fig. 1, Right Panel & Fig.S3)). FA values in aCR for the duloxetine study were higher than those for the desvenlafaxine potentially due to differing illness duration. ADC values in superior frontal gyrus, however, were lower than normal values, which did not change in duloxetine-treated patients. (b) In the desvenlafaxine study, FA values in frontal cortices that were comparable to normal values at baseline did not change in patients treated with desvenlafaxine, but these values deviated away from values in healthy participants for placebo-treated patients (Fig. 1, Right Panel & Fig.S3). ADC values in frontal cortices in placebo-treated patients were comparable to, but were higher in desvenlafaxine-treated patients than, normal values at baseline. Following trial, FA and ADC values declined in placebo-treated but not desvenlafaxine-treated patients, although these reductions were not statistically significant (p = 0.24 for FA; p = 0.65 for ADC).Fig. 1Significant 3-Way (Time-by-Treatment-by-Study) Interaction Effects assessing whether differing time effects of treatment (active medication or placebo) differed by study. The three-way (time-by-treatment-by-study) interaction effect on the brain microstructure was tested using repeated measures analyses applied at each voxel of the brain. We controlled for false positives using a cluster-size FWER with CDT = 2.0. Left Panel: Findings that survived FWER correction were color-encoded and displayed on axial slices of the template brain in the radiological convention. Slice locations in the Talairach atlas are labeled below. These findings show that the time-by-treatment effects differed between the two studies in white matter of the anterior corona radiata (aCR), posterior thalamic radiations (PTR), and superior corona radiata (sCR), and in gray matter of the superior frontal gyrus (SFG). The color bar encoding p-values is shown at the bottom. Right Panel: To understand significant 3-way interaction effects on the brain microstructure, we generated bar graphs for representative brain regions (labeled A & B). Blue bars show data at baseline, and orange bars show data on trial completion (TC). Data for healthy controls are presented as a gray horizontal line. In the desvenlafaxine study, FA and ADC values deviated away from values in healthy participants for placebo-treated patients, whereas FA and ADC did not change for desvenlafaxine-treated patients. In the duloxetine study, FA and ADC values did not change for placebo-treated patients, whereas FA and ADC normalized for duloxetine-treated patients. P-value = 9.0*10-5 (β = -143.68, t-statistic = -2.79, 95%CI = [-244.4, −42.9]) in the superior longitudinal fasciculus (region A) and p = 1.6*10-3 (β = -255.77, t-statistic = -2.96, 95%CI = [-425.08, −86.45]) in the superior frontal gyrus (region B) for the 3-way interaction effects. Bar graphs for significant 3-way interaction in regions 1, 2, and 3 are presented in Supplementary Figure Fig.S3. aCR = anterior corona radiata; pCR = posterior corona radiata; sCR = superior corona radiata; PTR = post thalamic radiations; SFG = superior frontal gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; TC = trial completion; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Significant 3-Way (Time-by-Treatment-by-Study) Interaction Effects assessing whether differing time effects of treatment (active medication or placebo) differed by study. The three-way (time-by-treatment-by-study) interaction effect on the brain microstructure was tested using repeated measures analyses applied at each voxel of the brain. We controlled for false positives using a cluster-size FWER with CDT = 2.0. Left Panel: Findings that survived FWER correction were color-encoded and displayed on axial slices of the template brain in the radiological convention. Slice locations in the Talairach atlas are labeled below. These findings show that the time-by-treatment effects differed between the two studies in white matter of the anterior corona radiata (aCR), posterior thalamic radiations (PTR), and superior corona radiata (sCR), and in gray matter of the superior frontal gyrus (SFG). The color bar encoding p-values is shown at the bottom. Right Panel: To understand significant 3-way interaction effects on the brain microstructure, we generated bar graphs for representative brain regions (labeled A & B). Blue bars show data at baseline, and orange bars show data on trial completion (TC). Data for healthy controls are presented as a gray horizontal line. In the desvenlafaxine study, FA and ADC values deviated away from values in healthy participants for placebo-treated patients, whereas FA and ADC did not change for desvenlafaxine-treated patients. In the duloxetine study, FA and ADC values did not change for placebo-treated patients, whereas FA and ADC normalized for duloxetine-treated patients. P-value = 9.0*10-5 (β = -143.68, t-statistic = -2.79, 95%CI = [-244.4, −42.9]) in the superior longitudinal fasciculus (region A) and p = 1.6*10-3 (β = -255.77, t-statistic = -2.96, 95%CI = [-425.08, −86.45]) in the superior frontal gyrus (region B) for the 3-way interaction effects. Bar graphs for significant 3-way interaction in regions 1, 2, and 3 are presented in Supplementary Figure Fig.S3. aCR = anterior corona radiata; pCR = posterior corona radiata; sCR = superior corona radiata; PTR = post thalamic radiations; SFG = superior frontal gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; TC = trial completion; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Two-Way interaction (Treatment*Time) effects In brain regions where the 3-way interaction was not significant, the Treatment*Time 2-way interaction indicated where changes in DTI measures differed significantly across treatments during the trials in both studies combined. In patients treated with active medication (duloxetine or desvenlafaxine), FA and ADC values were significantly lower than healthy control values at baseline and increased toward values for healthy participants in the insular cortex and lingual gyrus. In placebo-treated patients in these same regions, FA and ADC values declined further during the trial from already-low values at baseline (Fig. 2).Fig. 2Significant 2-Way (Time-by-Treatment) Interaction Effects assessing whether treatment (active medication or placebo) effects differed during trials. Findings are shown in regions where the 3-way interaction was not significant: that is, where treatment effects during trials did not differ by study. P-values for findings that survived cluster size FWER correction with CDT = 2.0 were color-encoded and displayed on the axial slices of the template brain in the radiological convention. Slice locations in the Talairach atlas space are labeled below. Change in FA and ADC values during both trials in medication-treated patients differed significantly from change in placebo-treated patients in the insular cortex (Ins), amygdala (Amy), and lingual gyrus (LG). To understand these significant 2-way interactions, we generated bar graphs at a few selected points (labeled A, B, C, & D). In medication-treated patients, FA and ADC values moved towards values for healthy participants in the insular cortex and lingual gyrus, whereas deviated away from values for healthy participants in the amygdala. In placebo-treated patients, FA and ADC values deviated away from values for healthy participants in the insular cortex and lingual gyrus, whereas moved towards values for healthy participants in the amygdala. P-value = 3.0*10-4 (β = 80.26, t-statistic = -2.91, 95%CI = [26.35,134.17]) in the insular cortex (region A), p = 1.0*10-3 (β = 112.51, t-statistic = 2.83, 95%CI = [34.76,190.25]) in the orbitofrontal cortex (region B), p = 1.0*10-4 (β = 197.98, t-statistic = 2.44, 95%CI = [39.11,356.86]) in the insular cortex (region C), and p = 1.3*10-5 (β = 215.37, t-statistic = 2.47, 95%CI = [44.94,385.79]) in the lingual gyrus (region D) for significant 2-way interaction effects. Ins = Insular Cortex; OFC = orbital frontal cortex; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; TC = trial completion; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Significant 2-Way (Time-by-Treatment) Interaction Effects assessing whether treatment (active medication or placebo) effects differed during trials. Findings are shown in regions where the 3-way interaction was not significant: that is, where treatment effects during trials did not differ by study. P-values for findings that survived cluster size FWER correction with CDT = 2.0 were color-encoded and displayed on the axial slices of the template brain in the radiological convention. Slice locations in the Talairach atlas space are labeled below. Change in FA and ADC values during both trials in medication-treated patients differed significantly from change in placebo-treated patients in the insular cortex (Ins), amygdala (Amy), and lingual gyrus (LG). To understand these significant 2-way interactions, we generated bar graphs at a few selected points (labeled A, B, C, & D). In medication-treated patients, FA and ADC values moved towards values for healthy participants in the insular cortex and lingual gyrus, whereas deviated away from values for healthy participants in the amygdala. In placebo-treated patients, FA and ADC values deviated away from values for healthy participants in the insular cortex and lingual gyrus, whereas moved towards values for healthy participants in the amygdala. P-value = 3.0*10-4 (β = 80.26, t-statistic = -2.91, 95%CI = [26.35,134.17]) in the insular cortex (region A), p = 1.0*10-3 (β = 112.51, t-statistic = 2.83, 95%CI = [34.76,190.25]) in the orbitofrontal cortex (region B), p = 1.0*10-4 (β = 197.98, t-statistic = 2.44, 95%CI = [39.11,356.86]) in the insular cortex (region C), and p = 1.3*10-5 (β = 215.37, t-statistic = 2.47, 95%CI = [44.94,385.79]) in the lingual gyrus (region D) for significant 2-way interaction effects. Ins = Insular Cortex; OFC = orbital frontal cortex; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; TC = trial completion; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Associations of symptom change with DTI changes In medication-treated patients, FA values normalized as symptom severity declined in the insula and orbitofrontal cortices and globus pallidus, and ADC values normalized in the lingual gyrus (Fig. 3 & Fig.S4). Consistently, decreasing symptom severity was associated with increasing AD values in the putamen and globus pallidus (Fig.S7), but with decreasing RD values in smaller regions across the brain, suggesting that FA provided significantly greater statistical power than AD and RD to detect treatment-related changes in white matter microarchitecture.Fig. 3Associations of Change in DTI Measures with Change in Symptom Severity from baseline to trial completion for medication-treated patients in the two studies. Left Panel: P-values for findings that survived cluster size FWER correction with CDT = 2.0 were color-encoded and displayed on the axial slices of the template brain in the radiological convention. Blue and violet colors show brain regions where FA or ADC increased, and red and yellow colors show brain regions where FA or ADC decreased, as HamD17 scores decreased for medication-treated patients during the studies. The color bar encoding p-values is shown at the bottom. Right Panel: We generated line graphs at selected regions (labeled A, B, & C) to understand how changes in ADC and FA values were associated with decreases in symptom severity. Blue line: changes in DTI measures with changes in symptom severity. Gray line: DTI measure in healthy controls. These line graphs show that in medication-treated patients, FA values in the insular and orbital frontal cortices normalized as symptom severity declined during the trials. On the other hand, ADC values increased but did not reach values for healthy participants (HC = 0. 74, gray line not shown) in the lingual gyrus as symptom severity declined. Note: HamD17 scores decrease from left to right of the X-axis to denote the decrease in symptom severity during the trials. P-value = 1.0*10-4 (β = -7.21, t-statistic = -3.23, 95%CI = [-11.58,-2.85]) in the insular cortex (region A), p = 1.0*10-3 (β = -3.89, t-statistic = -2.79, 95%CI = [-6.63,-1.16]) in the orbitofrontal cortex (region B), p = 2.4*10-4 in the globus pallidus (region C), and p = 1.6*10-5 (β = -4.2, t-statistic = -2.48, 95%CI = [-7.5,-0.89]) in the lingual gyrus (region D), for significant associations between change in symptom severity with change in DTI measures. Ins = insular cortex; OFC = orbital frontal cortex; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; HamD17 = 17-item Hamilton Depression Rating Scale; Med = medication-treated patients; HC = healthy controls; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Associations of Change in DTI Measures with Change in Symptom Severity from baseline to trial completion for medication-treated patients in the two studies. Left Panel: P-values for findings that survived cluster size FWER correction with CDT = 2.0 were color-encoded and displayed on the axial slices of the template brain in the radiological convention. Blue and violet colors show brain regions where FA or ADC increased, and red and yellow colors show brain regions where FA or ADC decreased, as HamD17 scores decreased for medication-treated patients during the studies. The color bar encoding p-values is shown at the bottom. Right Panel: We generated line graphs at selected regions (labeled A, B, & C) to understand how changes in ADC and FA values were associated with decreases in symptom severity. Blue line: changes in DTI measures with changes in symptom severity. Gray line: DTI measure in healthy controls. These line graphs show that in medication-treated patients, FA values in the insular and orbital frontal cortices normalized as symptom severity declined during the trials. On the other hand, ADC values increased but did not reach values for healthy participants (HC = 0. 74, gray line not shown) in the lingual gyrus as symptom severity declined. Note: HamD17 scores decrease from left to right of the X-axis to denote the decrease in symptom severity during the trials. P-value = 1.0*10-4 (β = -7.21, t-statistic = -3.23, 95%CI = [-11.58,-2.85]) in the insular cortex (region A), p = 1.0*10-3 (β = -3.89, t-statistic = -2.79, 95%CI = [-6.63,-1.16]) in the orbitofrontal cortex (region B), p = 2.4*10-4 in the globus pallidus (region C), and p = 1.6*10-5 (β = -4.2, t-statistic = -2.48, 95%CI = [-7.5,-0.89]) in the lingual gyrus (region D), for significant associations between change in symptom severity with change in DTI measures. Ins = insular cortex; OFC = orbital frontal cortex; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; HamD17 = 17-item Hamilton Depression Rating Scale; Med = medication-treated patients; HC = healthy controls; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) In placebo-treated patients, as symptom severity declined, FA values deviated further away from values for healthy participants in the posterior corona radiata and lingual gyrus, whereas ADC values deviated further away from values for healthy participants in the lingual gyrus (Fig. 4). Declining symptom severity was associated with increasing AD and RD values across cortical gray matter (Fig.S8), which are consistent with declining FA values.Fig. 4Associations of Change in DTI Measures with Change in Symptom Severity from baseline to trial completion in placebo-treated patients in the two studies. Left Panel: P-values for findings that survived cluster size FWER correction with CDT = 2.0 were color-encoded. Blue and violet colors show brain regions where FA or ADC increased, whereas red and yellow colors show brain regions where FA or ADC decreased, as HamD17 scores decreased for placebo-treated patients during the studies. The color bar encoding p-values is shown at the bottom. Right Panel: We generated line graphs at selected regions (labeled A & B) to understand how changes in FA values were associated with decreases in symptom severity. Blue line: changes in DTI measures with changes in symptom severity; Gray line: DTI measure in healthy controls. Line graphs show that for placebo-treated patients FA and ADC values deviated further away from values for healthy participants as their symptoms decreased in the lingual gyrus. Note: HamD17 scores decrease from left to right of the X-axis as symptom severity decreases during the trials. P-value = 9.0*10-5 (β = 13.7, t-statistic = 3.49, 95%CI = [6.02,21.38]) in the lingual gyrus (region A), p = 4*10-3 (β = 6.22, t-statistic = 2.26, 95%CI = [0.84,11.61]) in the posterior corona radiata (region B), and p = 1*10-4 (β 8.17, t-statistic = 3.34, 95%CI = [3.38,12.9]) in the lingual gyrus (region C) for significant associations between change in symptom severity with change in DTI measures. pCR = posterior corona radiata; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; HamD17 = 17-item Hamilton Depression Rating Scale; PBO = placebo-treated patients; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Associations of Change in DTI Measures with Change in Symptom Severity from baseline to trial completion in placebo-treated patients in the two studies. Left Panel: P-values for findings that survived cluster size FWER correction with CDT = 2.0 were color-encoded. Blue and violet colors show brain regions where FA or ADC increased, whereas red and yellow colors show brain regions where FA or ADC decreased, as HamD17 scores decreased for placebo-treated patients during the studies. The color bar encoding p-values is shown at the bottom. Right Panel: We generated line graphs at selected regions (labeled A & B) to understand how changes in FA values were associated with decreases in symptom severity. Blue line: changes in DTI measures with changes in symptom severity; Gray line: DTI measure in healthy controls. Line graphs show that for placebo-treated patients FA and ADC values deviated further away from values for healthy participants as their symptoms decreased in the lingual gyrus. Note: HamD17 scores decrease from left to right of the X-axis as symptom severity decreases during the trials. P-value = 9.0*10-5 (β = 13.7, t-statistic = 3.49, 95%CI = [6.02,21.38]) in the lingual gyrus (region A), p = 4*10-3 (β = 6.22, t-statistic = 2.26, 95%CI = [0.84,11.61]) in the posterior corona radiata (region B), and p = 1*10-4 (β 8.17, t-statistic = 3.34, 95%CI = [3.38,12.9]) in the lingual gyrus (region C) for significant associations between change in symptom severity with change in DTI measures. pCR = posterior corona radiata; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; HamD17 = 17-item Hamilton Depression Rating Scale; PBO = placebo-treated patients; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Symptom severity mediating treatment effects on tissue microstructure A decline in symptom severity significantly mediated treatment effects on tissue microstructure (Fig. 5). Longitudinal mediation analyses showed that active medication (duloxetine or desvenlafaxine) significantly reduced symptom severity upon trial completion (Fig.S5), and that the reduction in symptom severity in turn significantly increased FA or ADC values in white matter of the superior longitudinal fasciculus, anterior corona radiata, and genu and splenium of corpus callosum and in gray matter of the putamen. In contrast, a treatment-induced reduction in symptom severity significantly reduced FA in gray matter of the cingulate gyrus (Fig.S5).Fig. 5Two-Wave Longitudinal Mediation Analyses We conducted two-wave, longitudinal mediation analyses to test whether the decrease in symptom severity significantly mediated treatment effects on tissue microstructure. The p-values that survived cluster size FWER correction with CDT = 2.0 for the overall mediating effects were color-encoded and plotted on the axial slices of the template brain. The color bar at the bottom shows color coding for the p-values, with colors yellow and red show brain regions where FA or ADC values increased in patients with lower baseline symptom severity, whereas colors violet and blue show brain regions where FA or ADC values decreased in patients with lower baseline symptom severity. That is, a decrease in symptom severity during the trials significantly mediated treatment effects on tissue microstructure in white matter of superior longitudinal fasciculus, anterior corona radiata, posterior thalamic radiata, genu and splenium of the corpus callosum, and in gray matter of cingulate gyrus and putamen. To visualize and understand the longitudinal mediation effects, we generated scatterplots presented in supplementary figure Fig.S5. pCR = posterior corona radiata; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; HamD17 = 17-item Hamilton Depression Rating Scale; PBO = placebo-treated patients; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Two-Wave Longitudinal Mediation Analyses We conducted two-wave, longitudinal mediation analyses to test whether the decrease in symptom severity significantly mediated treatment effects on tissue microstructure. The p-values that survived cluster size FWER correction with CDT = 2.0 for the overall mediating effects were color-encoded and plotted on the axial slices of the template brain. The color bar at the bottom shows color coding for the p-values, with colors yellow and red show brain regions where FA or ADC values increased in patients with lower baseline symptom severity, whereas colors violet and blue show brain regions where FA or ADC values decreased in patients with lower baseline symptom severity. That is, a decrease in symptom severity during the trials significantly mediated treatment effects on tissue microstructure in white matter of superior longitudinal fasciculus, anterior corona radiata, posterior thalamic radiata, genu and splenium of the corpus callosum, and in gray matter of cingulate gyrus and putamen. To visualize and understand the longitudinal mediation effects, we generated scatterplots presented in supplementary figure Fig.S5. pCR = posterior corona radiata; LG = lingual gyrus; FA = fractional anisotropy; ADC = average diffusivity coefficient; HamD17 = 17-item Hamilton Depression Rating Scale; PBO = placebo-treated patients; yellow and red = positive associations; violet and blue = negative associations. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.) Sensitivity analyses The duloxetine and desvenlafaxine studies differed in trial duration and illness chronicity. We therefore conducted within-study (1) the time × treatment interaction effects and (2) associations between changes in DTI measures and changes in symptom severity to assess sensitivity of our main findings to between-study differences. Within-study analyses have fewer subjects than both studies combined, and therefore, have lower statistical power to detect true positive findings. The time × treatment effects observed in the duloxetine study alone (Fig. S9) and in the desvenlafaxine study alone (Fig. S12) were directionally consistent with those obtained in the pooled analysis (Fig. 2). The effects were smaller and weaker in the desvenlafaxine study (Fig. S12), which is consistent with its shorter trial duration, differing clinical response profile, and reduced power when analyzed independently For medication treated patients, the associations between changes in symptom severity and changes in brain measures in the duloxetine study (Fig. S10) were consistent with those observed in the combined patient cohort across both studies (Fig. 3). In contrast, these associations were substantially reduced or absent in the desvenlafaxine study (Fig. S13), likely reflecting reduced statistical power when analyzed independently. Conversely, for placebo treated patients, the associations between changes in symptom severity and changes in brain measures in the desvenlafaxine study (Fig. S14) aligned with those observed in the pooled analysis (Fig. 4). These associations were notably weaker in the duloxetine study (Fig. S11), again consistent with reduced statistical power and study specific differences when examined separately. Discussion Medication-Specific Effects We found evidence that medication effects on DTI measures over time differed significantly across the duloxetine and desvenlafaxine studies (i.e., the 3-way interaction of treatment*time*study was significant) for FA in deep white matter tracts (anterior and superior corona radiata, thalamic radiations, superior longitudinal fasciculus, and superior frontal gyrus) and for ADC in frontal cortical gray matter (Fig. 1). FA values that were higher than control values in anterior corona radiata at baseline normalized for patients treated with duloxetine but not placebo, whereas ADC values in frontal cortices that were comparable to control values at baseline declined in patients treated with placebo but not duloxetine (Fig. 1). In the desvenlafaxine study, DTI measures either did not change or they became increasingly abnormal, especially for placebo-treated patients (Fig. 1). Duloxetine and desvenlafaxine, although differ in potency and in their relative affinities for serotonin and norepinephrine transporters, both are SNRIs and therefore share the same core pharmacological mechanism of being dual reuptake inhibitors of serotonin and norepinephrine. The sparse but significant 3-way interaction effects potentially could be due to differences in trial durations, patient characteristics across the two studies, including differing illness duration and recruitment sites, or because of sampling differences in small sample sizes. Medication-specific effects on tissue microstructure in the dorsal brain regions may at least partly be explained by differing potency of duloxetine and desvenlafaxine in inhibiting the reuptake of serotonin and norepinephrine. Prior in vitro studies demonstrate that serotonin alters tissue microstructure by modifying dendritic arborization.(Kondoh et al., 2004, Rojas et al., 2017) Norepinephrine produces similar microstructural effects by increasing BDNF expression, enhancing CREB phosphorylation, and activating growth–associated proteins.(Chen and Russo-Neustadt, 2017) Although serotonergic and noradrenergic neurons project to large portions of the brain, including the limbic system, cognitive centers in the dorsolateral prefrontal cortex receive inputs only from noradrenergic neurons.(Joyce et al., 2024, Moret and Briley, 2011, Ramos and Arnsten, 2007, Stahl, 2003) Norepinephrine therefore modulates both emotion processing and executive regulation of cognition and motivation,(Moret and Briley, 2011) thereby enhancing cognitive flexibility(Yang et al., 2018) and emotion processing. Consequently, greater potency of duloxetine relative to desvenlafaxine in inhibiting norepinephrine transporters(Deecher et al., 2006) may confer the medication-specific effects on the tissue microstructure in the dorsal brain regions. Overlapping Medication Effects In brain regions where treatment effects did not differ significantly across studies (i.e., where significant 3-way interactions were absent), we assessed treatment effects combined across studies. Consistent with prior studies(Blood, A.J., Iosifescu, D.V., Makris, N., Perlis, R.H., Kennedy, D.N., Dougherty, D.D., Kim, B.W., Lee, M.J., Wu, S., Lee, S., Calhoun, J., Hodge, S.M., Fava, M., Rosen, B.R., Smoller, J.W., Gasic, G.P., Breiter, H.C., Phenotype Genotype Project on, A., Mood, D et al., 2010, Liao et al., 2013, Schmaal et al., 2020, van Velzen et al., 2020, Versace et al., 2010), DTI measures were significantly lower than control values in the insula cortex and lingual gyrus at baseline. Medication either normalized or increased those measures towards control values. In contrast, placebo caused DTI measures to deviate further away from control values in those regions (Fig. 2). Furthermore, as symptom severity declined, DTI measures normalized in medication-treated patients (Fig. 3) but deviated further from control values in placebo-treated patients (Fig. 4). Taken together, in placebo-treated patients a deviation further away from values for healthy participants is a neural compensatory response that helps them to manage symptoms. In contrast, as medications decrease symptom severity via not yet unidentified mechanism, this compensatory response is not required, enabling white matter microstructure to revert toward values for healthy participants. The absence of a 3-way interaction in these regions suggests that the mechanism of action for medication-induced normalization of DTI measures is likely similar for duloxetine and desvenlafaxine. Normalization of tissue microstructure in medication-treated patients but increasing abnormalities in placebo-treated patients in the same regions suggests that the DTI abnormalities at baseline may be compensatory, neuroplastic responses that help patients modulate the severity of their depressive symptoms. Medication treatment that successfully reduces depressive symptoms may then render neuroplastic compensation unnecessary, leading to normalization of DTI measures. Formal testing of this mediation pathway showed a medication-induced reduction in symptom severity significantly mediated treatment effects on tissue microstructure (Fig. 5 & Fig.S4), similar to our previous report of this mediation pathway for cortical thickness.(Bansal et al., 2018a) Two-wave mediation analysis therefore provided support for the possibility that symptom improvement may partially account for treatment-related changes in brain measures. Although mediation analysis is statistically consistent with the observed associations between treatment, symptom change, and imaging change, statistical analysis alone cannot establish a definitive biological pathway. Sensitivity Analyses The findings from the combined cohort showed time related effects that were consistent with those observed when each study was analyzed separately. As expected, the effects were less significant and had smaller spatial extent in the desvenlafaxine study, likely due to reduced statistical power and the use of two scanners. Nonetheless, the overall consistency across studies suggests that scanner differences in the desvenlafaxine trial, as well as the differing durations of illness and trial length across the two studies, had minimal impact on the main treatment related effects observed in the pooled cohort. Pathophysiological Implications Duloxetine normalized DTI measures in white matter tracts likely as a consequence of myelin reorganization or changes in extracellular matrix, thereby restricting the diffusion of free water molecules and increasing FA. The corona radiata comprise both ascending and descending axons to and from the cerebral cortex. It is continuous superiorly with the centrum semiovale and inferiorly with the internal capsule,(Chowdhury et al., 2010) which interconnects the cortex with the basal ganglia, thalamus, and limbic system(Qian and Tan, 2017, Tredici et al., 1990) to form the limbic-thalamo-cortical circuits that support mood regulation.(Catani et al., 2002, Sanjuan et al., 2013) Abnormalities in white matter of the corona radiata, posterior thalamic radiations, and superior longitudinal fasciculus(Aghajani et al., 2014, Bessette et al., 2014, Henderson et al., 2013, LeWinn et al., 2014) have been reported previously in MDD and PDD,(Vilgis et al., 2017) and especially in recurrent depressive episodes.(van Velzen et al., 2020) Our findings suggest that microstructure in these white matter pathways is normalized to the extent that medications reduce the severity of depressive symptoms. The shared effects of duloxetine and desvenlafaxine in normalizing low FA and ADC values at baseline in the insula cortex, globus pallidus, and lingual gyrus suggests that these medications reduce dendritic arborization and synaptic density in the cortical and subcortical gray matter. A reduction in dendritic arborization and synaptic density increases the diffusion of extracellular free water molecules, which in turn increases not only ADC, but also FA values because the reduced neuropil reveals a proportionally greater representation of radially oriented, intracortical axons.(Eaton-Rosen et al., 2017, Huang et al., 2013, Hüppi et al., 1998, Jespersen et al., 2011, Ouyang et al., 2019, Sizonenko et al., 2007, Wang et al., 2017) Normalization of gray matter microstructure is consistent with our previously reported thinning of abnormally thicker baseline cortex in medication-treated patients,(Bansal et al., 2017, Bansal et al., 2023) likely due to reduction in dendritic arborization and synaptic density. The insular and orbitofrontal cortices are key components of the limbic system,(Roxo et al., 2011) globus pallidus is a constituent in the limbic loop of the basal ganglia that processes emotion motivation, and reward-related behavior,(Aoki et al., 2019, Baumann et al., 1999, Kosel et al., 2007) whereas the lingual gyrus participates in processing emotional valence in visual stimuli.(Kesler-West et al., 2001, Kim et al., 2007) Abnormalities in structure,(Redlich et al., 2018) function,(Drevets, 2000, Lemke et al., 2022, Ravindran et al., 2009) and metabolism(Rao et al., 2011) of the limbic system have long been implicated in the pathogenesis of MDD and dysthymia. Our findings suggest that antidepressant medications normalize gray matter microstructure in the limbic system. Placebo Effects In placebo-treated patients of both studies, symptom severity decreased significantly likely due several contributing factors, including (1) the expectancy effect associated with taking a pill,(Rutherford et al., 2017) (2) the Hawthorne effect, whereby participants change behavior because they are being observed,(Berkhout et al., 2022) and (3) regression to the mean, as patients often seek treatment when symptoms peak and, consequently, improve spontaneously over time.(Hengartner, 2020, Rutherford and Roose, 2013) The placebo response rate, defined as at least a 50% reduction in HAMD-24 score and 2 point improvement on the CGI-I scale, was 26.9% in the desvenflaxine study and 25% in the duloxetine study. These rates are consistent with, and in fact lower than, the 35–40% response rates reported in a large meta-analysis of 256 RCTs of patients with MDD.(Furukawa et al., 2016, Rutherford and Roose, 2013). The decline in symptom severity was associated with a further reduction in low baseline FA values in white matter, as well as in ADC values of the lingual gyrus. Further lowering of FA and ADC values in gray matter of the lingual gyrus suggests an increase in dendritic arborization and synaptic density in response to placebo, consistent with our previous findings that thicker cortices at baseline thickened during administration of placebo.(Bansal et al., 2017, Bansal et al., 2023) Exaggeration of baseline abnormalities in DTI measures was proportional to the degree of improvement in their depressive symptoms. These findings collectively support our hypothesis that baseline abnormalities in DTI measures can be interpreted as neuroplastic compensation, which are further accentuated in placebo-treated patients. Other Contributors of Findings Our findings potentially could be influenced by differing variability in MRI measures and symptom assessments across treatment groups, nonspecific time effects in MRI-derived brain measures, and regression-to-the-mean improving symptom severity. We minimized or eliminated systematic, between-group measurement variability by collecting MRI data using identical pulse sequence parameters for every patient. In addition, all imaging data were processed blind to study identity, treatment assignment, and acquisition order. Symptom ratings were also collected under double blind conditions: patients were randomized to active medication or placebo, and both patients and clinicians remained blind to treatment assignment throughout the study. Scanner drift, developmental/aging effects, or test–retest variability could introduce nonspecific time related shifts in our measures. Furthermore, such effects will also limit our interpretation whether brain measures in patients normalized or deviated further away from values for healthy participants. However, such nonspecific time effects would be expected to influence both treatment arms similarly and, therefore, would produce parallel changes over time. In contrast, our analyses revealed a significant time × treatment interaction, indicating that the trajectories of MRI measures diverged between the medication and placebo groups. Thus, even if nonspecific time effects were present over these brief (10-week) trials, they were smaller than—and did not obscure—the differential longitudinal changes associated with treatment. Regression to the mean could have contributed to the observed symptom improvement. However, the MRI data did not follow a regression to the mean pattern. In placebo-treated patients, symptom improvement was accompanied by greater deviation of brain measures from healthy reference values. In contrast, in medication-treated patients, brain measures moved toward normal values as symptoms improved. This divergence suggests that the MRI changes reflect symptom linked state effects rather than a mere regression to the mean. Moreover, the opposing directions of MRI change—despite comparable symptom improvement across groups—indicate that distinct neurobiological processes likely underlie the trajectories observed in the two treatment arms. Scar Hypothesis of depression proposes that depressive symptoms exert neurotoxic effects on the brain (Allott et al., 2016, Hammar et al., 2022, Ronold et al., 2020), which leads to greater brain abnormalities with increasing symptom severity. If scarring is neuroplastic, then reduction in symptom severity may correspondingly reduce “scarring”. This framework could explain the pattern observed in medication-treated patients: baseline abnormalities potentially could be symptom-related “scarring,” which then diminishes as symptoms improve with the active treatment. The “scar hypothesis”, however, cannot explain the changes in brain measures in placebo-treated patients when symptom severity declined, given that their brain measures deviated further from values for healthy participants. In other words, the “scar” increased in placebo-treated patients even though the severity of their symptoms declined during the trials. Under the “scar hypothesis,” declining symptom severity should reduce scarring in both treatment arms. The opposite pattern of change in DTI measures across treatment arms despite declining symptom severity in both arms therefore challenges the scar hypothesis. Furthermore, our two-wave mediation analyses provides statistical support for our interpretation that changes in symptom severity mediated the association of treatment condition with changes in DTI measures, further supporting of interpretation of brain changes during the trials as representing compensatory neuroplasticity. Relation to Studies in MDD Prior studies have reported lower FA values in patients with MDD relative to healthy controls (Blood, A.J., Iosifescu, D.V., Makris, N., Perlis, R.H., Kennedy, D.N., Dougherty, D.D., Kim, B.W., Lee, M.J., Wu, S., Lee, S., Calhoun, J., Hodge, S.M., Fava, M., Rosen, B.R., Smoller, J.W., Gasic, G.P., Breiter, H.C., Phenotype Genotype Project on, A., Mood, D et al., 2010, Schmaal et al., 2017, Versace et al., 2010). Furthermore, naturalistic longitudinal studies tracking clinical course in MDD have shown neuroplastic thickening of the cortex in patients who remit but progressive cortical thinning in patients who do not remit (Phillips et al., 2012; Yuksel et al., 2018). These findings suggest that symptom‑linked neuroplastic processes—compensatory and maladaptive—may be present in both MDD and PDD. We therefore hypothesize that the use of SNRIs in MDD would induce white matter microarchitecture changes similar to those observed in PDD patients. Limitations Our findings should be considered in the context of several limitations. First, data from forty-one patients separately in each study may not have provided sufficient statistical power to detect small treatment effects on DTI-derived brain measures or the mediating effects of symptom change on tissue microstructure. We significantly increased statistical power to detect common treatment effects by combining data across the two studies. Second, small sample size can also lead to false positives by inflating effect sizes and instability in the data, especially in 3-way and mediation analyses. To control for false positives in our reported findings, we therefore employed rigorous cluster-size family-wise error rate (FWER) procedure. Replication in a larger cohort would be needed to confirm our findings. Third, patients who were acutely suicidal or who had medical and psychiatric comorbidities were ineligible to participate in these studies, thereby limiting the generalizability of our findings. Fourth, findings in the desvenlafaxine study may have been affected by DTI data collected on two different MRI scanners. We minimized this possibility by acquiring MRI data using identical pulse sequences on two identical 3 T GE scanners, and statistically by covarying for scansite in our analyses. Sensitivity analyses showed that our findings were invariant to including scansite as a covariate. Fifth, a mechanistic, compensatory interpretation of our findings is limited by data collection at only two timepoints limit, even though this interpretation is supported by our data and analyses. Studies in larger cohort are needed to confirm the compensatory mechanism for symptom improvement in patients with PDD. Sixth, significant 3 way interaction effects—reflecting differences in treatment effects between the duloxetine and desvenlafaxine studies—could arise from factors such as differing illness duration, trial length, clinical response profiles, or the use of two scanners in the desvenlafaxine study. Despite these between study differences, the absence of 3 way interaction effects across most of the white matter indicates that the treatment–related effects on white matter microstructure were statistically comparable across the two trials. Therefore, pooling the datasets afforded greater statistical power to detect treatment effects in brain regions with no study–specific divergence. Seventh, biological interpretations for changes in FA and ADC measures should be considered in the presence of other tissue characteristics that affect DTI-derived brain measures. In addition to dendritic arborization and synaptic density, FA and ADC in gray matter can be affected by several other tissue components, including the density and morphology of neuronal and glial cell bodies, glial processes, and unmyelinated intracortical axons (Radhakrishnan et al., 2022). Neuronal soma density or intracortical axon counts however are unlikely to change over a 12-week period, whereas glial processes, dendritic branching, synaptic density are highly dynamic and responsive to changes in neuronal activity over timescales of days to weeks (Kolb et al., 2003, Wong and Wong, 2000). Although combined changes in these tissue components likely contributed to the changes in DTI measures, changes in synaptic density and dendritic arborization likely contributed more to changes in DTI-derived measures over the short trial periods. Eight, in placebo-treated patients, deviation of FA and ADC measure further away from values for healthy participants could represent the continued need for compensatory response that help them cope with symptom severity. Alternatively, symptom severity changes in placebo-treated patients potentially could represent nonspecific trial effects, and deviations in brain measures further away from values for healthy participants could derive from the relatively small sample size or noise in the data. Conclusions The unique and common effects of duloxetine and desvenlafaxine on monoamine neurotransmitter systems are likely responsible for their spatially unique and common spatial effects on brain tissue microstructure. The contrasting findings that medication attenuates or normalizes abnormal values, whereas placebo further accentuates them relative to values for healthy participants suggest that abnormal tissue microstructure at baseline is at least in part a neuroplastic compensatory response to the presence of more severe depressive symptoms; medication then reduces symptoms, which thereby reduces the need for neuroplastic compensation. Placebo generally does not reduce symptoms to the same degree, which thereby sustains the need for ongoing neuroplastic compensation. These findings suggest that novel therapeutic interventions should target and enhance neuroplasticity to more effectively treat chronic depressive illness. Disclosures Drs. Peterson and Hellerstein received funding for the desvenlafaxine study from Pfizer, Inc., and for the duloxetine study from Eli Lilly & Co. Dr. Hellerstein has received research funding from Beckley PsyTech, Compass Pathways, MindMed, and Relmada Therapeutics. All other authors have no conflict of interests to state. CRediT authorship contribution statement Ravi Bansal: Writing – review & editing, Writing – original draft, Validation, Supervision, Software, Methodology, Formal analysis, Data curation, Conceptualization. David J. Hellerstein: Writing – review & editing, Funding acquisition, Conceptualization. Siddhant Sawardekar: Writing – review & editing, Formal analysis, Data curation. Ying Chen: Writing – review & editing, Data curation. Bradley S. Peterson: Writing – original draft, Funding acquisition, Conceptualization. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.